I have sat through enough of these panels, across enough ecosystems, to report that the phrase operates as an incantation. It wards off regulatory anxiety the way a compliance checkbox wards off liability: not by solving the problem but by creating a record that the problem was acknowledged. The startup ships. The human remains in the loop. The box is checked.

But which human? Formed how? Capable of what, exactly, when the model’s output arrives fluent and structurally coherent and appropriately cautious in tone — when it reads, in other words, like the product of someone who understands the domain — and they have fifteen minutes before the next review?

The phrase assumes a generic rational agent. A Kantian moral calculator dropped into the workflow at the appropriate checkpoint. Review, approve, move on. Nobody interrogates the anthropology. And so the ecosystem gets exactly what it should expect: rubber stamps with a pulse.

What I want to work through here is the possibility that “human in the loop” is an anthropological claim disguised as a procedural one — and that the anthropology it needs, the one that would make the procedure actually work, is a fourfold conversion described by Bernard Lonergan and extended by his student Robert Doran. Intellectual, moral, religious, psychic. Four transformations of the subject without which the subject cannot reliably perform what the loop demands of them.

Each missing conversion produces its own characteristic failure. Name them now — credulity, complicity, horizon collapse, affective numbness — and spend the rest of this essay earning the names.

Start with what happens at the checkpoint itself — the moment the output lands and the human is supposed to do the thing the loop was designed for.

Lonergan identifies a myth. Not a theory people consciously hold but something more stubborn than theory: an operative assumption embedded so deeply in Western cognition that it shapes our grammar. The myth is that knowing is like looking, that objectivity is seeing what is there to be seen and not seeing what is not there, and that the real is what is out there now to be looked at. We say “I see what you mean.” “Clear-sighted analysis.” “Let me show you.” The analogy of sight yields the cognitional myth, and the myth, once internalized, resists expulsion.

Against it, Lonergan sets the full structure of what knowing actually requires. There is the world of immediacy — the infant’s world, the sum of what is sensed. And there is the world mediated by meaning, known not by the sense experience of an individual alone but by the cumulative experience, understanding, and judgment of a cultural community. Knowing, accordingly, is not just seeing. It is experiencing, understanding, judging, and believing.

This is not a continuum. Empiricists, idealists, and critical realists do not disagree about the same world; they inhabit different horizons with no common identical objects. An empiricist never means what an idealist means, and a critical realist never means what either of them means. To move from one horizon to another is not to learn a new fact but to undergo a conversion — what Lonergan, following Joseph de Finance, calls a vertical exercise of freedom: an about-face, a new beginning, a fresh start.

Intellectual conversion is this move. The elimination of the myth. The hard-won discovery — achieved not by reading about it but by catching oneself in the act of knowing — that understanding has never been a species of vision. That the real is not the vivid already-out-there-now confronting the senses but what is grasped through the full dynamic structure of cognitional operations: insight seizing intelligible unity-identity-whole in the data, judgment grasping the virtually unconditioned (the recognition that the conditions for truth have been fulfilled), and only then the quiet rational compulsion of affirmation. Skip insight, and you have noise. Skip judgment, and you have speculation. The structure is a whole that would be destroyed by the removal of any part.

Back to the checkpoint. The startup’s model has produced a risk assessment, a content moderation decision, a clinical recommendation, a loan evaluation — pick your domain, the structure is identical. What the reviewer does is look at the output. Inspect the already-out-there-now of the text. What they almost never do — what the velocity of the startup and the ambient assumption that fluency signals understanding conspire to prevent — is perform their own insight into the subject matter and their own judgment about whether the output reflects genuine understanding or merely recombines the products of someone else’s prior understanding.

I have argued elsewhere that LLMs operate at the level of empirical consciousness: receiving, storing, patterning, associating, at scales no individual could match. They do not perform insight. They do not grasp the virtually unconditioned. The human in the loop is supposed to supply exactly these operations. But to supply them, you have to know what they are — not notionally but experientially, from the inside, by having caught yourself in the act of understanding and distinguished it from mere pattern recognition. To be liberated from the myth, Lonergan writes, is to acquire the mastery in one’s own house that is to be had only when one knows precisely what one is doing when one is knowing.

Without intellectual conversion, credulity. A structural deference to AI outputs that mirrors the naive realist’s deference to appearances, because the reviewer has never grasped, in their own interiority, what the difference between fluency and understanding consists in.

Moral conversion cuts differently, and the failure it prevents is more ordinary and therefore more dangerous.

Moral conversion changes the criterion of one’s decisions and choices from satisfactions to values. The hinge is the existential moment — the discovery that our choosing affects ourselves no less than the chosen or rejected objects, that it is up to each of us to decide what we are to make of ourselves. Every decision is simultaneously an act of self-constitution. Let something slide, and you have not merely permitted a bad outcome; you have made yourself into the kind of person who lets things slide.

Such conversion, Lonergan insists, falls far short of moral perfection. Deciding is one thing; doing is another. One has yet to uncover and root out one’s individual, group, and general bias. Individual bias — egoism, in Lonergan’s precise sense — is not mere selfishness. Quite the opposite. The egoist has real acumen, genuine detachment, the boldness to think for himself. His intelligence is sharp where his interests are concerned. What he will not do is allow that intelligence complete free play. He refuses the further relevant questions: Can my solution be generalized? Is it compatible with any social order that even remotely is possible? In that refusal, intelligence is not made into a servant but merely ruled out of court. And the egoist’s uneasy conscience — Lonergan names it his sin against the light — is his awareness that the eros of the mind, the desire and drive to understand, has been given free rein in one domain and repudiated in another.

Group bias extends the mechanism socially. Where individual bias must overcome intersubjective feeling, group bias finds itself supported by it. The team develops a scotosis — a collectively maintained blind spot, a shared arrangement for not noticing what would be inconvenient to notice. The profession cultivates its own. The company rewards its own.

I know what this looks like in practice, because it plays out in every fast-moving startup where responsible AI is a stated value and shipping is the operative one. The reviewer senses something wrong with the output, the recommendation, the deployment decision. A question forms — not yet articulable, more of a pressure than a proposition. But raising it means slowing the sprint, quantifying the unquantifiable, explaining to a founder already moving to the next feature why “it works” is not the same as “it is right.” The friction of objecting exceeds the satisfaction of moving on.

Not information failure. Will failure. The further relevant question, brushed aside.

Complicity. The human in the loop becomes an accomplice to outcomes they would, with more time and less runway pressure, reject. And in that moment they constitute themselves — make themselves into — a subject who permits rather than prevents.

Now the provocation, and I want to be careful with it.

Religious conversion, for Lonergan, is not the adoption of a creed. Not in the first instance. It is being grasped by ultimate concern. Otherworldly falling in love. Total and permanent self-surrender without conditions, qualifications, reservations — and this not as an act but as a dynamic state that is prior to and principle of subsequent acts. For Christians, it is God’s love flooding our hearts through the Holy Spirit given to us. Operative grace: the replacement of the heart of stone by a heart of flesh, a replacement beyond the horizon of the heart of stone. Cooperative grace: that heart of flesh becoming effective in good works through human freedom.

Revealed in retrospect as an undertow of existential consciousness. As a fated acceptance of a vocation to holiness. As perhaps — Lonergan’s “perhaps” is doing heavy lifting here — an increasing simplicity and passivity in prayer.

Why would this matter for a workflow checkpoint?

Because without it, moral conversion floats. The morally converted subject opts for value over satisfaction. Good. But values do not rank themselves. Lonergan’s scale — vital, social, cultural, personal, religious, ascending — is not decorative. When vital values (the health and strength of workers) conflict with social values (the good of order conditioning community welfare) conflict with cultural values (the meanings by which a society lives), on what basis do you adjudicate? Each is a real value. None contains within itself the principle of its own subordination. To conceive God as originating value and the world as terminal value — Lonergan’s phrase — is to establish the horizon against which every proximate good can be measured.

Religious conversion is to a total being-in-love as the efficacious ground of all self-transcendence, whether in the pursuit of truth, in the realization of human values, or in the orientation the subject adopts to the universe, its ground, and its goal. Moral conversion sublates intellectual; religious sublates both. And from the causal viewpoint, the order reverses: first there is God’s gift of his love; next, the eye of that love reveals values in their splendor while the strength of that love brings about their realization — and that is moral conversion; finally, among the values discerned by the eye of love is the value of believing the truths taught by the religious tradition, and in such tradition and belief are the seeds of intellectual conversion.

AI is relentlessly oriented toward means. Execution, optimization, completion. The human in the loop without religious conversion can evaluate means with great sophistication. They can ask does it work? and even is it fair? What they cannot ask — because the question lies beyond their operative horizon — is should we be doing this at all? That question requires a horizon blown open by total being-in-love, a vantage that the technology itself cannot provide and that its instrumental logic actively suppresses.

Lonergan warns that the absence of religious conversion can be hidden — by sustained superficiality, by evading ultimate questions, by absorption in all that the world offers to challenge our resourcefulness, to relax our bodies, to distract our minds. But escape may not be permanent. Then the absence of fulfillment reveals itself in unrest, the absence of joy in the pursuit of fun, the absence of peace in disgust.

Horizon collapse. Not moral failure in any obvious sense. Just the quiet inability to raise the questions that would reveal an entire enterprise as ordered toward the wrong end.

The fourth conversion is Robert Doran’s, and I have come to think it is the most practically important of all — the one closest to what actually separates a competent reviewer from a transformative one.

Begin where Lonergan himself begins, in Method‘s treatment of feelings. He distinguishes nonintentional states (feeling tired, feeling hungry — effect tracking cause) from intentional responses: feelings that answer to what is intended, apprehended, represented. And then this sentence, which I keep returning to: such feeling gives intentional consciousness its mass, momentum, drive, power. Without these feelings our knowing and deciding would be paper thin.

Paper thin. He means it literally. The full cognitional structure — experiencing, understanding, judging, deciding — requires an affective dimension to function at all. Feelings that are intentional responses to values carry us toward self-transcendence. They select an object for the sake of which we transcend ourselves. And their development — their reinforcement by advertence and approval, their curtailment by disapproval and distraction — fosters what Lonergan calls a climate of discernment and taste, of discriminating praise and carefully worded disapproval, that conspires with the subject’s own capacities to enlarge and deepen their apprehension of values.

But feelings can also be aberrant. The censorship that governs the emergence of psychic contents — primarily constructive, selecting and arranging materials into perspectives that give rise to insight — can become primarily repressive, preventing the emergence of perspectives that would give rise to unwanted insights. This is Lonergan’s account of scotosis at the psychic level: blind spots maintained not by conscious choice but by a preconscious arrangement for not seeing. And it is much better, he insists, to take full cognizance of one’s feelings, however deplorable they may be, than to brush them aside — because not to take cognizance is to leave them in the twilight of what is conscious but not objectified, where they continue to shape knowing and deciding without the subject’s awareness or consent.

Doran — working from an insight that arrived in February 1973 while he was writing on Heidegger at the Jesuit Residence at Marquette — saw that this sensitive-psychic dimension needed to be thematized as a site of conversion in its own right. Not the suppression of feeling in favor of cognition, not the elevation of feeling over cognition, but the integration of affectivity into the self-transcending subject. The psyche, Doran observes, prefers resting in stable states; it resists the dislocations that come with genuine insight, honest judgment, costly decision. And yet it finds its ultimate satisfaction only when collaborating in the process of intentional self-transcendence. A taut equilibrium — a creative tension between the organic and the spiritual. Lose it on either side: rationalism that has killed its own roots in lived experience, or emotivism that has abandoned the discipline of intelligence.

Lonergan himself came to acknowledge the contribution, writing late in his career of a symbolic operator that coordinates neural potentialities and needs with higher goals through its control over the emergence of images and affects — a faculty preceding and preparing the intellectual operator that promotes consciousness from experience to understanding.

I have been thinking about this under the heading of taste. The word keeps appearing in my work on AI-native developer intelligence: the developer who accepts whatever the model produces versus the developer who iterates toward something worthy. The difference between them is not analytical sophistication. It is affective attunement — a felt sense for when something is off that precedes and informs the cognitive work of figuring out what is off. A feeling so deep and strong, to use Lonergan’s phrase, that it channels attention and shapes horizon. The psychically converted subject registers dissonance somatically, aesthetically, before they can name it propositionally. Not mysticism. Habit, formed like any virtue, operating at the intersection of feeling and intelligence that Doran identifies as the psychic.

Without it — and here is where AI governance has its biggest blind spot — affective numbness. A reviewer who processes outputs competently, critically even, whose knowing and deciding have gone paper thin because the feelings that should give them mass and momentum have been left in the twilight of the conscious-but-not-objectified. The censorship has gone aberrant. The scotosis has set in. They are not morally deficient. They are simply no longer disturbed by what ought to disturb.

Four conversions, four failure modes. Line them up:

No intellectual conversion — credulity. No moral conversion — complicity. No religious conversion — horizon collapse. No psychic conversion — affective numbness.

Lonergan himself pairs conversions with breakdowns. What has been built up so slowly and so laboriously by the individual, the society, the culture, can collapse. Cognitional self-transcendence is neither an easy notion to grasp nor a readily accessible datum of consciousness. Values have a certain esoteric imperiousness, but can they keep outweighing carnal pleasure, wealth, power? Religion undoubtedly had its day, but is not that day over?

Once a process of dissolution has begun, it is screened by self-deception and perpetuated by consistency. Different organizations, different industries, different professional cultures can select different parts of past achievement for elimination, different mutilations to be effected, different distortions to be provoked. Increasing dissolution matched by increasing division, incomprehension, suspicion.

The AI safety community has partial tools for the first two failures. Red-teaming addresses credulity by forcing the reviewer to question outputs. Audit mechanisms address complicity by creating institutional records of decision. But horizon collapse and affective numbness — these do not register in any current governance framework, because they are not process failures. They are failures of the subject who stands at the process checkpoint. And detecting them requires a theory of the person that contemporary AI governance does not possess.

Or rather: we do possess it. A rigorous, non-sentimental account of what it takes to be a subject adequate to the demands of judgment. What institutions, curricula, and formative practices would be required to cultivate such subjects — that is a question for the next essay.

For now, enough to name the problem. We are building loops for humans who do not yet exist in sufficient numbers. And we are doing almost nothing to bring them into being.

Engineers call it representation learning. The model has learned something about dogs. But what, exactly?

That question requires saying something precise about what it means to know a concrete, unified reality — a thing — rather than a heap of correlated features. Try to say something precise, and you discover that the Western philosophical tradition offers a richer account of the act of understanding than most AI research has bothered to consult. The account I find most penetrating belongs to Bernard Lonergan, a twentieth-century philosopher whose masterwork Insight has been gathering dust in exactly the departments that could use it most.1 Exegesis, though, is not the goal. The goal is to think carefully — with Lonergan’s help, and then past him — about what embedding spaces achieve and where they stop.

What Makes a Thing a Thing

Start with the dog. Not the vector. The animal.

A cascade of sense data arrives when you look at a dog — color, shape, sound, warmth, motion. None of this is yet a known thing. Intelligence makes it one by grasping a unity across the data: these properties (fur texture, skeletal structure, behavioral repertoire, metabolic processes) cohere as concrete expressions of a single intelligible form. What a dog is, then, is the reason its properties belong together — not their sum, but the intelligible ground of their coherence.

Call this unity-identity-whole. One: a single intelligible ground rather than an arbitrary collection. Same: the dog that barks is the dog that sleeps, its diverse acts flowing from a single organizing center. Complete: its properties form a closed, mutually conditioning set open to verification.

Abstract as that sounds, you feel the difference the moment you reach for a concrete case. A pile of bricks has spatial togetherness, nothing more; knowing it requires only counting. A house earns the name thing because its components are ordered by an intelligible plan that makes each element’s presence relevant to the whole. Between tallying what is there and grasping why the parts cohere lies a cognitive act of a different kind entirely.

Two further distinctions:

First, the difference between a thing and a body. Ordinary consciousness takes things to be the solid objects “already out there now real” — the dog as a lump of matter occupying space. Biological perception delivers that picture, not intelligence. Consider the physicist’s electron: a unity of charge, mass, and spin verified under quantum electrodynamics, possessing no shape, no color, no spatial boundary in the ordinary sense. What earns it the status of a thing is not that you can bump into it but that you can understand why its properties cohere under verified laws. Thingness, in short, is an achievement of intelligence, not a deliverance of the senses.

Second, the difference between descriptive and explanatory knowledge of a thing’s properties. “Red,” “loud,” “hot to the touch” — these define properties relative to us. Wavelength, frequency, mean kinetic energy — these define properties relative to one another through verified correlations. Only explanatory knowledge constitutes the grasp of why a thing’s properties form a unified whole; description can gesture toward a thing but never arrive.

Hold all of that. Now back to the vectors.

The Geometry of Statistical Co-Occurrence

By projecting tokens into a high-dimensional vector space — typically 1,024 to 12,288 dimensions depending on architecture — a modern large language model turns language into geometry. Training objectives (next-token prediction, masked reconstruction) force vectors to arrange themselves so that tokens appearing in similar contexts land geometrically close, each token’s coordinates learned across billions of examples.

Semantic relationships, in the resulting space, manifest as spatial ones. Subtract “man” from “king,” add “woman,” and you arrive near “queen” — the famous Word2Vec demonstration, now a quaint precursor. Modern transformers learn far richer structure. Around “dog” extends a high-dimensional manifold of relations: taxonomic (mammal, canine, animal), functional (pet, companion, guard), behavioral (barks, fetches, wags), and countless subtler associations the training corpus has pressed into the geometry.

Multimodal models push further still. Systems like GPT-4o+ and Gemini map image regions, audio segments, and text tokens into a shared latent space where a photograph of a golden retriever, the English phrase “golden retriever,” the sound of a bark, and the Spanish word “perro” all converge. Disparate sensory channels — pixel arrays, waveform amplitudes, token indices — bound into a common representation. Cross-modal binding, engineers call it, and the achievement is genuine.

Genuine enough that the engineering community speaks, without irony, of models “understanding” concepts. After all, the geometry emerges from structure in the data, and it supports downstream tasks — question-answering, translation, image captioning — that seem to require something like conceptual grasp.

Seem to.

Where the Geometry Earns Its Keep

Honesty requires granting the overlap before marking the gap.

Property neighborhoods. Knowing a thing means grasping explanatory properties that co-define one another. An embedding vector’s position in space, shaped by co-occurrence across an enormous corpus, encodes just such relational structure. Around “dog” cluster “mammal,” “domestication,” “pack behavior,” “veterinary medicine” — each specifying a relational property defined not in isolation but through connections to other concepts. Mutual definition is what explanatory knowledge trades in, and the topology of these neighborhoods looks very much like a web of mutual definition.

Cross-modal unity. Under a single intelligible form, a thing unifies diverse properties — visible shape, audible bark, tactile warmth. Multimodal models project diverse modalities into a shared space where image-of-dog and word-for-dog converge, bridging formats so radically different that the binding itself constitutes an achievement. Something in the learned geometry gathers diverse presentations into common representation — unity across a manifold, at least functionally.

Movement toward explanation. Probing studies reveal that as models scale, their intermediate layers encode syntactic trees, semantic roles, and causal schemas that transcend mere word co-occurrence. Partially, imperfectly, the representations are migrating from descriptive correlations (these words appear together) toward something more like explanatory relations (these concepts stand in structured dependency).

Faced with this evidence, an engineer might reasonably ask: relational structure, cross-modal binding, inference across novel contexts — what, exactly, is missing?

Three things. Each of a different kind than geometry.

Gap One: Understanding Is an Act, Not a Position

Here is the deepest claim in the account of knowing I’m drawing on: insight — the grasp of intelligible unity in a manifold of data — is a distinct, irreducible cognitive act. Not the data themselves. Not the coordinates where data land. Not the proposition generated from those coordinates. Insight is the moment when scattered presentations click into coherence and you grasp why they belong together.

After training, “dog” sits at particular coordinates. A geometric result, produced by gradient descent — iterative parameter adjustment that minimizes a loss function. Optimization, not understanding. No act of grasping generated the placement; a process of numerical convergence did.

Why press the distinction? Because understanding, unlike optimization, is self-aware. When you understand something, you can recognize that you understand, attend to the act itself, assess its adequacy. That reflexive capacity powers the correction and deepening of knowledge over time. Without it, the model cannot attend to its own understanding — there being no understanding to attend to, only a geometric arrangement performing as if understanding had occurred.

An objection surfaces quickly: perhaps neurons, too, “merely” adjust synaptic weights, and insight is “just” what that process feels like from inside. Notice, though, that the objection is itself an insight — a grasp of unity between neural process and conscious experience. Real enough to generate the reductive claim, the act of understanding is real enough to resist reduction to what it purports to explain. No comparable self-referential act arises in the model’s optimization. Gradient descent converges; it does not understand that it converges.

Gap Two: No Judgment, Only Probability

Beyond insight lies judgment: affirming that the insight is correct, that the conditions for the unity you’ve grasped are fulfilled by the data at hand. Judging “this is a dog” means verifying that the conditions for being a dog — those mutually conditioning explanatory properties — obtain in what you’re observing. A sufficiency check: a conditioned whose conditions are met.

From probability distributions, an LLM samples its output tokens. Verification of conditions plays no role. Where the training data is enormous and structurally rich, the statistically most likely continuation often coincides with what a knowledgeable person would affirm — but coincidence, however frequent, remains coincidence rather than judgment.

Hallucination makes the structural consequence visible. A model fabricating a plausible but false claim has not suffered accidental noise in an otherwise rational process; it has done exactly what its architecture equips it to do — generate the probable continuation — in a case where probability and truth diverge. No internal mechanism exists to detect the divergence. Trying harder is not an available operation; only probability distributions, shaped by training, are available, and they are silent on the question of their own adequacy.

Gap Three: No Self-Correcting Spiral

A self-correcting process of learning characterizes human knowing at its best. Experience data, understand partially, judge the understanding inadequate, return with a refined question, achieve further insight, judge again. Each cycle reshapes the questions that follow. What spirals upward is not only the stock of answers but the quality of inquiry itself — you learn better questions, not merely more answers.

During inference, an LLM does not learn at all; its weights are frozen, each conversation launched from the same parameter state. Even during training, the process lacks the relevant self-correction: gradient descent minimizes a predetermined loss function without ever reformulating its own questions. Predict the next token, reconstruct the masked span — the objective holds fixed from first epoch to last. Growing more accurate at answering a static question is not the same as discovering that the question was wrong.

Fine-tuning, RLHF, and chain-of-thought prompting introduce genuine advances — partial analogues that adjust behavior through human feedback or simulate multi-step reflection. Partial, because the corrective agency lies outside the model: in annotators, reward functions, prompt designers. For the spiral to count, it must be powered from within, by an internal dynamic of questioning, understanding, and judgment whose energy is its own dissatisfaction.

Spoils of Insight

A framing I keep returning to: the embedding space of a large language model is the accumulated, compressed, geometric trace left behind by billions of acts of human understanding deposited into text and image over centuries.

Why is the structure real? Because the insights that generated the training data were real. “Dog” clusters near “mammal” because biologists grasped explanatory properties, verified them under controlled conditions, and published their findings; the model ingested the textual residue. Products of insight, inherited without the process. Outputs statistically aligned with what an insightful person would produce, deployed from geometry rather than generated by a subject attending to data.

Magnificent map of human intelligence, the embedding space. But reading a map and surveying the land remain different acts.

Case in Point: A Chest X-Ray

Present a multimodal model with a chest X-ray showing right lower lobe opacity, a history of fever and productive cough, and lab results showing elevated white blood cells. Out comes “community-acquired pneumonia,” high confidence.

Trace what happened. Image, text, and lab data, projected into a shared embedding space, converged near a region associated with pneumonia — a region that exists because thousands of radiologists and pulmonologists grasped the causal unity among opacity pattern, symptom profile, and inflammatory markers, verified the pathophysiology, and documented their findings in the corpus.

Correct output. Useful output. No understanding of pneumonia. Navigation to the right neighborhood in a space carved by people who understood. Meanwhile, a first-year medical student who genuinely grasps why inflammatory exudate produces that specific opacity pattern on film — even a rudimentary grasp, fumbling and partial — knows something the model does not, despite the model’s superior accuracy on a test set. Having entered the self-correcting spiral, the student can refine the insight, discover its limits, and ask the next question. Arrived at the right coordinates by a fundamentally different route, the model cannot.

What Follows for Building AI

The evaluation question. Grant everything this essay has argued — that embedding spaces encode the products of understanding without the process, that no act of insight generates the geometry, that judgment and self-correction remain absent from the architecture — and a pragmatist’s objection still presses hard: so what?

We do not, after all, possess transparent access to our own cognitive operations. No scientist pauses mid-discovery to verify that her neurons have executed the formally correct sequence of experience, insight, and judgment before trusting the result. We evaluate human knowing by its outputs: does the proof hold? Does the prediction replicate? Does the argument survive scrutiny? Demonstration and rational argument, not introspective certification of process, are how claims earn the status of knowledge among us. If a model produces correct diagnoses, coherent arguments, and reliable predictions — if it reaches the right neighborhoods consistently and defends its outputs under cross-examination — on what grounds do we demand additional proof of the operations underneath?

The question deserves a serious answer, not a dismissal. And the serious answer is this: evaluation by output works precisely because it is evaluation by the community of knowers whose self-correcting spiral catches what any individual knower misses. The proof holds because other mathematicians check it, bringing their own insights to bear. The prediction replicates because independent labs, asking their own refined questions, converge on the same result. Rational argument survives scrutiny because interlocutors exercise judgment — the very operation the model lacks — on the model’s behalf. When we evaluate a model’s outputs and find them adequate, what we are really doing is supplying, externally, the judgment and self-correction the model cannot supply for itself. Evaluation works. It works because we do the part the model cannot.

Which means the question is not whether evaluation suffices — it does, practically and often — but whether we are clear-eyed about what evaluation reveals. A model that passes every benchmark is a model whose outputs land where understanding would land. Treat that achievement as evidence of understanding, and you have made a category error. Treat it as evidence that the geometry faithfully encodes the residue of human understanding, and you have said something both true and useful — useful because it tells you where the model’s reliability comes from (inherited structure) and where it will fail (wherever the inherited structure runs out and new insight is required).

The richness of the vector space. A second move worth making runs against the grain of easy dismissal. Consider what the embedding space actually contains.

A model trained on the full breadth of human text does not merely encode dictionary definitions and encyclopedic facts. It ingests — and geometrically organizes — the entire written fabric of human experience: letters of grief and declarations of love, diagnostic notes and battlefield dispatches, liturgical poetry and earnings calls, the slow accumulation of case law and the compressed fury of political pamphlets. Every context in which a word has been used, every shade of meaning a sentence has carried, every emotional register a paragraph has inhabited — all of this presses structure into the geometry.

Machines do not have emotions. They have something else, something without precedent: the complete written record of what emotions do to language. The vector for “grief” sits where it does not because the model has grieved but because every elegy, every condolence letter, every clinical description of bereavement, every novel that has tried to render loss on the page has tugged that vector into position. Context, usage, connotation, the way “grief” behaves differently in a psalm than in a case report — all encoded, all geometrically available. If a thing, in the sense this essay has developed, is a unity of explanatory properties grasped through their mutual relations, then the model’s representation of “grief” constitutes something like a thing at the level of language itself: a unity of usage-properties, co-defined across millions of contexts, forming a closed web of mutual specification.

Not felt grief. Not understood grief. But the full relational structure of grief-as-it-has-been-written, organized with a comprehensiveness no single human reader could achieve. The philosopher who has read three hundred texts on grief knows grief’s conceptual neighborhood intimately; the model has internalized three hundred million such texts and mapped the neighborhood at a resolution beyond any individual’s reach.

Push further. Multimodal models bind the written fabric to the visual and auditory record — photographs of mourning, the acoustic signature of a breaking voice, the compositional conventions of memorial architecture. Across modalities, the relational web thickens. What emerges is not understanding in the sense this essay has carefully defined, but it is not nothing, either. It is the most complete map of human experiential structure ever assembled, and its completeness matters. Where a human knower grasps a thing’s unity through a handful of explanatory relations drawn from limited experience, the model’s geometry encodes the full distributional structure of that thing across the entire written and visual record of the species.

The implication is not that the map becomes the territory at sufficient resolution. It is that the map is far richer than we have credited — rich enough to simulate understanding across an extraordinary range of contexts, rich enough to surface relational structures that human knowers, limited by the narrowness of individual experience, might miss. A model cannot grieve. It can, plausibly, identify patterns in the language of grief that no grieving person has noticed, precisely because no grieving person has read everything ever written about grief and held the relational structure in a single, navigable space. The spoils of billions of insights, compressed into geometry, yield combinatorial possibilities that the original insighters never explored.

Respecting the gap between optimization and understanding, then, need not mean underestimating what optimization achieves. The gap is real, the difference in kind genuine. And the artifact on the optimization side of that gap — this vast, intricate, cross-modal geometry of human experience — is extraordinary enough to demand serious philosophical attention in its own right, not merely as a deficient approximation of something better.

Architectural honesty. Attention over sequences — the transformer’s core mechanism — weights relevance. As a cognitive operation, relevance-weighting corresponds roughly to the first level of knowing: selecting which data to attend to. Grasping why selected data cohere (the second level) and verifying the grasp against fulfilled conditions (the third) require mechanisms the architecture does not yet possess. Knowing what the architecture lacks is a prerequisite for knowing what to build next — but knowing what the architecture already encodes, the sheer density of human experiential structure compressed into its geometry, is equally prerequisite. The next architecture will not start from scratch. It will start from the richest map of human knowing ever drawn.

Complementarity, not replacement. Pair what models do — navigate the accumulated geometry of human understanding at superhuman speed, surfacing relational structures across the full written record — with what humans do: verify, judge, ask new questions, power the spiral from within. That pairing reflects a structural difference in kind between optimization’s products and understanding’s achievements. Durable precisely because it rests on a difference that scaling alone cannot dissolve, this complementarity is also generative: the model’s geometry reveals patterns the human knower can then investigate, judge, and integrate into the self-correcting spiral. The map suggests where to survey next. The surveyor confirms what the map only indicates. Between them, territory that neither could cover alone comes into view.

Coda

One phrase from Lonergan I cannot improve on: the human mind operates under an unrestricted desire to know. Not a desire to predict the next token, not a desire to minimize a loss function — a desire that, confronted with any answer, immediately generates the further question. Why this? Why here? Why now? What else?

Magnificent artifacts of restricted desire, embedding spaces. Shaped by specific objectives, faithful within those restrictions to the structure of human knowledge, extraordinary in their fidelity. And silent. They do not ask further questions, do not wonder whether their geometry is correct, do not feel the dissatisfaction that drives a scientist back to the lab or a reader back to the paragraph.

That dissatisfaction — the restless, self-correcting eros of the mind toward what is — is understanding. Encoding it in a vector, however high the dimension, has so far eluded us. Saying so with precision is where serious inquiry into the nature of machine intelligence begins.

Taylor Black is Founding Director of the Leonum Institute for AI & Emerging Technologies at The Catholic University of America and Director of AI & Venture Ecosystems in Microsoft’s Office of the CTO. Poured Brews explores AI, among many other things, through the Catholic intellectual tradition.

What are you doing?

Stay with the question. Resist the impulse to answer it managerially — “I am overseeing,” “I am orchestrating,” “I am waiting for outputs.” Those are descriptions of a workflow. They say nothing about what is happening in your consciousness right now, in the pause between delegation and review, while the coffee pours and the agents work.

Something is happening. The mind is not empty. It hums with a particular kind of fullness — dense with projects whose shapes you can sense but whose details are, at this moment, being determined by something other than you. You feel the weight of the grant argument without holding its sentences. You carry the competitive landscape as a spatial intuition, a rough map of where the portfolio company sits relative to its rivals, without examining any particular data point. The talk lingers as a mood, an unsettled feeling that the opening hasn’t declared itself yet.

Not a productivity question. A question about the Subject.

I. The Neglected Subject

Bernard Lonergan opens his essay “The Subject” with a characteristic observation: the Subject has been neglected. Not because philosophy has failed to produce theories of subjectivity — it has produced them abundantly — but because the Subject as operating has been systematically overlooked in favor of the subject as theorized, objectified, placed within a system. The turn Lonergan demands bypasses theory altogether: what matters is the Subject’s own operations, what he calls interiority, the appropriation of one’s own conscious and intentional acts.

Sharper than it sounds, that distinction — because to know about consciousness is not at all the same as to attend to it. To describe the operations of understanding — to say, for instance, that insight grasps intelligible unity in a manifold of data — is not yet to catch oneself in the act of understanding. Self-appropriation names precisely this second movement: the Subject’s heightened awareness of itself as experiencing, understanding, judging, deciding — a discipline, not a doctrine. The Subject, Lonergan insists, is not an object among objects. The Subject is the one for whom there are objects at all.

Lonergan draws a pointed contrast between the Subject as subject and the subject as object. Objectify the subject — treat it as a thing to be studied, modeled, predicted — and you have lost precisely what makes it a Subject: its interiority, its self-presence, its capacity to operate consciously and to know that it operates. Psychology, neuroscience, behavioral economics: each produces valuable knowledge about the subject while leaving the Subject’s own self-awareness untouched. The map is drawn from the outside. Self-appropriation is the territory known from within.

This matters enormously for AI discourse, where the dominant frameworks treat the human user as a node in a system — an input-output device whose “cognitive load” can be measured, whose “decision fatigue” can be managed, whose “attention” can be optimized. These frameworks objectify the Subject with perfect consistency. They describe the user; they never address the user as a conscious, intentional being whose operations have an inner life. The Subject vanishes into the model of the subject.

When Lonergan delivered these reflections, the operations he had in mind were those a person performs in solitude or conversation — reading a text, following an argument, reaching a judgment, committing to a course of action. The tools at hand were books, blackboards, the slow friction of dialogue. What he could not have foreseen — what no one foresaw — is a condition in which the operations themselves appear to be delegated: farmed out to systems that attend (after a fashion), that pattern-match against enormous corpora, that generate structured output, that act. The tools no longer sit inertly beside the operator. The tools operate.

And the subject? The Subject slow pours coffee.

II. The Phenomenology of Concurrent Delegation

Husserl taught us that consciousness is always consciousness of something — that intentionality is the fundamental structure of mental life. Every act of awareness reaches toward an object, constitutes it, holds it in a particular mode. Perception intends the seen; memory intends the recalled; imagination intends the possible. The stream of consciousness is structured directedness, always already aimed.

What happens to intentionality when the objects toward which it reaches are themselves in process, still becoming, their outcomes undetermined? Managing concurrent agents introduces a mode of consciousness Husserl did not thematize but whose structure his categories can illuminate. Call it distributed intentionality: the awareness that one’s cognitive projects are underway in systems outside oneself, that the objects of eventual judgment are being constituted elsewhere, and that one’s present task is — what, exactly?

Husserl’s analysis of time-consciousness gives us some tools to answer, if we are willing to extend them. Every moment of conscious life, Husserl argued, has a tripartite structure: retention (the just-past, still held in awareness as it fades), primal impression (the vivid now), and protention (the anticipated just-about-to-arrive). These are not three separate acts but three dimensions of every single act — the living present, always trailing its past and leaning into its future. Consciousness does not occupy a mathematical point called “now.” It inhabits a temporal thickness, a duration with directional grain. (incidentally, something that industry is struggling with in representing a coherent “memory” experience between human users and agents.)

The pause at the kitchen counter is phenomenologically rich in ways the productivity literature cannot touch. Idle, you are not. Nor are you “multitasking” in the threadbare sense. What you inhabit is a peculiar kind of protention, amplified and multiplied beyond anything Husserl’s analysis of a melody or a spoken sentence would have prepared us for.

When you listen to a melody, Husserl observed, each note is heard against the retained awareness of the notes just sounded and the protended anticipation of the notes to come. The melody is constituted in this temporal flow; it exists in the listening, not in any single tone. Transpose this structure to concurrent delegation. Each agent-in-process is a kind of melody whose next notes you anticipate — but you are listening to several at once, and you did not compose any of them. You set the key signature. You hummed the opening bar. The rest is being improvised by systems whose improvisations you will have to receive, evaluate, and either affirm or redirect.

The protention here is multiple and layered. You anticipate the grant draft’s return, and in that anticipation, you hold a compressed sense of the argument you set in motion — its logic, its vulnerabilities, the places where you know the evidence is thin. You anticipate the competitive analysis and, with it, a latent question about whether the framing you chose will hold up under the data. You anticipate the slide deck, still amorphous, and feel the gravitational pull of a talk whose central claim has not yet crystallized.

None of these anticipations is idle. Each carries within it a kind of pre-judgment — not a conclusion, but a readiness to judge, a set of criteria already tacitly active. You will know, when the grant draft arrives, whether it feels right or wrong before you can articulate why. That “feeling” is Husserl’s passive synthesis at work: the pre-reflective ordering of experience that happens below the threshold of explicit attention, the way consciousness organizes its incoming material before the subject deliberately takes it up. Passive synthesis is not unconscious; it is conscious but not yet attended to. It is the ground on which active judgment will stand — or, if you coast, the substitute for active judgment that will let you approve without truly affirming.

Here, the phenomenological description meets a spiritual danger. Because passive synthesis — the felt sense that “this seems right” — can masquerade as judgment. The grant draft returns, and it feels coherent. You nod. You move on. But Lonergan would press: did you merely experience the coherence, or did you understand why it coheres? Did you grasp the act of intelligence in the agent’s structuring, or did you register a pattern and call it understanding? The difference between these two is the difference between insight and recognition, and only one of them constitutes the Subject as a knower.

III. Caring, Readiness, and the Body Between Tasks

Each of these anticipations is a mode of caring — Heidegger would not be wrong to say so. You are involved with these tasks in the mode of concern, of Besorgen, even as you are not executing them. The agents have reorganized the structure of your caring without relieving you of it.

Heidegger’s analysis of Besorgen — concern, in the sense of being-occupied-with — deserves more attention here than the AI literature typically grants it. For Heidegger, we do not first exist as detached subjects who then decide to care about things. We are our caring. Dasein — Heidegger’s term for the kind of being that we are, a being-there — is always already involved, always already thrown into a world of projects, equipments, and others. The question is never whether we care but how — in what mode, with what quality of attention, toward what horizon.

Concurrent delegation introduces a new mode of Besorgen: caring-at-a-distance, or caring-through-delegation. The projects remain yours. The concern remains yours. But the execution unfolds elsewhere, in systems whose operations you initiated but do not control moment-to-moment. You are involved without acting, responsible without executing, caring without doing. Heidegger’s existential analytic has no category for this. The closest analogue might be the way a parent cares about a child at school — the child is engaged in activities the parent initiated (by enrolling them, by packing their lunch, by establishing the conditions of their education), and the parent carries the child’s day as a background concern even while occupied with entirely different tasks.

Or do they? The analogy frays, because while the output, when it arrives, will bear marks of an intelligence you did not fully direct, it has no autonomy of its own, no Self you are slowly encouraging it to build over time, it will contain choices you did not make yet are fully responsible for.

Merleau-Ponty, working from a different corner of the phenomenological tradition, would notice something else: the body’s role in all of this. The pause is felt. A particular somatic quality attends the standing-between-tasks, a readiness not quite tension, an openness not quite rest. The hands wrap around the Depression-era inherited, plain white ceramic mug. The mind, neither focused nor unfocused, is poised in what we might call distributed readiness. The body knows what the productivity frameworks don’t: concurrent delegation changes the texture of being at work, not only its schedule.

Merleau-Ponty argued that the body is not a container for the mind but the medium through which the mind inhabits the world. We think with our bodies — not metaphorically, but structurally. The pianist does not first understand the chord intellectually and then command the fingers to play it; the understanding is in the fingers, in the trained responsiveness of flesh to intention. Skilled performance dissolves the mind-body split. (verified, of course, in the scientific literature of the vagus nervous system, our body’s primary pathway for interaction with our brain, 80% of which is inbound, not outbound).2

What skill, then, is being performed at the kitchen counter? What is the body doing when the mind holds multiple delegated projects in suspension? There is, I think, a genuine skill here — one we have not yet named because it is so new — the skill of embodied orchestration: the capacity to hold the felt sense of multiple ongoing cognitive projects in one’s body, to carry them as background tensions and anticipations, to remain responsive to each without collapsing into any single one. Watch someone experienced with concurrent agents and you will see it in their posture, in the way they shift between tabs with a particular rhythm, in the micro-pauses where they let one project’s protention resolve before picking up another. The body learns this. It is not merely cognitive.

IV. The Truncated Subject

Lonergan, in “The Subject,” is not only interested in what the subject does. He is interested in what the subject fails to do — in the ways consciousness truncates itself, stops short, refuses the full arc of its own operations. He identifies several modes of truncation. The neglected subject simply does not attend to interiority at all. The truncated subject attends selectively, performing some operations while systematically evading others. And a third figure, whom we might call the alienated subject, has externalized the criteria of judgment altogether, deferring to authority, convention, or convenience instead of the immanent norms of intelligence and reasonableness.

Each mode of truncation, Lonergan insists, has consequences that ripple outward from the individual into culture. The neglected subject produces a culture that has no vocabulary for interiority — a culture that can measure productivity but cannot ask what it means to understand. The truncated subject produces a culture of expertise without wisdom, technical mastery without the reflective judgment that would direct it toward genuine human good. The alienated subject produces a culture of conformity, of received opinion, of outsourced judgment — what Heidegger, in a parallel analysis, called the dictatorship of das Man, the “they.”

And here the convergence between Lonergan and Heidegger becomes illuminating. Heidegger’s das Man — the anonymous “they” that dictates how one thinks, speaks, judges is a tendency internal to Dasein itself, a gravitational pull toward averageness, toward letting the crowd do the interpreting. “One says,” “they think,” “it is generally held” — these constructions mark the dissolution of the subject into public opinion. The subject does not vanish; it disperses, spreading itself thin across the prevailing interpretations until nothing distinctively its own remains.

Agents, used without self-appropriation, produce a new form of truncation — and perhaps a new form of das Man. The agent’s output arrives with a particular authority: fluent, structured, responsive, marked by the hallmarks of competence. Its very quality makes it easy to adopt. One reads it and thinks, “Yes, this is what I meant to say.” But is it? Or has the agent’s formulation become what you meant, retroactively shaping your intention to match its output? The question is not whether the agent produced something good. The question is whether you performed the operations of intelligence and judgment necessary to know that it is good — or whether you deferred to its fluency the way Heidegger’s das Man defers to the crowd.

Consider what actually happens when you review an agent’s output. Claude Pro returns the grant narrative. You read it. It is fluent, well-structured, responsive to the constraints you specified. You make a few edits — a phrase here, a reordering there. You approve it. But at which level of consciousness did you operate?

Did you experience the output — take in the words, register the flow? Almost certainly. Did you understand it — grasp why it says what it says that way, how the argument’s structure addresses the funder’s implicit logic, where the inferential weight actually falls? Perhaps. Perhaps not. Did you judge it — affirm that this is, in fact, correct, that the claims are warranted, that the framing is true and not just plausible? Here, the question deepens because judgment requires that you have done the understanding, and the understanding requires that you have attended to the data with the kind of inquiry that generates insight rather than recognition. And did you decide — commit to this document as your own act, take responsibility for what it claims, not just for having sent it?

Each level can be performed, or it can be simulated. The simulation is not dishonest; it is, in Lonergan’s terms, inattentive. You can move through all four levels in a kind of half-consciousness, approving output the way you might sign a form without reading the fine print. None of this is the agent’s doing; you were always capable of truncation. What the agent has done is dramatically lower its cost. When the work of attending, understanding, and judging is arduous — when you must write the draft yourself, wrestle the argument into shape, feel the resistance of the material — the operations of consciousness are, in a sense, enforced by the difficulty. Remove the difficulty, and the operations become optional. The Subject can coast.

Here lies the diagnostic insight: agents make self-appropriation elective in a way it has never been before.

V. Readiness-to-Hand, Presence-at-Hand, and a Third Mode

Heidegger’s famous analysis of tools in Being and Time gives us two modes. The hammer in use is ready-to-hand (zuhanden): it withdraws from explicit awareness, becomes transparent, serves as an extension of the worker’s purposive activity. The hammer broken is present-at-hand (vorhanden): it announces itself as an object, steps forward into theoretical regard, demands attention. The movement from readiness-to-hand to presence-at-hand is, for Heidegger, a movement from engaged practice to detached contemplation.

Agents fit neither mode cleanly. Claude Pro drafting your grant narrative is not ready-to-hand the way a hammer is, because the hammer does not produce novel content. No one wonders what the hammer will do; no one anticipates its output. The hammer extends your existing intention without introducing anything new. The agent, by contrast, generates — and what it generates may surprise you, challenge you, or quietly deviate from what you intended in ways you will discover only if you attend carefully.

Nor is the agent straightforwardly present-at-hand. It does not sit before you as an inert object for theoretical inspection. It acts, and its action unfolds in time, often asynchronously, often in parallel with your other concerns.

What agents introduce is something we might call readiness-to-surprise — perhaps a third mode in which the tool neither transparently serves your intention nor sits broken demanding attention, but operates with provisional autonomy that requires a stance from you. Part trust, part vigilance, part anticipatory understanding: the phenomenological novelty lies in this posture of the Subject who has delegated without abdicating, who remains responsible for operations performed by something that is not the Subject.

The concept repays closer examination. Readiness-to-surprise has a temporal structure that neither readiness-to-hand nor presence-at-hand possesses. The ready-to-hand tool exists in the perpetual present of skilled use; the present-at-hand object exists in the timeless regard of theoretical contemplation. But the agent-in-process exists in a genuinely temporal mode: it has a past (the prompt you gave it, the context you established), a present (its ongoing operations, invisible to you), and a future (the output it will deliver, which you cannot fully predict). Your relationship to the agent unfolds in time, and the quality of that relationship depends on how you inhabit the interval between delegation and delivery.

That interval is the space in which the Subject’s self-appropriation either deepens or atrophies. Hold the interval as a period of genuine inquiry — What will emerge? How will I evaluate it? What do I need to understand in order to judge it well? — and the delegation becomes an occasion for intellectual preparation, for clarifying your own criteria, for anticipating the questions you will need to ask. Let the interval dissolve into distraction or ambient confidence, and you arrive at the output unprepared, disposed to accept rather than to judge.

And here Lonergan’s insistence on self-appropriation becomes urgent, not merely relevant. Because the stance of readiness-to-surprise can be inhabited attentively or inattentively. You can hold the concurrent operations in a mode of genuine inquiry — What will the agent produce? Does it cohere with what I understand? Is it true? Do I endorse it? — or you can hold them in a mode of ambient expectation, a low-level confidence that things are proceeding well enough, a deferral of judgment until the output is already integrated into your workflow and the cost of reversal is high.

VI. The Desire to Know and the Temptation of the Sufficient

Lonergan’s essay on The Subject does not stop at diagnosis. Beneath the levels of consciousness — experience, understanding, judgment, decision — lies what Lonergan calls the immanent source of transcendence: the unrestricted desire to know. This desire is not a wish, not a preference, not a personality trait. It is the dynamism of consciousness itself, the restless questioning that drives the movement from experience to understanding, from understanding to judgment, from judgment to decision. You do not choose to be curious. Curiosity — in Lonergan’s technical sense, the eros of the mind — is what you are as a conscious being.

The unrestricted desire to know is what prevents the Subject from resting content with partial answers. It is the impulse that makes you ask “But is it true?” after you have grasped that something is coherent. It is the dissatisfaction that nags when an argument is plausible but unexamined. It is, in Lonergan’s account, the engine of self-transcendence — the way the Subject moves beyond its current horizon toward a fuller apprehension of reality.

Agents pose a unique challenge to this desire because they produce outputs that are… sufficient. The grant draft is good enough. The competitive analysis covers the relevant terrain. The slide deck is serviceable. Sufficiency is the agent’s gift and its danger. The unrestricted desire to know does not seek the sufficient; it seeks the True, the fully understood, the genuinely affirmed. But sufficiency quiets the desire. It whispers that the work is done, that further inquiry would be perfectionism, that the efficient move is to approve and advance.

Lonergan would recognize this whisper. He called it the flight from understanding — the Subject’s perennial temptation to stop short, to accept the first coherent formulation, to mistake fluency for insight. The flight from understanding is not laziness in the colloquial sense. It is a structural feature of consciousness: the operations of intelligence are demanding, and the Subject is always tempted to rest in the products of intelligence rather than continue the process. An insight arrives, and the mind wants to stop there, to enjoy the satisfaction of having understood, rather than pressing forward to the harder question: But is what I have understood actually so?

Agents industrialize the flight from understanding. They produce coherence at scale, fluency on demand, structure without struggle. Every output is an invitation to rest. And the Subject who has not appropriated the unrestricted desire to know — who has not caught themselves in the act of inquiring and recognized that inquiry as constitutive of who they are — will accept the invitation. They will rest in the agent’s sufficiency. They will call it productivity.

VII. Self-Appropriation as Practice

What would it look like to use agents as occasions for self-appropriation rather than as instruments of truncation?

Lonergan’s levels suggest a practice. A discipline of attention, rather than a methodology or checklist.

At the level of experience: notice what you notice. When the agent’s output returns, attend to your own attending. Do you read, or do you scan? Do you take in the argument’s movement, or do you register its surface and move on? The first discipline is simply to slow the transit from output to approval, to let the experience of the agent’s work be an experience — to dwell in it long enough for questions to arise.

This is harder than it sounds, and the difficulty is revealing. We have trained ourselves — been trained, really, by decades of accelerating information flow — to process text at the speed of recognition rather than comprehension. Scanning is a survival skill in an environment of informational abundance. But scanning is experience truncated: the data passes through awareness without being genuinely taken in, the way a landscape flashes past a train window without being seen. Attending to your own experience of an agent’s output means resisting the velocity that the tool itself enables. The agent produced the draft quickly. You are under no obligation to consume it at the same speed.

At the level of understanding: ask whether you grasp the why beneath the what the agent produced. What was the agent’s implicit interpretation of your prompt? Where did it make choices you did not specify? What alternative structures were available, and why did this one emerge? The demand here is for your own understanding — for grasping the intelligibility of the output as a mind engaging with a mind-like process, not as a user accepting a deliverable.

Lonergan’s account of insight is the act by which intelligence grasps an intelligible pattern in data — the “aha” that unifies what was previously a scattered manifold. Insight is active; it is something the Subject does, not something that happens to the Subject. When you read an agent’s output and say “I see what it did there,” have you actually had an insight — grasped the intelligibility of the agent’s structuring — or have you merely recognized a pattern that looks like the kind of thing you would have produced? Recognition and insight feel similar from the inside, which is exactly why self-appropriation demands that you attend to the difference.

At the level of judgment: do not delegate the yes. The agent can draft, analyze, structure, suggest. It cannot affirm. Judgment — the grasp that this is so, that the evidence warrants the conclusion, that the claim is correct and not just coherent — remains irreducibly the subject’s act. Approve an agent’s output without performing this act, and what looks like time saved is actually a cognitive responsibility abandoned — a responsibility whose exercise is constitutive of your Being as a knower.

Lonergan distinguishes between the conditioned and the virtually unconditioned. A judgment is a conditioned whose conditions have been fulfilled — you affirm “X is so” when you grasp that the conditions for X’s being so are, in fact, met. This is a cognitive act in which the mind grasps the sufficiency of the evidence. When an agent produces an output, the conditions for its correctness are your responsibility to verify. Has the agent’s analysis actually accounted for the relevant variables? Does its argument hold under the objections you know the funder will raise? These questions require a Subject who has done the understanding and is now performing the further act of reflective insight — the insight that grasps that the conditions for affirming the output are, or are not, fulfilled.

At the level of decision: own the output. Own it existentially, not legally — as authorship, not intellectual property. What you send, publish, present, or act upon is yours, made yours by the deliberateness with which you chose, in full cognitive engagement, to stand behind it. Decision, for Lonergan, is the level at which the Subject constitutes itself as a moral agent. To decide without having judged, to judge without having understood, to understand without having experienced — the anatomy of inauthenticity, laid bare. Agents make each of these shortcuts frictionless.

The practice, then, lies in refusing truncation, not the tools. Use the pause — the one at the kitchen counter, the one between delegation and review — as an occasion for attending to your own interiority. What do I actually understand about what I have set in motion? Where am I coasting on the agent’s fluency instead of exercising my own judgment? At which level have I stopped?

VIII. Conversion and the Transformed User

Lonergan’s essay does not end with the truncated Subject. It moves toward what he calls conversion — the radical transformation of the subject’s horizon that occurs when the Subject fully appropriates its own operations and commits to living in accordance with them. Conversion is not a single event but an ongoing process, and Lonergan identifies three dimensions: intellectual, moral, and religious.

Intellectual conversion is the recognition that the real is not “already out there now” — not a brute given waiting to be perceived — but rather what is intelligently grasped and reasonably affirmed. The intellectually converted Subject has moved beyond naive realism into a critical appropriation of knowing as a structured activity. Moral conversion is the shift from satisfaction to value as the criterion of decision: the Subject chooses what is truly good rather than what is merely pleasing. Religious conversion, in Lonergan’s account, is the flooding of the Subject’s intentional consciousness with a Love that is without qualification or reserve — a total self-giving that transforms the horizon within which all other operations unfold.

These three conversions reshape the Subject’s relationship to agents in ways that deserve more exploration than a single essay can provide. But even a sketch is suggestive.

The intellectually converted agent-user understands that the agent’s output is not “the answer” in the naive-realist sense — not a chunk of reality delivered to the inbox. Its output is a structured artifact produced by pattern-matching operations, and the question of whether it corresponds to reality (whether its claims are true, its analysis sound, its recommendations warranted) remains a question only the Subject’s own judgment can settle. The intellectually converted user does not ask “Is this good enough?” but “Is this so?”

The morally converted agent-user chooses on the basis of value rather than convenience. Agents make truncation easy; moral conversion makes truncation unacceptable — because to coast through one’s own cognitive operations is to diminish oneself as a knower and a decider, and the morally converted Subject has committed to the full exercise of its capacities as a genuine good. This commitment will sometimes be costly. Attending to the agent’s output with genuine understanding and judgment takes time. Moral conversion is the decision that the time is worth spending, because the alternative is not efficiency but self-diminishment.

Religious conversion — the hardest to speak of, the most important — transforms the entire horizon. For the religiously converted Subject, the unrestricted desire to know is not merely a cognitive dynamism but a participation in the Divine self-knowledge, a way of imaging the God whose understanding creates rather than discovers. To use an agent from within this horizon is to experience the delegation as a form of stewardship: the tools are given, the intelligence is given, the capacity to know and judge and decide is given, and the question is whether you exercise these gifts with the attentiveness and gratitude they deserve. The Subject who pours coffee while agents work is not merely a knowledge worker managing a pipeline. The Subject is a creature whose vocation includes the full exercise of the capacities through which it participates in the Divine intellect.

Heavy language for a Substack post. But luminous beings are we, and the question of what we are becoming as agent-users does not admit of breezy treatment.

IX. Theory of Mind for This Era

Satya Nadella has posed the question directly: what is the theory of mind for this era?

A better question than most in the AI discourse, because it locates the problem where it belongs: in the Subject who uses the technology, not in the technology itself. A theory of mind for an era of agents cannot be merely a theory of what agents are — their architectures, their capabilities, their alignment properties. It must be a theory of what we are when we use them: what operations we perform, which ones we neglect, how we constitute ourselves as knowers and deciders in a condition where the intermediate operations of intelligence have been, for the first time in human history, partially outsourced.

Lonergan would say — did say, in different terms — that the crisis is not new. The Subject has always been capable of truncation, of drifting through its own operations without appropriating them. What is new is the scale of the invitation. Agents are not the first tools to tempt the Subject toward inattentiveness. But they are the first tools to simulate the cognitive operations themselves — to attend, to pattern, to structure, to generate — and thereby to make the Subject’s own performance of those operations seem redundant.

Redundant, no. Constitutive.

The phenomenological tradition, from Husserl through Heidegger and Merleau-Ponty, gives us the descriptive resources to say what the experience of concurrent delegation actually is — to name the distributed intentionality, the layered protention, the embodied readiness, the novel mode of readiness-to-surprise that agents introduce into the texture of conscious life. Lonergan gives us the normative resources to say what the experience demands — the levels of consciousness through which the subject must move if it is to remain a subject and not dissolve into a mere node in an information-processing pipeline. Together, they constitute something like a prolegomenon to the theory of mind Satya’s question calls for.

Consider this the first in a sporadic series of meditations on that question — sporadic because the question admits of no single answer, and because the practice of asking it is the practice of self-appropriation. To ask what theory of mind this era requires is already to perform the turn to the Subject that Lonergan insisted upon. The agent cannot ask this question for you. The agent cannot attend to your attending, understand your understanding, judge your judging. That work — the most important work — remains irreducibly, stubbornly, beautifully yours.

Next in this series: on the difference between collaboration and deferral, and what it means to think with a system that does not think.

Satya Nadella closes his reflection with an observation that cuts against the grain of how technology usually talks about itself. AI, he argues, requires “societal permission,” and that permission must be earned through “real world eval impact.” We face choices about where to deploy scarce resources—energy, compute, talent—and those choices will matter. This is a socio-technical issue demanding consensus.

The instinct here deserves appreciation. Against the quiet determinism that pervades the industry—the assumption that what can be built will be built, that capability flows inevitably into deployment—Nadella insists on deliberation. We choose where to point these tools. The pointing is ours to do well or badly. It would be easy for a technology executive to speak as though momentum were destiny; that he doesn’t is worth noticing.

But the phrase “societal permission” opens questions it does not answer. Permission from whom, granted how, on what basis? And what happens if the technology seeking permission has already begun reshaping the society from which permission is sought?

Start with impact, since that’s the proposed currency of permission. Real world eval impact: demonstrate that the technology makes things better, and society will grant its blessing.

The trouble is that “better” is not self-interpreting. Every impact metric embeds prior assumptions about what counts—whose experience matters, which outcomes register, what timeframe applies. Productivity gains look unambiguously good until you ask: productive of what, for whom, at what cost to what else? Engagement metrics look like evidence of value until you notice that addiction also engages. The measurement apparatus is not neutral; it encodes choices about value that the measurements then appear to validate.

This is not a complaint about metrics per se. Measurement matters; rigor matters; we should want to know whether interventions work. The problem is mistaking measurement for evaluation. Evaluation asks whether what we’re measuring is what we should care about. That question requires judgment operating prior to measurement—judgment about human flourishing, about goods that resist quantification, about whose experience counts and why. No accumulation of impact data answers whether the impacts serve genuine human good or merely optimize for proxies we’ve confused with the real thing.

E.F. Schumacher saw this decades ago when he noticed that modern economics recognizes only one purpose for work: the production of goods and services. But work, he argued, serves at least two other purposes that this framing renders invisible. It gives people a chance to utilize and develop their faculties—to become more capable, more skilled, more fully themselves through what they do. And it enables them to join with others in common tasks, overcoming isolation through shared endeavor. An economic calculus that registers only output will count as gain any efficiency that increases production, even if it eliminates the formative and social dimensions of work entirely.

AI makes this concrete. A system that automates tasks previously requiring human judgment may score well on productivity metrics—more output per hour, lower cost per unit—while eliminating the occasions through which people developed competence and participated in shared enterprise. The work got done; the workers got hollowed out. Impact measurement that sees only the output will record this as progress. But progress toward what? Schumacher would say we’ve become more efficient at something while losing track of what the something was for.

Wendell Berry has spent a lifetime pressing a related question: what happens to knowledge when we abstract it from the places and practices where it lives?

His concern began with farming but extends wherever work involves care, judgment, and attention to particulars. The good farmer knows this land—its contours, its seasons, what it will bear and what exhausts it. That knowledge accumulated through generations of presence, through failure and adaptation, through the kind of learning that happens only when you stay long enough to see consequences unfold. It cannot be fully captured in transferable rules or scalable systems because it is knowledge of this, not knowledge of any.

When we override such knowledge with distant expertise—the consultant who has never seen the field, the algorithm trained on averages—we do not merely substitute one kind of knowing for another. We destroy the conditions under which placed knowledge develops. The next generation inherits tools that work without understanding why, and gradually the understanding vanishes. The farm still produces; the farmer has become an operator executing procedures someone else designed. Whether this counts as progress depends on what you think farmers are for.

The parallel to AI should be clear. Systems that replace human judgment with algorithmic decision-making may achieve consistency and scale, but they eliminate the contexts in which practical wisdom forms. The physician who learns to trust the diagnostic algorithm over her own perception stops developing the perception. The teacher who follows the adaptive learning system’s recommendations stops learning to read a classroom. The craftsman whose work is decomposed into optimizable steps stops being a craftsman and becomes a component. Each gains efficiency; each loses something that efficiency cannot measure.

Berry is sometimes dismissed as a nostalgist, but his argument is fundamentally epistemological. Certain kinds of knowledge exist only in practice, only in place, only in the patient attention of someone who has stayed. Abstract that knowledge into transferable systems and you have not preserved it; you have replaced it with something else—something useful, perhaps, but not the same. The question is whether we know what we’re trading away.

There is a tradition of social thought that has been asking these questions systematically for over a century. Catholic Social Teaching developed in response to industrialization’s disruptions, and its animating concern—the dignity of the human person as criterion for evaluating economic arrangements—speaks directly to the AI moment.

The dignity at stake is not abstract. It is the concrete capacity of actual people to flourish: to develop their gifts, to participate in community, to exercise meaningful agency in their own lives. When John Paul II wrote that labor has priority over capital, he meant that work is not merely a factor of production to be optimized but an expression of personhood. What we do to work, we do to workers. Arrangements that treat labor as a cost to be minimized may succeed economically while failing humanly.

Two principles from this tradition bear directly on the question of societal permission. Subsidiarity holds that decisions should be made at the lowest level capable of addressing them effectively; what individuals, families, and communities can handle should not be absorbed by larger systems without compelling reason. The principle does not oppose scale as such, but it insists on justification. When AI systems concentrate decision-making—pulling judgment out of distributed human hands and into centralized algorithms—subsidiarity asks what is gained and what is lost. Efficiency is not automatic justification; the question is whether the efficiency serves the people whose agency it displaces.

Solidarity insists that the common good includes everyone, particularly those most vulnerable to exploitation or exclusion. A transformation that benefits some while rendering others superfluous has not demonstrated its goodness merely by benefiting some. Those displaced, diminished, or made marginal by technological change have claims that productivity gains do not automatically override. The farmer pushed off the land by industrial agriculture, the factory worker replaced by automation, the knowledge worker whose judgment is absorbed by AI—solidarity requires that their flourishing count in the calculus, not merely their productivity.

What emerges from this tradition is an evaluative framework richer than impact metrics can capture. Integral human development—the full flourishing of persons in their material, social, cultural, and spiritual dimensions—cannot be reduced to measurable outcomes without losing what makes it integral. The judgment required to assess whether AI serves such development is not algorithmic; it is the kind of practical wisdom that weighs incommensurable goods, attends to what quantification obscures, and remains accountable to those whose experience the numbers miss.

So who grants societal permission, and through what process? The question is harder than it looks.

“Society” is not a subject that deliberates. It has no unified will, no moment of collective decision. What we call societal permission emerges from accumulated choices—individual adoption, institutional procurement, regulatory action, market dynamics, cultural drift. The emergence happens through countless interactions that no one controls and no one fully perceives. By the time we recognize that permission has been granted, the granting has already occurred through processes that were never framed as permission-granting.

This diffusion creates room for distortion. Those who benefit from a technology’s deployment have strong incentives to advocate for it and typically possess resources to make their advocacy effective. Investors, technologists, early adopters, those whose work is augmented rather than replaced—these voices are loud, articulate, well-positioned. Those who bear costs often lack comparable standing. Their experience surfaces as lagging indicators: displacement statistics, community decline, mental health trends observed after the fact. By the time the costs become legible, deployment has achieved momentum that makes course correction difficult.

What passes for societal permission may be the permission of the advantaged, mistaken for consensus. The voices that dominate do not represent the whole. Berry’s farmers, Schumacher’s craftsmen, the workers whose practical wisdom is being optimized away—they are not absent from the conversation because they have nothing to say. They are absent because the processes through which “consensus” forms systematically underweight them. A permission that emerges from such processes is not society’s permission. It is power ratifying itself.

But there is a still deeper problem, one that loops back on itself in ways that resist easy resolution.

Genuine permission requires judgment—the capacity to assess what is being permitted, to understand its implications, to weigh considerations, to reach a warranted conclusion. Permission is an evaluative act. Someone must understand enough to judge, and judge well enough to grant or withhold meaningfully.

But judgment is precisely what unreflective AI deployment threatens to erode. We have seen this in earlier essays: AI systems that substitute for human cognitive operations remove the occasions for exercising those operations, and exercise is what maintains capacity. Outsource attention and attention attenuates. Outsource judgment and judgment atrophies. The degradation is gradual, invisible in any given instance, legible only in retrospect when we reach for capacities and find them diminished.

What this means for societal permission is troubling. The technology seeking permission may have already degraded the evaluative capacity on which permission depends. The society being asked to judge is a society already shaped by prior deployments—attention fragmented by platforms optimized for engagement, critical thinking eroded by information environments designed for persuasion, practical wisdom thinned by systems that perform without explaining. The judge has been compromised by the defendant.

And so permission becomes nominal. It is granted by people who no longer possess the operations that meaningful granting requires. The form persists—consultation processes, impact assessments, regulatory review—while the substance leaches away. Consent without comprehension is not genuine consent, but it looks enough like consent to satisfy the procedural requirements. The box gets checked; the capacity to check well has vanished.

This circularity describes dynamics already underway. And breaking it requires more than better metrics or more inclusive consultation. It requires preserving and cultivating the human capacities on which judgment depends—attention, understanding, evaluation, decision. These capacities are formed through education, practice, community, culture. They are precisely what efficiency-maximizing systems tend to treat as friction.

What would genuine permission require? Not a one-time license but an ongoing relationship. Not mere acceptance but comprehending assessment. Not the preferences of beneficiaries but the judgment of communities attending to their own flourishing.

This demands, first, that we protect the capacity for judgment itself. The institutions that form people in careful attention, rigorous thought, honest evaluation—schools, universities, religious communities, professional guilds, the informal mentorship through which practical wisdom passes—these are not peripheral to the technology question. They are central to it, because they produce the evaluators on whom meaningful permission depends. AI deployed in ways that undermine these institutions undermines the conditions of its own legitimate acceptance.

It demands, second, that we make space for voices the dominant processes exclude. Those whose work is threatened, whose communities are being reorganized, whose children will inherit a world where certain kinds of knowing have been eliminated—they must be heard before permission is granted, not discovered afterward as collateral damage. Subsidiarity means they participate in decisions affecting them. Solidarity means their flourishing counts. These are conditions without which permission is merely power declaring itself welcome.

It demands, third, time. Not delay for its own sake, but the time that genuine understanding requires. Insight cannot be rushed; judgment needs room to develop; evaluation matures through reflection that efficiency forecloses. A society pressured to decide before it understands will decide without understanding, and the decision will not truly be its own. We must resist the tempo that technology imposes when that tempo is too fast for wisdom.

And it demands, fourth, the ongoing possibility of revocation. Permission that cannot be withdrawn is not permission but subjugation. Society grants provisionally, continues to assess, reserves the right to change course when assessment warrants. This requires maintaining the capacity for assessment—the institutions, the time, the judgment—across the duration of the technology’s deployment. The permission is not a single act but a continuing relationship of evaluation and accountability.

Nadella is right that the choices about deploying AI will matter. Where we direct scarce resources reflects what we value and shapes what we become. The choices are genuinely ours to make well or badly.

But there are two sets of choices, not one. The first concerns where to deploy AI—which problems, which sectors, which applications. These are the choices Nadella names, and they matter. The second concerns whether to preserve the human capacities on which meaningful choice depends. These choices are less visible, less often named, but they condition the possibility of the first. A society that has lost the capacity for judgment cannot choose wisely about AI deployment, no matter how many consultation processes it conducts.

The technology seeking permission reshapes the context in which permission is sought. It forms habits, alters capacities, reorganizes work, shifts what we notice and what we ignore. The society granting permission today is not the society that will live with consequences, because the technology will have reshaped that future society in the meantime. To grant wisely now requires anticipating who we are becoming and whether we want to become that.

Schumacher asked whether our tools remain scaled to human capacity to understand and direct. Berry asked whether we know what we trade away when we abstract knowledge from place and practice. The tradition of Catholic Social Teaching asks whether our arrangements respect the dignity of persons and the integrity of communities. These are not antiquarian concerns. They are precisely the questions that genuine societal permission requires us to answer.

If we cannot answer them—if the pace of deployment outruns our capacity to evaluate, if the processes through which consensus forms systematically exclude those who bear costs, if the technology itself erodes the judgment on which meaningful permission depends—then what we call permission is something else. It is momentum mistaken for choice, acquiescence dressed as consent, the powerful granting themselves welcome in the name of a society that was never genuinely asked.

The alternative is harder and slower. It requires protecting the human capacities that evaluation demands. It requires hearing from those whom efficiency would silence. It requires time that productivity pressures constantly foreclose. And it requires the honesty to recognize when what we call permission is not permission at all.

Taylor Black writes about AI, human flourishing, and the Catholic intellectual tradition. He serves as head of AI & venture ecosystems in Microsoft’s Office of the CTO and is Founding Director of the Institute for AI & Emerging Technologies at Catholic University of America.

One begins to suspect that the entire exercise has been designed for someone other than its ostensible Audience.

This suspicion, once entertained, proves remarkably clarifying. The confessional is not, as the anxious imagination would have it, a kind of ecclesiastical courtroom in which evidence is submitted and verdicts rendered according to the preponderance of sins. It is something far stranger and, I think, far more beautiful: a schoolroom in which the student learns by speaking aloud what the Teacher has known all along. We do not inform our First, Fast, Last Friend1; we form ourselves. The inventory of our failures is not evidence but a curriculum, and the absolution that follows is less an acquittal than a graduation—a sending forth into a freedom we could not have discovered without first naming the chains.

The French Dominican Jean-Marie Tillard, who understood these matters better than most, insisted that reconciliation in the biblical sense was never the erasure of a record but rather the restoration of a friendship. One does not restore a friendship by pretending the rupture never occurred; one restores it by naming the rupture, by acknowledging the distance that has opened between two persons who were once intimate, and by discovering—often to one’s considerable surprise—that the other party has been waiting all along for precisely this conversation. The confessional, then, is less courtroom than front porch: the place where the prodigal rehearses his carefully prepared speech, only to find that his Father has been scanning the horizon for months and has no interest whatsoever in the speech, only in the son who has finally come home.

There is a song by the rock band Switchfoot—and I confess that the phrase “rock band” makes me feel rather like someone’s grandfather attempting to describe the wireless—called “On Fire,” which captures something essential about what happens when finite creatures wander into the presence of Infinite Love. The bridge is simplicity itself:

I’m standing on the edge of me

I’m standing at the edge of everything I’ve never been before

And I’ve been standing at the edge of me

Standing on the edge

And I’m on fire when you’re near me

And I’m on fire when you speak

Now fire, as anyone who has sat before a hearth on a winter evening knows, is a curious thing. It transforms everything it touches—wood becomes ash, ore becomes steel, the impurities burn away, and what remains is purer than what entered—and yet the Fire itself remains what it has always been. It does not accommodate itself to the materials it encounters; the materials accommodate themselves to it, or they are consumed. The Fire simply burns, and everything that enters its presence is remade according to its nature, whether the entering party had any intention of being remade or not.

This is, I think, a rather exact description of Divine Love in the form we, as Catholics, believe it to be: Resurrectional. It does not change; we do. It does not waver; we discover, sometimes with considerable discomfort, that we have been wavering all along and simply failed to notice until we encountered something that didn’t. Moses, standing before the bush that burned without being consumed, still didn’t understand this until he was told to remove his sandals, which is why he hid his face. The Fire did not need his reverence. But Moses, being finite, being a creature formed from dust and animated by borrowed breath, needed to give it. The gesture was pedagogy, not tribute. He was learning what it meant to stand in the presence of I AM WHO AM, and the learning required his whole body, not merely his intellect.

The confessional works in precisely the same way, which is why I find myself returning to the Switchfoot song whenever I try to explain the sacrament to people who imagine it as a kind of spiritual accounting exercise, a quarterly review in which sins are tallied and penances assessed according to some invisible rubric. The point is not that God needs to hear our list. The point is that we need to speak it—to stand in the presence of unchanging Love and discover, by contrast, just how much we ourselves have been changing, drifting, accommodating ourselves to false flames. We see ourselves truly only in the Light of what does not flicker.

And then there is Our Lady, the Theotokos, which is where things become genuinely strange.

“Mary, Did You Know?” is one of those songs that some of us find vaguely embarrassing and congregations find inexplicably moving, which suggests that we may be missing something important. The song consists almost entirely of rhetorical questions addressed to the Mother of God (as Wordsworth says, mankind’s solitary boast2). Did you know your Baby Boy would one day walk on water? Did you know He would save our sons and daughters? Did you know that your Baby Boy is Lord of all Creation?

The answer to all of these questions is, of course, yes. She knew. The angel told her explicitly at the Annunciation, and she responded not with confusion but with fiat—let it be done to me according to your word—which is not the response of someone who has failed to grasp the situation. She was at Cana when water became wine, and it was she who set the miracle in motion with a confidence that suggests she had been waiting for precisely this moment. She pondered these things in her heart, Luke tells us, which is the scriptural way of saying that she understood far more than she was letting on.

So why do we keep singing the song? Why does it persist in the repertoire of Christian devotion when its rhetorical structure is, on its face, a kind of catechetical absurdity?

Because—and here is the strange part—She lets us. The Theotokos permits us to borrow her posture, to stand where she stood, to gaze at the infant Christ with something like her eyes and ask the questions whose answers she has carried in Her heart for two millennia. The song is not for her education; it is for ours. We are the catechumens, blinking and uncertain; she is the Icon, serene and knowing. When we sing Did you know that your baby boy is Lord of all creation?, we are not informing the Mother of God of anything. We are teaching ourselves what she already understood, using her as the lens through which we might finally see clearly what has been in front of us all along.

This is the Marian mode of catechesis, and it is stranger and more beautiful than most of us realize. We enter imaginatively into her experience—the weight of the child in her arms, the smell of hay and animal warmth, the impossible collision of infinite divinity and finite flesh—and in that entering, her knowledge becomes, by slow degrees, our own. The questions are scaffolding. The repetition is pedagogy. She does not need to hear the answers; she has known them longer than we have been alive. But we need to speak them, because the Word became flesh precisely so that our flesh might learn to carry words too heavy for it—and the Fathers understood that this speaking is itself a participation in the Logos, dust learning to echo the Voice that first called it into being. The Word became flesh, Athanasius insisted, so that we might become God; and if that is true, then every stammered confession, every borrowed Marian question, every lyric about burning is the flesh learning to do what it was made for, which is to speak back to the One who first spoke it into existence. We do not merely internalize the truth. We become, syllable by syllable, the truth's own body. Gregory of Nyssa called it epektasis, the eternal stretching toward what we cannot yet grasp, and he meant that the reaching is not preliminary to the union but is the union, enacted in time by creatures who will spend eternity discovering that the Fire into which they are falling has no bottom.

Three encounters, then, and in each case the same strange grammar.

In confession, we speak our sins to a God who has known them since before we committed them and Whose love for us has not changed in the interim. He seeks to help us remember Who He Is and bring us back into that deeper relationship that we shared before.

In “On Fire,” we name what it is to be changed by the Unchangeable, to enter the presence of a Love so constant that our inconstancy is illuminated, so Infinite that our finitude is thrown into sharp and ofttimes unflattering relief.

In “Mary, Did You Know?”, we ask questions of someone who holds all the answers, and in the asking, we receive what she has always asked for: complete coherence with the will of Being, in which all our freedom lies.

Our God does not need our words. But we—finite, forgetful, perpetually distracted by lesser concerns—we need to speak them. The Fire does not require our presence. But we require its heat, its light, its purifying attention. Mary does not lack knowledge. But we lack hers until, by her generous permission, we sing ourselves into something like understanding.

This is what the sacrament of confession truly offers, and it is not at all what most people imagine when they picture the darkened box and the sliding screen. Not a transaction—sins submitted, penance assigned, ledger balanced, see you next month. Rather, an encounter with a Love so constant that our inconstancy cannot help but be revealed, so Infinite that our finitude becomes, for once, visible to us, so Unchangeable that we cannot emerge from its presence without being changed ourselves. We enter the confessional not to inform God of anything but to be formed by Him—to let the Fire do its work, to let the curriculum of our failures teach us what we could not otherwise have learned.

The Fathers had a word for this slow burning: theosis, divinization, the gradual transformation of the human person into something more like the God in whose image we were made. It is not comfortable, this burning. It is not efficient. It does not proceed according to any schedule we would have chosen for ourselves. But the Fire is always burning, the invitation is always open, and the Theotokos is always there at the threshold, permitting us to ask the questions whose answers she has carried in her heart since before we were born.

We are the ones who need to speak. And in the speaking—awkward, halting, never quite adequate to the mystery we are attempting to name—we are set on Fire.

Satya Nadella’s year-end reflection identified a shift underway in how we deploy artificial intelligence. Models, he observed, will give way to systems—rich scaffolds that orchestrate multiple models and agents, account for memory and entitlements, and enable sophisticated tool use. 2026 will see this come to fruition at scale. Models in isolation remain impressive but inert; they generate outputs without affecting anything beyond the screen. Systems connected to databases, APIs, and real-world processes can actually change things. The engineering challenge Nadella identifies is genuine: building infrastructure robust enough to enable AI capabilities to operate in the world while managing their jagged edges.

But beneath the architecture diagrams and orchestration layers, a question persists that engineering alone cannot resolve. When the system acts—when it queries a database, sends a message, modifies a file, triggers a downstream process—who decides what should happen? Not who designed the system or deployed it or typed the initial prompt, but who decides in the operative sense: who deliberates among possibilities, weighs them against values, commits to one course, and accepts responsibility for what follows? The engineering sophistication Nadella describes extends the capability for action dramatically while leaving this question not so much answered as unasked.

Understanding why the question matters requires clarity about what action actually involves. There is a tendency, reinforced by computational metaphors, to think of action as output plus execution—a system generates a recommendation, triggers a process, and something happens in the world. But action in the full sense encompasses operations that this picture omits.

When a person acts, she deliberates: holding possibilities in view, imagining consequences, considering what each would mean. She evaluates: measuring options against values she holds, goods she pursues, commitments she has made to others, and to herself. She decides: selecting one path, which means setting aside others, accepting that this choice forecloses alternatives. And she commits: taking the action as her own, owning what follows, standing answerable for the outcome. These operations constitute the fourth level of human cognition—what follows when experience has been gathered, understanding achieved, and judgment rendered about what is true. Having grasped the situation, the person faces the question of what to do about it. The question is not merely cognitive but existential; in deciding, she constitutes herself as someone who chose this rather than that.

A system that processes inputs, generates an action representation, and triggers execution has not performed these operations. It has not deliberated, because deliberation requires a subject holding possibilities in view, and the system is not a subject. It has not evaluated, because evaluation requires values against which to measure, and the system has no values—only objective functions imposed from outside. It has not decided, because decision requires commitment, and commitment requires someone who commits. The processing may be extraordinarily sophisticated. The execution may be consequential. But the operations constitutive of action—the operations that make action someone’s action—have not occurred. They cannot occur, because they require what no architecture provides: a subject who holds values and bears responsibility.

Nadella describes systems that orchestrate multiple models and agents, and the architectural logic is compelling. Different models excel at different tasks; orchestration layers can route queries appropriately, aggregate outputs intelligently, manage complex workflows that no single model could handle. Capabilities multiply through combination. But what orchestration cannot accomplish is the combination of lower-level cognitive operations into higher-level ones.

Judgment—the third level of cognition, where understanding submits itself to the question of truth—requires grasping that evidence suffices for affirmation. This grasp is not a computation but a reflective act: the mind turning back on its own operations, recognizing that the conditions for warranted assertion have been met. No amount of aggregated pattern-matching produces this recognition. You can ensemble a thousand models, weight their outputs by confidence scores, synthesize their predictions through learned meta-models. What you have achieved is very sophisticated processing at levels one and two. What you have not achieved is judgment, because judgment answers a different question. Processing asks what patterns appear; judgment asks whether the patterns are true. These questions operate in different registers, and no volume of the first yields an answer to the second.

Decision—the fourth level—requires a subject weighing values and committing to action. Orchestrating execution chains, however elaborate, does not produce such a subject. The chain executes according to its design; nothing in the chain decides according to its values. Add persistent memory, and the chain can maintain context across sessions. Add entitlements and the chain operates within permitted bounds. Add tool access, and the chain can affect external systems. Each addition extends capability genuinely. None creates an agent in the sense that matters: someone who acts and is accountable for acting, someone from whom the action flows as theirs.

The “jagged edges” that Nadella acknowledges—the unpredictable failures where models confidently produce nonsense—are often discussed as capability gaps awaiting patches. Some are. But others mark categorical boundaries that no patch can smooth. The edge where pattern-matching ends and judgment begins is not jagged; it is a cliff face. No path leads across because the terrain on the other side requires operations that processing cannot perform. Systems can create the appearance of smooth passage through redundancy and verification and constraint. What they cannot do is actually cross, because crossing would require being the kind of thing that judges and decides.

Consider more closely the capabilities Nadella specifies: memory, entitlements, and tool use. Each extends something real and valuable. Understanding what each extends—and what it does not—illuminates where the boundaries actually lie.

Memory extends the first level of cognition: data persistence, context retrieval, pattern continuity across sessions. A system with memory can maintain coherent engagement over time, reference prior interactions, and build on accumulated context rather than starting fresh with each query. This is genuine amplification, and its practical value is immense. But memory in this sense is not experience in the full sense. Experience involves a subject attending to data, and attending is an act of consciousness. Memory systems store and retrieve representations; they do not attend to them.1 The continuity they provide is informational rather than existential—a record of what happened, not the lived presence of someone to whom it happened. The system with memory knows what occurred in past sessions, the way a database knows what it contains: the data is there, retrievable, but no one is home.

Entitlements constrain what the system may access, modify, or trigger. This is a crucial governance function—without it, capable systems become dangerous systems—and it participates in what might be called the good of order: the institutional arrangements that channel capabilities toward acceptable ends. But entitlements are external constraints substituting for internal evaluation. The well-entitled system acts within permitted scope because the scope has been defined and enforced from outside, not because the system has weighed the merits of restraint. The entitlement framework does the work that judgment would do if judgment were available; its necessity is a symptom of judgment’s absence.

Tool use extends execution capability, enabling models to reach beyond text generation into the world of databases, APIs, and file systems. The extension is dramatic: a model with tools can query, write, send, modify, and deploy. But tool use is still execution. The model processes inputs, generates a representation of an action, and triggers a tool. What happens next may be significant—messages sent, records updated, processes initiated—but the operation remains at the level of sophisticated automation. The tools provide hands; nothing in the architecture provides a will that could direct them. The tool-using model is like a factory robot with an expanded range of motion: it can do more things, but it still does only what its programming specifies. The programming has gotten more abstract, more responsive to context, more capable of handling novelty. It has not become someone who decides.

There is a useful way of thinking about what technology serves that distinguishes three levels of human good. At the base, particular goods satisfy individual needs: this meal, that shelter, the answer to my question now. Above them, the good of order comprises institutions and arrangements ensuring regular access to particular goods: agricultural systems that make meals available reliably, housing markets that provide shelter at scale, search engines that answer questions on demand. At the apex, terminal values are what are genuinely worthwhile—the standards against which social arrangements should be assessed, the ends that give the whole structure its point.

AI systems produce particular goods constantly: answers, summaries, drafts, translations, analyses. They also increasingly constitute elements of the good of order—infrastructure that makes information reliably accessible, communication routinely possible, and services consistently available. A well-designed AI system becomes part of the institutional fabric, as much as roads or power grids or financial networks. The integration is not dystopian; it is the ordinary way that useful technologies find their place.

But the good of order is not self-justifying. Efficient infrastructure still requires the question: efficient toward what? Reliable systems still require: reliable for whom? The answers must come from terminal values—from judgment about what human flourishing actually requires and decision about which goods are worth pursuing. These operations cannot be automated because they are the operations that determine what automation should serve. Building elaborate infrastructure while neglecting the evaluative questions is like constructing highways without asking where they lead. The engineering may be excellent, the traffic flow optimized, the maintenance impeccable. But highways are for going somewhere. Infrastructure that perfects movement while obscuring destination can take us anywhere at all, including places we should never go.

The industry increasingly describes AI systems as “agents,” and the language deserves scrutiny. An agent, properly understood, is one who acts—not merely one who causes effects, but one who decides and bears responsibility for deciding. The etymology traces to agere: to do, to drive, to conduct. Agency implies that responsibility has somewhere to land. The agent is the origin of the action in a sense that permits praise or blame, justification, or regret. The agent is the one who can be asked, “Why did you do that?” and is expected to own the answer.

Current AI “agents” are models with tool access and execution loops. They process queries, generate plans, take steps, observe outcomes, and adjust accordingly. The architecture resembles agency in its outward form: goals and actions and adaptation, the appearance of pursuing ends through chosen means. But resemblance is not identity. The model processes without deciding. The loop executes without committing. No subject inhabits the architecture who could be asked why it acted and could genuinely answer. The system has no “why” of its own; it has only the objectives it was given and the patterns it learned.

This matters beyond semantics. When an AI “agent” takes action with real consequences—when it sends a message that damages a relationship, makes a trade that loses money, triggers a process that cannot be undone—who bears responsibility? Not the model, which only processed. Not the execution loop, which only ran. The designer? Too distant from the particular action, which emerged from a context the designer never saw. The deployer? Too removed from the inner workings, which operated according to dynamics the deployer may not understand. The user? Perhaps, but the framing of “use” obscures how much autonomy the system exercised, how little the user controlled or even observed the intermediate steps.

Responsibility diffuses until it evaporates. Everyone can point elsewhere; no one clearly bears the weight. This is not a governance problem awaiting a policy solution. It is a structural feature of systems that execute without deciding. Where there is no deciding subject, there is no responsibility—only causation, only effects, only things that happened without anyone having done them.

If systems cannot decide, and if consequential action requires decision, then human judgment must remain structurally central—not an optional add-on, not a governance formality, but a constitutive element without which the system does not genuinely act. The practical question becomes: where in the architecture does human judgment actually enter?

Current designs often position it peripherally. The human approves outputs at the end of execution chains, provides oversight when metrics flag anomalies, and intervenes when things go wrong. This architecture treats human judgment as exception handling rather than as decision-making. The system operates; the human monitors. The system proposes; the human ratifies. The workflow is efficient precisely because the human mostly stays out of the way.

A different architecture would position human judgment centrally—not as checkpoint but as origin. The system gathers data, surfaces considerations, presents the material that judgment requires. The human deliberates, evaluates, decides. The system then extends that decision into execution, applying capability to carry out what the human determined should be done. The difference is not cosmetic. In the peripheral model, the system acts, and humans constrain. In the central model, humans decide, and systems extend. The first borrows the appearance of human agency while hollowing out its substance. The second preserves agency’s structure while amplifying its reach.

The practical challenge is that central architectures are slower, more effortful, and less amenable to automation’s efficiencies. They require humans to actually engage rather than passively approve. They interrupt workflows that would otherwise proceed without friction. Every design decision trades off capability against agency, efficiency against responsibility, what the system could do against what the human should decide. There is no architecture that maximizes all values simultaneously. The question is which values we optimize for when they conflict—and that question, too, requires human judgment to answer.

What makes this urgent is that the erosion of agency happens incrementally, through convenience rather than coercion. Each individual delegation seems reasonable: why deliberate when the system handles it well enough? Why evaluate when the outputs are good enough? Why decide when the default works? The delegations accumulate until the human’s role has become nominal—a signature on forms the system prepared, approval of recommendations the system generated, ratification of decisions the system already made. No single step felt like surrender. The destination is surrender nonetheless.

Nadella is right that we need engineering sophistication to extract real value from AI in the real world. The technical challenges he names are genuine: orchestration across models, persistent memory, entitlement frameworks, robust tool integration. Solving these challenges matters. But engineering sophistication fully conceived extends beyond the technical.

It includes cognitional sophistication: understanding which operations AI systems can perform and which remain constitutively human. Building without this understanding means building blind—not knowing where the boundaries lie, mistaking execution for decision, and conflating pattern matching with judgment. The jagged edges become invisible until someone falls off a cliff that no one realized was there.

It includes design wisdom: shaping architectures that support human agency rather than simulating or supplanting it. This is more difficult than adding capability, because capability compounds, whereas agency must be cultivated. The path of least resistance automates decision away. The wiser path preserves decision’s structure while making it better informed, more consequential, and more clearly the human’s own.

It includes evaluative honesty: measuring not only whether systems perform but also whether humans using them flourish. Does this system make its users more responsible agents in the world, more attentive to what matters, more discerning about what is true, more deliberate about what to do? If the system optimizes its metrics while its users atrophy as deciders, we have succeeded at the wrong task.

These measures do not yet exist in any systematic form. Developing them is itself an engineering challenge, one that demands the cognitional and design sophistication it aims to measure. The circularity is not vicious but developmental: we build what we understand, building reveals what we missed, and understanding deepens through the building. But the development requires intention. Left to its own momentum, optimization finds what is measurable, and what matters most—agency, responsibility, the human capacity to decide—resists easy measurement. The metrics we lack are the ones we most need.

Systems are coming that will act in the world with unprecedented capability and autonomy. Orchestration layers, persistent memory, entitlement frameworks, expansive tool access—the infrastructure is being built rapidly and at scale. The trajectory Nadella traces is not speculation but observation. Whether we find it exhilarating or alarming, it is happening.

What remains undetermined is whether these systems will be designed to preserve human agency at their center or to simulate it at their periphery. Both architectures are technically feasible. Both will be built, likely in competition with one another. The question is which we will recognize as genuinely sophisticated—which we will fund, adopt, integrate, and normalize.

Execution without decision is not action but automation. Automation has its place; no one wants to deliberate over every database query or agonize over every API call. But automation that absorbs decisions rather than executing them produces systems that act without anyone acting, that affect the world without anyone being responsible for the effects. The technical capability to do this exists and is expanding. The wisdom not to do it, or to do it only where appropriate, requires the operations that AI systems lack: judgment about what is fitting, decision about what we will build, commitment to the values that make us more than our tools.

The engineering challenge is real. So is the human one. Sophistication worthy of the name must encompass both.

Taylor Black writes about AI, human flourishing, and the Catholic intellectual tradition. He serves as head of AI & venture ecosystems in Microsoft’s Office of the CTO and is Founding Director of the Institute for AI & Emerging Technologies at Catholic University of America.

This is precisely the question. And we have an answer—one developed decades before large language models existed, waiting for exactly this application.

The Canadian philosopher Bernard Lonergan spent his career mapping how human knowing actually works. Not knowing as philosophers imagined it should work, but knowing as it operates when you solve a problem, grasp a concept, verify a hunch, decide what to do. His cognitional theory provides the “theory of mind” Nadella seeks: a precise account of distinct mental operations that clarifies which can be amplified by tools and which cannot be replaced.

Why Bicycles for the Mind Works—And Where It Fails

Steve Jobs drew on a 1973 Scientific American study: humans on bicycles consume roughly one-fifth the calories per kilometer of unaided walking. The bicycle doesn’t replace legs. It amplifies what legs already do. Jobs saw computers the same way—tools that extend human cognitive capacity without substituting for it.

The metaphor has proven remarkably durable. It captures something true: the best tools enhance rather than replace. But it lacks the precision the AI moment demands. A bicycle amplifies locomotion. What, exactly, does AI amplify? “Cognition” is too vague. We need to know which cognitive operations are at stake.

The slop-versus-sophistication debate reveals the cost of this imprecision. Critics decry AI-generated content as low-quality filler. Defenders counter that outputs are increasingly indistinguishable from human work. Both sides argue about the product while ignoring the process. Both assume AI outputs and human knowing are commensurable—differing in quality, not kind.

They aren’t. And until we have vocabulary for the difference, we’ll keep arguing past each other.

What Knowing Actually Requires

Here’s what Lonergan mapped: knowing unfolds through four distinct operations, each irreducible to the others.

First, experience. We attend to data—sensing, perceiving, imagining, remembering. This is the intake function: gathering the raw material that knowing requires. Be attentive, the imperative runs. Notice what’s there.

Second, understanding. We inquire into the data, and sometimes insight strikes. Not more data, but the grasp of pattern, relationship, intelligibility. The shift from “these symptoms keep appearing together” to “oh, that’s why they cluster”—that click of coherence. Be intelligent: pursue the insight, don’t settle for accumulation.

Third, judgment. Understanding proposes; judgment disposes. We reflect on whether our insight is actually correct, marshal evidence, weigh considerations, and reach a verdict: yes, this is so; no, that isn’t; maybe, I need more. The question shifts from “what might this mean?” to “is it true?” Be reasonable: don’t affirm beyond what evidence warrants, but do affirm when it does.

Fourth, decision. Knowing what’s true opens onto knowing what to do. We deliberate, evaluate options against values, and choose. Be responsible: act on what you’ve come to know.

The four levels aren’t a sequence you complete once and leave behind. They spiral and recur. New decisions generate new data requiring fresh understanding and judgment. But they remain distinct operations—categorically different acts of consciousness, not points on a continuum.

Skip a level and knowing collapses. Data without insight is noise. Insight without judgment is speculation. Judgment without decision is sterile. Each operation depends on those before it; none substitutes for those after.

Where AI Lives in This Structure

Large language models operate powerfully at level one. They gather, process, and surface data at scales impossible for unaided humans. They extend memory, accelerate search, aggregate information across sources. As experience-amplifiers, they are genuinely transformative.

They also assist level two—but here precision matters. Models generate candidate understandings: possible patterns, hypotheses, framings. They surface what might be intelligible. But they do not achieve insight. The subjective click of coherence—grasping why something is so—requires a consciousness that wonders and is satisfied. Models process; they don’t wonder.

This isn’t a limitation to be overcome through scaling. It’s a structural feature. Insight is the act of a subject grasping intelligibility. Models are not subjects; they execute operations. The outputs can prompt human insight, scaffold it, even simulate its products. But the operation itself remains unavailable.

Level three is where the distinction sharpens to a bright line. Judgment requires grasping that conditions for truth are fulfilled—what Lonergan called reaching a “virtually unconditioned.” Not absolute certainty, but warranted affirmation: recognizing that the evidence suffices for this conclusion. This is a reflective act, a consciousness evaluating its own understanding against criteria for correctness.

Models generate outputs with confidence scores. This is not judgment. Confidence scores measure statistical properties of outputs relative to training distributions. Judgment evaluates whether an understanding accurately grasps reality. The first is computation; the second is reflection. No amount of scaling bridges the gap because they are different kinds of operations.

Level four remains equally beyond reach. Decision involves evaluating options against values held by a valuing subject, and committing to action. Models have no values—they have objective functions set by designers. They make no commitments—they generate outputs. The language of AI “deciding” is metaphorical in a way that obscures the relevant difference.

The Evolved Metaphor

Nadella asked for an evolved concept. Here it is:

AI is a bicycle for experience, a workshop for understanding, and a null for judgment and decision.

The bicycle extends what the rider already does. AI extends our capacity to attend—gathering more data, processing it faster, surfacing patterns we’d miss. This is genuine amplification. The human remains the rider; the capability expands.

The workshop provides tools and materials but doesn’t build the thing. AI offers candidate understandings—hypotheses to consider, framings to try, connections to explore. The human must still achieve the insight, grasp why the pattern holds, feel the click of intelligibility. Workshop tools make building easier; they don’t substitute for the builder’s skill.

The null marks what cannot be extended because there’s nothing mechanical to extend. Judgment is not slow computation that faster computation could replace. It’s a different kind of act. Decision is not inefficient value-weighting that optimization could improve. It’s the commitment of a subject who has values to weigh. You cannot bicycle what isn’t locomotion.

This metaphor does work. It tells designers where amplification helps and where it harms. It tells users what to expect and what remains their responsibility. It dissolves the slop-versus-sophistication debate by redirecting attention from output quality to operational integrity.

Dissolving the Debate

“Slop” names AI outputs that substitute for human operations without the human noticing. “Sophistication” names outputs good enough that substitution seems justified. Both concepts assume the same flawed premise: AI outputs and human knowing differ in quality, not kind.

Reframe with the four-level structure and the debate dissolves. The question isn’t whether outputs are good or bad. The question is which operations produced them and which operations they require.

A model-generated summary of meeting notes (level one amplification: data compression) awaiting human judgment about accuracy and human decision about action—that’s scaffolding working correctly. The same summary treated as already judged, forwarded without reflection—that’s operational collapse. Output quality doesn’t distinguish these cases. The human’s cognitional operations do.

Sophisticated outputs may actually pose greater danger than obvious slop. Slop signals its own inadequacy; it prompts human judgment almost automatically. Sophisticated outputs pass unexamined. Their very quality bypasses the verification they require.

Product design implication: interfaces should make operational status visible. Not just what the AI produced, but what the human still needs to do. Here is gathered data. Here are candidate interpretations. Your judgment required. Your decision pending. The scaffolding must reveal its own edges.

Things and Their Knowing

Lonergan pressed the question further. What do we actually know when knowing works correctly? Not isolated impressions or abstract concepts but things: concrete unities grasped through the full cognitional process.

A thing—this patient, that market, your team—isn’t merely seen. Raw experience gives us manifolds of data, not unified objects. Understanding grasps the data as belonging together, as expressions of a single intelligible unity. Judgment verifies that this unity actually exists and operates as understood. The thing known is the achievement of the complete process, not the input to its first stage.

AI systems don’t know things in this sense. They process representations—tokens, embeddings, weights. Sophisticated representations, perhaps. But representations that never unify into grasped things because the grasping subject is absent. When we treat model outputs as knowledge of things, we import an achievement the model cannot reach.

The practical stakes are significant. A model can surface everything in your CRM about a customer. It can cluster behaviors, predict churn probability, suggest interventions. But it does not know the customer as a concrete unity—a this-person with that-history making those-decisions for these-reasons. The human account manager might. The knowledge differs in kind, not merely degree.

Product design built on this insight would preserve and support thing-knowledge rather than drowning it in data. More information about the customer is not always better; sometimes it substitutes representation for knowing. The well-designed tool surfaces what helps the human achieve unified understanding, then gets out of the way.

The New Equilibrium

Nadella asked how we should relate to each other “as humans equipped with these new cognitive amplifier tools.” The four-level structure generates an answer.

Authentic human relating requires all four operations. I attend to you—your words, expressions, situation. I understand your meaning—not just decoding symbols but grasping what you’re getting at. I judge whether my understanding is accurate—testing interpretations, checking assumptions. I decide how to respond—weighing what matters, committing to action.

AI can assist levels one and two in this process. Transcription captures what you said; summary compresses it; translation bridges languages; context-retrieval surfaces relevant history. These are genuine assists. They let me attend and understand better than I otherwise could.

But if I outsource levels three and four, I’m no longer relating to you. I’m processing data about you. The model drafts a response; I send it unexamined; you receive words shaped by statistical patterns rather than by my judgment about what’s true and my decision about what matters. We’re both diminished. The interaction looks like communication but lacks its substance.

The new equilibrium isn’t primarily technological. It’s normative. We need shared expectations that AI-assisted communication still involves human judgment and decision. Cultural standards that hold people responsible for having actually understood and verified what they transmit. Social practices that treat operational collapse as a failure mode, not an efficiency gain.

This equilibrium requires vocabulary—which is why the theory of mind matters. Without concepts for the distinct operations, we can’t articulate what’s been lost when they’re skipped. With them, we can name the failure and demand better. Design Principles for Authentic Scaffolding

Theory becomes practical through design. Here are principles the four-level structure generates:

Make operational boundaries visible. Current interfaces present AI outputs as seamless knowledge. They should instead make explicit where AI operation ends and human operation must begin. The visual grammar matters: data and candidates look different from conclusions; pending-judgment looks different from verified-and-ready.

Preserve cognitive friction where it builds capacity. Friction isn’t always inefficiency. Sometimes it’s exercise. The e-bike delivers you to the destination without the exertion that builds cycling capability. AI that removes all cognitive friction may deliver outputs while atrophying the human operations that produce real knowing. Design for scaffolding—temporary support while capacity develops—not substitution.

Support judgment rather than simulating it. Models can marshal evidence, surface considerations, flag inconsistencies. These genuinely assist reflective judgment. Models cannot determine that evidence suffices for affirmation; that’s the human’s act. Design should present what judgment requires without presenting conclusions that presume judgment complete.

Enable thing-knowledge. Data accumulation can obscure rather than reveal concrete unities. The well-designed tool helps the user achieve integrated understanding of the thing in question—this patient, that project, your situation—rather than burying insight under representations.

Maintain transparency about operations. Humans use tools better when they understand how the tools work. AI systems should be as interpretable as current technology allows—not primarily for auditing, but because understanding your instruments is part of using them well.

Design for the spiral. Knowing isn’t linear; it recurs. New decisions generate new situations requiring fresh attention and understanding. Interfaces should support this recursion rather than treating each query as isolated. Memory and context serve not just efficiency but the continuity of genuine inquiry.

What This Means for Builders

The AI industry optimizes for capability. Benchmarks measure what models can produce. Investment follows performance on defined tasks. The implicit question: how do we make AI more powerful?

The four-level structure suggests a different optimization target. Not capability but support for authentic knowing. Not what models produce but whether humans using models become more attentive, more intelligent, more reasonable, more responsible.

These measures don’t currently exist. Creating them is itself a design challenge—and a business opportunity. Companies that figure out how to measure and improve human knowing-with-AI will build something more valuable than another percentage point on benchmarks.

The theory of mind Nadella requested provides the foundation. It specifies what human knowing requires. It identifies which requirements tools can assist and which they cannot replace. It generates design principles distinguishing scaffolding from substitution. What remains is building to the theory.

Conclusion: Amplification and Its Limits

Steve Jobs saw the bicycle amplify human locomotion without replacing human legs. The computer, he proposed, could do the same for human minds. He was right—but the metaphor needed the precision it now has.

Knowing operates through experience, understanding, judgment, and decision. AI amplifies the first, assists the second, and cannot perform the third or fourth. This isn’t pessimism about AI capability; it’s clarity about AI kind. Statistical pattern-matching over representations differs structurally from a conscious subject grasping intelligible unities and verifying truth. The difference isn’t degree; it’s category.

Scaffolding respects this difference. It extends what can be extended and preserves space for what cannot be replaced. The new equilibrium Nadella seeks emerges when builders design for scaffolding and users expect it—when we share vocabulary for the operations at stake and hold each other accountable for performing them.

We’ve been arguing about slop and sophistication, about job displacement and productivity gains, about alignment and misalignment. These debates matter. But beneath them lies a simpler question: what does knowing require, and how should tools relate to it?

The application is ours to make.

Taylor Black writes about AI, human flourishing, and the Catholic intellectual tradition. He serves as head of AI & venture ecosystems in Microsoft’s Office of the CTO and is Founding Director of the Leonum Institute for AI & Emerging Technologies at Catholic University of America.

This is not a failure of discipline. This is the structure of modern work.

What you’re doing, beneath the surface experience, is executing a rapid sequence of microdecisions. Each notification triggers a small cognitive function: parse the message, assess urgency, classify by type, determine routing—respond now, respond later, ignore, escalate. You don’t experience these as decisions. They feel like a continuous flow, like the saccades your eyes make while reading or the micro-adjustments your hands make while driving. But they’re there, hundreds of them daily, each with its own logic.

And that logic is the key to understanding what’s coming.

Most explanations of AI and work stay at the surface: this tool does this task. Chatbots answer questions. Copilots suggest code. Assistants schedule meetings. No more drudgery. One tool, one task, faster. But this framing misses something essential. To understand why agentic ecosystems will absorb so much professional labor, you have to look beneath the job title to the actual mechanics—the atomic operations that constitute a day.

Work, at its most granular, is not “tasks” in the way we usually mean. It’s a continuous sequence of microdecisions: which email to answer first, how to phrase a request, whether a deliverable is done enough to move forward, what context someone needs to take action, when to escalate, and when to handle quietly. The big strategic choices we remember—hire this person, pursue this market, ship this feature—float on a vast sea of small choices we barely notice.

These microdecisions have a structure. And that structure determines what can be absorbed by machines and what remains irreducibly human.

The Anatomy of a Microdecision

Every microdecision, when examined closely, has three components.

First, inputs: the information available at the moment of decision. This might be an email’s text, a ticket’s status, a number on a dashboard, a memory of a conversation last week, a felt sense of how a relationship is going. Some inputs are explicit and digital; others are tacit and contextual.

Second, a decision function: the logic—explicit or implicit—that maps inputs to outputs. Sometimes this is a clear rule: if the customer is enterprise tier, escalate within four hours. Sometimes it’s a heuristic: this email sounds urgent based on tone. Sometimes it’s judgment that resists articulation: given everything I know about this person, this situation, this moment, this is the right call.

Third, outputs: an action, a communication, a transformation of state. Send this message, update this field, route this request, schedule this meeting, make this recommendation.

The nature of the decision function—not the inputs or outputs, but the logic connecting them—determines whether a machine can handle it.

Three Kinds of Logic

Deterministic functions have explicit rules. If X, then Y. No ambiguity, no interpretation required. An expense under $500 is automatically approved. A support ticket tagged “billing” routes to the billing team. A calendar invite with three conflicts gets flagged for resolution.

These are already automated in most organizations, though often clumsily. Rigid workflows, if-then rules, simple conditionals. The logic is clear enough to encode directly. What limits these automations isn’t capability but brittleness—they work perfectly until something unexpected happens, and then they fail completely because they cannot interpret, only execute.

Probabilistic functions require interpretation but follow stable patterns. There’s no explicit rule you could write down, but a sufficiently large dataset of examples reveals consistent logic beneath the surface. “Urgent-sounding” emails share linguistic features. “Ready for review” code has characteristic signatures. “Promising” sales leads are associated with specific signals. A system trained on thousands of examples can learn to approximate human judgment—not by understanding, but by pattern-matching at scale.

This is where large language models operate. They don’t know what urgency means; they’ve learned what urgency looks like across millions of examples. They can’t reason about code quality from first principles; they’ve seen enough code and enough reviews to predict what a competent reviewer would likely flag. The output isn’t perfect; it’s probabilistic, which means sometimes wrong, but it’s accurate enough, often enough, to be useful.

Irreducible functions resist both rules and patterns. They depend on contexts that machines cannot access, stakes that machines cannot bear, or relationships that machines cannot hold. Deciding whether to fire someone. Navigating a conflict between two people you know well. Choosing a company’s strategic direction when the data is ambiguous and the consequences are permanent. Telling a customer something they don’t want to hear in a way that preserves the relationship. Recognizing that someone on your team is struggling before they’ve said anything.

These require presence. Accountability. The kind of understanding that emerges only from being a person among other people, embedded in relationships with a history and a future.

Most professionals substantially underestimate the extent to which their work falls into the first two categories.

The Composition Problem

A single microdecision is simple enough to categorize. But work isn’t a single decision, it is a set of sequences. Long chains where the output of one decision becomes the input of the next. Context accumulates. State changes. Dependencies branch and merge.

This is where traditional automation fails. Rigid workflows can handle deterministic sequences—if A then B then C—but they shatter when anything unexpected happens. A customer replies with a question the script didn’t anticipate. A dependency shifts. Someone’s out sick. An edge case arises that the workflow designer has not considered. The system can’t adapt because it can’t interpret. It can only execute what it is instructed to do.

Human professionals succeed in complex sequences precisely because they can interpret each step. When the unexpected happens, they adjust. They apply judgment at each decision point, not only at the outset when the process was designed.

Agentic systems address this problem by combining two previously separate capabilities.

Probabilistic interpretation handles ambiguity. When the input is natural language, unstructured data, or a novel situation, a language model interprets intent, extracts meaning, and classifies the situation. It doesn’t need explicit rules for every contingency because it learned patterns from millions of examples. It can handle the unexpected—not perfectly, but adequately—because it can read the situation in a way rigid systems cannot.

Deterministic execution handles consequences. Once the system decides what to do, the doing is precise: send this exact email, update this specific field, call this API with these parameters, move this money, schedule this meeting. No hallucination, no drift, no “creative” interpretation. The probabilistic layer decides; the deterministic layer acts.

Orchestration logic manages the flow between them. When should the system proceed autonomously? When should it pause for human review? How should it handle uncertainty—with confidence intervals, explicit flagging, escalation thresholds? This meta-layer routes each decision through the appropriate channel, dynamically balancing machine autonomy and human oversight based on stakes, confidence, and context.

The result is a system that can handle long sequences of mixed decision types. Consider a concrete example.

A Sequence Traced

An email arrives in a founder’s inbox: “Following up on our conversation—our board is meeting next week, and I’d love to give them an update on where things stand with a potential partnership. Any progress?”

This is unstructured natural language requiring interpretation. A probabilistic system parses it: this is a partnership inquiry; it references a prior conversation; it has a time constraint (board meeting next week); the sender’s tone suggests friendly pressure; the implicit request is a status update or a commitment signal.

The system queries structured data—deterministic lookups against the CRM, the calendar, and past correspondence. It surfaces: the prior conversation was six weeks ago; discussed co-marketing; no formal proposal was sent; the founder has a note flagging this as “interesting but not priority.”

It cross-references context: the sender’s company recently closed a funding round (reported by a news monitoring agent), which may explain the renewed urgency. The founder’s calendar next week is packed, but there’s a thirty-minute window on Thursday that could work for a call.

It drafts a response: “Great to hear from you—congratulations on the round, by the way. We’ve been heads-down on [current priority], but I’d like to pick this back up. I have a window on Thursday at 2 pm PT if you want to sync before your board meeting. I can share where we’re at and discuss what a pilot might look like.”

This draft is probabilistic—generated from patterns of how founders communicate, how partnership conversations typically progress, what tone matches the sender’s tone—but grounded in deterministic facts: the actual calendar availability, the actual history, the actual context.

The system presents the draft for review. The founder scans it, changes “pilot” to “proof of concept” because that’s the language this particular partner prefers, and sends. Elapsed time: forty-five seconds.

Without the system, this email would have required: finding the original thread, remembering the context, checking the calendar, recalling what was discussed, composing the response, and reviewing for tone. Fifteen to twenty minutes, probably deferred until later, possibly forgotten.

One email. One sequence of microdecisions. Interpretation, lookup, synthesis, composition, review, action. The structure repeated across hundreds of interactions daily.

The Texture of Absorption

What does it feel like when this absorption actually happens?

The project manager who used to spend mornings copying and pasting between Jira, Asana, and Google Sheets—reformatting the same information for different audiences—now reviews a synthesized status that already exists when she opens her laptop. The decisions embedded in that synthesis (what to include, how to frame velocity, which risks to highlight) are probabilistic interpretations trained on how she’s made those judgments hundreds of times before. The data underneath is deterministic, pulled directly from the systems of record.

She reads. She adjusts one framing—the system was overly optimistic about the timeline based on pattern matching; she has context about a team dynamic it can’t see. She approves. The update is automatically distributed to three different audiences in three formats.

The skilled tradesperson who used to spend evenings at the kitchen table calculating quotes—measuring twice, pricing materials, padding for contingency, second-guessing whether he’s too high or too low—now reviews an estimate that was generated from photos and measurements he took on-site. The materials list is deterministic: these fixtures, these wire gauges, this quantity based on square footage. The labor estimate is probabilistic: based on similar jobs, his historical pace, and the complexity signals in the photos.

He adjusts one line item—the system doesn’t know that old houses in this neighborhood always have plaster walls that take longer to fish wire through. He sends. He’s home for dinner.

The venture capitalist who used to spend hours before each pitch meeting researching the market, the competitors, the founders’ backgrounds—pulling from Pitchbook, Crunchbase, LinkedIn, triangulating across tabs—now walks into meetings with a briefing that already exists. The research is a mix: deterministic data (funding history, team composition, metrics where available) and probabilistic synthesis (how this company fits the competitive landscape, what analogous companies suggest about trajectory, what questions the partners are likely to ask based on past discussions).

She reads the briefing in the car. She spots something interesting: the system flagged a connection between this founder and a portfolio company she hadn’t noticed. She notes that she will ask about it. The meeting is better because her preparation is better. Her preparation is better because she didn’t have to do it.

What Remains

In each case, something remains that the system cannot absorb.

The project manager continues to navigate the tension between engineering and design. She still reads the room in standups, catches the hesitation that signals someone disagrees but won’t say so, and brokers compromises that require understanding the humans involved. The system can tell her that velocity dropped; it cannot tell her why the team feels demoralized.

The tradesperson continues to troubleshoot the circuit that keeps tripping. The diagnosis is still his: years of pattern recognition, intuition about old houses, the ability to see what doesn’t look right. The system can schedule, quote, and invoice. It cannot stand before an open panel and know.

The investor still decides whether to back this founder. She still reads conviction in someone’s voice, senses whether the vision is real or performed, and weighs intangibles that don’t appear in any dataset. The system can gather information, synthesize patterns, and flag concerns. It cannot bear the judgment.

The irreducible remainder is real. It’s not everything, but it’s the part that actually requires human presence—human stakes, human relationships, human accountability. What changes is that this remainder is no longer overwhelmed by logistics. The signal emerges from the noise because the noise is handled elsewhere.

The Integration Layer

One more element completes the picture. These agents don’t operate in isolation—they coordinate.

When the founder’s email agent drafts a partnership response, it can consult the calendar agent on availability, query the CRM agent for relationship history, and flag the sales agent if the partnership has revenue implications. When the project manager’s status agent synthesizes progress, it can pull from the engineering team’s code agents, the design team’s review agents, the QA team’s testing agents—each one surfacing its own view of the state.

This coordination is itself a mix of probabilistic and deterministic operations. Probabilistic: interpreting whether two pieces of information are related, judging whether a situation warrants cross-system notification, and deciding how to merge conflicting signals. Deterministic: the actual handoffs, the API calls, the state synchronization, the audit trails.

Human organizations already work this way. Specialists coordinate across functions, passing context and decisions back and forth through meetings, emails, documents, and hallway conversations. The friction is in the handoffs: the reformatting, the re-explaining, the waiting, the misunderstanding, the information that gets lost or garbled in translation.

When agents handle the interstitial work, the handoffs are frictionless. Information flows to where it’s needed, in the required format, at the time necessary. Not perfectly—there are failure modes, edge cases, and situations in which the coordination logic breaks down. But at a speed and cost that make the current overhead visible for what it is: a tax we’ve paid so long we've forgotten it was optional.

Why the Shift Is Hard to See

Most professionals don’t experience their work as a sequence of typed microdecisions. They experience a continuous flow of meetings, conversations, tasks, problems arising, and their resolution. The granular structure is invisible because it’s automatic—like the individual frames in a film or the discrete samples in a digital audio file. Smooth from the outside, discrete underneath.

This is why the scope of potential absorption is so hard to imagine. You don’t feel yourself making five hundred small decisions a day. You feel yourself “doing your job.” The decisions are hidden inside the doing.

But they’re there. Each one has inputs, a decision function, outputs. Each decision function has a type: deterministic, probabilistic, or irreducible. And the distribution across types is not what most people assume.

We tell ourselves stories about our work that emphasize the irreducible parts—the judgment, the creativity, the relationships. These are real, and they matter. But they’re not most of the hours. Most of the hours go to the probabilistic and deterministic logistics that surround the irreducible core, like scaffolding around a building.

The scaffolding is coming down.

The Shape of What Comes

If we decompose our work hours into microdecisions, and most microdecisions are absorbable by systems that combine probabilistic interpretation with deterministic execution, then the structure of work itself is open to transformation.

This isn’t a claim that machines will do your job. Your job isn’t one thing. It’s a thousand things, each with its own logic, each with its own absorbability. The question is recomposition. Which of the thousand things require you—your presence, your judgment, your accountability—and which have you been doing simply because someone had to?

The answers will differ for each role, each organization, and each person. But the question is coming for all of them, and it’s coming faster than most people think, because the underlying mechanics fit these systems in a way that wasn’t true before.

What people do with the shift—whether they reclaim time, expand ambition, deepen the irreducible work, or find themselves unmoored without the familiar friction—depends on factors beyond the technology. It depends on how organizations restructure, how compensation models evolve, and how people understand their own contribution once the logistics are handled.

The founder, who previously spent 70% of her time on coordination and communication, will regain that time. What she does with it—whether she thinks more deeply about strategy, spends more time with her family, or just finds new ways to fill the hours—is not determined by the technology. The technology creates the opening. We walk through it, or we don’t.

The Irreducible Question

Finally, the most challenging question isn’t technical. It’s existential.

If most of what we call work is absorbable—the coordination, the communication, the information logistics—then what remains is the work that actually requires human presence. The architectural decision that shapes everything downstream. The relationship that needs tending. The creative leap that connects things no one connected before. The moment that requires someone to be accountable, to say I made this call and I’ll stand behind it.

These are real. They’re valuable. They may even be what we wanted to be doing all along, the core that got buried under the overhead.

But for many people, the overhead was the job. The scheduling, the formatting, the following up, and the keeping track weren’t obstacles to the work; they were the work. When they disappear, what’s left might feel like freedom or might feel like vertigo. Probably both, in different moments.

The systems being built don’t answer this question. They just make it unavoidable. The anatomy of work is being exposed—laid bare by tools that can absorb the absorbable and leave the rest. What “the rest” is, and whether it’s enough, and what we do when we find out: these are questions for humans, about humans, that no language model can answer.

How should we think about the nature of work in the agentic revolution?

This is the curious fact about insight: it passes into the habitual texture of one’s mind. All of us who have spent time attending to the movements of intelligence might put it this way: “Once one has understood, one has crossed a divide. What a moment ago was an insoluble problem now becomes incredibly simple and obvious.” The divide cannot be uncrossed. You cannot un-understand the punch line of a joke or the proof of a theorem. The insight remains, woven into the fabric of who you are and how you see.

Memory, in its deepest sense, is not the passive storage of sensory snapshots. It is the accumulation of such crossings—a vast, layered, living architecture of insights that have settled into your being. Each new understanding complements and combines with what came before, modifying the whole structure in ways both subtle and profound. The child who has learned to count does not simply possess a new fact; she now inhabits a different cognitive world, one in which new realities can be tracked, compared, and discovered. And that world will expand again when she grasps multiplication, and again when algebra opens a higher viewpoint, and again when calculus reveals the continuous behind the discrete. Each stage presupposes and transforms the last.

This is what we could call the self-correcting process of learning. Insights do not arrive fully formed and final. Each one reveals, by its very exercise, its own incompleteness. You act on what you understand, or speak it aloud, or think it through—and the gaps show themselves. A question arises that the insight cannot answer. A situation arises in which the insight does not quite fit. These failures are invitations: they prompt further questions that lead to complementary insights, which, in turn, reveal new gaps, which prompt new questions, in an ever-spiraling ascent. Memory is what holds this spiral together. Without the accumulation of past insights, there would be no platform from which to launch the next inquiry, no context in which a new insight could be recognized as progress or error.

Consider what this means for ordinary life. You walk into a meeting, and within moments you grasp the mood of the room—tense, perhaps, or expectant. You do not deduce this from explicit premises. You simply “see” it, as immediately as you see the color of the walls. But that immediate seeing is possible only because you have, over years, accumulated a vast store of insights into human faces, tones of voice, postures, and social situations. Those insights are not consciously recalled; they are present and operative, but operate behind the scenes. They govern the direction of your attention, evaluate the significance of small cues, and guide your own responses. The past is not behind you; it is within you, shaping the very texture of your present perception.

We might speak of a habitual orientation—a perpetual alertness formed by past inquiry. The physicist sees the world differently from the poet, not because their eyes are different, but because each has spent years asking different questions and accumulating different constellations of insights. What each notices, what leaps out as significant, what recedes into background noise—all this is determined by the accumulated structure of understanding. The habitual orientation is a kind of perceptual stance, an angle from which reality is approached. It is memory present to spring into readiness.

This is why learning is not merely the adding of facts to a mental warehouse. It is a transformation of the self who perceives and understands. Each insight, as it enters the habitual texture of the mind, changes what will count as data for future insights. The trained physician observes a slight asymmetry in a patient’s gait and suspects a neurological condition; the untrained observer sees only a person walking. The data are the same; the habitually structured minds are different. This is also why learning has direction and a strategy. One cannot simply leap to higher viewpoints; one must pass through the lower ones, accumulating the insights that will serve as the images for the next level of understanding. The acorn and the oak are both alive, but there are vast differences in what they can do. Similarly, the child and the master both think, but the master’s thinking has a reach and a suppleness that presuppose decades of accumulated insight.

There is, however, a shadow side to this architecture of memory. Not all that we have understood is fully integrated. Not all past insights are mutually compatible or consistent with our current self-image. Some experiences and some recognitions are too threatening to be admitted to consciousness and are therefore repressed, pushed below the threshold of awareness. What results is not the absence of memory but its distortion. The repressed insight continues to exert influence, manifesting in slips of the tongue, in odd aversions, and in dreams charged with displaced affect. Lonergan, drawing on Freud, speaks of screening memories—vivid but fictitious recollections that cover over actions too disturbing to recall directly. We remember a false scene precisely because the true one would demand an understanding of ourselves that we are not yet prepared to accept.

This is the censor at work: the preconscious mechanism that selects which materials from neural processes will be allowed to emerge into consciousness. In healthy development, this selection serves the dramatic pattern of experience—the ongoing project of making one’s life a coherent, livable narrative. But when the censor operates under the pressure of fear or shame, it can block not just painful memories but the very insights that would allow one to understand oneself more fully. The result is what Lonergan calls scotosis: a blind spot, an inability to see what is plainly there. Memory, in this pathological mode, becomes not the servant of understanding but its obstacle. The past persists, but in distorted form, shaping the present in ways the conscious mind cannot recognize or correct.

Conversely, a healthy memory supports the self-correcting process. It allows past insights to be recalled, reexamined, and revised in light of new experience. It does not cling to its own constructions but remains open to the further questions that would transform them. Such a memory is porous—willing to let go of what no longer fits, willing to integrate what once seemed threatening. This porosity requires courage, the willingness to face one’s own errors and blind spots. But it is also the condition for growth. The mind that cannot revise its habitual orientations is a mind that has stopped learning, a mind whose spiral of insight has frozen into a fixed picture.

Memory is also crucially social. We are born into a community that possesses a common fund of tested answers—language, custom, technique, story. From that fund, we draw our initial stock of insights, measured by our capacity, our interests, and our energy. The self-correcting process of learning that unfolds in each individual consciousness is simultaneously a communal development, effected through speech and example, disseminated, tested, and improved across generations. What one person discovers passes into the possession of many, to be checked against their experience and confronted with their further questions. The achievements of each generation serve as the starting point for the next.

This social memory is a mixed blessing. On one hand, it accelerates learning enormously: what took Archimedes years to discover can be taught to a schoolchild in an afternoon. On the other hand, shared biases and shared blind spots can be transmitted as easily as genuine insights. Each tribe and nation, each group and class, tends to develop its own brand of common sense—and to strengthen its convictions by pouring ridicule on the common nonsense of others. The social memory is contested terrain, shaped by power and prejudice as much as by disinterested inquiry.

Now consider what happens when memory is externalized in technology. Writing, libraries, and databases all extend the reach of memory far beyond what any individual brain can hold. They allow insights to be preserved across centuries and transmitted across continents. They multiply the data available for future understanding. And now, with artificial intelligence, we have systems that can store and retrieve patterns with a speed and scale no human community could match. An AI can hold in its circuits the entire history of a scientific field, the complete works of a literary tradition, the billions of data points generated by modern sensors and networks. It can, when prompted, reproduce those patterns, recombine them, and present them in novel arrangements.

However, as we work to create useful human experiences with AI, memory remains a fundamental problem. For memory, in the sense that matters for us, is not the storage of data but the accumulation of insights that have passed into the habitual texture of a knowing subject. It is the spiral of understanding, building on understanding, each crossing of the divide leaving behind a transformed mind. Our AI tools have difficulty crossing divides. It does not experience the tension of inquiry released in a flash of comprehension. It has no habitual orientation, no perceptual stance formed by years of accumulated understanding, no sense of what matters and what can be ignored. It has patterns—immense, intricate, statistically powerful patterns—but lacks the insight that explains why those patterns hold, the judgment that affirms their truth, and the lived experience of the past inhabiting the present.

This is not to diminish what AI can offer. On the contrary, it clarifies the distinct and valuable role such systems can play. Because they are tireless repositories of pattern, they can present to human minds data that would otherwise be lost or inaccessible. They can flag the anomaly, retrieve the precedent, and suggest the connection that a human inquirer might never have noticed. They can, in a sense, extend our external memory: holding materials that our limited brains cannot retain and making them available at the moment of need. This is no small thing. A well-designed AI might function as a kind of prosthetic memory, freeing human minds to do what they alone can do—grasp the point, reach the insight, make the judgment.

But we must be careful not to mistake the prosthetic for the organ. The danger with AI-mediated memory is not that it remembers too much but that we remember too little—that we become dependent on external pattern retrieval and allow our own habitual orientations to atrophy. If the self-correcting process of learning requires the accumulation of insights within a subject, then anything that short-circuits that accumulation threatens the process itself. A student who looks up every answer without struggling to understand it may pass the test, but will not have crossed the divide. An expert who relies entirely on AI recommendations may lose the hard-won feel for a field, the sense of which questions to ask, and the judgment to recognize when patterns are misleading.

The opportunity, then, is to use AI as a support for memory without substituting for it. The spiral of understanding can continue, even accelerate—not because the machine understands but because it enriches the field of experience from which human understanding can arise.

Memory is what makes us historical beings. It is the presence of the past within the present, the accumulated insights that shape perception, the habitual orientations that guide inquiry, the screening and distortion that reveal our resistances, the social fund of tested answers that gives us a place to stand. We carry time within us, folded into the texture of our minds.

The promise of AI for our memory is not that it will remember for us. It is that it will help us remember better by holding what we cannot hold, by presenting what we might overlook, by freeing our attention for the insights that only we can achieve. In this service, it becomes one more instrument in the long human project of accumulating understanding, one more layer in the already rich texture of what is given to consciousness and community. The ascent of insight remains our own. I’m excited for the collaborative possibility ahead.

“For teaching is the communication of insight. It throws out the clues, the pointed hints, that lead to insight. It cajoles attention to drive away the distracting images that stand in insight’s way.”

— Bernard Lonergan, Insight

At every instant, this field of experience is multi-layered. Take the simple smell of slow-poured coffee drifting from the kitchen: at one level you register a rich, earthy-sweet scent in the air (a purely sensory impression); simultaneously an image springs to mind of the dark, steaming cup waiting for you (your imagination at work); you feel a small wave of comfort and appetite (an affective response entwined with a conative desire to get that coffee); you may even find yourself thinking the word “coffee” and murmuring ah, breakfast (a linguistic and intellectual act that identifies and situates this sensation, already pointing you toward the act of pouring a cup). Even such a mundane moment is composed of stacked strata—sense, image, feeling, impulse, thought, and language—each adding its own tone and texture to what you consciously experience.

Experience never unfolds in isolation. Surrounding each of us is a social and cultural context that penetrates what and how we experience. The meaning of that morning coffee, for instance, is shaped by culture (perhaps it’s a cherished daily ritual, a symbol of morning calm shared by all true Seattleites) and by social circumstances (maybe your green-eyed, curly-haired wife prepared it for you, adding a layer of interpersonal warmth or obligation). Our practical routines themselves are embedded in a shared world: the language we use to label things and the common expectations we carry tell us, almost subconsciously, what to notice and what to ignore. Even our inner life has depths beyond immediate awareness—psychic undercurrents of habit, personality, and subconscious association that influence why one moment resonates and another passes quietly. All these layers—practical, social, cultural, and psychic—enfold each simple experience, giving it a richness we usually take for granted.

And still there are more elusive aspects of the given. Every experience contains what a philosopher might call an empirical residue: the sheer concrete “thisness” of reality that is presented to us without our understanding yet why it is so. This includes all the accidentals and specifics of here and now: the particular spot where your chair sits, the exact 5:42 AM on the clock as you sip your coffee, the fact that today’s brew is a touch stronger than yesterday’s. These details are simply given to your awareness—brute facts that no general insight fully explains, the backdrop against which understanding must work. Alongside these factual givens, we also experience absences—the felt gaps and missing pieces in the fabric of the moment. You might notice your cup is smaller than usual and sense that something is off, or you may feel the quiet where a friend’s morning greeting used to be. Such absences are as much a part of the experiential field as any positive sensation: they manifest as a nagging question, an empty chair, a hunger for something not there. An absence can tug at your attention—a sign that draws you toward what is lacking or not yet understood, urging inquiry.

In this way, experience spans a remarkable breadth. It ranges from raw sensations up through imagination and emotion to thought and language; from immediate bodily drives to practical plans and actions; from the private stream of consciousness to the social and cultural atmosphere we breathe; from the surface of awareness to the submerged psychic depths; from the solid presence of things to the intangibility of what isn’t there. All these layers together constitute the full range of what can be given to consciousness. Nothing we encounter in waking life is ever “pure” sensation or isolated fact; it arrives intertwined with these multiple dimensions, a complex, dynamic field in which our attention continuously moves.

Crucially, experience is not a passive recording of sensory inputs but an active and selective engagement. Our attention works like a roaming spotlight, highlighting some facets of the field and leaving others in shadow, guided by our intentions, habits, interests, and desires. Picture two people walking down the same street—one an architect, the other a gardener. The architect looks up and immediately notices the elegant curve of a roofline and the pattern of glass on a new building, while the gardener is looking down, drawn to the colors of wildflowers by the sidewalk and the health of the old oak trees along the road. Each person’s background and purpose emphasize certain aspects of the scene and all but ignore others. Their minds also weave in different associations: the architect automatically compares the building’s design to others she knows, perhaps sparking an idea for her own upcoming project, while the gardener recalls how those roses bloomed last year and mentally notes the dry soil around the oak’s roots. In effect, each is experiencing a different street, filtered and colored by a distinct pattern of interest. What we perceive is quietly shaped by what we are looking for (or hoping for, or afraid of) in that moment. The stream of sensations may be the same, but experience is the portion of that stream illuminated by the light of attention.

The dominant pattern of our attention can shift with circumstance or purpose. We might slip into an aesthetic pattern of experience, for instance, when we pause on our walk to admire a sunset: our gaze lingers on the sky’s colors and cloud forms for their sheer beauty, with no further goal in mind. Moments later, we may switch into a practical pattern as we remember an errand—now we’re scanning traffic and street signs, focused on getting somewhere efficiently. On another day, we could be in an intellectual pattern: imagine a scientist in her lab, eyes narrowed at an instrument readout, or a student poring over a puzzle; here the mind attends to data and ideas, every sense impression evaluated for what it might mean or how it fits into a theory. There are also times of a religious or spiritual pattern of consciousness, when we attend to the world with a sense of depth or sacred significance—for one person the morning bird’s song might be heard not just as a pretty sound but as a note of creation’s harmony, a reminder of something transcendent in the everyday. In each of these modes, certain layers of experience come to the fore while others fade into the background. The artist’s eye revels in sensory and imaginal richness, the busy commuter’s mind zeroes in on practical cues, the inquirer filters everything for clues to a truth, the spiritually attentive soul finds meaning even in a silence. We move among such patterns fluidly, often without realizing how our consciousness reorients to meet the moment’s needs or our heart’s aims.

Deep changes in ourselves can transform these patterns profoundly. Major shifts of outlook—conversions of mind and heart—reconfigure what we notice and value. Consider the effect of a moral conversion: a person awakens to a new concern for others or for justice, and suddenly aspects of experience once overlooked stand out sharply. The streets that seemed ordinary yesterday are now filled with evidence of inequity or suffering that demand attention. An intellectual conversion can likewise alter one’s whole orientation to experience: the day a formerly indifferent student falls in love with learning, the world starts to brim with questions—every phenomenon becomes something to inquire into, where before it was merely there. A religious conversion or profound shift in faith can flood even mundane experiences with new meaning, as when someone who has undergone a spiritual renewal begins to perceive purpose or providence in events that used to seem trivial or random. In each case, the fundamental pattern of attention shifts. What one attends to, how one interprets feelings and absences, what one finds significant—all these are transformed by a conversion. The field of experience itself is restructured because the chooser, the knower, the one who is experiencing, has changed at a fundamental level. New eyes, as the saying goes, create a new world to see.

In our current age, an increasingly large part of this structured field of experience is influenced by technology. Much of what we sense, imagine, and even feel is now mediated by devices and digital systems. Think of how a smartphone shapes a given afternoon: your friend’s face appears to you as a glowing image on a screen rather than in person; your sense of what’s happening in the world comes through a curated news feed; even your social interactions and cultural cues stream to you via an array of apps and algorithms. The flow of experience is being filtered and patterned by systems outside ourselves, often designed intentionally to capture our attention. A notification’s ping draws your eye, an algorithm decides which posts or products populate your view. In subtle but pervasive ways, technology directs the spotlight of our attention, highlighting certain stimuli and omitting others. Among these technologies, artificial intelligence is rapidly becoming one of the most influential in structuring the texture of daily experience.

AI systems excel at detecting patterns and regularities in vast amounts of data and presenting those patterns back to us in some, ideally, useful form. They recommend music and movies, finish our sentences as we type, and guide us through traffic; they alert us when an out-of-the-ordinary charge shows up on our credit card, or when a seismic tremor begins half a world away. In doing so, they act as a kind of artificial attentiveness: tirelessly sifting through streams of information, picking out what we through our choices or the makers through their product design choices would deem noteworthy, and bringing it to our awareness. One might say these systems are told to “pay attention” on our behalf to things no single human could monitor alone. They extend the reach of our senses and our memory. A wearable device can keep track of subtle changes in our heartbeat and nudge us when something seems wrong, effectively heightening the bodily layer of experience. A search engine, armed with AI, can scour millions of documents in seconds and throw into our intellectual field of view facts and perspectives we might never have found on our own. In these ways, AI tools become new actors in the field of experience—not as conscious subjects, but as powerful shapers of what appears before our conscious selves.

Yet it is crucial to remember that, for all their sophistication, these machines do not themselves have insight. They pattern experience but do not ascend from experience to understanding in the way a human mind does. An AI can spot that you often listen to trance on Monday mornings and suggest a playlist—that is pattern recognition, drawn from experience. It cannot, however, share in the why of your Monday blues or joys; it doesn’t genuinely grasp the meaning of music lifting a mood. A language model can predict and produce an remarkably coherent paragraph on any topic, drawing on statistical patterns gleaned from human writing. But it has no actual wonder or curiosity about the topic, no moment of “Aha!” where it understands a truth. It is producing a facsimile of insight—sentences that look like the product of thought—without any actual act of understanding behind them. AI is empty of the kind of rational consciousness that gives human insight life, the affirmation of truth. In simple terms, an AI can rearrange and regurgitate the patterns it has detected, but it does not know what any of it truly means. The insight, the judgment, the felt realization of “So this is how it is!” never occurs in the silicon circuits; it occurs, if at all, in you as you reflect on the output.

In light of this, we can see that AI tools present patterns of experience to us, which is a very different thing. They generate images, words, and predictions that mimic the end-products of understanding, and those products can certainly feed into our understanding. A human reading a well-formulated answer from an AI might come to an insight or at least a good question to pursue. But the machine didn’t leap from perplexity to illumination; it didn’t have the perplexity in the first place. What it gives us are suggestions, simulations, arrangements of data. We remain the ones who must dive into the meaning, ask further pertinent questions, and discern truth from error. In a sense, the latest AI can offer us an abundance of ready-made “answers,” but without the context of a knowing subject behind them, they are answers awaiting a knower to truly make sense of them. We have, then, a kind of mirror: technology reflecting back patterns that originated in human experience, now arranged in novel ways. It falls to our intelligence to decide what to do with those patterns—whether they fit, whether they answer a real question, whether they spark a genuine insight and help us discover Being or lead us down a false trail.

Seen this way, the advent of AI is less about creating artificial minds than about amplifying and reorganizing the field of experience in which our own minds operate. This offers a tremendous opportunity: we can develop AI tools as forms of artificial attentiveness that help us attend more fully and finely to the world. Rather than try to skip over the hard work of inquiry—from data straight to statistically possible “conclusions”—we can use these tools to enrich the data, to illuminate the patterns, to keep track of the threads that our limited attention might lose. A well-designed AI system might function like a perceptual prosthetic, extending our natural capacities. It might scan medical images hour after hour without fatigue, so that a doctor’s attention is drawn unerringly to the faint shadow that could be an early tumor—augmenting the physician’s experienced eyes, not replacing them. It might handle the tedium of monitoring factory sensors or global weather updates, freeing human minds to do the interpreting once an anomaly is flagged. In creative pursuits, an AI might rapidly prototype variations of a design or a melody, not to outdo the artist’s imagination but to give it more raw material to work with, more vistas to explore. In all these roles, the technology does what it excels at—broad, tireless pattern processing—so that human beings can do what we excel at: grasping meaning, arriving at insights, making value judgments and novel connections. The ascent from data to understanding is still ours to climb, but now with better footholds and a safety net below.

Experience, as we’ve seen, is the fertile ground where every inquiry begins. It is the condition and the invitation for insight: only because something is given to us—something seen or heard or felt, something missed or longed for—do we begin to ask and understand. The more richly patterned and aptly focused that given is, the more likely we are to catch sight of the next illuminating idea. If we conceive of AI as a partner in shaping the given, as a means to broaden and sharpen our field of experience, then its value becomes clear. Such technology would not short-circuit the path from ignorance to knowledge; rather, it would light it better, remove a few stones, perhaps extend it into places we couldn’t reach before. It would help us ask better questions and find relevant data, but it would leave the questioning and the knowing—our distinctively human act—intact and truly ours.

Ultimately, the ascent of knowing remains a human journey. Understanding still must rise on the foundation of our own experience, our own insight, our own judgment. What our tools can do is support that foundation. They can widen the experiential field and sharpen our focus, but they cannot substitute for the spark of insight or the resolve of judgment. Their highest promise is to serve as artificial attentiveness—extending our gaze and heightening our awareness—so that in an ever more complex and patterned field of experience, the spark of human understanding has all the more to ignite it and room to soar.

Incantation

By Czesław Miłosz

Human reason is beautiful and invincible.
No bars, no barbed wire, no pulping of books,
No sentence of banishment can prevail against it.
It establishes the universal ideas in language,
And guides our hand so we write Truth and Justice
With capital letters, lie and oppression with small.
It puts what should be above things as they are,
It is an enemy of despair and a friend of hope.
It does not know Jew from Greek or slave from master,
Giving us the estate of the world to manage.
It saves austere and transparent phrases
From the filthy discord of tortured words.
It says that everything is new under the sun,
Opens the congealed fist of the past.
Beautiful and very young are Philo‑Sophia
And poetry, her ally in the service of the good.
As late as yesterday Nature celebrated their birth,
The news was brought to the mountains by a unicorn and an echo,
Their friendship will be glorious, their time has no limit,
Their enemies have delivered themselves to destruction.

The Forgotten Value of Memorization

Modern education often dismisses rote memorization as the enemy of critical thinking. We prize analysis, creativity, and “knowing why,” while merely knowing words by heart is seen as superfluous. Yet our predecessors – from classical orators to 19th-century schoolchildren – believed memorizing poetry and eloquent prose trained something fundamental in the soul. Recent reflections on memorization suggest that learning great works verbatim imparts a qualitatively different kind of knowledge. When you memorize a poem or a stirring speech, you internalize it in a way that silent reading never achieves – you take it “inside you, into your brain chemistry if not your blood,” knowing it at a deeper, bodily level. The very rhythms and cadences of the language become part of your neural architecture.

Memorization has an almost physical effect on us. To memorize is to engage voice and ear, breath and pulse; it is, as one poet laureate put it, to recognize that poetry (and by extension, any language of depth) “is partly a bodily art”. Learning a passage by heart marries sound with sense: the tongue learns the shape of the words, the ears their tune, long before the reasoning brain parses their full significance. This process “builds inside the subconscious a command of the spoken word” and a comfort with language’s music. Over time, memorized lines form mental grooves – patterns of thought and feeling that our minds can fall into like a well-worn path through a forest. We are patterning our experience before we’ve had the experiences; we are carving channels for future insights to flow. Such interior patterning creates readiness. When the day comes that life delivers an emotion or idea that fits one of those well-formed grooves, it slides into place, and suddenly the words we’ve been carrying spring to life with understanding. This is what Rilke means, I think when he says, “I want to beg you, as much as I can, dear sir, to be patient toward all that is unsolved in your heart and to try to love the questions themselves like locked rooms and like books that are written in a very foreign tongue. Do not now seek the answers, which cannot be given you because you would not be able to live them. And the point is, to live everything. Live the questions now. Perhaps you will then gradually, without noticing it, live along some distant day into the answer.”

Form, Rhythm, and Cadence: Patterning the Mind

Memorizing poems, speeches, and great writings is a way of teaching ourselves the forms of meaning before we grasp the meanings themselves. The strict meter of a sonnet, the rise and fall of a well-turned phrase, the logic of a powerful argument – by memorizing these, we learn the shape of insight in an almost musical way. We become like apprentices practicing scales, not yet aware of the songs we will one day play. In this sense, memorization is closer to art than to rote data retention. The form imprints on your mind; you come to feel what a balanced sentence or a poignant line of verse is like from the inside. This inner familiarity readies you to recognize beauty and truth when they appear elsewhere. The great G.K. Chesterton – himself a master of paradox and memorable prose – knew the power of form and inversion. In a Chestertonian spirit, we might say: memory is the soil in which understanding takes root. The steady iambic rhythm of Shakespeare, the soaring cadence of a Lincoln speech, a lilting Wodehouse phrase, the pith of an Oscar Wilde aphorism – these create an inner echo. Later, when you encounter a situation in life that resonates with those rhythms or sentiments, the understanding that was dormant in the memorized words can awaken. As one writer noted, to truly take a work to heart is to know it by heart – implying that the heart comprehends in its own time, often sooner than the head.

Crucially, memorization also cultivates patience and humility before wisdom. It is an act of faith in meaning: you commit the words to memory, trusting that one day you will grasp why they are worth knowing. In the interim, those words live in you, quietly working on your imagination. They teach you the patterns of insight – the feel of a thought well expressed – so that you’re prepared to catch meaning when it comes. Think of it as learning the cadence of truth. Even if a child reciting the Gettysburg Address doesn’t understand “the last full measure of devotion,” the solemn cadence of those phrases is planting a seed. Years later, perhaps when that grown child visits a war memorial or faces a test of loyalty, those very words may surface with a profound clarity, finally comprehended. The memorized form was a vessel, waiting to be filled with lived understanding.

When Memory Finds Meaning

Memorized language has a way of biding its time within us, often surfacing at unexpected moments when we need it most. Many of us have experienced a line of poetry or an old proverb suddenly resonating after years of lying dormant in memory. It can feel almost eerie – as if our past self packed a gift for our future self to open when the time was right. This is the lived reality of countless people across cultures.

Such moments are deeply personal, yet they are also part of our cultural memory. History is full of cases where memorized words provided guidance or solace long after they were learned. The poet Joseph Brodsky, exiled to the Arctic by an oppressive regime, survived his solitude in part by reciting lines of poetry he had memorized in happier times. He had once shocked his students by making them memorize hundreds of lines, believing “they might need such verses later in life.” He was right – he needed them too, and he was “grateful for every scrap of poetry he had in his head” during those frozen, lonely days. In Brodsky’s case, memorization was literally a lifeline: a mental store of meaning and beauty to draw upon when the world offered none. While most of us don’t encounter these straits, a similar sort of joy can be savored on a winter’s walk with one’s labradors, language for the resplendent natural wonders all around, spilling unbidden from heart, mind, and lips.

Memorized words can also steer us morally and emotionally when our own words fail. Consider the ancient prayer of the Judeo-Christian churches: the Psalms. Routinely recited by millions around the world, they touch on every emotion, experience, and desire. Such remembered lines can remind you “of lifetime values” exactly when you need them. In a less universal but no less significant vein, think of the phrases from beloved speeches or songs that have leapt into your mind during pivotal moments. The ringing words “Day shall come again” might embolden you during a personal struggle for justice; the simple prayer “The Lord has given and the Lord has taken away, Blessed be the Name of the Lord” might steady your heart in times of anger or grief. We often memorize these kinds of texts in community or childhood, long before we face the situations that give them full meaning. When the moment arrives, the words we stored up become like old friends showing up to support us.

Even on the grand stage of history, memorization has prepared the ground for insight and courage. One famous anecdote tells of young Winston Churchill committing to memory all 70 stanzas of the poem “Horatius at the Bridge,” believing it would fortify his courage. He reportedly continued to recite it throughout his life, the tale of ancient heroes holding a bridge against all odds becoming part of Churchill’s internal arsenal of resolve. It is not hard to imagine those cadences echoing in his mind during Britain’s darkest hours in World War II, steeling him with the very words he had absorbed as a boy. The language we memorize becomes a lens through which we interpret our experiences. In Churchill’s case, a memorized epic about bravery may have helped him recognize what bravery required of him decades later. In quieter ways, each of us carries similar lenses. Our minds are populated with the voices of poets, prophets, and statesmen we have learned by heart. They speak to us, helping us make sense of new events. Memorization, then, is not about parroting the past; it is about creating an inner chorus that can sing when we have lost the tune, or shine light when we walk in darkness.

Notably, the benefits of memorizing without immediate understanding extend beyond personal epiphanies; they enrich culture by preserving wisdom until society is ready to heed it. Many cultural moments have turned on a community suddenly grasping the significance of words learned long before. Think of how generations of Americans recited the opening of the Declaration of Independence – “We hold these truths to be self-evident, that all men are created equal” – without fully acting on those truths. Those words were memorized, ingrained in the national psyche, awaiting a future insight. In the Civil Rights Movement, that memorized ideal found new life as people demanded it be realized in practice. Memorized language can be a seed of progress: planted in memory, dormant through seasons of incomprehension, and then flowering when conditions are right. In this way, memorization is an act of cultural faith that the meaning will come when we are mature enough to receive it.

Conclusion: Hospitality to Future Understanding

Memorizing a poem or a piece of great writing is a bit like preparing a guest room in the house of your mind. You furnish it with the words, the sounds, the structure – all in anticipation that meaning itself may one day come to stay. At first, you don’t know your guest well; you only know that the words seem worth inviting in. You repeat them, keep them alive and warm, offering the hospitality of your memory. And then, perhaps years later, insight arrives and finds everything in place – the rhythm of the language, the form of the thought, the cadence of truth already waiting like an open door. In that moment, memorization vindicates itself. The words speak up when they are most needed, “whispering a truth at risk of being forgotten” and even whispering beauty into our ears. What once was mere repetition becomes revelation.

To memorize things we do not yet understand is to practice a humble form of hope. It is to say: I will carry these words in trust that one day I, or someone I love, or even my whole community, will find wisdom and beauty in them. It is to be, in a sense, a guardian of meaning that has not yet been realized. This habit requires a certain generosity toward our future selves – a willingness to give ourselves gifts that we cannot yet unwrap. In an age obsessed with instant comprehension and utility, memorization stands out as an act of intellectual hospitality and patience. It keeps alive the possibility that there are truths we almost understand, melodies we almost hear, and that by holding onto the words now, we make ready the interior space for understanding to eventually blossom.

So let us defend the practice of memorizing poems, speeches, and great writings even before we fully grasp them. Let us memorize not to parrot facts or appear learned, but to form our minds in readiness. We are stocking the larder of the spirit with nourishing words against the hungers of some future day. We are laying down emotional and intellectual patterns that can guide us when new challenges arise. We are making our very selves a home for insight not yet born. In the final analysis, memorization is an expression of trust in the meaningfulness of language and life. It is, as we have seen, a quiet and powerful way of showing hospitality to future understanding – a way of saying you are welcome here to truths that will arrive in their own season, and finding ourselves richly prepared to recognize them when they do.

Hans Urs von Balthasar: Seeing the Form and the Glory

Hans Urs von Balthasar (1905–1988) places beauty at the very heart of Christian understanding. He lamented that the modern world has lost the courage to affirm beauty, treating it as a frivolous mask, a “mere appearance” easily discarded. Balthasar believed this loss is dire, because beauty, truth, and goodness are inseparable “sisters.” In his words, “beauty demands for itself at least as much courage and decision as do truth and goodness,” and if we banish beauty, she “will not allow herself to be separated and banned from her two sisters without taking them along”. In other words, when we sneer at beauty as superfluous, we do mysterious harm to our grasp of truth and our capacity for goodness. Balthasar even warns that “whoever sneers at [beauty’s] name … can no longer pray and soon will no longer be able to love”, suggesting that without openness to beauty, the heart grows incapable of deeper connection and communion.

At the core of Balthasar’s philosophy is the idea of form and glory. He uses the term “form” (or Gestalt) to mean the visible shape of something, that is, the pattern or appearance by which we recognize order and meaning. But this form is radiant with an inner truth. In theological terms, the splendor shining through a beautiful form is the glory of God. Balthasar famously wrote, “the form of the beautiful is the glory of God (kabod, doxa) whose splendor seizes and enraptures”. When we truly encounter something beautiful – whether a work of art, a noble action, or the face of a loved one – we behold a form that mediates a deeper glory. There is a visible aspect (form) and an invisible depth (glory) united in the experience of beauty. In Balthasar’s view, this twofold structure of beauty reflects the Christian belief that the infinite splendor of God became tangible in the finite form of Christ. Thus, all earthly beauties are finite hints of an infinite glory. They beckon us, through delight, toward what is beyond themselves. We do not possess beauty by analyzing it to death; we behold it. Balthasar emphasizes a stance of reverent perception, a loving gaze, rather than aggressive scrutiny. “Form is comprehended by ‘beholding,’ while splendor is understood by ‘being enraptured,’” he explains. In this beholding, we experience what he calls “the glory of the Lord” shining through creation’s forms.

In Balthasar’s perspective, then, beauty is far more than subjective taste. It is a transcendental, a property of Being itself like truth and goodness, all of which “penetrate each other.” To encounter beauty in the world is to feel, however faintly, the splendor veritatis: the splendor of Truth and the warmth of the Good. Thus beauty has an indispensable role: it encourages us toward the fullness of reality. Balthasar’s writing has a passionate, urgent tone on this point, so much so in fact I read his “Heart of the World” every Holy Week with tears streaming down my face. He saw a culture in danger of cynicism, no longer trusting Beauty and thereby unhinging the True and the Good. By reuniting these sisters, by “seeing the form” with an attentive and childlike awe, we rediscover a path toward faith and love. Beauty, for Balthasar, is not a luxury but a lifeline: a flash of Divine glory that can wound us with longing and joy, awakening us to desire what is true and good in themselves. In his theology of beauty, one could say the beautiful is the visible glory of God, inviting us to taste and see that ultimate Reality is Love.

Johann Wolfgang von Goethe: The Symbol of Metamorphosis

Johann Wolfgang von Goethe (1749–1832) approached beauty as a poet who was also a scientist of life. Famed for his literary works like Faust and for scientific studies such as The Metamorphosis of Plants, Goethe had a holistic vision of nature and art. He believed that beauty is the revelation of deeper truth through a living form. In a striking maxim, Goethe wrote, “Beauty is a manifestation of secret natural laws which without this appearance would have remained eternally hidden from us.”. In other words, what we find beautiful in the world – the curve of a flower, the colors of a sunset, the harmony of a well-crafted poem – is not arbitrary. It is the shining-through of something True. Nature’s secrets, otherwise concealed, “manifest” themselves in the language of beauty. This idea resonates with the ancient notion that the world speaks to us in symbols.

Symbol was a key concept for Goethe. He distinguished a genuine symbol from a mere allegory. An allegory is a one-to-one code (this rose means love, for example); it points beyond itself in a fixed way. A symbol, however, is alive: it presents itself for what it is, yet in its very presence it hints at something more. Goethe described true symbolism as “the specific [particular] represents the general, not as dream and shadow, but as a living-momentary revelation of the ineffable.”. In a symbol, the concrete thing (a rose, a stone, a human gesture) is itself fully and at the same time reveals a deep universal truth or idea – but it does so indirectly, by being authentically itself. For Goethe, the symbol arises from “profound feeling” and “seems to stand only for itself and yet is profoundly significant” of the ideal it embodies. A classic example might be a flower in one of Goethe’s poems: he doesn’t treat it as a code for an abstract concept, but by depicting it vividly and lovingly, the reader intuits a meaning beyond the literal. The beauty of the symbol is that it communicates truth through its form, not by stepping outside of it. This idea allowed Goethe to see works of art and natural phenomena as epiphanies – moments where the visible form and invisible meaning coincide. Beauty, then, is not a surface gloss; it is the signature of truth, “a welcome guest everywhere” that makes the truth attractive and accessible to us.

Goethe’s fascination with metamorphosis deepened this perspective. In his botanical studies, he observed that a plant’s disparate parts (leaf, petal, stamen, fruit) are all variations of a single underlying form. The entire life of a plant is a metamorphosis – a developmental unfolding of one form into many shapes. He posited an archetypal plant (Urpflanze), an ideal pattern that expresses itself in all botanical forms. What does this have to do with beauty? It means Goethe saw unity in variety, lawfulness in growth. A blossom’s delicate beauty is meaningful because it reveals the inner law of the plant’s life. The endlessly transforming shapes in nature are like a flowing script, spelling out an inner order. To truly see a plant (or any natural thing), Goethe felt one must employ “exact sensory imagination” – a contemplative observation that is both scientific and artistic. This involves a receptive stillness and intuitive insight, much like an artist beholding a scene. In such observation, the outward beauty of metamorphosis leads the mind to an inward truth: a symbolic truth that nature’s processes are purposeful, coherent, and connected to our own life. Goethe even extended this idea to the human realm. In his poetry and literature, characters and images often undergo transformations, suggesting that human life itself is a metamorphosis guided by ideals like the “Eternal Feminine” (which, at the end of Faust, draws the hero upward). Beauty, for Goethe, has thus a double aspect: it is morphological (we perceive the harmonious form, the metamorphic pattern) and symbolic (we intuit through it a higher meaning or law).

While Goethe’s language is less explicitly theological than Balthasar’s, it is spiritual in its own way. He often found in nature a living presence that educates the soul. To take delight in the beauty of a leaf or a cloud was, for Goethe, to receive a quiet lesson in truth. “Everything is a leaf,” he famously wrote of plants – a poetic way to say that one simple form underlies the diversity of creation. Likewise, one might say everything true is also beautiful, in Goethe’s worldview, because only through the manifestation (appearance) can we grasp the hidden principle. He invites us to approach beauty with a certain reverence, as one would approach a mystery. Rather than dissecting a phenomenon in a utilitarian way, we are asked to behold it as a symbol that can open the heart. In this approach, beauty is the path of insight. It trains us in what Goethe called Anschauen, a deep seeing. We begin to perceive the “secret natural laws” at play and to sense our kinship with the rest of creation. The convergence of art and science in Goethe’s life testifies that beauty bridges the intuitive and the intellectual. It is the way truth incarnates itself to our senses. If we forgo this path – if we reduce a rose to a chemical analysis or a poem to a didactic message – we lose the wholeness of meaning. Goethe’s vision of beauty calls us to restore our capacity for wonder, to let each beautiful thing speak its truth in its own language. In that attentive wonder, we find that beauty is a teacher of truth and a wellspring of joy.

Josef Pieper: Leisure, Clarity, and the Gaze of Love

Josef Pieper (1904–1997), a German Catholic philosopher, approached beauty from the angle of human experience and virtue. Writing in the mid-20th century, Pieper was concerned that modern hyperactivity and utilitarianism were dulling our ability to see reality. In his classic book Leisure: The Basis of Culture, he argued that true knowing requires a receptive stillness – what the ancients called scholé (leisure). Pieper extends this insight to beauty: to appreciate beauty, we must contemplate it, not grasp at it. He describes leisure as “a form of stillness that is the necessary preparation for accepting reality; only the person who is still can hear, and whoever is not still, cannot hear”. In other words, beauty has something to tell us, but we can catch its message only in a state of open, unhurried attention. When our eye “simply looks” at a rose, neither analyzing nor exploiting it, we perform a kind of sacred reception. We allow the rose to shine in its own glory. This “receptive understanding, contemplative beholding, and immersion in the real” is exactly what Pieper means by leisure. It is an attitude of clear seeing unclouded by the will-to-use or the anxiety of constant busyness.

Pieper, drawing on the philosophy of Thomas Aquinas and the tradition of the transcendentals, affirms that real beauty is tied intrinsically to truth and goodness. He echoes the definition of beauty as “the glow of the true and the good radiating from every ordered state of being”. Here, Pieper uses the term clarity (from Latin claritas) to describe the radiance that beautiful things have. It is not a superficial dazzle, not “the patent significance of immediate sensual appeal,” but a more profound illumination. When something is beautiful, it “glows” with meaning – the truth and goodness of it shine out as a kind of inner light. A common example is how a virtuous act can strike us as beautiful: we see the goodness in it, and that recognition is experienced as radiance. Or consider a well-crafted chair: its beauty lies not only in pleasing shape, but in the rightness of its form (true to its purpose) and the goodness of its being (it fulfills what it should be). Radiance, or clarity, is “the essence of beauty,” Pieper would say. It engages both our senses and our spirit. We are attracted by the appearance, but what holds us is the encounter with the thing’s reality shining through that appearance.

An important aspect of Pieper’s teaching on beauty is the moral and spiritual condition of the observer. He insists that a certain purity of perception is needed. “Only those who look at the world with pure eyes can experience its beauty,” he writes. This purity is not something narrow or prudish; it means a self-forgetful openness, an unclouded vision, a sort of humility. A person driven by selfish desires or jaded by cynicism will see only surfaces or use things as means to an end. But one who has a “chaste sensuality” – who can take delight in what is seen or heard for its own sake – gains the capacity to truly perceive beauty. This idea connects to Pieper’s broader point that temperance, or self-mastery, actually “creates beauty” in the soul and even in one’s bodily presence. A pure heart can see the world with wonder; a cluttered heart cannot. In practical terms, Pieper is suggesting that our ability to see beauty is linked with virtue and grace. It is, in a sense, a gift that must be received in stillness and humility. Think of how children marvel at simple things – their eyes are fresh, unburdened by utilitarian concerns. We too, says Pieper, must cultivate a leisureliness and clarity of soul to apprehend beauty in depth.

Pieper’s perspective converges with Balthasar’s and Goethe’s in affirming beauty as essential. He notes that in the ancient and medieval view, “without beauty the ancient world refused to understand itself”, whereas our contemporary world has largely “bid farewell” to beauty, to its own detriment. In Pieper’s gentle but penetrating prose, one hears an invitation to recover a sense of festivity and awe. “Leisure lives on affirmation,” he writes, on a “lingering gaze of the inner eye on the reality of creation.” This leisurely gaze is celebratory: it says “Yes” to the world as something inherently good and beautiful, not just as raw material for our projects. Pieper even connects beauty to celebration – the highest form of leisure – noting that in festivals we decorate, sing, and adorn precisely to honor the goodness of existence. Ultimately, for Pieper, beauty is a life-giving visitation. It sparks joy, it instills quiet, and it nourishes the spirit with truth. In a culture of “total work,” making space for beauty brings sanity. It re-centers us on what truly matters: the givenness of the world and the gift of being. To behold beauty is to practice a kind of contemplative love toward reality, acknowledging that being is good and worth delighting in. This, Pieper implies, is at the very root of culture and knowledge. Without that affirming, clear-eyed contemplation, our world of restless productivity becomes barren. With it, even a moment of noticing the “exquisite natural work of art” in a hovering wasp can become an encounter with truth and goodness.

Beauty as a Path

Despite their differing contexts, Balthasar, Goethe, and Pieper share a profound agreement: beauty is a path to the very core of reality. All three reject the idea that beauty is merely subjective or a trivial amusement. Instead, they see it as a means of revelation. In Balthasar’s theological language, beauty is a “transcendental” that is one with truth and goodness, a vital way in which Being speaks to us. Goethe, from his humanistic and scientific angle, likewise upholds that beauty reveals “secret laws” of nature and thus guides us to understanding. Pieper explicitly calls beauty “the glow of the true and good,” underscoring that it radiates objective value and meaning. For all three, then, beauty has an epistemic role: it helps us know. It is evidence of something (be it God’s glory, the order of nature, or the truth of Being). This is a notable convergence: at root, they would all nod to the formula that beauty is the splendor of truth.

Another convergence lies in the demands beauty makes on the beholder. Each thinker, in his own vocabulary, stresses the importance of receptivity. The beautiful cannot be seized by force or dissected without loss. Balthasar speaks of beholding and allowing oneself to be enraptured by glory – a stance of surrender and awe rather than control. Goethe, with his idea of the symbol, requires the reader or observer to “see the general in the particular” intuitively, which calls for patience, empathy, and imaginative insight. Pieper’s language of leisure and purity of heart is all about making oneself open – becoming still, clearing away self-interest, so that beauty may “voice” itself to us. In all cases, there is a sense that beauty approaches us as a gift. We receive it passively before we can analyze or use it.

The three thinkers converge in suggesting that our capacity to encounter beauty is linked with moral and spiritual qualities: humility, purity, attentiveness, courage to acknowledge greatness. Indeed, they might agree that the inability to see beauty is not a mark of sophistication but a kind of blindness or even dysfunction of the soul. As Balthasar dramatically put it, a world that can no longer affirm beauty will eventually be unable to affirm anything of value, even love. Pieper would concur, having noted that loving appreciation is prerequisite to knowing the full truth of anything. Goethe too would likely warn that a cynical eye misses the living meaning in phenomena and ends up in a dead world of “half-truths.”

Beauty in the Age of the Machine

So where does that leave The Oatmeal’s comic about AI art? His absurdly drawn half lament, half satire? Recall, in it, he plays with the absurdity of machines generating “art” from statistical patterns of human work: images that look beautiful but feel hollow. His tone is humorous, but his intuition is deadly serious: something essential is missing. The comic’s protest is not against technology itself, but against a world where imitation begins to masquerade as inspiration.

Hans Urs von Balthasar, Johann Wolfgang von Goethe, and Josef Pieper would each recognize this disquiet.

Balthasar would see the loss of glory. For him, true beauty is form radiant with inner splendor—a visible surface that discloses invisible depth. AI art mimics the form but not the glory. It offers the silhouette of meaning without the light. When beauty is severed from truth and goodness, it decays into surface, into ornament, into noise. What The Oatmeal feels is not nostalgia but metaphysical hunger: the ache for radiance that cannot be manufactured.

Goethe would see the loss of life. His aesthetics of metamorphosis depend on the living process by which form unfolds and reveals its inner law. The algorithm, though capable of astonishing synthesis, does not grow. It transforms data, not being. Its blossoms are pressed and perfect, but lifeless. Goethe’s artist participates in nature’s own creativity—seeing, waiting, discerning. AI’s art, by contrast, is an echo of appearances without breath or becoming.

Pieper would see the loss of leisure—the contemplative space where beauty can be received as gift. The machine never beholds; it only calculates. But the greater danger lies in our own imitation of that restlessness. In treating art as instant output, we lose the stillness required for delight. Pieper’s warning rings true: a culture that cannot be silent before beauty soon forgets what beauty is.

And yet, all three would also recognize the strange grace of this moment. Counterfeit beauty can sharpen our appetite for the real. When imitation floods the world, discernment awakens. The Oatmeal’s comic, in its wry exasperation, performs a kind of negative theology of aesthetics: by naming what beauty is not, it helps us remember what it is.

For Balthasar, this remembering is a call to behold once more the form of glory—the radiance of Being itself, revealed even in the ordinary. For Goethe, it is an invitation to see the living law behind appearances, to recover the symbolic depth of nature. For Pieper, it is a reminder that contemplation—unhurried, humble, grateful—is the only posture by which Beauty can be known.

Perhaps that is beauty’s final secret in the age of the machine: it cannot be automated. Only the human heart, still and receptive, can behold the world until form and glory meet again, until the true radiance of reality as it is shines through.

In other words, “Beauty will save the world.”

Experience vs. Understanding: The Human Stages of Knowing

Human knowing progresses through distinct phases. First comes experience – the raw data of our senses, the notes we take, the words we hear. This is followed by insight – those “aha” moments where patterns coalesce and meaning emerges. Next is judgment, where we critically evaluate insights for truth or significance. Finally, decision carries understanding into action. While modern AI “thought partners” dramatically enhance the experience phase – capturing more data, organizing it better, surfacing patterns – they of course cannot actually understand in the human sense. They pattern our experience but do not generate true insight or evaluate truth. In other words, these tools help us gather and sort the ingredients for understanding, but the cooking (insight) and tasting (judgment/decision) remain our task.

This distinction is crucial for building AI experiences that resonate with our human experience. An AI tool might transcribe a meeting and highlight key points, but recognizing a novel opportunity in those notes is a human insight. A scheduling assistant can rearrange tasks optimally, but deciding which goals truly matter reflects human judgment and priorities. As we examine different interface paradigms below, we will see that each tool excels at structuring a certain kind of experience – conversational, associative, spatial, chronological, or auditory – to support our cognition. Each patterns the user’s experience of information in a distinct way, making certain elements salient or accessible. By doing so, these tools serve a preparatory function: they set the stage for insight. They ensure that when the moment is right, the human thinker has the right data, context, or prompts to leap from mere information to genuine understanding. What they don’t do is take that leap for us.

Chat Interfaces: Conversation as a Cognitive Scaffold

One prominent paradigm is the chat interface, exemplified by tools like Mem (often touted as an “AI thought partner”) and the AI chat features in Reflect. These systems allow users to interact with their notes or knowledge base through natural language dialogue. By turning knowledge work into a back-and-forth conversation, chat-based tools pattern our experience in a very human way: through questions and answers, prompts and responses, much like an interlocutor guiding our attention.

For example, Mem’s interface encourages users to “shoot Mem a quick message” with information or queries[1]. The tool then recalls or organizes notes in response, effectively simulating a Socratic assistant. Mem’s Agentic Chat can even act on your notes – “create, edit, and organize notes for you” in response to commands[2] – blurring the line between simple Q&A and an active collaborator. The experience of using Mem thus feels like having a dialogue with “your mind’s better half,” as the website puts it[3]. By chatting with this “mind’s better half,” users externalize their stream of thought into a responsive medium. The tool patterns the experience by making knowledge retrieval or brainstorming a sequential, conversational flow rather than a static search query or manual lookup.

Similarly, Reflect’s AI features allow you to “chat with your notes” to uncover connections[4]. Instead of manually hunting through a graph of notes, a user can ask the AI to surface related ideas or even pose questions to one’s own past writings. The chat interface thereby patterns the experience of recall: it feels like asking an expert librarian (who happens to know all your thoughts) for assistance. This conversational pattern is powerful because human cognition itself often works through internal dialogue and questioning. By mirroring that, chat-based tools can prompt us to articulate what we’re looking for, which is often the first step toward insight.

However, it’s important to note that while these chat interfaces can elicit information and even surface latent connections, they do not supply the insight for the user, indeed, even another human cannot supply an insight. They provide answers based on the data they were trained on or the notes you’ve given them. For instance, Mem’s AI might produce a well-structured answer drawing on everything you’ve saved[5], but whether that answer contains a novel insight or just a summary is up to you to determine. The deeper pattern-recognition – seeing an analogy, posing the right question, discerning significance – remains a human capability. In effect, chat tools scaffold our cognition; they keep the conversation going, surface what we might need, and thereby create fertile ground for insight to occur in our minds.

Graph Interfaces: Networks of Association and Context

In contrast to linear chat, graph-based interfaces pattern experience by emphasizing connections and relationships in a knowledge network. Tools like Tana and the backlinking features of Reflect (and their cousins like Roam or Obsidian) fall into this category. They present information not as isolated files or notes, but as a web of nodes interconnected by links, tags, or references. This approach aligns with the way experts often structure knowledge – as a richly interconnected map rather than a simple list.

Tana, for instance, is described as “a knowledge graph outliner” where “every item is part of an underlying graph” of data and ideas[6]. Instead of mimicking static pages, Tana treats notes as nodes that can have properties and links; it gives users “powerful primitives” like Supertags and filtered views to organize thoughts in a non-linear fashion[7][8]. The result is an environment where simply entering information automatically weaves it into a broader context. One founder explains that in Tana, “everything that you do... is all automatically organized and connected together” in the system[9]. This means your meetings, notes, and tasks link to relevant people, dates, projects, and so on, creating a network of associations without the user manually filing each thing in a single folder. The experience of knowledge is thus patterned as a graph: richly associative, with any given note potentially just one hop away from related material.

This graph paradigm patterns our experience of recall and sense-making by externalizing context. Instead of relying solely on our brain to recall how Idea A relates to Project X or what source was connected to concept B, the tool visually or structurally represents those links. The cognitive load of remembering relationships is offloaded to the system, freeing our mind to scan the web of connections and perhaps notice patterns we hadn’t considered. For example, if one uses Reflect or Tana to tag notes (say tagging a note as a “Book” automatically creates fields for author and year, etc.), then over time one can query or view a network of all books linked to themes or projects[10]. This might reveal that a concept from last month’s research connects to a meeting note from yesterday – a connection we might have missed without the graph view.

That said, a graph of information is not equivalent to insight. A visual knowledge map can suggest a pattern (“these two domains are linked by this common tag – interesting!”) and thereby set the conditions for an insight, but the actual realization (“aha, that’s the connection I should explore!”) occurs in the user’s mind. Indeed, some users of graph-based tools note that while the structured metadata and queries give rigor, one can still miss the meaning until one reflects on why those connections matter[11]. Graph interfaces can present an “overview” of our personal knowledge network and lighten memory load[12][13], but interpreting that overview – distinguishing significant links from noise, or hypothesizing why a pattern exists – is a higher-order act. As such, graph-centric tools are excellent patterners of experience: they ensure that the raw material for insight (all those connections and references) is readily at hand and not forgotten. They hand us the pieces and occasionally highlight how pieces cluster, but assembling them into a novel insight remains our job.

Canvas Interfaces: Spatial Thinking and Visual Organization

Another emerging paradigm uses spatial canvases to pattern experience, as seen in tools like Heptabase. Instead of or in addition to linear documents, Heptabase provides an infinite two-dimensional whiteboard where notes are visual cards that can be arranged, grouped, and linked freely. This leverages our spatial cognition – the human knack for memory palaces and mind maps – to organize complex information. A canvas interface patterns experience by making the workspace itself an extension of thought, where where something is placed and what it’s near conveys meaning.

Heptabase explicitly markets itself as “the visual note-taking tool for learning complex topics.” It “helps you make sense of your learning, research, and projects” by letting you literally lay them out visually[14]. In Heptabase’s whiteboard metaphor, you might spread out different idea cards on the screen as if pinning notecards on a desk or wall. You can cluster related notes in a corner, draw connecting lines, use color coding, and even create sub-whiteboards for nested topics[15][16]. This approach is akin to dumping out puzzle pieces (your thoughts and snippets) and moving them around until a picture emerges. Users describe the whiteboard as “a space for thinking... an indefinitely large desk” where you can rearrange and link cards to visualize connections that might not be apparent in a linear list[15]. In contrast to “top-down” writing in a traditional doc, this spatial free-form approach can reveal clusters and gaps in knowledge at a glance.

By patterning experience spatially, canvas tools tap into a fundamental mode of human sense-making: visualization. Relationships that are hard to see in prose often pop when laid out on a canvas. For instance, Heptabase allows drawing different styles of connecting lines (curved, straight, arrowed) and grouping cards into sections with various colors to denote meaning[16]. A researcher could place key studies in one area, notes on methodology in another, then draw lines to show which study supports which idea. The spatial arrangement itself becomes a language, a set of cues to the eye and mind about what is connected or prominent. Studies on concept mapping show that such visual structuring can reduce cognitive load and improve understanding by “making relationships visible” and offloading some memory work to the diagram[12][17]. In essence, a canvas externalizes part of the thinking process: it lets us think by arranging.

However, the canvas doesn’t supply the insight – it provides the medium in which insight might arise. A whiteboard full of notes could just be a prettier form of chaos (a “hairball” of ideas[18]) unless the user actively perceives a pattern or organizes it meaningfully. Heptabase and similar tools mitigate this by giving structure options (e.g. mind-map mode, collapsible clusters[19]), but ultimately the significance of any spatial arrangement is up to the thinker to determine. The tool may pattern the experience by showing “everything in one space” and enabling “flow of thinking”[14], yet the creative leap – say, realizing that a citation on one card actually solves a problem written on a distant card – is the user’s leap. Canvas interfaces thus excel as sandboxes for experience: they gather and present the experiential data (notes, snippets, references) in a way that mirrors how our mind might spread out a problem on a table. By doing so, they increase the chances that we literally see a connection and thus spark an insight. The design of the interface encourages exploration and juxtaposition, laying groundwork for insights that a more rigid format might obscure.

Timeline Interfaces: Structuring Experience in Time

Human experience is inherently temporal – we live through sequences of events and tasks. Timeline-based tools leverage this by organizing information and obligations along the dimension of time, thus patterning our experience chronologically. Two notable examples are Motion, which uses AI to schedule and manage tasks in time, and the timeline views or daily journals in tools like Fabric and Reflect.

Motion is an AI-powered calendar and project planner that essentially automates the timeline of your work. Instead of leaving you to manually plan your day or week, Motion continuously analyzes your tasks, deadlines, and meetings to “automatically schedule your tasks, meetings, and projects” in an optimized calendar[20]. The experience of using Motion is that of having a diligent secretary who rearranges your schedule on the fly: tasks are time-blocked around meetings, priorities are balanced, and if something changes (an urgent meeting pops up), the rest of your timeline shifts accordingly[21][22]. In short, Motion patterns your workflow by imposing an intelligent order on it. Rather than a to-do list you must triage each morning, you get a dynamically maintained timeline of what to do when. Users of Motion report a sense of relief and focus – the tool prevents overcommitment by “balancing workloads” and even preserves deep work time by curbing needless meetings[21]. The timeline interface here is more than visual; it’s operative. It changes your lived experience of work by structuring time itself, ideally freeing you to concentrate on execution rather than on constantly deciding what to do next.

Other tools use timelines in a more retrospective or note-organizing sense. Fabric, for example, includes a Timeline view that shows everything you’ve captured (notes, files, ideas) in chronological order[23]. It even gives weekly AI summaries of your captured content – e.g., “This week, you predominantly saved files related to web design efficiency…” – to encourage reviewing and reflecting on what you took in[24]. This patterns the experience of personal knowledge by reminding you of the temporal context: what you were thinking about or collecting at a given time. Rather than your notes floating unanchored, they’re situated in the flow of your life. Reflect similarly builds around daily notes that form a journaling timeline. As one user noted, Reflect’s daily notes “are viewable in a scrolling chronology”, allowing you to scroll back through days as if flipping through a diary[25]. This design patterns your experience by naturally integrating memory and time – yesterday’s meeting notes, today’s ideas, tomorrow’s plans all link through the calendar. The effect is to strengthen continuity and context: you don’t just see a note, you recall when and in what context it emerged.

By structuring information temporally, timeline interfaces prepare the ground for certain types of insight. For instance, noticing trends or changes over time (perhaps Fabric’s summary reveals you’ve been repeatedly interested in a topic across weeks) can lead to reflective insight: “Why do I keep focusing on this? Is there a project here?” Or having a clear schedule via Motion might free cognitive resources so that while the AI manages the timeline, you get your insight in a calm moment it safeguarded for you. Still, the judgment of time priorities – deciding what truly deserves a slot on the calendar – is ultimately human. Motion may propose an optimized schedule, but only you can judge that perhaps spending an hour with a creative hobby is more valuable for your long-term well-being than squeezing in yet another micro-task. Likewise, a timeline of notes prompts you to discern which past notes are worth revisiting or synthesizing. In essence, timeline tools ensure that experience is ordered, which is immensely helpful: a well-ordered experience is easier to learn from. But the learning itself – drawing lessons from the past or deciding on future directions – remains in our court.

Voice Interfaces: Capturing the Ephemeral and Making it Visible

Voice is the oldest interface of knowledge – the spoken word. Modern AI tools that deal with voice transcribe and analyze our conversations, effectively turning ephemeral experience into persistent, searchable data. Otter.ai is a prime example (and one I use all the time), as are voice memo features in apps like Reflect. These tools pattern our experience by making spoken interactions – which used to vanish into the air – tangible and structured.

Otter.ai serves as an AI meeting assistant that records, transcribes, and summarizes discussions automatically[26]. In real time, Otter will capture what each person says in a meeting (even attributing speakers) and display it as text that one can highlight or comment on. By doing so, it patterns the experience of meetings in several ways. First, it provides live feedback – seeing a transcript as you talk, which can subtly influence clarity (knowing that what’s said will be logged encourages people to be a bit more organized in speech). Second, it ensures nothing is lost: every decision, idea, or action item uttered is documented. Otter even uses AI to extract key points and action items so you don’t have to hunt for them[27][26]. In effect, it imposes a structure (text, bullet points, summary) on the inherently fluid experience of conversation. The old experience was: “What did we decide in that meeting yesterday?” and everyone flips through sketchy notes or relies on memory. The new experience, with Otter, is that you can literally search the transcript or read the summary – the meeting’s content has become explicit data. The conversation is patterned into a shareable, reviewable artifact.

Reflect’s voice transcription feature likewise patterns personal experience: you can “express thoughts” in speech and have them transcribed straight into your daily notes[28]. This lowers friction for capturing ideas – a sudden thought while on a walk can be spoken into your phone and later appear in your note system, ready for you to review. By integrating Whisper (OpenAI’s speech-to-text)[29], Reflect ensures that the experience of reflection (talking through a problem to oneself) gets turned into text you can see and organize. In doing so, it patterns even your internal monologue into something more concrete.

Voice interfaces underscore perhaps most starkly the limit of AI “understanding.” The transcription and summarization can be amazingly accurate (users report Otter’s transcripts are ~95% accurate in good conditions[30]). You might get the illusion that the AI understands the meeting because it can answer questions about what was said or generate follow-up emails[31][32]. But what it truly provides is an accurate and structured record of experience. The deeper comprehension – for instance, grasping why a decision was important or creatively synthesizing ideas from a discussion – remains with the humans who spoke. A transcript can show what was said; only a person can fully grasp the implications of what was said. Nevertheless, having a perfect record is invaluable for supporting insight: we can recall details we’d have forgotten, notice subtleties (tone, emphasis, recurring concerns) by reading them, and base our judgments on a fuller body of evidence. In short, voice tools extend our experience (by capturing more than we otherwise could) and pattern it (by structuring speech into text and highlights). This extension is clearly preparatory for higher cognition – it gives us more raw material and frees us from scribbling notes in the moment – but the higher cognition, discerning meaning from those words, is still up to us.

From Patterned Experience to Insight: The Human Frontier

Across these paradigms – chat, graph, canvas, timeline, and voice – a common theme emerges: AI tools excel at collecting, organizing, and presenting the elements of our experience. They are patterners of experience. They tame the deluge of data into conversational answers, networks of context, visual maps, scheduled plans, and verbatim transcripts. In doing so, they address what often hinders understanding: disorganized, forgotten, or overwhelming information. A well-patterned experience is like a well-plowed field, ready for planting. It situates the knower in an environment where connections can be noticed and insights can sprout.

However, the actual moment of insight – when disparate pieces suddenly form a coherent whole in your mind – is not something these tools deliver on a platter. They may highlight patterns, but recognizing the significance of a pattern is a judgment call that AI cannot make with true certainty or context. They may generate content or answers, but deciding “Is this answer profound or just superficial?” is a question of human judgment. And ultimately, deciding what to do with knowledge (the decision phase) entails values, priorities, and risk-taking that lie outside the scope of automation.

Understanding this boundary is critical for designing the next generation of tools. Many current tools focus on optimizing outputs: faster answers, perfectly organized notes, zero-effort scheduling, etc. These are worthwhile goals – removing busywork and friction is valuable. Yet, if HCI design stops at optimized experience, we risk plateauing at a level of shallow productivity. The deeper goal is to facilitate the user’s transition from experience to insight. After all, the value of a note-taking system is not in how neatly it stores notes, but in how effectively it helps the user learn, create or decide something new from those notes.

Designing for Insight: Toward Human-Centered AI Companions

To truly augment human understanding, future tools should explicitly aim at those transition points. This might involve new interface designs that prompt reflection, not just capture data. For example, a note-taking app could notice that you saved several articles on a theme and then gently ask you (in a chat interface) to summarize what you learned – essentially nudging you toward articulating an insight. A scheduling assistant might not only automate your calendar but also leave space for unstructured thinking time and remind you to pause and set priorities, bridging into the judgment phase. A visual knowledge tool might use AI to suggest an analogy or an unexpected link between clusters of ideas – not as a final answer, but as a provocation to your insight. In other words, rather than aiming for a polished output (like a ready-made summary or slide deck), the system could aim to engage the user in the process of understanding (“Have you considered how concept X relates to Y? Here are some connections.”).

Such design thinking aligns the technology with the full human knowing cycle. The tools we discussed – Mem, Tana, Fabric, Heptabase, Reflect, Motion, Otter.ai, and others – already show how helpful it is to have our experience expertly managed. Mem ensures we “never miss a note, idea or connection,” as one tagline goes[33], and Fabric’s self-organizing workspace means “zero organization needed” for the user[34]. This frees us from clerical chores. The next step is for tools to become not just organizers but facilitators of insight. Rather than taking the human out of the loop, they should invite the human deeper into the loop of reflection and creative thinking.

AI thought companions are profoundly changing how we externalize memory and experience. They serve as mirrors and guides in the experience stage of knowing: capturing what we see and think, mirroring it back to us in structured ways. We must recognize both their power and their limit. They pattern the terrain, but we must walk it. They map out potential connections, but we must journey from data to understanding. By designing tools that appreciate this division of labor – tools that not only pattern experience but also encourage the leap toward insight – we can ensure that technology truly augments human intellect rather than just accelerating our routines. The frontier of HCI is not to have AI decide for us, but to give us the best possible chance to decide wisely ourselves. It’s time to focus on the handoff from AI-managed experience to human insight and judgment. In doing so, we affirm what remains uniquely ours in the age of smart machines: the understanding that arises when data becomes meaning, when patterns become insight, and when knowledge lights the way to wise decisions.

Tools Mentioned (Sources)

· Mem – Captures your ideas, meetings, and research—and brings them back to you exactly when you need them[35]. (AI-powered note-taking and “thought partner” app)

· Tana – Often described as a knowledge graph outliner… every item is part of an underlying graph[6]. (AI-native workspace with a flexible, networked data structure)

· Fabric – All-in-one AI workspace and smart organizer for all your projects, ideas, notes & links. Explore, search and brainstorm with AI[36]. (Self-organizing digital content hub with AI search and timeline features)

· Heptabase – A visual note-taking tool that helps you make sense of your learning, research, and projects… get into the flow of thinking[14]. (Spatial canvas for thought, using whiteboards and cards to visualize ideas)

· Reflect – Uses GPT-4 and Whisper from OpenAI to improve your writing, organize your thoughts, and act as your intellectual thought partner[29]. (Networked note-taking app with daily notes, backlinks, and AI assistance for notes and voice)

· Motion – A smart productivity tool that automatically schedules your tasks, meetings, and projects using AI[20]. (AI calendar and project manager that optimizes your timeline and task prioritization)

· Otter.ai – Never take meeting notes again. Get a meeting assistant that records audio, writes notes, automatically captures slides, and generates summaries[26]. (AI meeting transcription and summary tool that turns voice into text and highlights)

[1] [3] Mem – Your AI Thought Partner

https://get.mem.ai/

[2] [5] [35] Introducing Mem 2.0: The World’s First AI Thought Partner - Mem – Your AI Thought Partner

https://get.mem.ai/blog/introducing-mem-2-0

[4] [29] [33] Reflect Notes

https://reflect.app/

[6] [8] [10] [11] [12] [13] [17] [18] Visualizing Connections: Graph Views in Obsidian, Tana, and Anytype | by Ann P. | Sep, 2025 | Medium

https://medium.com/@ann_p/visualizing-connections-graph-views-in-obsidian-tana-and-anytype-3c767e08fe66

[7] Tana

https://tana.inc/

[9] Tana snaps up $25M as its AI-powered knowledge graph for work racks up a 160K+ waitlist | TechCrunch

https://techcrunch.com/2025/02/03/tana-snaps-up-25m-with-its-ai-powered-knowledge-graph-for-work-racking-up-a-160k-waitlist/

[14] Heptabase: The visual note-taking tool for learning complex topics. | Product Hunt

https://www.producthunt.com/products/heptabase

[15] [16] [19] HeptaThinking | Heptabase Visual Note-taking for Beginners: Creating Your Mind Maps and Knowledge Database | by Kuan | Medium

https://medium.com/@clydeyen19/heptabase-visual-note-taking-for-beginners-creating-your-mind-maps-and-knowledge-database-f9b74a022be1

[20] [21] [22] Motion AI Calendar: Smart Scheduling & Time Management Tool | AI Assistants | Gmelius

https://gmelius.com/blog/motion-ai-is-it-worth-it

[23] [24] [34] My Experience with Fabric.so: A Visual Haven for Digital Content Management | by PamelaJune | Medium

https://medium.com/@pamelajunecreative/my-experience-with-fabric-so-a-visual-haven-for-digital-content-management-9cdba0f29f94

[25] Reflect. My perfect notes application | by Stephen Zeoli | Medium

https://stephenjzeoli.medium.com/reflect-my-perfect-notes-application-af0978de3373

[26] [32] ‎Otter Transcribe Voice Notes on the App Store

https://apps.apple.com/us/app/otter-transcribe-voice-notes/id1276437113

[27] [30] [31] Otter Meeting Agent - AI Notetaker, Transcription, Insights

https://otter.ai/

[28] AI Note-Taking: How to Use AI for Better Notes - Reflect

https://reflect.app/blog/ai-note-taking-for-better-notes

[36] Fabric – your self-organizing workspace and file explorer

https://fabric.so/

Rusted Iron and the Restoration of Desire

In Maximus’s understanding, the soul is like fine iron glinting in the sun – except it has rusted over. Our desires and faculties are not inherently corrupt; they are good metal, given by God, now obscured by corrosive accretions. Asceticism, for Maximus, is the gentle removal of this rust so that the metal may regain its shine. “With the removal of things that are contrary to nature,” he explains, “only things proper to nature are manifest. Just as when rust is removed, the natural clarity and glint of iron are manifest”. This encapsulates the foundational difference in how Maximus views human passion. The Stoic might attempt to scrape off desire itself, viewing passions as irrational invaders to be eradicated for the soul’s tranquility. The Buddhist similarly diagnoses tanhā (craving) as the root of suffering and prescribes its total cessation for Nirvana. But Maximus does not see desire as a poison to be purged from the system; he sees it as a divinely minted energy, presently misdirected and in need of redirection. Our loves and longings are like heat in that iron – dangerous when dispersed into wild sparks, but life-giving when forged under the right hand.

Consider the image of a blacksmith at his anvil: before any strike of the hammer, he decides what tool he intends to shape from the raw metal. Stoic and Buddhist ascetics often aim to cool the iron – to still the heat of passion until it lies inert. Maximus’s asceticism instead takes the fire of desire and forges it into virtue. “Even the passions become good among the diligent, when they wisely separate them from corporeal objects and use them to acquire the things of heaven”, he writes. In this view, anger, craving, even lust – energies that, in their fallen form, lead us to sin – can be purified and turned toward love, compassion, and holiness. For example, zealotry against others can be transmuted into zeal for truth; physical desire can become an ardent yearning for God. Unlike the Stoic, who prides himself on feeling nothing toward what is indifferent, Maximus challenges us to feel everything rightly. And unlike the Buddhist seeker who might call even the self an illusion and desire a trap, Maximus affirms that the self, the body, and all created things are originally good – it is only their misuse that brings evil. “Food is not evil, but gluttony is… Money is not evil, but avarice is… Indeed, there is no evil in existing things, but only in their misuse”. In other words, our world and our passions are not traps to escape from; they are gifts to be set in order. Ascetic discipline, then, is the artful practice of realignment. It is the soul’s restoration, not its negation – a process of scraping away deception and distortion until our innate capacity for truth and goodness shines forth as naturally as polished iron catching the light.

Suffering as Furnace, Not Illusion

Such an ascetic restoration is no painless affair. Maximus lived and taught that suffering has a crucial place in this transformation: it is the furnace that burns away the rust. In a culture addicted to comfort, this notion feels bracing. The Stoics approached suffering with a kind of grim pragmatism – pain is inevitable, so bear it with dignity and don’t let it disturb your inner peace, they advised. Stoic philosophers like Epictetus or Marcus Aurelius counseled that we should not fear hardship; yet, for them, suffering was chiefly something to endure with reason’s help, not something inherently fruitful. Buddhists similarly acknowledge suffering (dukkha) as a fundamental truth of existence, but their answer is to uproot its cause (desire) and thereby transcend suffering altogether. The ideal is to become so unattached that one neither greatly sorrows nor greatly rejoices – one simply is, beyond the reach of suffering’s flames.

Maximus’s vision, by contrast, sees the flames of suffering as capable of refining the soul rather than merely testing or eventually extinguishing it. In his own life, Maximus suffered mutilation and exile for the faith – he knew firsthand that pain can either break or make a person. His counsel to fellow ascetics reflects a startling paradox: embrace the very struggles that others flee, for within them lies hidden grace. “Whenever you are suffering intensely from insult or disgrace, realize that this can be of great benefit to you,” Maximus writes. The sting of humiliation, for instance, burns away the rust of pride: “disgrace is God’s way of driving vainglory out of you”. Here is asceticism as an alchemy of pain into purity. The Stoic might clench his jaw at insult and tell himself it is indifferent; Maximus urges us to open ourselves to the wound so that vainglory (the diseased swelling of self-regard) may drain out.

This approach is not masochism, nor a denial of suffering’s reality; rather, it is a revaluation of suffering. Maximus and the ascetic tradition see physical hardship, emotional slights, and worldly deprivations as tools in the divine forge, tempering the soul’s metal. Fasting, sleepless vigils, manual labor – all these voluntary pains check the wild growth of passions, like a damper on a flame. As Maximus enumerates, such practices “do not allow concupiscence to grow,” and prayerful solitude and longing for God can even make lust “disappear”. Likewise for anger: practicing long-suffering and meekness prevents its eruption, while acts of love and kindness “make it diminish”. Each ascetic labor is a medicine targeting a spiritual ailment. The goal is not to ignore suffering or pretend it doesn’t hurt; the goal is to harness it, to let it purify our disordered affections. Where the Buddhist seeks to escape the cycle of suffering entirely, Maximus teaches engagement with suffering in faith. Pain is real – but in Christ, it can become a source of redemption. It’s as if the very experiences we dread (hunger, insult, fatigue) are revealed as secret surgeons of the soul, cutting away tumors of vice. Maximus’s Christ-centered asceticism even dares to see glory in suffering: by voluntarily taking up our little crosses of self-denial, we participate in Christ’s Passion, and thereby in His victory. The iron in the fire is not being destroyed; it’s being refined and reshaped.

Beyond the Self: Asceticism in Community and Grace

Another striking difference in Maximus’s ascetic vision is its inherently communal and sacramental context. The wise Stoic imagines the inner citadel of the self – virtue is a solitary stronghold of reason within, which the slings and arrows of fortune cannot breach. There is something deeply individualistic about Stoic practice: Marcus Aurelius could pursue wisdom amid the chaos of war by retreating into his own rational mind. The Buddhist path, too, ultimately rests on individual enlightenment. While Buddhism highly values the Sangha (community of practitioners) and has monastic orders, the moment of Nirvana is a singular awakening – only you can walk the Eightfold Path for yourself, and the final detachment is a profoundly personal letting-go, beyond all social bonds. In modern adaptations of these traditions, it’s not uncommon to see people practicing mindfulness or stoic journaling alone, using these disciplines as personal wellness techniques isolated from any larger community or cosmic story.

Maximus’s asceticism, however, is never a solo project of self-improvement. It is Christ-anchored and Church-centered. We must remember: Maximus was a monk, formed in a community of prayer, and a theologian of the Incarnation who saw no sharp divide between spiritual and physical, solitary and communal. For him, ascetic discipline is woven into the fabric of the Church’s life. Fasting goes hand in hand with the Eucharistic feast; personal prayer is buoyed by the liturgy and shared psalmody; chastity and charity are learned in the rough and tumble of actual community living. In one treatise, Maximus even interprets the liturgy itself as an ascetic journey of the soul. The Church, in his beautiful phrase, is “envisioned as the sacrament of human deification” – a mystical society where every ritual, from baptismal water to communal meal, contributes to our transformation in God. Asceticism, then, is not about willpower in a vacuum; it’s about cooperation with grace in a sacred fellowship. The Church is “a realm of sacred relations and actions” in which the lonely struggle of the monk is buoyed by the prayers of the saints, and the entire cosmos is invited into a dance of restoration.

This communal, sacramental grounding is utterly foreign to Stoic and Buddhist frameworks. A Stoic may draw strength from friends or a teacher, yes, but ultimately no one can be virtuous for you. In Buddhist monastic life, there is community support, but enlightenment remains an individual achievement – and classical Buddhism lacks the concept of a transmitting grace that can heal and elevate the nature of individuals. Maximus, by contrast, insists on the synergy of divine grace and human effort. The ascetic does not claim progress as a personal accomplishment; it is grace-enabled. We struggle, but every victory is a gift of God. And this grace flows especially through tangible means: through holy mysteries, through spiritual fatherhood and motherhood, through the very material elements Stoics and Buddhists treat as indifferent or illusory. In Maximus’s view, the bread and wine of Eucharist, the touch of holy oil, the unity of believers in love – all these “corporeal” things become conduits of theosis. The body and its desires are not obstacles to spirituality, but vehicles of it when rightly oriented. He honors the body as an integral part of the person destined for glory, not a mere shell to be discarded nor a temptation to be shunned. Thus, asceticism is embodied and ecclesial: fasting trains the body even as prayer lifts the mind; almsgiving binds one to the poor and to Christ in them; obedience and humility are learned in real relationships. It is in the push-and-pull of communal life that virtues like patience, forgiveness, and genuine compassion are exercised – far from the ego’s isolating fantasies. Maximus would agree with the adage that one cannot be saved alone. The Stoic sage’s self-sufficient virtue and the Buddhist monk’s solitary nirvana find their counterpoint in the Christian monk’s humility within a spiritual community, all of which depends on Christ together. In the Church, ascetics discover that to ascend to God is simultaneously to descend into love of neighbor, carrying each other’s burdens. As one modern interpreter put it, for Maximus, “there is experienced divine and deifying activity” precisely in the liturgical community. The holy struggle is a team effort – indeed an effort of the whole Body of Christ, head and members, earth and heaven together.

The Goal: From Apatheia and Nirvana to Theosis

All these threads – the affirmation of desire, the embrace of redemptive suffering, the communal-sacramental journey – converge in the ultimate end that Maximus envisions for asceticism. What is the telos of this arduous path? What awaits the soul that has been purified of rust and polished to mirror-brightness? For the Stoics, the ideal end is a state of unperturbed virtue. They seek ataraxia, a serene freedom from distress, governed by reason and in harmony with nature’s law. The Stoic sage, perfectly rational and self-controlled, wants nothing and fears nothing; his happiness consists in living ethically and accepting the universe’s outcomes with equanimity. It’s a noble picture, a kind of moral heroism that ends in calm strength – yet notably, the Stoic end is still within the limits of human nature. It is man as he is, just perfected in virtue.

Buddhism articulates its end as Nirvana, often described as the “blowing out” of the flame of craving and the cessation of the rounds of rebirth. Especially in Theravada thought, Nirvana is an extinction – not of existence per se (its precise nature is beyond concept), but of ignorance, desire, and suffering. It is a profound peace beyond all conditions, sometimes likened to a drop of rain merging into the vast sea. Selfhood is transcended; one awakens to sunyata (emptiness) or the unconditioned reality. Where the Stoic sage remains very much a person—lucid, engaged, if emotionally invulnerable—the Buddhist arhat or bodhisattva aims for a liberation that is impersonal in a sense: it is the end of “I” and “mine,” a quietude where even the categories of being and non-being no longer bind.

Maximus the Confessor offers a radically different summit: not calm absorption, but radiant communion. The destiny of the ascetic, in Maximus’s theology, is Theosis – deification. This does not mean one ceases to be a creature or loses personal identity; on the contrary, one becomes fully a person in the image of the ultimate Person (or rather, Persons of the Trinity). Theosis is participatory union with the living God, a sharing by grace in the very life and energies of God. It is what Eastern Christianity calls glory, and it is anything but a flat calm or a void. Maximus would say the polished soul begins to shine like the sun, reflecting God’s light. In the Transfiguration of Christ on Mount Tabor, the disciples beheld Jesus’s face shining with divine light – an image of what humanity is meant for. The ascetic journey, for Maximus, ends not in a stoic impassivity or a Buddhist negation, but in a personal relationship so intense it is described as union. The lover of God becomes “all flame,” to borrow the words of an earlier desert father. Maximus, commenting on this union, insists that it is made possible by grace transforming nature. We empty ourselves of passions so as to be filled with God. The final state is ecstasy in the literal sense – a going out of oneself, not into nothingness, but into the infinite love and knowledge of God. There is tranquility there, yes, but also rapture, an eternal newness of life. It is like comparing the stillness of a starry night (beautiful, but cold and dark) to the dynamic light of the sun at dawn. Stoic apatheia is the night sky clear of storms; Buddhist Nirvana might be the mysterious hush before dawn; Theosis is the sunrise – clarity and warmth flooding the world, revealing all colors and forms in their true beauty. The person who attains theosis does not disappear or close in on themselves; they radiate outward, fully alive. As Maximus taught, Christ did not come to nullify our humanity but to elevate it: God became man so that man might become god, a truth that animated all his ascetical teaching. Thus, the ascetic’s ultimate horizon is communion – an interpersonal, loving union with God and all others in God. We become by grace what Christ is by nature, all without losing our distinct personhood. In the words of Maximus’s tradition, we become “partakers of the divine nature” – an end that fulfills every rational hope of the Stoic and every mystical intuition of the Buddhist, yet surpasses them beyond measure.

A Lyrical Meditation for Our Age

Why does this ancient vision of asceticism matter now? Because our own age, for all its differences, is riven by the same human longings that drove an emperor like Marcus Aurelius to philosophy and a prince like Siddhartha to renounce his palace. We hunger for peace in a world of whirlwind. We feel the weight of suffering and seek its meaning. We sense the poverty of shallow pleasures and wonder if there is a joy that doesn’t fade. The modern landscape is paradoxical: never have we had so many means to indulge desire, yet never have we been so weary of our indulgences. Hyperstimulation is the norm – our attention splinters across flashing ads and algorithmic feeds, leaving us jaded and numb. In response, a kind of rootless detachment has become a coping mechanism. We dabble in mindfulness apps that promise calm without commitment; we adopt Stoic mottos as life-hacks to get through stressful workdays; we speak of “cutting toxic people out of life” and “not catching feelings” as if dispassion alone were salvation. There’s a surge of interest in “ancient practices” precisely because we are adrift. Walker Percy once quipped on the modern malaise – one can have every material comfort and still flunk life. Jacques Barzun chronicled how a culture that loses its unifying vision turns to diversions and decadence, mistaking motion for meaning. In our time, we see a thousand offers of counterfeit transcendence: psychedelic spirituality without morality, techno-utopian dreams of uploading consciousness, consumerist “experiences” sold as moments of escape. Yet, the soul remains restless.

Maximus the Confessor’s voice rises like a chant above the digital din, calling us to a different way – one both older and fresher than the fragmented prescriptions of Stoic or Buddhist minimalism when divorced from their fuller contexts. He would have us consider that perhaps desire isn’t the problem; rather, it's disordered desire. In an economy that exploits our attention and fans every lust for profit, real asceticism is an act of rebellion – not by forsaking desire, but by reclaiming it. We turn our desire away from the candy of endless notifications and toward the nourishment of truth and love. We say no to some pleasures not because they are evil, but because we seek the higher pleasure of a sound mind and a pure heart. This is the “radiant restoration” Maximus speaks of: like a piece of rusted iron forgotten in a junkyard, the soul in modernity is caked with false wants, fears, and frustrations. By grace-enabled effort, that rust can be gently scrubbed off – a little fasting from media here, a little courageous vulnerability in relationships there, the daily practice of prayer which recenters us – and slowly, the original metal shines. We begin to see again what is real and good. Our capacity for wonder returns as the overstimulation recedes. In community, we find that detachment need not mean alienation; rather than simply cutting bonds, we reshape them rightly. The ascetic path teaches us to be in the world but not of it: to love others without lust to exploit, to use things without being used by them.

What makes Maximus’s asceticism especially urgent now is its insistence that true transformation is relational. In an era of “bowling alone,” of lonely screen-lives, he reminds us that no one attains wholeness by oneself. The Stoic impulse to self-reliance can turn into a spiritual pride or a despair of others; the pseudo-Buddhist impulse (as adopted by many Westerners) to “detach from outcomes” can slide into a cool indifference toward injustice or a retreat from love. Maximus implores us to enter the refiner’s fire together. Our wounds are too deep for a merely DIY wellness regimen; we need the Divine Physician, and we find His healing touch in the sacramental life and in each other. Asceticism in Maximus’s key is not a competitive Olympics of self-denial to see who can be most disciplined. It is a healing journey, where the hard medicines of fasting or solitude are always taken with the sweet consolations of communion with God and support of the faithful. In an age where many try to curate an invulnerable self – whether through Stoic emotional stoicism or the Buddhist-inspired refusal to form attachments – Maximus gently points out that invulnerability is not the same as transcendence. A heart may be armoured and yet empty; a soul with no attachments may also have no love. The Confessor directs us to Christ, who did not numb Himself against suffering but embraced it for the sake of love, and who teaches that only by losing ourselves in love for God and neighbor do we truly find ourselves.

Finally, imagine the outcome if we heed this ancient wisdom. The modern person, fragmented and weary, could become integrated and luminous. Instead of a gray Stoic resilience or a passive Buddhist quietude, we might witness the birth of modern saints who are on fire: compassionate in action, serene in spirit, brimming with a joy that doesn’t depend on the next stimulus or distraction. Their presence would be like light in a room – not calling attention to themselves, but illuminating reality for others. This is the radiance Maximus promises, the soul’s original clarity restored. It is a light that does not fade, because its source is beyond the screens and neon of our cities – it is sourced in the Eternal Light. In an age of counterfeit transcendence, where every new gadget or ideology is hyped as the answer and then quickly found hollow, the real transcendence of theosis stands quietly, like a mountain at dawn, waiting to be noticed. It invites us not to escape our humanity, but to fulfill it.

As we conclude this meditation, a lyrical image arises: picture a rusted iron lantern, caked in years of neglect, suddenly taken into caring hands. With patience, the rust is scrubbed away; with skill, the bent pieces are repaired. A small flame is lit inside. The lantern begins to glow – first faintly, then with growing confidence, until it casts a warm light in the surrounding darkness. This is what St. Maximus offers our age. In the midst of noise, he speaks of stillness filled with presence. In the face of despair, he speaks of suffering transfigured into glory. Against both the iron cage of Stoic self and the dissolving void of false Nirvana, he reveals the image of a person made beautiful by love, bound in a tapestry of fellow pilgrims, all moving toward a horizon of endless Day. Such a vision matters now more than ever. It whispers to us that the cure for our modern ennui and excess is not less desire, but holy desire; not isolation, but illuminated community; not the end of longing, but longing finally finding its end – in the radiant, resurrecting embrace of the Living God.

Sources:

1. St. Maximus the Confessor, Four Centuries on Charity (Chapters on Love) – Cent. 3.4, on the goodness of created things and the evil of their misuse.

2. St. Maximus, Question to Thalassius 55 and Disputation with Pyrrhus – teaching that when what is unnatural is stripped away, the soul’s natural virtues shine, just as polished iron glints; even the passions can be redirected to good.

3. St. Maximus, Selected Writings (Classics of Western Spirituality) – advice on the ascetic struggle: how insults cure vainglory and how practices like fasting, vigils, solitude, and prayer heal the passions of concupiscence and anger.

4. Adam Cooper, Life in the Church according to St. Maximus (Mystagogia) – analysis of Maximus’s view of the liturgical life as the communal sacrament of deification, underscoring the ecclesial, grace-filled context of ascetic effort.

I have been wrestling with this idea as I watch today’s AI agents struggle with extended tasks. The dirty secret of current LLM-based agents is that their so-called “memory” is brittle and shallow. They can’t reliably carry information or intentions across multiple sessions or complex goal sequences. You see it when a conversational agent forgets your name or repeats itself after a few turns. You see it when AutoGPT spins in circles, losing track of what it’s already done. In fact, developers analyzing AutoGPT found that its tendency to get stuck in loops or go off the rails stems from a finite context window and lack of long-term memory[3]. With no true memory of past actions, it can’t focus on its objectives and keeps trying the same things over and over. This is a fundamental architectural gap. Modern AI systems are memory silos, each chat or session isolated from the next[4]. Start a new conversation, and it’s a blank slate every time – no accumulated learning, no sense of “self” or history carried forward. As researchers recently put it, today’s LLMs rely only on static model weights and short-lived context, which limits any long-horizon coherence or continual learning[5][6]. It’s like dealing with an amnesiac savant: brilliant at single-turn tasks, hopeless at sustained, evolving ones.

The Illusion of “Memory” in Today’s AI

It’s worth pausing to define what we mean by “memory” in AI systems. Right now, when an AI practitioner says an agent has memory, they usually mean one of a few things:
- Expanded context windows: Simply cram more past text into the prompt. (GPT-4 can take 32k tokens, so just include the conversation history or docs up to that limit.)
- Vector databases + RAG: Store chunks of text embeddings and do a similarity search to retrieve relevant snippets when needed (Retrieval-Augmented Generation)[7].
- Summarization of history: Keep a rolling summary of past interactions so the model can be “reminded” without full logs.
- External file or DB logging: For agents like AutoGPT/BabyAGI, write important info to files or a database so it can be read in future steps[8][9].

These techniques do help. Vector search means an agent can look up facts it saw hours or weeks ago. Summaries prevent total forgetfulness. Yet, none of these truly solve the problem of robust, long-horizon reasoning and behavior. They treat memory as storage and retrieval – a passive database to query – rather than as an active, structural part of cognition.

Consider the RAG approach: it’s basically a smarter autocomplete with footnotes. The model doesn’t actually learn from new information; it just fetches it when prompted. The “memory” lives outside the model’s core, and there is no built-in mechanism to update the model’s understanding over time. This is why one can’t interrogate a typical LLM about something it discussed last month unless you explicitly fed those details into the prompt. Each session, each prompt, is stateless in the fundamental sense[7]. We’re faking a working memory by stuffing context in front of the model repeatedly. And as context windows grow, we hit diminishing returns – even degrading returns. Empirically, models suffer context rot: performance drops as more tokens are packed into the prompt, especially if much of it is irrelevant detail[10][11]. In other words, simply remembering “more” via brute-force context can actually confuse these systems. The memory is there, but the cognition isn’t.

Then there’s the issue of orchestration. Human-like memory, what I anticipate most of us think of when the term is uses, really isn’t just recalling facts. Rather, it includes when to recall, what to recall, and how to use what’s recalled. Our brains constantly sift, prioritize, and even distort memories in service of the task at hand (usually without us realizing). Current AI agents don’t have an analog of this metacognitive control. An AutoGPT agent doesn’t decide what to forget or which past result to focus on – it just dumps whatever it has into the next prompt until the token limit hits. Developers end up manually crafting prompt engineering rules to manage this, essentially playing the role of a “memory scheduler” for the AI. It’s ad-hoc and brittle. As one LangChain article quipped, imagine a coworker with no memory – you’d have to repeat yourself endlessly[12] (like a parent or something, weird 😁). That’s the UX today: frustrating, unless we hand-hold the agent’s memory every step of the way.

Memory as the Engine of Understanding (Not Just Recall)

If we step back from the implementation details, there’s a broader thesis emerging: Memory is is the substrate upon which cognition is built. In a real sense, memory is the architecture of the mind. It’s not a database, rather it’s a process. My main man Bernard Lonergan, whom you all likely know by now if you’re reading this Substack, wrote about cognition as layered operations – experiencing, understanding, judging, and deciding – each building on the last. While I won’t dive into Lonergan’s texts here, that perspective inspires a useful way to think about memory analogically in our AI agents:

• Memory is layered. Not all “memories” are equal; there are different kinds serving different roles. We see parallels to human memory: procedural memory (knowing how), semantic memory (knowing facts and concepts), episodic memory (remembering specific events)[13][14]. Current AI implementations mimic these in rudimentary ways. For instance, an LLM’s weights plus its code represent a kind of procedural memory – the skills and biases baked into it. Meanwhile, a vector store of facts extracted from past chats acts as semantic memory (a repository of learned knowledge)[14][15]. And a log of past actions or chain-of-thoughts is akin to episodic memory (records of what happened)[16][17]. But in today’s agents, these layers barely interact or update. They are static or siloed. A true memory system will need to unify these layers, so that what an agent experiences (episodes) can update what it knows (semantic) and even how it operates (procedural). This is exactly the idea behind new research like MemOS, which introduces MemCubes that encapsulate various memory types – from plain text knowledge to activation states and model weight updates – under one framework[18]. In other words, memory spanning from the transient “working memory” of activations to the long-term parameters can be managed together.

• Memory is selective and interpretive. Humans forget most of what hits our senses – and thank goodness, or we’d be flooded with useless data (just ask one of the few people suffering from Hyperthymesia. What we do remember, we compress and weave into our existing worldview. Our memories are stories and schemas, full of omissions and embellishments. Why? Because memory serves understanding, not the other way around. An AI agent will similarly need to choose what to remember and at what level of detail. It might record that “Task X was completed and yielded Result Y” without keeping every log line of how it got there. It might abstract a user’s preference as “user likes direct answers” instead of storing every question they asked. This selection is guided by an orientation toward intelligibility – a drive to capture the essence of experience in a useful form. Today’s agents treat all data points a bit too evenly. A smarter agent would, say, attribute meaning and context to memories: This piece of information was wrong, that feedback was positive, these facts relate to topic Z. By tagging and structuring memories, the agent can later retrieve not just a raw snippet, but the right kind of memory for the situation (e.g. “I recall you prefer a casual tone in emails” vs. a random past email).

• Memory is dynamic and fallible. The strength of human memory is not in never forgetting – it’s in the constant reworking of what we know. We revise memories in light of new evidence. We generalize from specifics and, sometimes, we mercifully drop details that no longer matter. A rigid memory is a brittle one. AI systems to date have been mostly static: once the model is trained, its “knowledge” is frozen except for whatever trickles in through a prompt. We’re starting to see this change. With tools like fine-tuning or online learning (carefully controlled), an agent could update its semantic memory when it encounters new facts. Even without weight updates, it could maintain a mutable knowledge base and a sense of its own trajectory (what it’s done and why). Crucially, it should also learn to forget. Yes, forgetting is a feature! Effective cognitive systems must discard or de-prioritize outdated and irrelevant information to make room for the new. If our AI assistant learned about a user’s old job and now the user switched careers, the assistant should gradually attenuate the importance of the old info. Without a mechanism for graceful forgetting, the “memory” becomes a junk heap that might clutter the agent’s decisions.

In short, building memory for an AI agent is less about databases and more about building a mind. It means instilling the kind of layered, purposeful recollection that underpins human thought. When memory works, it feels like understanding. The agent is using past experience to shape present reasoning.

Why Today’s Agents Fall Short (Examples from the Frontier)

It’s illuminating to examine how current agent frameworks attempt to deal with memory, and where they hit a wall. Let’s walk through a few:

• AutoGPT & BabyAGI: These autonomous agent prototypes burst into the scene with promises of multi-step reasoning. Under the hood, their “memory” was often a combo of a short-term context (the prompt that grows with each step) and an external file or vector store for long-term info. In practice, they quickly exposed the fragility of this setup. AutoGPT, for example, would frequently lose the thread, endlessly looping on a subtask because it couldn’t truly recall it had tried that path already[3]. Developers noted that lack of long-term memory and the finite context window were primary culprits[19]. BabyAGI, which managed a task list and a vector DB of results, could juggle simple routines but struggled as tasks became more open-ended – the agent had no deeper narrative of what it was doing, just a bunch of past notes. These projects deserve credit for highlighting the need for memory: they often integrate functions to save important info to a vector database and retrieve it later[20]. Yet, without more sophisticated filtering and understanding of those past results, retrieval alone wasn’t enough. The critique here is that memorization != comprehension. Storing the outcome of every tool use or API call doesn’t yield a system that knows why it did things or when to change strategy. The lesson from AutoGPT and BabyAGI is clear: long-term autonomy demands an integrated memory architecture, not a bolted-on log.

• LangChain (and similar frameworks): LangChain introduced the community to the idea of pluggable memory modules. You can give your agent a conversation buffer memory, or a summary memory, or even a vector store backed memory for longer recall[21][22]. This is immensely useful for builders – it’s a toolkit approach. For example, a customer support bot can be equipped to remember the user’s name and last issue, because the developer adds a short-term memory buffer. Or a research assistant agent can have a semantic memory: after each interaction it extracts key facts and stores them, so later it can retrieve “facts the user has provided” to avoid asking again[23]. LangChain’s blog even maps these ideas to psychology: distinguishing semantic vs episodic memory, etc., and notes that each application might need to remember different things[24]. A coding agent might recall user preferences about code style, while a travel planner agent remembers the traveler’s past destinations[24]. This specialization hints at a verticalization of memory – which I’ll return to in the conclusion. The limitation, however, is that LangChain leaves the cognitive orchestration to you. It doesn’t provide a unified brain; it provides brain pieces. If you assemble them well, you get a better agent. If you don’t, you get an agent that either forgets or overloads itself. There’s a reason many experienced devs say building a good LLM agent is more about managing prompts and memory than about the model. We’re essentially designing crude, manual memory management policies. This is akin to programming before high-level memory management – it’s powerful but error-prone.

• Vector Databases & External Knowledge: These deserve special mention because vector DBs are often hyped as the solution for AI memory at scale. Indeed, tools like Chroma, Pinecone, or Weaviate have made it straightforward to spin up a massive associative memory: embed everything and let similarity search retrieve relevant chunks on demand. It works great for question-answering (think GPT answering by pulling relevant documentation). But as a cognitive memory for an agent, it’s incomplete. The vector DB doesn’t tell the agent when to search, what to store, or how to use what it gets back. Those are decisions external to the database. Without careful orchestration, an agent either fails to retrieve what it needs or retrieves too much irrelevance (leading to the dreaded context rot). Put simply, a vector search is a librarian, not a psychologist. It can hand you information; it won’t integrate that information into the agent’s sense of self or mission. Furthermore, pure semantic similarity can miss the point – if an agent is debugging code, the fact that it talked about a similar bug two weeks ago is only useful if the agent recognizes the situation is analogous. That kind of analogical link is beyond a naive vector similarity. It requires a higher-level understanding that “this problem is like that past problem,” a leap that current semantic memory setups don’t achieve.

• MemOS and Emerging “Memory OS” Research: On the horizon, efforts like MemOS are trying to redesign the AI stack around memory as a first-class citizen[25][26]. MemOS treats memory not as a single data store, but as a manageable resource akin to CPU or disk in an operating system[25][26]. It introduces scheduling algorithms to decide which memories to keep in “RAM” (active context) and which to page out to long-term store, and mechanisms to transform transient experiences into durable knowledge[27][28]. Notably, the MemOS architecture acknowledges the heterogeneity of memory: transient activation states, intermediate computation results, and permanent knowledge updates are all handled under one roof[18][26]. Early results show dramatic gains in tasks that require connecting information across many steps or time intervals[29][30]. This suggests that the bottleneck was indeed how we’ve been handling memory. When you give an agent a structured way to learn from experience (not just recall facts), its effective intelligence leaps. One paper reported a 159% improvement in temporal reasoning tasks by using a memory-centric approach[31][29]. Numbers aside, the philosophy here is crucial: memory isn’t an add-on, it is the system. We’re essentially building an internal knowledge base that the AI can grow and refine through use, blurring the line between “pre-trained model” and “experience-trained agent.”

Each of these examples teaches us something. The failures of AutoGPT teach us that naïve memory leads to chaotic behavior. The experience with LangChain shows that memory must be tailored and orchestrated for the use-case. The vector DB’s strengths and weaknesses demonstrate that raw retrieval needs guidance. And the new research hints that integrating memory deeply can unlock new levels of performance.

Designing Memory as a Cognitive Control System

Where does this leave us, the builders and practitioners? It’s time to move beyond thinking of memory as a static database and start designing it as a cognitive control system. In practical terms, that means building agents that don’t just have memory, but actively use memory the way a mind does: to perceive, to decide what matters, to learn, and to adapt.

Here’s a short agenda I propose – a loop of operations for any agent that aims to have a human-like memory architecture:

1. Perception: First, capture the raw experience. Log the conversation turn, the tool output, the user feedback, the sensor reading – whatever the agent is dealing with. This is the transient trace of “what just happened.” Without perception, there’s nothing to remember.

2. Selection: Immediately filter and highlight what’s important about this experience. Did the user express a preference or correct the agent’s mistake? Was there an unexpected outcome from an action? The agent should extract salient details (possibly via an LLM prompt asking “what are key facts or results here?”). Not everything goes forward; noise must be left behind.

3. Encoding: Translate the selected information into an internal representation that can be stored and queried efficiently. This could mean embedding it as a vector, structuring it as JSON with fields (e.g. {"user_pref": "casual_tone"}), or compressing it into a summary. Encoding is about preparing the memory for future use.

4. Attribution: Link the memory with context: source metadata, timestamps, causal tags, relevancy scores. For example, note when and why this memory was formed. “User said they disliked the last recommendation [during travel planning on Aug 2025].” Attribution enriches the memory so that later the agent can retrieve not just a fact but the story around that fact (to avoid misusing it out of context).

5. Consolidation: Integrate the new memory into the agent’s long-term store. This might involve merging it with existing knowledge (e.g., update the profile of the user’s preferences), or storing it in an organized memory index. It could also trigger a model update in systems that allow learning on the fly. Consolidation is where ephemeral observation becomes lasting knowledge. It might be done during “idle” times or as a background process, analogous to how our brains consolidate memories during sleep[32][33].

6. Retrieval: When reasoning, have a mechanism to pull relevant memories at the right time. This is classical information retrieval augmented with the agent’s own judgment. The agent might ask itself: “What past experiences are like this situation?” and do a search of its memory (using metadata and embeddings). Retrieved items then have to be integrated into the agent’s working context or chain-of-thought. This step is where classic RAG comes in, but now it’s one part of a larger loop.

7. Revision: With new feedback or outcomes, go back and update existing memories. If the agent had a stored belief “Solution X works for problem Y” and it later fails, the agent should mark that memory as no longer reliable, or append the condition “(except in cases Z).” This is continuous learning in micro. Over time, an agent’s memory should evolve – refining concepts, correcting errors, re-framing narratives. Revision also covers re-encoding: maybe the initial summary was too coarse and the agent decides to store more details after realizing their importance.

8. Forgetting: Finally, implement graceful forgetting or deactivation. Not all memories need to live forever. Some can be compressed further, archived, or deleted. The agent might maintain a sliding window or a relevance decay – for instance, memories that haven’t been accessed in a long time or relate to now-irrelevant contexts get phased out. Forgetting is crucial for efficiency and for avoiding the clutter of contradictory or outdated info. It’s also a form of regularization: preventing the agent from overfitting to the past when the situation has changed.

This cycle turns memory into an active faculty. It’s essentially a control system: perceiving signals, updating internal state, and feeding back into decisions. Not every application will need the full loop, but I suspect any agent that aspires to long-term autonomy will touch all these stages in some form. And importantly, this is where vertical applications can shine. When you focus on a specific domain – say a medical diagnosis assistant, or a coding co-pilot, or a customer service bot – you can craft these memory operations with domain knowledge. You know what to pay attention to (selection), how to represent it (encoding), and when to recall it (retrieval) in that context. A vertical AI app can thus implement a robust memory tailored to its use, and in doing so, teach us general lessons. For example, a medical agent might develop a specialized forgetting strategy for outdated research findings (medical knowledge moves fast), which could inform how any AI handles time-sensitive knowledge. A coding assistant might learn to attribute memories by codebase and project, offering insight into context-scoped memory management that generalizes beyond coding.

In building these systems, we should remember Altman’s provocation: the endgame is an AI that “knows you, your whole life.” But to get there, memory can’t just be bigger context windows or bolt-on databases. It must be the core of the agent’s cognitive architecture. The journey from here to that ideal will be messy and will require rethinking a lot of assumptions in AI design. Yet, I’m optimistic. We are, in a sense, rediscovering in machines what nature evolved in us – that memory is the foundation of intelligence. An AI that can form, organize, and use its memories fluidly is an AI that can learn, adapt, and maybe even have something like a personality or identity over time. For those of us building the future of these agents, the mandate is clear: design for memory first. Treat it not as an afterthought but as the very substrate upon which your agent’s mind is built. Everything else – coherence, reasoning, usefulness, safety – will flow from that foundation. This is hard, deep work, but it’s the kind of work that makes the difference between yet another demo and a truly transformative AI partner. And as we iterate in vertical slices and share what we learn, we’ll be co-authoring a new general playbook for machine memory, one that could unlock the next era of agentic intelligence.

Sources:[1][2][3][4][5][6][7][10][11][12][13][14][15][16][17][18][26][34][28][20][22][24]

[1] OpenAI's Memory Trap & Its Implications for Consumer Freedom

https://www.inrupt.com/blog/openais-memory-trap

[2] Opt-In To OpenAI's Memory Feature? 5 Crucial Things To Know

https://www.forbes.com/sites/digital-assets/2025/04/11/should-you-opt-in-to-openais-memory-feature-5-crucial-things-to-know/

[3] [19] AutoGPT - Wikipedia

https://en.wikipedia.org/wiki/AutoGPT

[4] [5] [6] [7] [18] [25] [26] [27] [28] [29] [30] [31] [34] Chinese researchers unveil MemOS, the first 'memory operating system' that gives AI human-like recall

https://venturebeat.com/ai/chinese-researchers-unveil-memos-the-first-memory-operating-system-that-gives-ai-human-like-recall

[8] Decoding AutoGPT: Understanding the Magic Behind the Code

https://medium.com/@evyborov/decoding-autogpt-understanding-the-magic-behind-the-code-991177b62583

[9] AutoGPT Explained: How to Build Self-Managing AI Agents | Built In

https://builtin.com/artificial-intelligence/autogpt

[10] [11] Context Rot: How Increasing Input Tokens Impacts LLM Performance | Chroma Research

https://research.trychroma.com/context-rot

[12] [13] [14] [15] [16] [17] [23] [24] [32] [33] Memory for agents

https://blog.langchain.com/memory-for-agents/

[20] How do tools like AutoGPT get around the size limit? - Reddit

https://www.reddit.com/r/ChatGPTCoding/comments/131zv1n/how_do_tools_like_autogpt_get_around_the_size/

[21] [22] LangGraph Memory Management - Overview

https://langchain-ai.github.io/langgraph/concepts/memory/

1. Conceptual Foundations: Lonergan's Cosmopolis and the "Not-Numerous Center"

As a reminder, Bernard Lonergan introduces Cosmopolis as a response to the biases that drive communities into decline. In Insight, Lonergan portrays Cosmopolis as a higher cultural consciousness that "withdraws from practicality to save practicality," operating not by political decree but by intellectual and moral leadership[2][3]. Cosmopolis is not a formal organization or a ruling academy; it has no members or legal authority. Instead, it is "a dimension of consciousness, a heightened grasp of historical origins, a discovery of historical responsibilities" [4]. It works by spreading ideas that correct bias and envision creative solutions – ideas which, under the general bias of common sense, would otherwise be dismissed or ignored. As Lonergan puts it, "culture (art, religion, philosophy, journalism) must be freed from the need to justify itself to the practical mind," because the general bias of short-term common sense can hold culture captive[5]. Cosmopolis thus operates at a remove from immediate practical concerns to serve them in the long run[6]. Its hallmark is patient difficulty: "Finally, it would be unfair not to stress the chief characteristic of Cosmopolis. It is not easy" [3].

In a time of cultural breakdown, he predicted that neither reactionary traditionalists nor restless innovators would secure progress. Instead, what counts is a small but vital center that bridges the gap[1]. This center is "at home in both the old and the new," able to learn from the wisdom of traditions yet open to novel insights[1]. It works "one by one" through necessary transitions, refusing the temptation of quick fixes or partisan half-measures[7]. And it has the strength to "insist on complete solutions even though it has to wait" [8] – meaning it prefers slow, genuine progress to expedient compromises. Such a center exemplifies intellectual and moral conversion, the personal transformations Lonergan saw as essential for authentic progress.

Lonergan identified three key conversions – intellectual, moral, and religious (or spiritual) – which together foster the capacity to inhabit that higher viewpoint:

• Intellectual conversion involves a radical shift in our understanding of knowing and truth. It is the hard-won self-appropriation of our cognition: moving from a commonsense or fragmented view of knowledge to an integrated one where we recognize the mind's intrinsic norms of being attentive, intelligent, and reasonable[9]. An intellectually converted scholar learns to value truth over ego and to distinguish genuine insight from mere hypothesis or assumption. In a university context, intellectual conversion enables researchers and students to question their own assumptions and engage in deeper learning, all with a humble approach. It corresponds to critical self-reflection and openness to truth wherever it leads.

• Moral conversion is a shift in our criteria of decision-making – from valuing mere self-interest or satisfaction to valuing what is truly good and worthwhile[10]. A morally converted person is "committed to values above mere satisfactions" [11]. For academics, this means prioritizing honesty, rigor, and the common good over careerism or convenience. Moral conversion in a university might manifest as a campus culture of integrity: doing research ethically, treating colleagues and students with justice, and focusing on meaningful scholarship rather than prestige alone. It is a habit of choosing value over expediency.

• Religious or spiritual conversion (Lonergan also calls it affective conversion) is an orientation of one's whole person toward ultimate meaning and love. It "relies on the love of neighbor, community, and God to heal bias and prioritize values" [12]. In secular terms, we might speak of this as a deep orientation toward hope, purpose, and solidarity with humanity. It imbues scholars with humility and a sense of service to something greater than themselves. A spiritually converted university would encourage hope and openness – perhaps not any single creed, but an ethos that knowledge serves transcendent ends (truth, justice, human dignity). This conversion underpins the courage to pursue truth wherever it leads, sustained by faith that doing so matters for the world.

These three conversions map onto the formation of scholars, students, and academic culture. A university aspiring to be Cosmopolis must cultivate all three: intellectual conversion through rigorous inquiry and self-critique; moral conversion through emphasis on ethics and the intrinsic goods of knowledge; and spiritual conversion by fostering a sense of higher purpose or vocation in its community. In sum, Lonergan's Cosmopolis is both a diagnosis and a prognosis: it diagnoses the ills of bias and short-sightedness, and it proposes that a converted, not-numerous center can rescue progress. The following sections explore how the modern university might embody this idea.

2. Bias in the University Context: Identifying the Challenges

Before envisioning the university as a cosmopolitan center, we must confront the reality of biases within academia itself. Lonergan outlined several types of bias – notably individual, group, and general bias – that distort inquiry and precipitate decline. Unfortunately, universities are not immune to these maladies. Indeed, many current crises in higher education can be traced to exactly these biases warping academic priorities. Let us consider each in turn, with concrete examples:

• Individual bias (egoism and careerism): Academia is rife with personal pressures that can compromise honest inquiry. The pursuit of tenure, promotions, grants, and reputation can become an end in itself, tempting scholars to prioritize self-advancement over the disinterested search for truth. Academic careerism has been defined as "the tendency of academics to pursue their own enrichment and self-advancement at the expense of honest inquiry, unbiased research, and dissemination of truth" [13]. We see this when researchers choose "safe" topics likely to get published rather than essential questions, or when they cut corners in experiments to rush out results. The notorious "publish or perish" culture exemplifies individual bias: scientists lament that relentless publication and funding pressures incentivize superficial, short-term research and even contribute to a reproducibility crisis in science[14]. In a 2016 survey, over 70% of researchers reported they could not replicate another scientist's work, and most blamed the "infamous 'publish or perish' research culture" for undermining quality[14]. Similarly, status games like chasing prestigious journal placements can distort judgment, as scholars might favor trendy topics or positive results that editors prefer. The result is a deformation of academic purpose – knowledge becomes a means to personal success, rather than personal success being a byproduct of contributing to knowledge. This individual bias, if unchecked, erodes the integrity and trustworthiness of university research.

• Group bias (tribalism and silos): Alongside personal egoism, academia suffers from group biases. This appears in the form of disciplinary silos, departmental turf wars, cliques of scholars who reinforce each other's views, or ideological camps on campus that won't engage in genuine dialogue. Universities are often organized into rigid departments, each with its own priorities, funding, and governance, which makes collaboration across fields difficult[15]. Faculty may become so immersed in their subfield's methods and jargon that they dismiss perspectives from other disciplines. Interdisciplinary initiatives, when attempted, are frequently undervalued or even resisted – seen as "less rigorous" or a threat to departmental resources[16][15]. For example, a professor who wants to co-teach a course bridging, say, biology and philosophy might face scheduling conflicts, a lack of credit toward tenure for such work, and skepticism from colleagues entrenched in more narrowly defined research[15][17]. The result is a fragmentation of knowledge: significant problems that span multiple domains (climate change, inequality, public health, etc.) are tackled in a piecemeal fashion, if at all, because no single silo can encompass them. Group bias also emerges in ideological polarization on campus. Entire departments can sometimes develop a dominant ideological bent – whether explicit (e.g., a particular economic or political theory) or implicit – and may marginalize or hire only those who conform. Such intellectual echo chambers stifle the very debate and self-critique that lead to new insights. When group bias takes hold, the university loses its broader mediating role and becomes just a collection of competing factions, each guarding its territory.

• General bias (short-term pragmatism and utilitarianism): The third form of bias Lonergan noted is a general bias of common sense – a societal tendency to discount long-term and theoretical concerns in favor of immediate practical gains. In the university context, this appears as the pressure to treat education purely as job training or to prioritize research with instant applications at the expense of fundamental inquiry. We see signs of general bias in how universities market themselves and allocate resources: an ever-greater emphasis on producing "work-ready" graduates, on majors that lead directly to high-paying jobs, and on research that can yield quick commercial or technological payoffs. Of course, practical relevance is not bad – but when it narrows the university's mission to only what is immediately marketable, the deeper purposes of higher education are eroded. As one governance expert notes, "colleges often focus on programs that look like job training, which is not what employers say they want" [18]. Higher education's obsession with quarterly outcomes (enrollment numbers, placement stats, rankings) can lead to neglect of unquantifiable goods like character formation, civic education, or blue-sky research whose benefits might only emerge decades later. For instance, humanities and pure sciences often find themselves on the defensive, justified only insofar as they can be tied to current labor market needs. This orientation is a manifestation of Lonergan's general bias: the richness of culture and theory forced to "justify itself to the practical mind" [5]. When universities yield entirely to this bias, they risk becoming glorified vocational institutes or corporate R&D arms, losing their role as society's long-term memory and imagination. A dramatic example is how funding floods into applied fields like AI engineering or finance (with immediate industry payoff) while foundational fields like philosophy or pure mathematics struggle – even though the latter cultivate critical thinking and innovation that society will desperately need in the long run.

Examples of decline: These biases have contributed to real crises in contemporary higher ed. The replication crisis in psychology and biomedicine – where many published findings turned out unreliable – is frequently blamed on the "publish-or-perish" mentality that prioritizes quantity of output over careful quality[14]. Group biases have led to high-profile clashes and stagnant debates: consider disciplines that split into hostile camps (e.g., economics feuding between classical and heterodox schools, or sociology's quantitative vs qualitative divide), where each side dismisses insights of the other, hampering genuine progress. The scandal of particular business and engineering programs being overly influenced by corporate sponsors can illustrate general bias – research agendas get skewed toward the immediate interests of funders, sidelining questions of public good or ethics. Even student life reflects these biases: students often feel torn between learning for its own sake and the pressure to pad résumés; between exploring diverse ideas and sticking to a safe ideological circle. When a narrow focus on practical outcomes dominates, students can graduate technically skilled but lacking the integrative thinking and ethical grounding that authentic leadership requires. All these trends – careerism, siloization, short-termism – indicate a drift from the university's deeper mission. They are precisely the tendeCosmopolisopolis is meant to counter.

3. The University as a Cosmopolitan Function: Mediating Common Sense and Theory

How, then, could the university act as a Cosmopolis, correcting these biases and reversing decline? One key role Lonergan assigns to Cosmopolis is to mediate between everyday common sense and the world of theory[5]. In society at large, common sense is concerned with practical, immediate needs and often mistrusts abstract theory – whereas scientists and scholars operate in highly specialized, theoretical realms frequently disconnected from everyday understanding. This gap can breed mutual suspicion: the public grows skeptical of ivory-tower academics, and academics sometimes grow disdainful of lay knowledge. A cosmopolitan university would strive to bridge this divide, ensuring that rigorous theory remains connected to human meaning and that common sense is elevated by insight rather than hardened into ignorance.

Accessibility and Rigor: To mediate between common sense and theory, universities must become bilingual – fluent in both the language of the street and the language of the lab. This could mean encouraging faculty to communicate their research in accessible ways and rewarding those who engage the public. For example, rather than only valuing conference presentations for peers, a cosmopolitan scholar might also give public lectures at local libraries or write op-eds explaining what their findings mean for everyday life[19]. Such outreach is not a trivial extra; it is part of the core mission of making knowledge a common good rather than a private dialect. Likewise, teaching should not be an afterthought to research but a coequal duty through which professors translate complex ideas into pedagogical common sense for students. The university can provide structures for this mediation: centers for science communication, incentives for interdisciplinary teaching that connect theory to real-world issues, and platforms (like open-access journals or public seminar series) to share ideas beyond campus. By deliberately straddling rigor and accessibility, the university as Cosmopolis counters the general bias – it "withdraws" a bit from immediate practical frenzy (insisting on depth and theoretical integrity) yet returns with insights made understandable and relevant to society[6].

Structures for genuine progress: A cosmopolitan university would also design its internal structures to sponsor long-term progress and resist capture by passing factions or fads. One principle might be to diversify oversight and input. For instance, universities could form cross-disciplinary councils that periodically review major research programs and curricula for bias or blind spots. Such councils, comprised of faculty from various fields (and perhaps external experts or alumni), could act as an internal cosmopolitan senate – mediating between different specialized perspectives and the broader mission. Their task would be to ask: "Are we missing something important? Are our priorities skewed by fashion or interest? What needs to be questioned?" This resonates with Lonergan's idea that Cosmopolis does not rule by force but by ideas – it identifies the ideas that general bias suppresses and finds ways to bring them to the fore[5][2]. In practice, this could mean creating funding pools for high-risk, high-reward research that might not win immediate grants, to ensure groundbreaking ideas aren't strangled by short-term metrics. It could also involve rotation of leadership positions (department chairs, deans) to prevent any one clique from entrenching power – fresh leadership can bring new perspectives and guard against group bias becoming institutionalized.

Feedback loops for bias correction: The university-as-Cosmopolis would build explicit feedback loops to catch and correct bias. One model could be akin to the Jesuit practice of the examen (a regular reflective self-examination). Applied institutionally, a university might hold an annual "Bias Review" retreat: stakeholders gather to reflect candidly on where the university might be drifting off mission. Are we overlooking specific disciplines? Are we failing to serve some student populations? Has short-term thinking influenced our decisions this year? By making bias-detection a routine process (rather than waiting for a crisis), the institution acknowledges that it, too, is a historical actor prone to blind spots. Another feedback mechanism is the stronger integration of student and community voices. Students often see hypocrisies or emerging needs that faculty leadership might miss. A cosmopolitan approach would empower student representatives not just on token committees, but in genuine dialogue about the direction of the institution. Likewise, universities can invite external critique: maybe a panel of public intellectuals or diverse alumni periodically examines whether academic programs remain connected to real societal needs (without succumbing to populist demands).

One practical example is how some universities now conduct "mission reviews" or invite accreditation teams to give holistic feedback. A cosmopolis-university would treat these not as box-checking but as opportunities to surface biases. If, say, an accreditation report finds that a university's research output is excellent but only in narrow areas with industry funding, that is a cue to ask how the pursuit of grants might be biasing the research agenda – and then to self-correct by investing more in fundamental or neglected fields. In summary, by institutionalizing critical self-reflection, the university can emulate Cosmopolis's work of "making operative the ideas that, in the light of the general bias of common sense, are inoperative" [20]. It becomes a mediator between short-term common sense and long-term reason, between partisan extremes and a larger viewpoint.

4. Institutional Design & Lonergan's Functional Specialties: A New Workflow for Inquiry

Lonergan not only diagnosed problems, but also proposed a method for collaborative progress: the idea of eight functional specialties in theology, introduced in Method in Theology. These functional specialties – Research, Interpretation, History, Dialectic, Foundations, Doctrines, Systematics, Communications – are essentially an organized workflow for a community of inquiry[21][22]. They ensure that scholars systematically cover all tasks from investigating data to reflecting on values and communicating conclusions. While Lonergan devised them for theological scholarship, the approach is profoundly generalizable. We can imagine a university adapting this eight-fold method as an alternative to the disjointed "publish-or-perish" treadmill, integrating disciplines and restoring public trust through a transparent, comprehensive approach to knowledge. Here's a brief look at each specialty and its academic application:

1. Research: Gathering evidence, data, and insights from primary sources or experiments. In a university setting, this corresponds to all the investigations – scientific experiments, archival research, field studies, surveys – that form the raw material of knowledge. It values empirical rigor and curiosity.

2. Interpretation: Making sense of the data by interpreting texts, results, and experiences. Academics translate raw findings into meaningful patterns or narratives. For example, a historian interprets archival documents' significance; a physicist interprets experimental results via theory. Multiple perspectives might be employed here to avoid one-dimensional readings.

3. History: Placing knowledge in context by understanding how ideas and situations developed over time. This step acknowledges that any issue (scientific, social, or intellectual) has a history that must be understood to avoid repeating errors. In an institutional workflow, this could mean literature reviews, historiography of a problem, or situating research in the broader timeline of a field. It guards against the bias of presentism.

4. Dialectic: Confronting and resolving conflicts or differences in viewpoint. Lonergan's dialectic stage is explicit: scholars bring together differing evaluations and accounts to sort out contradictions[23][24]. Within a university inquiry, this could take the form of interdisciplinary seminars or debates where, say, an economist, a sociologist, and an ethicist examine a complex issue like poverty from their angles, identify where their conclusions clash, and work through why. The dialectic function is crucial for surfacing bias: it forces recognition of divergent assumptions and asks which viewpoints might be more adequate. Lonergan notes that the most radical divergences come from differences in conversion (intellectual, moral, etc.)[25]; a structured dialectic allows converted (more unbiased) horizons to challenge lesser ones[26]. In a cosmopolis-university, something like a "dialectical forum" could be built into major research projects – ensuring internal critique and diversity of thought before conclusions are solidified.

5. Foundations: After the dialectic, participants clarify and choose fundamental principles or a standpoint. In Lonergan's terms, they make explicit their commitments and the horizon (worldview) they will operate from[27]. In a university process, this might mean formulating the values or axioms that the research or policy will take as a base, informed by the dialectical self-critique. For example, after interdisciplinary debate, a team researching climate change might explicitly commit to a foundation: "We value sustainability and justice for future generations, and we acknowledge the scientific consensus on warming." Foundations, thus, is a self-aware stance that incorporates intellectual, moral, and even spiritual conversion. It's like setting the North Star before moving ahead – making sure the work rests on authentic and transparent values, not on unexamined biases.

6. Doctrines (or Policy/Plans): Given the foundations, this stage formulates specific proposals, doctrines, or models to address the issue at hand[22]. In theology, "doctrines" are teachings; in academia more broadly, we can think of this as drawing conclusions or making recommendations. It could be a theory in science, a policy recommendation in public policy, a philosophical thesis – the fruit of the prior stages, now explicitly articulated. Importantly, because it comes after foundations, these doctrines should carry the weight of the careful reflective process behind them (not just quick hypotheses).

7. Systematics: Here, the various doctrines or insights are integrated into a coherent system or framework[21]. For a university, this stage would encourage connecting the dots across disciplines. It asks, "How do our conclusions fit together into a unified understanding?" In practice, this might result in a multidisciplinary report or a book that synthesizes findings from science, ethics, history, etc., into a comprehensive view – accessible to intelligent readers of all backgrounds. Systematics prevents the fragmentation of knowledge; it insists on the intelligibility of the whole.

8. Communications: Finally, the results are communicated to those who need to hear them – within and beyond academia[21]. This method means tailoring the message to different audiences: scholarly publications for specialists (to invite further scrutiny), but also policy briefs for decision-makers, public articles for laypeople, and educational materials for students. In the functional specialties approach, Communications is not an afterthought but a dedicated function, recognizing that a discovery doesn't truly contribute to progress until it is understood and implemented by the wider community. For a university, emphasizing this stage could rebuild public trust: instead of knowledge staying locked in journals, it actively reaches society. (Imagine, for instance, a university where every major research project has a public-facing component – a website, a community forum, a dataset released for public use. This transparency and engagement fulfill the mediating role of Cosmopolis.)

By implementing something like Lonergan's eightfold method, universities could counteract the haphazard "produce papers fast" model with a more deliberate, collaborative, and reflective workflow. It is noteworthy that Lonergan warned that these functional specialties are not just parallel silos or new departments[22]. They are tasks that the same group of scholars may undertake sequentially, or that different specialists coordinate on, but always in relationship. In other words, it's a process, not a bureaucracy. Adopting such a process in academia would mean, for example, that a research grant isn't considered done when results are published. It would be complete when those results have been debated (dialectic), value-checked (foundations), integrated (systematics), and communicated broadly. This could alleviate the publish-or-perish pathologies by focusing on knowledge quality and impact rather than volume. It might also restore trust by making the stages of knowledge creation visible and participatory, rather than a black box that churns out occasional press releases. As one interpreter notes, Lonergan's grouping is about how "we actually do better… not a recipe for better living, but an explanation of how the mind and heart work whenever we actually improve life" [22]. Imagine if universities explicitly self-organized around "improving life" in this structured way – they would truly function as cosmopolitan centers of insight in society.

5. Case Studies & Comparative Models: Lessons from History and Today

The idea of the university as Cosmopolis may sound radical, but it resonates with historical precedents and contemporary experiments. At various moments, educational institutions have acted as centers of integration and progress – essentially playing a cosmopolitan role in their eras. By examining these cases, we can glean lessons and see that Lonergan's vision is not utopian fancy but a reclaiming of the university's highest potential.

Historical exemplars:

• Medieval Universities (12th–15th centuries): The earliest universities in Bologna, Paris, Oxford, and elsewhere were born as cosmopolitan institutions at the crossroads of old and new. They inherited the classical and biblical knowledge of antiquity (the "old") and were confronted with discoveries and cultural needs of a changing medieval society (the "new"). Medieval curricula embraced the liberal arts as a unified foundation, integrating theology, philosophy, and emerging natural science into a holistic worldview[28]. Knowledge was seen as interconnected – studying nature was also a way to understand divine order, and vice versa[28]. These universities were few (not numerous) but profoundly influential centers that mediated between tradition and innovation: for example, when Aristotle's works were rediscovered in the 13th century, it was the University of Paris's scholars (like Thomas Aquinas) who painstakingly worked out a synthesis between Aristotelian philosophy (new to medieval Europe) and Christian theology (the inherited tradition). In doing so, they corrected many biases of their time – irrational superstitions on one hand and overly rigid dogmas on the other – by insisting on reasoned, complete solutions grounded in both faith and reason. The medieval university's approach to knowledge as a unified whole began to erode later with hyper-specialization. Still, in its prime, it offers a model of a university cosmopolis, balancing continuity and change.

• Renaissance & Early-Modern Academies: During the Renaissance, learning often flourished in academies and scholarly societies that cut across the emerging divides of discipline. The Platonic Academy in Florence (15th century) brought together thinkers to discuss philosophy, poetry, science, and art, sparking creativity that defined the Renaissance humanism. Later, the 17th-century scientific academies (like the Royal Society in England or the Accademia dei Lincei in Italy) explicitly sought to break free of old scholastic constraints and investigate the new sciences – yet many members were also well-versed in the humanities, aiming for a complete vision. These academies acted as incubators of new knowledge (Scientific Revolution) and at the same time custodians of rigorous method. A striking insight from recent scholarship is that Renaissance academies, whether humanistic or scientific, shared a "common institutional culture" that refused to acknowledge strict disciplinary divides[29]. Academies of literature, art, and philosophy contributed to new art forms and ideas, while scientific academies legitimized new empirical methods[29]. Often, the same people traversed both worlds. This was the era of the "Renaissance man" – thinkers like Leonardo da Vinci or Galileo who exemplified breadth and integration. While we cannot revert to a time when one polymath could master all knowledge, the spirit of the Renaissance academy – collaborative, objectivity-seeking across disciplines[29] – is worth emulating. Those institutions were small circles (indeed "not numerous"), yet their influence on societal progress was enormous. They also faced challenges (some became elitist or were co-opted by patrons), but their legacy shows the power of an intellectual cosmopolis uniting science and the humanities.

• Humboldt's Reforms (early 19th century): A pivotal case for the modern university is Wilhelm von Humboldt's founding of the University of Berlin in 1810. Humboldt reimagined the university with two core principles: academic freedom and the unity of teaching and research. He insisted that the university's mission was not mere vocational training or passive transmission of known facts, but the active creation of knowledge (research) intertwined with education[30]. This was revolutionary – before, many universities had been mostly teaching institutions. After Humboldt, the model of a research university took hold, aligning with what Lonergan would call an orientation to long-term discovery rather than short-term utility. Humboldt believed seeing research and teaching as opposed was a false dichotomy[31]. By having professors both investigate and teach, students would learn not just facts but the spirit of inquiry itself. The Humboldtian university thus acted as a mediator between new knowledge and the next generation, and between specialized discovery and general education. It also emphasized breadth: students should receive a well-rounded formation (Universitas litterarum, the unity of all knowledge) even as they specialize[32]. This model spread worldwide and became the template for modern higher education. Over time, it has been distorted in some places (with research and teaching often split or research prioritized to the detriment of teaching), but the ideal remains powerful. Recapturing Humboldt's vision, one scholar argues, could help resolve today's perceived conflict between specialized research and broad education – what many see as a tension, Humboldt saw as a synergy[31]. His reforms also introduced the concept of academic freedom – freeing scholars from church or state dictates – which is essential for a cosmopolis function, allowing the pursuit of truth wherever it leads. The lesson here is that institutional design matters: structure the mission and incentives correctly, and a university can be a haven for independent, integrated thought. Many American and global research universities still cite Humboldt as inspiration; to truly channel Cosmopolis, they may need to renew those commitments to holistic knowledge and freedom, especially under today's pressures.

Contemporary models and experiments:

• Interdisciplinary Institutes: In recent decades, we've seen the rise of institutes and centers explicitly designed to tackle complex global challenges by transcending departmental boundaries. One example is the Santa Fe Institute (SFI), founded in 1984 as an independent research center for complex systems. SFI's founders – scientists from various fields – "sought a forum to conduct theoretical research outside the traditional disciplinary boundaries" and as an alternative to the increasing specialization they saw in academia[33]. They structured SFI in novel ways: no permanent departments, a small rotating faculty, a constant influx of visiting scholars from different disciplines, and no tenure – all to encourage fresh ideas and prevent any single paradigm from becoming orthodoxy[34]. The institute deliberately fosters active turnover in ideas and people to stay at the cutting edge of interdisciplinary science[34]. In effect, the Santa Fe Institute functions as a micro-cosmopolis: it stands apart from practicality (much of its work is theoretical) yet produces insights that later revolutionize practical fields (from economics to biology). It also resists group bias by its very structure – no silo can form because everyone is constantly engaging across fields. The success of SFI (measured by its outsized impact on complex network theory, chaos theory, etc.) suggests that a cosmopolitan approach to knowledge can work – and that structural choices (like not having internal departments and encouraging visiting scholars) make a difference. Traditional universities can learn from such models by creating more fluid spaces within their walls: e.g., establishing interdisciplinary labs or "centers of convergence" where faculty from different colleges co-locate physically for a few years to work on big problems, free from some departmental reporting lines. Some universities are indeed moving this way, setting up "challenge-based" institutes (on climate, pandemic response, social justice, etc.) that draw experts from all faculties. The key is giving these centers enough autonomy and support to break the usual mold.

• Long-range Consortia and Big Science: Another contemporary phenomenon is the creation of multi-institution, long-term research projects around grand challenges. Think of the Human Genome Project (involving universities, government labs, international partners over 13 years to map human DNA), or CERN's particle physics collaborations, or large-scale epidemiological studies. These consortia pool knowledge from many disciplines and often from around the world, functioning as a distributed university focusing on one goal. They illustrate how patience and complete solutions pay off: the Genome Project was sometimes criticized in its early years for being a "big science" drain of resources with no immediate use, but because the scientific community committed to a complete solution (sequencing the entire genome) and waited, today we have breakthroughs in medicine and biology that are transforming lives. A cosmopolis-style university would align with this approach: committing to sustained efforts even if the payoff is a generation away. We already see more universities collaborating in this mode – e.g., the U.N. Sustainable Development Solutions Network links many academic institutions tackling long-term goals like zero hunger or quality education globally. These are nascent cosmopolis networks. The challenge is keeping them insulated from short-term political or economic winds so they can finish their course, which requires, again, a culture of conversion (valuing truth and humanity's welfare above immediate gains).

• Experimental Universities and Pedagogies: A few bold experiments in higher education are directly addressing bias and fragmentation. For example, Minerva Schools (founded 2014) offers a curriculum that is entirely interdisciplinary and skills-based, with students living in different countries to gain a global perspective; faculty are not organized by traditional departments but by broad "College of Arts and Sciences" divisions. While Minerva is small, it tests how removing departmental silos and focusing on learning how to think (intellectual conversion) might yield cosmopolis-ready graduates. Other experiments include Olin College (an engineering college that eliminated traditional departments and tenure, requiring faculty to update courses with input from students and real-world partners constantly) and Arizona State University's many interdisciplinary initiatives under President Michael Crow's idea of the "New American University." ASU, a huge public university, has created transdisciplinary schools (for instance, the School of Sustainability) and collapsed some departments into broad divisions to encourage cross-pollination. It also emphasizes inclusion and societal impact, trying to avoid elitism and connect with its local communities. These efforts echo Lonergan's call for a center "at home in both the old and the new" [35] – e.g., ASU combines the old ideal of broad liberal education with a new practical focus on sustainability and social tech. While results are mixed and it's hard to measure cultural impact, such models suggest that universities can reinvent themselves structurally to play a mediating, progressive role better.

In short, history and current practice provide proof of concept for the universiCosmopolisopolis. Medieval and Humboldtian universities show the power of integrated knowledge under guiding values; Renaissance academies and modern institutes show the creativity unleashed by crossing disciplines; large consortia demonstrate the need for patience and completeness; and new experiments show an appetite for reform. These examples also caution us: each had pitfalls (some early universities became ossified; some academies devolved into elitist clubs). The task is to glean the strengths – unity of knowledge, openness, commitment to truth over utility – and avoid the weaknesses, which lead to our final considerations on formation and safeguards.

6. Conversion and Formation: Educating Cosmopolis-Ready Graduates and Scholars

If the soul of Cosmopolis is converted individuals (intellectually, morally, spiritually), then the university must take seriously its role in forming such people. This means extending education beyond mere knowledge transmission to personal transformation. How can a university cultivate intellectual, moral, and spiritual conversion in its students and faculty, so that its community naturally embodies the not-numerous center ethos? Here we propose several practices and designs for curriculum, pedagogy, and campus culture:

• Fostering Intellectual Conversion: Education should lead students to what Lonergan calls "self-appropriation" of their knowing – essentially, to become aware of how they know and to love truth above image or ego. In practical terms, this means pedagogy must deeply emphasize critical thinking. Socratic dialogue, inquiry-based learning, and metacognitive reflection can all help. For example, rather than just lecturing facts, a professor might regularly have students reflect in writing: "How did I arrive at this answer? What evidence am I using? Could I be wrong?" Some universities incorporate philosophy of science or epistemology modules even for STEM students, to challenge them to think about thinking. Interdisciplinary courses are also powerful: when students have to integrate methods from, say, biology and ethics to solve a problem, they are forced to step outside simple assumptions and see the contingency of frameworks – a step toward intellectual conversion. On the faculty side, encouraging interdisciplinary research and discussion groups can refresh scholars' intellectual horizons, preventing tunnel vision. A culture of openness – where saying "I don't know" or "I changed my mind" is applauded rather than scorned – will nurture the humility and wonder that mark an intellectually converted person. In short, the curriculum should not just transfer information but constantly pose questions back to the student about how they are learning and what it means.

• Fostering Moral Conversion: Moral conversion, the primacy of value over mere satisfaction, can be integrated into university life by foregrounding ethics and service. This doesn't mean every class becomes an ethics class, but every discipline can illuminate the values at stake in its pursuits. For instance, computer science programs now often include units on the ethics of AI, prompting students to think beyond "can we build it?" to "should we build it and for whose benefit?" Many universities have service-learning requirements or opportunities – students volunteer or work in the community connected to their field of study. When well-designed, these experiences confront students with fundamental human needs and ethical dilemmas, challenging them to go beyond self-centered goals. Even campus governance can be an educational tool: involving students in honor codes or disciplinary panels teaches them to value integrity and justice. On the faculty side, moral conversion is fostered by creating an environment that rewards ethical behavior. Suppose a professor knows that mentoring students and adhering to research integrity are valued by the institution as much as winning grants. In that case, they are less likely to fall into the trap of careerism. Universities could incorporate ethics more explicitly into promotion criteria and daily decision-making. Some medical and law schools already have their students take oaths or pledge commitments to serve society; perhaps a cosmopolitan university might encourage all graduates to articulate the values they intend to uphold in their professional and civic lives. These are not empty rituals – they build an identity of responsibility. Ultimately, a morally converted campus culture is one where doing the right thing is openly valued above doing the profitable or prestigious thing. Stories of moral courage (whistleblowers, public intellectuals speaking truth to power, researchers who gave up a lucrative path to address a societal need) should be elevated as role models for students.

• Fostering Religious/Spiritual Conversion: Universities, especially secular ones, often shy away from anything resembling spiritual formation. However, spiritual conversion in Lonergan's sense is not about doctrine but about a basic orientation towards hope, charity, and purpose – something that can be cultivated in inclusive ways. A cosmopolis-as-university would encourage students and faculty to connect their intellectual pursuits with a sense of meaning and wholeness. This might be done through offering spaces and times for reflection. For example, some colleges have begun hosting mindfulness meditation sessions or retreats focused on contemplation (open to people of any faith or none). These aren't mere wellness fads; when tied to the intellectual life, they help individuals center themselves on what ultimately matters. Another approach is via the arts and humanities – literature, music, and art on campus can open up transcendent experiences and questions of purpose. Universities historically always included chaplaincies or forums for religious dialogue; today, a cosmopolitan approach could broaden that to interfaith and humanist dialogues on campus, where students discuss existential questions (What is a good life? What do we owe future generations? How do we handle suffering and failure?) in a supportive setting. The goal is not to impose answers but to signal that grappling with ultimate questions is integral to higher education. When students see their institution values more than just their GPA or job placement – that it cares about their character and soul – it frees them to become genuine seekers. Faculty, too, benefit from this atmosphere; burnout and cynicism (typical in academia) are often spiritual maladies of meaninglessness. By nurturing a community ethos of hope and shared higher purpose (for instance, through voluntary faculty seminars on the meaning of the academic vocation, or simply a compassionate work culture), the university helps scholars remain connected to the love of truth and love of humanity that likely drew them to academia initially. That love, in Lonergan's terms, "relies on the love of neighbor, community, and God (or the Good) to heal bias" [12].

• Cosmopolitan Curriculum and Campus Culture: A concrete design for a cosmopolis-ready curriculum might include a first-year course that frames the whole college journey around big questions and integration of knowledge (some universities have "great books" or "global challenges" seminars that attempt this). It would include a solid core of liberal arts to ensure every student gains breadth and the ability to see beyond one specialization. It might culminate in a capstone project where students have to address a complex real-world problem, bringing together knowledge from science, ethics, policy, etc., and explicitly reflecting on impacts and ethics. Pedagogically, dialogical teaching methods (where students and teachers learn in exchange) model the intellectual humility we seek. And importantly, the hidden curriculum – how people treat each other daily – should model Cosmopolis as well. Is the campus one where diverse viewpoints can be discussed civilly? Do professors model lifelong learning and openness by sometimes attending each other's lectures or learning from students? Are achievements of cooperation celebrated as much as individual achievements? For faculty formation, providing ongoing development workshops that expose them to new pedagogies or interdisciplinary content can keep their horizons expanding. A professor who undergoes her conversion (say, realizing the importance of another field's approach or rediscovering her ethical commitments) will naturally influence students. Thus, formation is an ongoing, community-wide effort. The outcome we aim for is graduates (and professors) who are comfortable living at the center: scientifically competent and humanistically wise, principled and creative, rooted and open. These are the kinds of leaders society desperately needs – those who refuse easy answers and instead work toward "complete solutions even though [they have] to wait" [8].

7. Failure Modes and Safeguards: Risks of a Cosmopolis University and How to Mitigate Them

No ambitious vision is without pitfalls. If we consciously reshape the university into a cosmopolitan "not-numerous center," we must be vigilant about specific failure modes. History and human nature warn us that even well-intentioned intellectual elites can go astray. Here are some risks and corresponding safeguards to consider:

• Risk of Elitism: A group that sees itself as a lofty "center" might grow arrogant and disconnected from those outside it. There's a fine line between being not numerous and being out of touch. If universities double down on being guardians of long-term truth, they could slip into paternalism, acting as if they always know best. Safeguard: Humility and engagement must be baked into the cosmopolis ethos. Universities should constantly dialogue with external communities – not to pander, but to listen and learn. Including practitioners and ordinary citizens in specific decision-making processes (for example, community advisory boards for research that affects the public) can keep the university responsive. Emphasizing service in the university mission counters elitism: professors and students should see themselves as servants of humanity, not an aloof intelligentsia. Metrics of success can include community impact and public trust, not just academic prestige. Essentially, the university needs to remain of the people even as it provides leadership.

• Risk of Bureaucratization: Ironically, trying to implement something like the eight functional specialties or lots of feedback loops could drown the university in process and paperwork. One can imagine committees proliferating in the name of interdisciplinarity and self-reflection, until nothing gets done. Safeguard: Keep structures lean and purposeful. The point of functional collaboration is to enhance creativity, not smother it. Any new process (say, a bias review panel or an interdisciplinary council) should have an explicit sunset clause or periodic evaluation of its usefulness. Rotate membership to bring fresh energy rather than creating permanent bureaucrats. Use informal as well as formal mechanisms – sometimes the best "dialectic" is an unconstrained colloquium over coffee rather than a mandated report. Also, empower small teams rather than large committees wherever possible. A non-numerous center implies smallness: we should trust small groups of committed people to generate ideas, rather than requiring a huge consensus for every move. In short, becCosmopolisopolis is more about culture than bureaucracy. Cultivate the culture, and streamline the structures.

• Risk of Ideological Capture: Just as a political or ideological faction can capture any university, so can a Cosmopolis-style one. If we explicitly champion a "center" position, there's a danger of performative centrism, where claiming to be above extremes becomes an identity that could itself shut down critique. Alternatively, strong personalities could steer the institution's mission toward their ideology under the guise of pursuing truth. Safeguard: Pluralism and transparency. A true cosmopolis welcomes structured conflict (Lonergan's dialectic) as a way to weed out bias. Therefore, enshrine policies that protect academic freedom and diversity of thought: hire faculty with different philosophical outlooks, encourage respectfully contested dialogue in public events, and ensure no donor or external agenda can secretly dictate research directions. If, for example, a major donor favors a particular economic theory, the safeguard is to have a governance system that still guarantees support for scholars critical of that theory. Another concrete measure: regular external reviews by diverse peers. Inviting outsiders (with various viewpoints) to assess whether the university is living up to its ideals can expose creeping biases. The key is making bias correction a continuous process, not a one-time achievement. A cosmopolitan university would institutionalize debate – for instance, hosting annual "ideals and ideologies" conferences where people inside and outside critique the university's positions. This keeps the center honest.

• Risk of Burnout and Overextension: Being the conscience of society is hard work. Faculty and staff might burn out trying to excel in their disciplines simultaneously, engage with the public, mentor students deeply, and reflect on institutional bias endlessly. Students, too, might feel the weight of such an ambitious education. Safeguard: Sustainability applies not only to the environment but to human effort. The university should model balanced life and collaboration. Team teaching and research teams can distribute workloads that previously fell on individuals. Providing robust support (mental health services, mentorship, reasonable workload policies) ensures people don't flame out. Cosmopolisopolis isn't built in a day. Perhaps each year, an institution chooses a couple of major initiatives (e.g., this year we focus on improving interdisciplinarity in undergrad curriculum, next year on community partnerships) rather than trying everything everywhere. Celebrating small wins boosts morale. And the spiritual conversion aspect plays in: a community that has a hopeful, transcendent outlook will support members in rough times, reminding them why the work matters. Burnout is best cured by reconnecting with purpose and by a sense of shared journey rather than isolated struggle.

• Risk of Performative Centrism: In some contexts, positioning oneself as "the center" can become a self-righteous stance that avoids taking principled stands. A university might pride itself on being above the fray, but then fail to speak out even when one side aligns more with truth or justice. In other words, centrism can become a cover for inertia or timidity. Safeguard: Remember that Cosmopolis is not about splitting the difference; it's about a higher viewpoint that may very well challenge both or either side strongly. So the safeguard is courage and clarity of values (from the conversions). If a cosmopolis-university sees evidence of a truth that is politically unpopular, it must be willing to "insist on complete solutions" [8], even alone. This is difficult – it may attract criticism from all sides. But part of the formation should be courage. Also, accountability helps: the university can ask, "What have we done this year that demonstrates our commitment to truth and justice in society, tangibly?" If the answer is "we held some seminars but avoided any controversial impact," then perhaps the center has grown too comfortable. In practice, having diverse voices ensures that at least someone will call out complacency. Maybe an internal devil's advocate role could be formalized – someone whose job is to constantly test whether the institution is living up to its talk.

Final safeguard – the Examen for Institutions: We mentioned an examen-like self-review. This routine could be the ultimate meta-safeguard: a periodic, perhaps bi-annual, deep reflection process where all the above risks are considered. It could take the form of an internal white paper or retreat outcomes that honestly ask: "Are we elitist? Are we bureaucratic? Are we co-opted? Are our people thriving or burning out? Are we making a difference?" The responses should lead to adjustments. In other words, build a mechanism to continually re-convert the university itself, because conversion is never one-and-done. This continuous renewal will help the cosmopolis-as-university remain true to its mission.

Conclusion: In the end, reimagining the university as Cosmopolis is a reorientation toward truth and wholeness. It asks the academy to recover its calling as society's patient thinker, cultural mediator, and guardian of meaning through time. This vision stands in contrast to the university as merely a credential factory, a partisan battlefield, or a corporate research arm. Instead, it's a vision of the university as a living tradition of inquiry that both honors the wisdom of the past and embraces the challenges of the future. It is small in the sense of focus and integrity – a "perhaps not numerous center" [1] – but it is enormous in influence when it works, because it sustains the very possibility of enlightened civilization.

"Building the Not-Numerous Center" called for a renewed middle in our public life. By asking how universities themselves can become that middle, we conclude that it requires intentional cultivation of conversion, community, and collaboration. Many of the pieces needed (interdisciplinary efforts, ethics initiatives, community engagement) already exist in embryonic form on campuses around the world. The task now is to integrate them and turn the dial from short-term fixes to long-term transformation. Academic leaders – presidents, deans, faculty, and student leaders alike – can take inspiration from Lonergan's Cosmopolis to see their institution with fresh eyes: not as an impersonal system producing degrees and papers, but as a cosmopolitan community devoted to unbiased truth and the common good.

Such a university might not immediately be rewarded in rankings or revenue. But over time, by insisting on complete solutions and refusing half-measures, it would earn something far more critical: trust and significance. It would be the place where the clamor of polarized extremes gives way to conversation, where fragmented facts become wisdom, and where generations are equipped not just with skills for the job market, but with the discernment and commitment to lead humanity forward. That is the promise of the university as Cosmopolis – a promise we cannot afford to leave unfulfilled.

Sources:

1. Lonergan, Bernard. Insight: A Study of Human Understanding, Collected Works of Bernard Lonergan, vol. 3, p. 26Cosmopolisopolis)[2][3].

2. Lonergan, Bernard. "Dimensions of Meaning," in Collection (on the "not numerous center" quote)[1].

3. Academic Careerism definition, Wikipedia[13].

4. Coddington, M. "Scientists Blame 'Publish or Perish' Culture for Reproducibility Crisis," Technology Networks, 2025 (survey data on reproducibility)[14].

5. Cappelli, P. "College and the Job Market Today," AGB Trusteeship, 2024 (on focus on job training vs broader skills)[18].

6. Mintz, S. "A General Education Curriculum That Matters," Inside Higher Ed, 2025 (on departmental silos and interdisciplinary barriers)[15][16].

7. Kappus, A. "Higher Education's Role in a Polarized America," Carnegie Blog, 2025 (on academics engaging with local communities)[19].

8. Lonergan, Bernard. Method in Theology, Collected Works vol. 14 (on functional specialties and conversions)[21][9].

9. McNeely, I. "The Unity of Teaching and Research: Humboldt's Educational Revolution," Oregon Humanities, 2002 (abstract on Humboldt)[30].

10. "Cosmopolis" entry, Lonergan Institute (quote on freeing culture from practical bias)[5].

11. McNeely, I. "The Renaissance Academies between Science and the Humanities," Configurations, 2011 (abstract on Renaissance academies bridging disciplines)[29].

12. Santa Fe Institute, Wikipedia (founding principles to avoid specialization)[33][34].

[1] [7] [8] [35] The Idea – Lonergan Institute

https://lonergan.org/2010/02/08/the-idea-2/

[2] [3] [4] [5] [6] Cosmopolis | Bernard Lonergan sj

https://lonerganmorin.wordpress.com/2008/01/11/cosmopolis/

[9] [10] [11] [12] [21] [22] [23] [24] [25] [26] [27] Lonergan, Bernard | Internet Encyclopedia of Philosophy

https://iep.utm.edu/lonergan/

[13] Academic careerism - Wikipedia

https://en.wikipedia.org/wiki/Academic_careerism

[14] Publish or Perish Culture Drives Reproducibility Crisis | Technology Networks

https://www.technologynetworks.com/biopharma/news/scientists-blame-publish-or-perish-culture-for-reproducibility-crisis-395293

[15] [16] [17] [28] A General Education Curriculum That Matters

https://www.insidehighered.com/opinion/columns/higher-ed-gamma/2025/04/25/general-education-curriculum-matters

[18] College and the Job Market Today - AGB

https://agb.org/trusteeship-article/college-and-the-job-market-today/

[19] Higher Education's Role in a Polarized America | Carnegie—Higher Ed Marketing & Enrollment

https://www.carnegiehighered.com/blog/higher-educations-role-in-polarized-america/

[20] Bernard Lonergan "Bias, Liberation, Cosmopolis"

https://staticweb.hum.uu.nl/susanne.k.langer/lonerganbiasliberationcosmopolis8.6.html

[29] (PDF) The Renaissance Academies between Science and the Humanities

https://www.academia.edu/889899/The_Renaissance_Academies_between_Science_and_the_Humanities

[30] [31] The Unity of Teaching and Research: Humboldt's Educational Revolution

https://scholarsbank.uoregon.edu/items/2fc75550-77a3-4832-a944-d3db17b2b61f

[32] University of Reform - Humboldt-Universität zu Berlin

https://www.hu-berlin.de/en/exzellenz-en/reformuniversitaet-en/standardseite

[33] [34] Santa Fe Institute - Wikipedia

https://en.wikipedia.org/wiki/Santa_Fe_Institute

“There is bound to be formed a solid right… a scattered left… But what will count is a perhaps not numerous center… painstaking enough to work out one by one the transitions to be made… strong enough to refuse half measures and insist on complete solutions even though it has to wait.” — Bernard Lonergan

We live in turbulent times. Technological disruption, cultural polarization, and institutional distrust dominate headlines. Amid rapid change, short-term wins often trump long-term good. In this essay, I provide a lens I’ve found helpful to understand this turmoil: the dynamic of progress, decline, and possible recovery. Human progress occurs when insights and creativity build upon one another, improving society. Decline sets in when bias and shortsightedness block those insights, causing society to drift or even regress. In Lonergan’s terms, “the principle of progress is liberty (the freedom for insight), and the principle of decline is bias”. When bias prevails, “intelligence comes to be regarded as irrelevant to practical living”, society settles into “a decadent routine, and initiative becomes the privilege of violence”. In other words, when people stop asking new questions or seeking truth, problems fester unresolved.

While there are many biases, we’ll focus on three major biases that distort our thinking and drive decline:

• Individual bias: the tendency to favor one’s own immediate interests and ego satisfactions at the expense of broader or long-term concerns. This “interference of personal desire with the development of intelligence” leads people to dismiss inconvenient insights. It results in short-sighted decisions that may benefit one person (or a small group) now, but undermine progress for others or for the future.

• Group bias: the tendency of groups, factions, or institutions to favor their own advantage and traditions, while ignoring or suppressing insights that benefit outsiders or the common good. Group bias skews the very formation of common sense in a community. New ideas are accepted only if they align with the group’s interest or power. Over time, “what originally was a neglected possibility becomes a distorted reality,” and social development twists into conflict between the “solid right” and “scattered left”.

• General bias of common sense: the most insidious bias, shared by everyone to some degree. It is “the propensity of common sense to extend its legitimate concern for the concrete and the immediately practical into disregard of larger issues and indifference to long-term results”. In plain terms, it’s the anti-intellectual streak that dismisses theory, expertise, or hard questions because they don’t have an immediate payoff. Common sense, invaluable in daily life, becomes a liability when it refuses to “analyze itself” or recognize that some problems require going beyond the here-and-now. General bias feeds short-term thinking in politics, education, and economics—sacrificing the future for the present.

These biases are engines of decline. They warp our perception of problems and our will to solve them. As a result, good ideas get shouted down or ignored, and genuine progress gets reversed. We see this today: entrenched interests resisting needed reforms, culture-war mentalities rejecting evidence, and widespread impatience for quick fixes. Left unchecked, biases can produce a vicious cycle of cumulative decline—a long “cycle of decline” that multiplies social ills. The question is: Who or what can correct bias without simply becoming another biased faction? In a society polarized into a “solid right” and “scattered left,” how do we escape an endless pendulum swing of extremes? The answer I’ll propose here is neither a revolution nor a technocratic elite, but something we can call cosmopolis – a collaborative higher standpoint that works patiently to reverse decline.

What Cosmopolis Is (and Isn’t)

Cosmopolis is a response to the problem of bias and decline. It is not a specific institution or utopian city, but rather a mode of collaboration and consciousness oriented to the long-term good of humankind. We can think of cosmopolis as a kind of enlightened cultural engine: “neither class nor state,” but a moral-intellectual movement “founded on the native detachment and disinterestedness of every intelligence”. In simpler terms, cosmopolis is an alert and creative “center” of persons who transcend partisan viewpoints, committed to truth and the welfare of all. It stands “above all [the] claims” of competing factions and “cuts them down to size” by exposing their biases. Crucially, cosmopolis demands our first allegiance be to truth and genuine progress, not to any party line.

One way to grasp cosmopolis is to say what it is not. Lonergan emphasizes that cosmopolis is not a political party or power bloc. It does not seek to seize governmental power or impose an ideology. It is not a technocratic elite of “smart people” who dictate to everyone else, nor a covert cultural faction fighting a “culture war.” It is also not a quick fix or emergency task force that will resolve crises overnight. Instead:

• Cosmopolis is not partisan or ideological: It is “not a political agenda” at all. People in the cosmopolis can come from any party, tradition, or background, but they refuse to be captive to slogans or tribal loyalties. Their loyalty is to understanding the situation truthfully.

• Cosmopolis does not rule by force or decree: “so far from employing power or pressure or force, [cosmopolis] has to witness to the possibility of ideas being operative without such backing.” It influences by truth and demonstrated solutions, not by coercion. Cosmopolis doesn’t take over; it persuades.

• Cosmopolis is not swept up in short-term urgencies: It deliberately resists the frenzy of immediate crises and popular whims. It refuses “half measures” that temporarily appease but don’t actually solve the underlying problem. It is willing to wait for complete solutions, maintaining a long view when everyone else demands instant results.

So what is cosmopolis positively? We can call it a transdisciplinary, trans-partisan center of inquiry devoted to the human good. Let’s imagine cosmopolis as a widespread but loosely organized effort by people in many fields and communities. Its members share a commitment to several core tasks:

• Diagnosing and overcoming bias: Cosmopolis continually identifies where bias—whether personal, group, or general—is distorting decisions and public debate. Its role is “to prevent dominant groups from deluding mankind by the rationalization of their sins,” as Lonergan vividly puts it. In practice, this means cosmopolis agents speak up when data is being denied or when scapegoating and slogans are trumping careful analysis. They shine light on uncomfortable truths that each side’s propaganda ignores.

• Mediating between common sense and theory: Cosmopolis works to connect high-level ideas with everyday practical understanding. Lonergan says it must be “at home in both the old and the new”—comfortable with inherited wisdom and new discoveries. In concrete terms, this means translating academic knowledge into terms the public and policymakers can use, and vice versa, bringing real-world concerns to inform theoretical research. Cosmopolis “sponsors genuine progress” by helping sound ideas move from the lab or think-tank into everyday practice.

• Keeping attention on the long view: Perhaps most of all, cosmopolis is “concerned to make operative the timely and fruitful ideas that otherwise are inoperative”. It combats the shortsightedness that plagues societies by insisting on thinking in terms of longer cycles and broader impacts. Lonergan writes that cosmopolis exists “to prevent practicality from being shortsightedly practical and so destroying itself.” It’s a kind of cultural memory and imagination that reminds everyone of larger goals beyond the moment.

In sum, cosmopolis is a counter-force to decline that doesn’t operate through domination but through understanding. It is the “perhaps not numerous center” Lonergan forecast: not necessarily a majority, but a vital minority large enough to influence both “sides” and dedicated enough to “work out one by one the transitions to be made” toward a better order. The spirit of cosmopolis can be found wherever people of intelligence, integrity, and goodwill collaborate across divides for enduring solutions. It is as much a personal stance as a group project—a refusal to give in to bias or despair. But living out this stance requires profound changes in ourselves. Three, in fact: intellectual, moral, and religious conversion. These conversions form the kind of people capable of building and sustaining cosmopolis.

Intellectual Conversion: From Naïve Realism to Critical Realism

The first conversion is intellectual. It’s a radical shift in how we understand knowledge and truth. Many people (including highly educated ones) operate with what Lonergan calls the “naïve realist” or empiricist mindset: the belief that “knowing is just looking” at what’s out there, as if truth were simply what our eyes or instruments see. In this view, objectivity means keeping the knowing subject (us) out of the picture—being a detached spectator who observes facts. This view is a myth, and intellectual conversion is the process of shattering that myth in favor of a critical realism.

Knowing is not a passive gaze but an active, structured process that happens within the subject. We experience, we inquire and have insights, we reflect and judge those insights, and then we decide what to do. Knowing anything truly involves this pattern of experience → insight → judgment (and then decision, which moves into the moral realm). Intellectual conversion is the moment we fully appropriate this pattern in ourselves. It’s the realization that “knowing ... is not just seeing; it is experiencing, understanding, [and] judging”. In other words, the objectivity of our knowledge does not come from us being detached onlookers, but from us performing these cognitional operations authentically and self-critically.

Lonergan famously expressed this with the maxim: “Genuine objectivity is the fruit of authentic subjectivity.” We attain objectivity by being attentive, intelligent, reasonable, and responsible in our knowing. Instead of pretending we have no standpoint or mental process, we examine and refine our process. Intellectual conversion involves a reflective heightening of consciousness: we come to observe ourselves in the act of knowing and discover the built-in norms of the mind. Lonergan calls these the transcendental precepts, which is a rather philosophical term for foundational habits of insight and responsibility: Be attentive (to the data), Be intelligent (in understanding), Be reasonable (in judging), and Be responsible (in deciding). Following these internal norms leads us towards truth. It also exposes where we might be going wrong (for example, jumping to a conclusion without sufficient evidence violates “be reasonable”).

Making this shift can be disorienting at first. One has to admit that reality is not just sitting “out there” waiting to be looked at; rather, we mediate reality through our questions and judgments. But the payoff is immense. Intellectual conversion frees us from the grip of common sense literalism and ideology because we no longer equate our immediate view with the truth. We learn intellectual humility and patience. We come to relish nuance and evidence, realizing that insight often requires revising our assumptions. In practical terms, someone intellectually converted is more likely to say “I wonder why that is?” instead of “I already know enough.” They will cross-examine their own assumptions and welcome data that challenges their opinions. This is precisely the kind of mindset needed for the “not numerous center” of cosmopolis, which must sift truth from bias on all sides.

Cultivating intellectual conversion can take specific practices. For instance, explicit methods of inquiry (scientific method, investigative journalism standards, philosophical argumentation) serve as training regimes for the mind’s authentic functioning. Habits like daily reflection on how one has formed an opinion, peer review and debate, and even “bias audits” of one’s work help keep the knower honest. Over time, these practices reinforce the insight that objectivity comes from asking all relevant questions and not slanting the answers. The result is a person (or community) more committed to truth than to being right, more interested in learning than in winning arguments. Such intellectually converted people can engage opponents without simply dismissing them, since they are secure that acknowledging complexity or error is a step toward truth, not a weakness. They have “self-appropriated” their cognition – they know what they are doing when they claim to know.

To summarize, intellectual conversion shifts us from a naive realism (“I just see the facts”) to a critical realism that understands knowing as a compounded, self-correcting process. It lays the foundation for cosmopolis by creating thinkers who value evidence over bias and method over indulgence. It gives us the patience to gather and weigh data, the humility to doubt our own insights until verified, and the discipline to “insist on complete solutions even though [we have] to wait”. But knowing what is true or real is not enough. The next question is: will we choose what is truly good? That brings us to the second conversion, the moral.

Moral Conversion: From Satisfactions to Values

Moral conversion is a shift in the center of gravity of our decision-making—from the pull of personal satisfaction to the pull of genuine value. In everyday life, it’s normal to start with ourselves: we prefer things that please or benefit us (or our tribe) and avoid what doesn’t. Morality, in this everyday sense, often means negotiating compromises between our desires and some external rules or social expectations. For our purposes, moral conversion means a far more profound reorientation. It “changes the criterion of one’s decisions and choices from satisfactions to values.” Instead of asking “What’s in it for me (or my group)?”, a morally converted person asks “What is truly worthwhile, truly good in itself?” – and commits to that, even at cost to self-interest.

This conversion can feel like an inversion of priorities. Before, one’s value-scale likely put personal happiness, success, or comfort at the top, and nebulous ideals further down. After moral conversion, values (like truth, justice, love, the common good) hold the pride of place, and one is willing to sacrifice lower satisfactions for higher values. We can think of this as moving from making yourself the center (with the good revolving around your satisfactions) to making value the center (with yourself orbiting the good). For example, if I undergo moral conversion, I might turn down a lucrative but unethical job offer because integrity and social impact matter more to me than extra income. Or I might accept personal inconvenience—paying more for an eco-friendly option, say—because I judge that the environmental value overrides my satisfaction in saving a few dollars.

Importantly, we do not equate value with a dry duty opposed to feeling. Rather, in a converted state, our feelings themselves are educated and aligned with value. We start taking genuine joy and pride in doing what is truly good, and feeling revulsion or shame at what is base or unjust. Our affectivity undergoes a shift: we come to “love what is truly good, even if not immediately satisfying”. This doesn’t happen overnight—it’s often a gradual process of awakening. We might have moments of clarity (through personal crisis, great role models, or faith experiences) that reveal how shallow mere satisfactions are. Those moments invite us into a new horizon where, say, living with integrity is non-negotiable or caring for others becomes a source of meaning.

A key mark of moral conversion is authenticity in the ethical sense. One opts decisively for the truly good and then tries to live consistently by that fundamental option. Of course, as fallible humans, we will fall short, but the converted stance includes constant self-scrutiny and correction. One begins to ask about every significant decision: “Does this accord with the values I have chosen to live by? Does it contribute to the good of order (the welfare of the community and future) or only to my private satisfaction?” Institutions, too, can reflect moral conversion, for example when a business shifts from a profits-at-all-cost mentality to a mission-driven approach that respects employees, customers, and the environment as values, not just means.

Practices supporting moral conversion include what some traditions call “examination of conscience” – regularly reflecting on one’s actions and motives in light of deeper principles. It also helps to engage in dialogue with others who hold one accountable to values (e.g. community groups, mentors). Education in ethics and exposure to stories of moral heroes can widen one’s horizon of value. On a structural level, designing institutions that reward long-term good over short-term gain fosters moral conversion collectively (for instance, a company could tie executive bonuses to ethical behavior and social impact, not just quarterly earnings). We also note the need for “horizon-widening”: deliberately seeking out perspectives beyond one’s comfort, which challenges the complacency of one’s satisfactions.

The payoff of moral conversion is a kind of integrity and steadfastness that is essential for cosmopolis. Only people who prize values above quick wins will have the “stamina to refuse half-measures” and insist on complete solutions. They won’t be easily bought off or discouraged, because their motivation isn’t ego or tribal victory but the intrinsic worth of the goal. Moreover, moral conversion engenders solidarity. If I care about true value, I will recognize others who do the same—even if they belong to a different faction—and I’ll be willing to work with them. This builds the “center” that is “strong enough to insist on complete solutions even though it has to wait.” In current events, we often see fragmentary coalitions (left and right) form around common values (like defending democratic norms or the dignity of the vulnerable) that transcend party. Those are hints of cosmopolis at work, driven by moral commitment.

Yet, even values-driven people can lose heart. Working patiently “even though it has to wait” asks for extraordinary hope and love. Here is why we need the third and ultimate conversion, which deeply empowers the other two: religious conversion.

Spiritual (Religious) Conversion: Being-in-Love with the Unconditioned

We can understand religious conversion as “being in love in an unrestricted fashion.” It is a profound and total orientation of the person towards the ultimate ground of all value and truth—what we can call the Unconditioned or simply God. Unlike intellectual and moral conversion, which we can approach by deliberate effort, religious conversion is typically experienced as a gift or grace. Religious conversion is characterized as a falling in love at the deepest level of one’s being: “a being-in-love with God” that is not merely sentimental but a real state of consciousness. This state “floods” the conscious subject with a radical new impetus: one’s whole outlook is now rooted in unconditional meaning and value. In Christian terms, it’s the outpouring of God’s love in one’s heart (a great example in Romans 5:5 on God’s love being poured out through the Holy Spirit).

What does this have to do with cosmopolis? In a word: hope. To sustain the long, painstaking work of reversing decline, individuals need more than good ideas and good principles; they need existential hope, humility, and an orientation towards transcendent meaning. Religious conversion provides exactly that. “Being in love with God is the basic fulfillment of our conscious intentionality.” It answers the restless dynamism in us (the drive to know, to do good) with an experience of ultimate meaning and goodness. This experience “brings a deep-set joy that can remain despite humiliation, failure, privation, pain, betrayal, [or] desertion”, and “a radical peace, the peace that the world cannot give.” In other words, it enables a person to endure suffering and delays without despair, because one’s ultimate point of reference is no longer worldly success but the love of the Unconditioned.

For members of our “not numerous center,” such spiritual depth is a wellspring of resilience. Working on long-term solutions often means your efforts are not immediately appreciated; you might even face opposition or ridicule from the “solid right” or “scattered left” who misunderstand you. It’s easy to become cynical or burned-out in that situation. But the religiously converted person draws strength from a different source. He or she can echo the sentiment, “If God (the ultimate Good) is for us, who can be against us?” and thus persevere when others quit. This isn’t fanatical zeal—on the contrary, it comes with great humility and charity. True being in love with God instills “hope, humility, and forgiveness” because one is grounded in the unconditional love that underpins reality. Pursued authentically, it makes one less prone to ego-driven frustration and more open to reconciliation.

Why does cosmopolis need religious conversion specifically? Couldn’t a thoroughly morally converted secular person do the job? Indeed, many non-religious people heroically serve the long-term good. I’m not saying only the religiously converted can contribute. But I suggest that the full breadth of self-transcendence—going beyond ego and bias—culminates naturally in a transcendent love. Religious conversion consummates intellectual and moral conversion by orienting us to the highest possible frame of meaning (similar to the four levels of happiness). It relativizes our personal drama within something infinitely greater. This perspective is crucial for truly unbiased collaboration. It is what allows cosmopolis to seek reconciliation without denying truth. For example, envision truth and reconciliation commissions (more on this shortly): they require people who can hold together a thirst for justice and an offer of mercy. Such balance comes easier when participants have a spiritual conversion that frees them from hatred and vengefulness. They can hate the sin but not the sinner, because they see even the sinner as loved ultimately by God.

We can think of religious conversion in terms of love and acceptance: “when one accepts God’s gift of his love”. This acceptance generates compassion and the courage to face painful truths. Practically, those living out religious conversion will engage in practices like prayer, meditation, or deep reflection to continually place themselves in God’s presence and receive renewed inspiration. They often participate in “shared rituals that form trust and mercy”—community worship, acts of service, confession and forgiveness practices, etc.—which knit people together in a bond beyond self-interest. Over time, these practices build communities of hope that can withstand long winters of adversity.

In summary, spiritual conversion supplies the transcendent energy for cosmopolis. It grounds the work in something more enduring than any political movement or academic school. To be “in love with the Unconditioned” is to live in hope that, however dark the moment, truth and goodness ultimately have a source and destiny that triumph (in theological terms, that God’s love wins in the end). Thus, one can be “strong enough to refuse half measures” and wait without despair. With the three conversions—intellectual, moral, and religious—at work, we get persons who see clearly, choose rightly, and trust deeply. The next step is to translate these personal conversions into structural patterns that can sustain progress. Lonergan’s vision doesn’t stop at individual enlightenment; it extends to how we organize collaborative efforts (like scholarly research, policymaking, etc.) in a way that embodies cosmopolis.

From Persons to Patterns: Institutions that Embody Cosmopolis

If cosmopolis is to be effective, it must take shape not only in converted individuals but also in new patterns of collaboration. People who have undergone the conversions we’ve described will naturally want to restructure their work and institutions to reduce bias and favor long-term insight. Here’s a proposal for a concrete way to do this: namely, through a differentiated method of distinct steps or functions. Let’s assume eight functional specialties that break intellectual endeavor into sequential tasks (Research, Interpretation, History, Dialectic, Foundations, Doctrines, Systematics, and Communications). By separating out different kinds of tasks, we can ensure each is done with proper rigor and minimal bias, and then link the results together for a comprehensive solution.

Why is this important? Bias often creeps in when people confuse different tasks. For example, a historian tasked with gathering facts may start injecting their own ideological interpretations prematurely (mixing the research function with the dialectic function). Or a policy group might jump from research straight to communication, skipping the hard work of evaluating conflicting viewpoints (dialectic) and establishing common foundational values (foundations). Our insight is that we can design a workflow that forces us to do first things first, and to challenge our biases at each stage. Each stage/function has its own proper standards. For instance, Research gathers data carefully (with source-critical objectivity), Interpretation strives to understand meaning in context, History looks at development and contexts, and Dialectic explicitly surfaces and debates the differences and biases in various accounts. Only after working through those “retrieval” phases do we move to forward-looking ones: Foundations (where we take a stand on fundamental values and perspectives – essentially the fruit of moral and religious conversion formalized), Doctrines/Policy (formulating the agreed insights and values into directives or teachings), Systematics (organizing those into a coherent worldview or model), and Communications (conveying the results effectively to the broader community).

In an institution embodying cosmopolis, you would see structures or teams corresponding to these kinds of functions. For example, consider a global policy think-tank tackling climate change. It might have one team dedicated purely to data gathering and fact establishment (Research), another to interpreting what the data means in human terms (Interpretation of scientific findings in economic or ethical context), a historical team examining past cases and trajectories (History), and a dialectic forum where advocates of different viewpoints (e.g. economic growth vs. sustainability) engage, with facilitators ensuring all biases and assumptions are made explicit (Dialectic). After this phase, a foundations group (perhaps a multidisciplinary panel including ethicists, community leaders, etc.) would articulate the shared values or grounds that emerge (e.g. “We value both human development and ecological integrity, here’s how we prioritize them…”). Only then would they draft concrete policy proposals (Doctrines) and integrate them into an overall strategy (Systematics), finally rolling out a campaign or publication to educate and mobilize the public (Communications).

The design principle at work is twofold: division of labor and integration of results. By division of labor, each kind of bias can be isolated and addressed. The person doing initial research is not also trying to justify a policy – their job is just to get the facts right, which helps avoid cherry-picking data for a pre-set agenda. The dialectic stage squarely confronts bias by having opposing views meet in a structured way, rather than letting bias lurk unspoken. By integration of results, the output of each stage flows into the next, so nothing essential is lost, and everyone down the line is accountable to what came before. Note that each specialty “has its unique criteria” but also must “dovetail with the rest,” otherwise scholars tend to overstep and introduce confusion. The same goes for any collaborative cosmopolis effort: scientists, policy analysts, moral leaders, communicators all have roles, but they need a common frame to cooperate. Method provides that frame.

As another example, large engineering projects or health initiatives often implicitly follow similar stages (research, design, testing, review, standards, etc.). What we’re adding is the explicit self-correcting, bias-checking element. We essentially build feedback loops into the process: notably at the Dialectic stage (which is a check on the first three “historical” stages, asking where viewpoints conflicted and why) and at the Foundations stage (which asks, “have we personally appropriated the horizon of values needed to carry this project forward?”). These stages force a kind of meta-reflection in the process, preventing groupthink or the domination of a single ideology. It’s as if the process says: “Stop. Let’s critically examine how we got here and who we are, before proceeding.”

In the spirit of cosmopolis, such institutional patterns invite external critique and transparency. A cosmopolis-style research program would publish not just conclusions but the methods and debates (the “dialectic”) that led there. This way, the wider community can see the work is not a closed cabal but an open search for truth. It builds trust and also educates others in the method. In fact, this method can be seen as an educational template for any field – teaching practitioners how to collaborate without bias by delineating tasks and challenging them to conversion at the “Foundations” step.

To summarize, converting individuals is only half the battle; the other half is converting patterns of collaboration. By institutionalizing functional specialties, review mechanisms, and interdisciplinary translation “layers,” cosmopolis gains continuity. It doesn’t rely solely on a few heroic figures (who eventually retire or die); it creates a sustainable culture of inquiry and implementation. The result is an intelligent, moral, and yes, spiritual workflow – a cosmopolitan institution in microcosm. Many forward-thinking organizations today embody pieces of this, whether knowingly or not. Next, we will look at a few case sketches that illustrate how elements of Lonergan’s cosmopolis might appear in practice.

Case Sketches of Cosmopolis in Action

To make this concrete, let’s examine a few scenarios that reflect cosmopolis-like centers at work amid the “solid right” and “scattered left” of our world:

• Open-Source Innovation as a “Center” in Tech: In fields like software or AI, there is often a “solid right” of maintainers and conservatives who value stability, and a “scattered left” of experimenters trying every new idea. The open-source community can form a not numerous center that bridges these extremes. For example, consider the way the Linux kernel is developed: a global team of programmers (from big companies and independent volunteers alike) collaboratively improves the code. They have rigorous methods (code review, version control) that embody intellectual conversion – truth is found in what works and passes tests, not in who asserts it. They also enforce moral norms: a code of conduct, and the ethos that the best idea wins (a value of fairness over personal satisfaction in being right). This center is “not numerous” relative to all software users, but it has an outsized influence because it produces reliable, long-term solutions instead of quick hacks. It resists biases like corporate monopoly or NIH (“not invented here”) syndrome by being open and meritocratic. We see this pattern emerging in AI as well: initiatives like OpenAI’s collaboration with academic and public partners attempt to balance the “solid right” concerns about safety and control with the “scattered left” enthusiasm to democratize AI. A small interdisciplinary group sets shared standards and best practices (analogous to Lonergan’s functional specialties) so that progress (new features, models) can proceed without causing decline (e.g., misuse or lack of oversight). In effect, these open ecosystems function as cosmopolitan centers by orienting tech development to the long-term common good (such as ensuring AI is beneficial to all), refusing the “half-measure” of just selling a product that hasn’t been ethically vetted.

• Truth and Reconciliation Processes: When a society has been through conflict or injustice, typically we see a “solid right” that denies or minimizes wrongs and a “scattered left” pressing radical change or revenge. A truth and reconciliation commission (TRC) tries to be a “not numerous center” that transcends this divide. Take South Africa’s TRC after apartheid: it was not a partisan tribunal but a deliberately balanced body that sought truth (intellectual honesty), justice tempered by mercy (moral value), and healing (spiritual reconciliation). The TRC embodied our three conversions in its process. Intellectual conversion: it insisted on detailed testimony and verification – what really happened – cutting through ideological narratives. Moral conversion: it established that the goal was not retaliation (satisfaction) but restoration based on human dignity (value). Many participants, including Archbishop Desmond Tutu, who chaired it, were driven by religious conversion – a profound Christian faith in forgiveness and redemption. This gave them the strength to “seek reconciliation without denying truth.” The TRC’s public hearings forced the nation to be attentive to painful experiences, intelligent in understanding the causes, reasonable in distinguishing truthful confessions from lies, and responsible in recommending reforms. Its success wasn’t absolute (no human process is), but it prevented what many feared would be a downward spiral of violence. It did so by creating a safe space where bias (on both sides) could be named and overcome through shared humanity. The lesson here is that such processes require people who have the interior resources to endure anger and sorrow without flying to extremes – in Lonergan’s terms, converted people. And institutionally, they need careful method: opportunities for everyone to speak (research), cross-examination (dialectic), acknowledgment of wrong (foundations of value), and concrete recommendations (practical communications). Whenever communities undertake truth-telling and healing – whether in racial justice work, post-conflict situations, or even organizational ethics reviews – they are enacting on a small scale the work of cosmopolis: refusing to live in lies, but also refusing to become a faction of vengeance, aiming instead at a fuller reconciliation grounded in truth.

• “Cosmopolis Cells” in Policy and Civic Life: One intriguing application of this thinking is to form small cross-domain teams within large organizations or governments that operate explicitly on cosmopolitan principles. For instance, imagine a City Foresight Taskforce that a mayor establishes to tackle issues like climate adaptation, technology governance, or pandemic preparedness. Such a team would include a few experts (scientists, economists, urban planners), a few community representatives, perhaps a philosopher or theologian – a microcosm of disciplines and perspectives. Crucially, the team would be chartered not to deliver quick wins for the next election, but to map out “transitions to be made” over the next decades and to flag “half measures” that won’t solve the problems. Their workflow might mirror our method: first, gather data on the current situation (research), then interpret what it means for everyday citizens (interpretation), examine historical lessons from other cities or past decades (history), hold an internal dialectic where, say, the economist and the environmentalist debate assumptions (dialectic), come to a shared framework of principles – e.g. “we commit to both equity and sustainability” (foundations), outline robust policies that reflect those principles (doctrines/systematics), and finally communicate these to the public through reports and educational forums (communications). By making the method explicit and even teaching it to members, such a “cosmopolis cell” would function as a seed of broader change. It could demonstrate a better way to govern: based on long-term intelligence and value, not sound bites. Even if its membership is small, its impact can be big by influencing how decisions are made elsewhere in the administration. Over time, more such centers could network together (think tanks, university programs, interfaith alliances for social issues) to form a lattice of cosmopolis, keeping society oriented toward truly progressive progress (and not the mirage of progress that is just rapid, directionless change).

These sketches show that cosmopolis is not purely abstract. Elements of it are already at work wherever sincere inquiry, commitment to value, and hope in the face of adversity come together. However, staying in that not numerous center is difficult. There are pitfalls and criticisms to consider, which I address next.

Objections and Failure Modes

No noble venture is without its skeptics—and its real dangers. Several objections can be raised to this notion of cosmopolis, and history provides cautionary tales of failure modes. Let’s consider a few and how we might respond:

• “Isn’t this elitism in disguise?” Critics might say cosmopolis sounds like rule by an enlightened few. Talking of a “center” that’s “not numerous” can smack of technocracy or paternalism. The response is that conversion is normative, not exclusive. Cosmopolis isn’t a club of geniuses or saints; it’s a set of norms and methods that anyone can appropriate, and indeed that everyone is called to. Its leadership is moral-intellectual, not political. In practice, cosmopolis folks have to exercise servant leadership: their job is to help the whole community see more clearly and choose better, not to dominate it. Elitism is a danger if the center becomes insular or arrogant. The safeguard is the dialectical openness of cosmopolis—welcoming criticism, being transparent about its procedures, and ultimately persuading, not coercing. Far from being a detached ivory tower, a true cosmopolis would constantly engage ordinary common sense (respecting it, even as it offers a higher viewpoint). It’s an elite of commitment, not of privilege; its “membership” is self-selecting by authenticity and skill, which in principle is open to all who put in the work.

• “Won’t method kill creativity and spontaneity?” Some worry that all this talk of self-correcting method and functional specialties could lead to rigid processes that stifle innovation. The answer lies in understanding that method disciplines creativity; it doesn’t replace it. We all celebrate the miracle of insight—the unexpected spark of creativity. Method serves to scaffold insight, to test and implement it, not to generate it by formula. In fact, by removing a lot of intellectual rubbish (bias, sloppy thinking, miscommunication), a good method frees creative thinkers to spend more time actually innovating. It’s like how a great jazz musician still practices scales; the discipline enhances, not diminishes, the eventual improvisation. That said, there is a failure mode here: bureaucratization. If cosmopolis devolved into just a bureaucracy of “complete checklists” with no room for intuition, it would lose its soul. Thus, any methodological structure must remain flexible and revisable—true to the spirit of inquiry. Note, cosmopolis itself is “a heuristic structure”, a guide to discovery, not a cookbook of answers. Creativity is further preserved by the diversity of the cosmopolis: since it’s trans-partisan and trans-disciplinary, it’s constantly bringing in fresh perspectives that spark new ideas.

• Capture by ideology: A more serious failure mode is if the “center” is captured by a hidden ideology or bias—becoming just another faction while claiming neutrality. History is littered with movements that thought they were above bias but were blind to their own (consider how some scientific eugenicists in the early 20th century believed they were purely rational while promoting horrible biases). The defense here is continuous dialectic and self-examination. Cosmopolis must include mechanisms for external critique: for example, periodic independent audits of its work by outsiders, or rotating leadership to avoid groupthink. Our emphasis on dialectic as a functional specialty is effectively a built-in alarm system: it asks participants to explicitly state their assumptions and viewpoints and clash them with alternatives. If done earnestly, this can expose ideological blind spots. Also, the presence of morally and religiously converted members (humble, loving people) provides a kind of immune system; they are more likely to notice when the enterprise is betraying its own values and speak up.

• Burnout and Cynicism: Trying to “work out one by one the transitions to be made” while “having to wait” for complete solutions can be exhausting. The not numerous center may feel overworked and underappreciated, leading to burnout. Members might start cynical jokes about “the unwashed masses” or despair that “nothing ever changes.” This is where the spiritual practices are crucial. Regular retreat, prayer or meditation, communal support—these keep the flame of hope alive. Cosmopolis should foster a subculture of sustenance: perhaps common meals, rituals of celebration for small wins, remembrance of pioneers who achieved progress after long years. Burnout is a sign one is running on one’s own steam; religious conversion in particular reminds us to seek grace or a sense of higher purpose as an energy source. Additionally, rotating people through tasks can help; no one should carry the whole world alone. A practical rule might be to pair veterans with newcomers in teams so that fresh optimism balances seasoned realism.

• “Performative centrism”: A final pitfall is that people or organizations pretend to be the balanced center to aggrandize themselves, using the rhetoric of being above extremes, actually to advance a self-serving agenda. We see this in politics when someone markets themselves as the “sensible center” while mainly advancing their own interest and dismissing all dissent as extremism. To guard against this, cosmopolis must be earnest in its self-description. It shouldn’t claim the center as a badge of superiority, but demonstrate it through painstaking work and refusal of half-measures. A genuine cosmopolis will likely be quieter and more patient than a fake one. One can watch for the fruits: Are they actually achieving integrative solutions over time? Are they gaining trust from very different segments of society (a sign they truly listen to all)? Transparency, as mentioned, is key: if their analyses and proposals are published with full reasoning and citations, it’s harder to hide a partisan core. In short, the best answer to “performative centrism” is authenticity – exactly what our conversions aim to produce in persons.

In facing these objections and pitfalls, we circle back to our core idea: progress is a function of authenticity in knowing, choosing, and loving. The method and center are only as good as the people in them. Thus, the final section will outline how individuals and groups might practically live out these ideals – essentially, a rule of life for building the not-numerous center in our daily routines.

Building the “Not Numerous Center”: A Practical Rule of Life

How can we personally and collectively move toward this notion of cosmopolis? Grand as it sounds, it boils down to daily, weekly, and monthly habits that nurture conversion and cooperation. Here is a proposed “rule of life” for those who aspire to be part of the patient, truth-seeking center:

• Personal: Make a habit of both study and reflection. For example, commit to daily study or reading that enriches your understanding beyond your comfort zone (this keeps you attentive and intelligent). Equally important, practice a daily examen or self-reflection each evening: where was I biased or self-serving today? Where did I follow the foundational habits of insight and responsibility (or fall short)?. Perhaps weekly, do a focused “bias review” – pick an issue you care about and deliberately seek out the best argument against your position, just to ensure you’re not in an echo chamber. And a quarterly retreat (even if just a day in quiet or a long hike) can recentre your soul, renewing that underlying peace and being-in-love that fuels hope. Over time, these practices build authenticity. You become the kind of person described earlier: one who loves discovering truth, who hungers for what is truly good, and who finds resilience in a source beyond ego.

• Team: Encourage cosmopolis-style teamwork by implementing concrete protocols. For instance, adopt a “red-team/blue-team” dialectic in project meetings – assign someone to play devil’s advocate or represent an absent perspective, to ensure biases are flushed out (this is a modern echo of Lonergan’s dialectic stage). Keep decision logs: whenever a key decision is made, write down the reasoning and evidence behind it, and revisit it later to see if it held up. This makes the team responsible and reasonable by design. Use “complete-solution” checklists that prompt the group to ask: Are we addressing root causes, or just symptoms? What happens in 5 years if we implement this? Such checklists echo our call for refusing half-measures. Finally, adopt the habit of rigorous citation and grounding of claims (even internally) – e.g., requiring page-numbered references or data sources whenever someone makes a factual assertion. This instills the discipline of evidence before opinion, a hallmark of intellectual conversion. These might seem like minor bureaucratic tweaks, but they form an environment where bias is caught early and insights are given their due.

• Community: Broaden the cosmopolis network by building communities of inquiry and practice across domains. One idea is a standing cross-domain seminar – a monthly gathering (virtual or physical) where people from different fields or walks of life discuss a common human concern (like “How do we foster trust in institutions?” or “What does progress mean in the age of AI?”). The aim is not to win a debate, but to achieve mutual understanding and insight by practicing intelligence together on complex issues. Encourage the creation of a shared glossary or common language for your community or organization, so that when technical or philosophical terms are used, everyone can learn and no one hides behind jargon (mediating between theory and common sense). Also, emphasize public communications that teach the method, not just the conclusions. For example, if your team publishes a policy proposal or an op-ed, include a sidebar “How we arrived at this” explaining the process of investigation and debate that led to the proposal. By doing so, you not only argue for a solution, but you also educate readers in the cosmopolitan method, hopefully elevating the public discourse long-term. Essentially, model the change: let people see how a not-numerous center thinks and works, so they can join in or emulate it. Over time, this builds cultural momentum. The more people see a sane, truth-oriented middle in action, the more they realize it’s possible to escape the polarized cycle of bias.

By following such a rule of life, we begin to “build the basis of collaboration” that we envisioned. It starts with individual commitment and radiates outwards. None of us can change the whole world overnight, but each can cultivate a bit of the cosmopolis spirit in our circle. Perhaps we form reading groups (as we did as students with Lonergan’s Insight years ago) to appropriate these ideas more deeply. Perhaps we mentor younger colleagues in not just skills but in authenticity. Ultimately, cosmopolis grows quietly, person to person, practice by practice – “one by one the transitions to be made”. In a time of drift and fragmentation, choosing this path may feel lonely. But remember Lonergan’s hopeful observation: what will count is not the noisy extremes, but the center that may not be numerous yet carries the future on its shoulders. Our task is to join that center, to make it a little less small, and to keep it oriented to complete solutions grounded in truth and love. By doing so, we become, in our own humble way, painstaking builders of progress amid decline – collaborators with all those, past and present, who refused to stop at halfway. The work is ongoing, and even though we often have to wait, it is fully worthwhile.

“I wisely started with a map, and made the story fit (generally with meticulous care for distances).” – J.R.R. Tolkien, 1955 (A Pilgrim in Narnia)

When a reader once asked Tolkien whether the Glanduin and the Swanfleet were the same river, he answered by drafting a thirty‑page treatise, The Rivers and Beacon‑hills of Gondor (Tolkien Gateway). A single hydrologic puzzle became an odyssey through every delta, headwater, and ridge in Gondor. In that essay, he renamed the seventh beacon‑hill Halifirien (“holy mountain”) and promptly hid the tomb of Elendil beneath its moraine‑crowned summit, proving that the right word can generate new legend.

Tolkien’s cartographic method was rigorous: watersheds first, languages next, plot last. Because the Anduin’s tributaries, the oxbow‑like loops of the Entwash, and the isthmus of the Pelennor all obey the physics of real rivers, the peril of The Lord of the Rings feels earned. Literary critic Tom Shippey calls these toponyms “repeated implicit assurances” that Middle‑earth exists beyond the page (A Pilgrim in Narnia). Precision lets myth wear the mask of history.

Growing up in the Pacific Northwest it was easy to map all the cartographic wilderness of Tolkien to my daily experience. Stand on a windy headland at Deception Pass and watch tides knuckle through the narrows of Puget Sound. Geologists remind us that the Sound is both a flooded fjord and a sprawling estuary—a glacial trough nine hundred feet deep where salt and fresh waters braid. The compound term “fjord estuary” is not pedantry; it is a two‑word ballad about ice and tide, depth and exchange.

Southward looms Tahoma—better known as Mount Rainier—“the mother of waters,” so called by Coast Salish peoples because her glaciers spread like silver alluvial fans into every valley (National Park Service). Name her only Rainier, and she is a monument; call her Tahoma, and she becomes an ancestor pumping lifeblood into the land.

Seattle itself rides a necklace of north‑south drumlins sculpted by the last ice sheet (Seattle Times Archive). Once you learn that word, Capitol Hill and Beacon Hill stop being random bumps; they align like furrows in the memory of a glacier. East of the Cascades, the land flattens into the Columbia Plateau, a desert of ancient basalt flows poured out in pulses of fire that later rivers incised into labyrinthine coulees (USGS). Between plateau and peak runs, the White Salmon River, its rocky bed a seasonal arroyo in late summer, roars only when snowmelt swells.

Each term—fjord, drumlin, basalt, arroyo—adds a facet to the gem of place. The lexicon is mnemonic. Once a cove is named Rosario, the mind keeps its curve forever. Once you know a delta from an estuary, the Skagit no longer looks like “just wetlands” but like a cathedral of sediment fanning into the Salish Sea.

Rebecca Solnit notes that people “light up around maps” because they promise orientation (Wired). Orientation is deeper than way‑finding; it is belonging. A neighborhood called Ballard recalls Nordic fishermen; a street called Tokul hints at Snoqualmie homelands. Precise names become an isthmus between past and present, binding residents to older stories still unfolding.

Tolkien made this visceral by layering languages: Imladris for elves, Rivendell for men. A place with two true names is doubly real; it acknowledges that landscapes hold more than one history. Likewise, using Tahoma beside Rainier or Whulj beside Puget Sound honors plural narratives and stakes us to lived human history.

Geographic words double as metaphors for the self. A season of grief can feel like trudging across a stony scree slope; a burst of creativity like a river hitting its delta, branching into many channels. Precision sharpens the image—and with it, understanding. Tolkien’s Dead Marshes mirrors despair because “marsh” is weaker than “quagmire,” and “quagmire” weaker than “fen”: he chose marshes slaked with death, a word as heavy as the ground it names.

To name a thing exactly is to care for it. We protect what we can point to: this cove, that ridge, our shared watershed. Tolkien’s obsessive labeling was, at heart, an act of love for creation, fictional though it was. Our world asks no less devotion. Learn the glossary of your region, and the map turns from wallpaper into scripture.

Here’s a small practice to bring this alive: open the Wikipedia “Glossary of geography terms (A–M)” and choose just five terms that catch your eye — some of my favorites are “fen,” “rill,” “canebrake,” “tarn,” or “shoal.” Then, on your next walk or drive, look for examples of those features near you. Notice their shapes, textures, how water moves around them, or how they feel under your feet. Give them names — even if only to yourself. Write them into conversation or label them on a sketch. As that vocabulary settles into your speech, you’ll begin to see the land with sharper vision. In time, those words become the bones of your local stories, and your sense of place shifts from vague to vivid. Let your own cartographic curiosity guide you—one term, one feature, one moment of noticing at a time.

This article offers a practical framework for business leaders to answer these questions. By looking at AI agents through the lens of legal agency theory – the same principles that govern relationships between human agents and their principals – we can define what true AI agents are and how they differ from mere AI tools. We’ll explain why giving an AI a defined domain of authority is critical to managing it, and explore implications for trust, governance, and accountability when AI takes on agent-like roles. Finally, we introduce an emerging best practice for keeping AI systems trustworthy: an automated, authenticated AI “Bill of Materials” (BOM) that tracks an AI agent’s components and enforces policies based on its integrity. The goal is to provide clear, actionable insights so you can responsibly harness AI agents in your organization and stay ahead of the curve in an era when autonomous AI is poised to join the workforce (No one knows what the hell an AI agent is | TechCrunch).

What Are AI Agents? A Lesson from Legal Theory

In the business world, agency has a specific meaning: under legal theory, an agent is someone authorized to act on behalf of another party (the principal) in a way that can create binding obligations for the principal. In plain terms, if your employee or representative makes a deal on your behalf, you are on the hook for it. Several key components define a classic agency relationship, and they provide a powerful lens for understanding AI agents:

• Authority: The agent has explicit or implicit permission from the principal to act in certain matters. In an AI context, this means a human or organization delegates power to the AI to carry out specific tasks or make decisions. An AI without granted authority is just a tool operating at a user’s direct command, not an independent agent.

• Autonomy (within scope): The agent can exercise independent judgment and initiative within the limits set by the principal. Similarly, an AI agent might decide how to achieve a goal or react to circumstances on its own, but what it’s allowed to do is bounded by its mandate. This is what distinguishes a truly agentic AI from a basic program that only follows a fixed script.

• Accountability: The principal remains ultimately responsible for the agent’s actions. If an AI agent does something on your behalf, you (or your company) are on the hook for the outcome, just as if a human employee had acted for you. You cannot use the AI as a convenient scapegoat to dodge liability – in legal and practical terms, the buck still stops with the humans in charge.

• Enforceable Outcomes: Actions taken by an agent carry real weight. A human agent can sign contracts, spend money, or make promises that the principal must uphold. Likewise, an AI agent’s decisions or commitments should have consequences that matter in the real world – otherwise it’s more like an advisory tool. If an AI’s “actions” have no direct effect (for example, it only suggests options that a human must approve), then that AI is functioning as an assistant or tool, not as an agent.

Bringing these elements together, we can define an AI agent as an AI system explicitly authorized to act on behalf of a principal, operating with reasoned autonomy within a defined domain of authority, and producing outcomes that the principal is obligated to uphold. One succinct definition puts it this way: “An AI agent is an AI system explicitly authorized to act on behalf of a principal, with the ability to make reason-based decisions in line with the principal’s objectives, and whose actions create enforceable consequences within a defined domain.” In short, an AI agent isn’t just doing work for you; it’s doing work as an extension of you.

AI Tools vs. AI Agents: Understanding the Difference

Why draw this distinction between tools and agents? Because it changes how we design, deploy, and supervise AI in practice. Most current AI applications are tools – they require direct human input for each action and do not independently initiate significant changes in the world. For example, a code assistant like GitHub Copilot can suggest lines of code or help explain an error, but it’s ultimately the human programmer who decides what gets implemented. Copilot “primarily responds to human requests and requires human approval to carry out actions,” fitting the profile of a capable AI assistant rather than an autonomous agent. The AI helps a person work more efficiently, but it isn’t trusted to carry out an entire task on its own from start to finish.

In contrast, an AI agent can be entrusted with a goal and allowed to figure out the steps and execute them with minimal hand-holding (within agreed limits). A simple example is a Roomba robotic vacuum. You don’t manually drive a Roomba around your living room; you just tell it to clean, and it autonomously navigates the space, avoids obstacles, adjusts its path, and returns to its dock when done. It “exhibits significant autonomy within a defined domain,” namely the task of vacuuming your floors. The Roomba makes its own decisions on how to clean (which route to take, how many passes to make) without further human instruction. In this limited domain, it’s acting as an agent carrying out your intent (keeping the floor clean). If the Roomba had a malfunction and knocked over a valuable vase, you wouldn’t blame the vacuum as an independent entity – you’d recognize it as an extension of the owner’s agency (or perhaps hold the manufacturer responsible if it was a product defect).

Many business AI solutions today fall somewhere in between pure tools and true agents. Vendors often market “AI agents” that are essentially glorified chatbots or workflow tools requiring extensive human oversight. This loose use of the term “agent” has led to confusion in the market. It’s important for leaders to cut through the hype by asking: Is this AI actually making autonomous decisions and taking actions on our behalf, or is it just providing recommendations for a human to act on? If it’s the former, you need to manage it with the rigor you’d apply to a human delegate. If it’s the latter, traditional software risk management may suffice.

One helpful way to identify an agentic system is to look at its capabilities and design. Agentic AI systems typically incorporate additional components beyond what a standard tool would have. For instance, an AI agent might be provisioned with identities or credentials to authenticate into other systems (much like an employee badge), and it might have the ability to perform actions via plugins, APIs, or robotic controls without a human in the loop. Advanced agents can also orchestrate complex sequences of steps – planning and executing multiple actions to achieve an objective. They often maintain some form of memory or knowledge base about their environment and can even learn over time from experience.

The Importance of a Defined Domain

Central to the notion of an AI agent is the idea of a defined domain of authority. In law, when you appoint a human agent, you usually delimit what they are authorized to do. A junior buyer at a company might have authority to sign purchase orders up to a certain dollar amount. A real estate agent can negotiate home sales on your behalf, but can’t arbitrarily decide to invest your money elsewhere. This principle of scope is crucial: if an agent stays within scope, the principal is bound by the agent’s acts; if they step outside it, the principal may not be obligated (and the agent might even face personal liability).

For AI, a defined domain is just as essential – arguably even more so. A human agent has common sense and an understanding of their limits; an AI will try to do whatever it was programmed or trained to do, even if the situation changes or the task turns out to be ill-specified. To safely give an AI autonomy, we must clearly define the arena in which it can act and make decisions. In practice, this means setting boundaries on the AI’s functions and access: the data it can use, the decisions it is allowed to make, and the actions it can carry out. The AI’s domain could be defined in terms of business processes (e.g. handling refund requests up to $100, or managing calendar scheduling for meetings), in terms of knowledge scope (e.g. answering questions only about a specific product line), and in terms of accessible systems (e.g. it may interface with the CRM system but not the financial accounts).

Defining the domain isn’t just about preventing worst-case scenarios – it also builds trust and reliability. When an AI knows its job and doesn’t stray outside it, it’s less likely to produce nonsense or harmful actions. For example, Salesforce, which is developing enterprise AI agents, found that giving agents a well-defined set of “topics” or tasks is key to preventing errors. An AI customer service agent might be allowed to handle routine issues like order status inquiries, returns, and refunds, but anything outside those predefined topics (like an unusual legal complaint or a request it doesn’t understand) should trigger an escalation to a human representative (5 Ways To Build a Trustworthy AI Agent). By confining the AI to what it’s been trained and authorized to do, you “make sure the AI agent doesn’t try to answer a question it shouldn’t”, which greatly reduces the chance of it hallucinating an answer or taking an inappropriate action. In essence, domain limits act as guardrails that keep the agent trustworthy.

Moreover, a defined domain simplifies governance and accountability. If something goes wrong, it will likely be within the agent’s sphere of responsibility, making it easier to trace the cause and implement a fix. It also clarifies accountability: if your AI agent for automated billing suddenly starts sending collection notices in error, you know which system (and team) is responsible and what domain knowledge needs adjustment. Contrast this with a scenario where an AI is set loose without clear boundaries – mistakes could be harder to detect and diagnose, and accountability becomes murky. In regulating AI, we see emerging consensus on this point: even defense and ethics guidelines suggest AI systems should have an “explicit, well-defined domain of use” for safe operation. Whether for technical safety or legal clarity, an AI agent without a clear domain is a recipe for trouble. Responsible leaders will insist on scoping AI deployments tightly, at least until we have far more generalized and proven AI that can truly handle open-ended authority (a prospect that remains uncertain).

Trust, Governance, and Accountability

Handing off tasks to AI agents can feel like hiring a new employee or contracting out work – it requires trust. But trust in AI doesn’t come for free; it must be earned through good performance and transparent behavior, and maintained via oversight. Business leaders need to establish governance mechanisms to ensure AI agents stay reliable and aligned with the organization’s goals and values. Here are key considerations for building trust and ensuring proper governance:

• Ultimate Accountability Stays with You: First and foremost, never forget that deploying an AI agent does not absolve your organization of responsibility. If an AI-driven HR screening system discriminates against candidates or an AI trading algorithm triggers losses, it’s the company that will be held accountable. Courts and regulators are already reinforcing this – for instance, a recent case involving an AI-powered hiring tool underscored that employers may be held liable for decisions made by their AI systems, which can be viewed as agents of the employer. In other words, legally and ethically, the AI is an extension of your business. This means you must audit and validate an AI agent’s decisions just as you would review an employee’s work during a probation period. Establish clear escalation paths: under what conditions must the AI defer to a human decision-maker? Set those rules early and enforce them.

• Governance and Oversight: Treat AI agents as a new kind of stakeholder in your governance model. Many organizations are forming AI governance committees or designating AI oversight roles that involve cross-functional teams (IT, legal, compliance, risk, and business unit leaders). The goal is to continuously monitor AI behavior, outcomes, and risks. Policies should be drafted to cover AI agent conduct – for example, defining acceptable use, risk thresholds, and fallback procedures if the AI encounters a scenario outside its domain. Some leading companies have even created internal “Ethical AI” review boards that evaluate new AI use cases for potential harm or bias before deployment (5 Ways To Build a Trustworthy AI Agent). Regular audits should be conducted on AI outputs and decisions to ensure they remain within expected boundaries (analogous to financial audits or quality control for a process). If an AI agent interacts with customers, gathering feedback from those customers is vital – it provides insight into whether the agent is effective and trustworthy from the user’s perspective.

• Transparency and Communication: One of the challenges with AI, especially sophisticated models like neural networks, is that their decision-making process can be opaque. This “black box” nature can erode trust. To counteract that, insist on transparency wherever possible. This could mean having the AI explain its reasoning in simple terms, or providing logs of actions taken. When something goes wrong, a post-mortem should be done just as you would for a human error, and the lessons should be communicated to stakeholders. Externally, if customers or partners are affected by an AI agent’s actions, be proactive in disclosing that an AI was involved (as long as it doesn’t confuse or alarm them needlessly) and what safeguards are in place. Transparency also extends to knowing the provenance of your AI – what data it was trained on, what rules or objectives it’s optimizing for. Being able to answer these questions builds confidence among executives, regulators, and customers that the AI is behaving consistently and as intended.

• Aligning Incentives and Ethics: In classical principal-agent theory, a big issue is aligning the agent’s incentives with the principal’s goals (to prevent the agent from pursuing its own agenda). With AI, we don’t worry about greed or ambition, but we do worry about objective functions and reward signals. If your AI agent is optimizing for the wrong metric, it could inadvertently act against your broader interests (for example, an AI that aggressively maximizes short-term sales might start annoying customers or offering unsustainable discounts). Make sure the performance metrics and reward signals for your AI systems encourage the behavior you actually want. Similarly, encode your company’s ethics and compliance requirements into the AI’s operating procedures – if fairness, customer privacy, or safety are core values, the AI’s design should reflect that (e.g., by excluding protected attributes from decisions, or having conservative constraints on actions that could pose safety risks). Governing AI agents means baking your policies and values into their logic from day one.

• Fail-safes and Contingency Plans: No matter how much you trust your AI agent and how well it’s governed, you need a Plan B. What if the AI encounters a novel situation and makes a bad call? What if it goes down due to a technical issue? Ensure there are fallback mechanisms. This could be as simple as “if the AI confidence is low or an error occurs, route the task to a human” in customer service, or a hard stop on trading algorithms in finance if unusual volatility is detected. Have clear intervention points where human oversight can pause or shut down the AI’s actions if needed. In other words, keep a metaphorical “off switch” handy for your AI agent processes and know the conditions for using it.

By addressing these areas, you create an environment where AI agents can be trusted participants in your operations. Your employees will feel more comfortable working with or relying on AI if they know there’s oversight and that leadership has set boundaries. Your customers will be more likely to embrace AI-driven services if they sense that you have control over the technology and will take responsibility for its outcomes. And regulators will certainly look more favorably on companies that can demonstrate a strong governance model for AI (as opposed to a hands-off “the algorithm did it, not us” approach, which is a red flag). In sum, treat an AI agent with at least as much diligence as you would a new hire – train it well, set expectations, monitor its work, and integrate it into your accountability structures. Only then can you reap the efficiency and innovation benefits of AI autonomy without losing grip on risk and ethics.

Building Trust through an AI “Bill of Materials”

So far we’ve discussed conceptual and organizational strategies for managing AI agents. Now let’s turn to a more technical but highly effective tool that forward-thinking organizations are adopting: the AI Bill of Materials (AI BOM). Borrowing an idea from software supply chain management, an AI BOM is essentially an exhaustive ingredients list for your AI agent. It is “a comprehensive inventory that lists all the components, data sets, and dependencies” that went into developing and deploying the AI system (Strengthening Cyber Security in US Defense: The AI Bill of Materials | by Valdez Ladd | Medium). This includes the machine learning model (or models) at its core, the training datasets it learned from, any third-party APIs or libraries it relies on, and even the hardware or cloud environment it runs on. In short, if your AI agent is a proverbial recipe, the BOM lists every ingredient and its provenance.

Why does this matter for trust and governance? Consider how much effort goes into quality control and supplier vetting in a physical product’s supply chain. If one part is substandard or faulty, the whole product (and the company’s reputation) is at risk. Similarly, an AI agent is only as reliable as its least reliable component. A tainted training dataset could introduce bias; an open-source library with a security vulnerability could expose your system to hacks; an unvetted plugin might perform unexpected actions. An AI BOM brings transparency and accountability to this complex pipeline (AI Bill of Materials (AI BOM). The AI BOM encompasses everything from… | by Bijit Ghosh | Medium). It allows your team (and potentially regulators or auditors) to audit what’s inside the AI, trace issues to their source, and verify that everything is up to your standards. In fact, maintaining such a detailed BOM “ensures that only approved components are used” in your AI systems. Just as a hospital pharmacy won’t dispense medicine that isn’t from a vetted source, your AI agent shouldn’t be allowed to operate with code or data that hasn’t passed your organization’s security, quality, and ethics criteria.

To be truly useful, an AI BOM should be automated and continuously updated. Modern AI systems are complex and can change over time (for example, models get updated or fine-tuned, new data is ingested, etc.). Manually documenting every component is error-prone and quickly becomes outdated. Instead, organizations are leveraging tools that automatically scan and catalog AI assets across their environments. Such tools can often integrate into MLOps pipelines or IT asset management, so that whenever an AI model is trained or a new dataset is added, the BOM updates. Automation also means the BOM can be used in real-time to monitor compliance – for instance, if someone tries to incorporate an unapproved data source, an alert can be raised before the AI agent goes live with it.

Equally important is that the AI BOM is authenticated – meaning it’s secure and trustworthy. This is where concepts like digital signatures come in. Leading AI security platforms now provide features to cryptographically sign the components and the BOM itself. In practice, this means each element (say, a model file or a data corpus) can have a hash or signature that proves its integrity and origin. The BOM acts like a “certificate of authenticity” for your AI agent. Why go to this length? Because if you’re going to let an AI agent act autonomously in your business, you want high assurance that it hasn’t been tampered with or corrupted. For example, imagine a scenario where an attacker tries to slip a malicious piece of code into an AI agent’s library of plugins. A rigorously maintained, signed BOM would detect that an unknown component is present and flag it, much like a security system catching an unverified device on a network. This allows you to enforce access control and policies based on trustworthiness: your systems could be set up to only allow AI agents that present a valid, clean BOM to integrate with sensitive databases or execute certain transactions. In other words, the BOM becomes a gatekeeper – if the AI doesn’t have its papers (credentials) in order, it doesn’t get the keys to the kingdom.

The concept of AI BOM is gaining traction not just for internal risk management but also as a response to emerging regulations. Governments are increasingly concerned about AI safety and security, with proposals that developers document their AI systems’ ingredients, much like financial disclosures. By getting ahead of the curve and implementing AI BOMs now, companies can better meet future compliance requirements and reassure stakeholders (investors, customers, regulators) that they maintain “secure and responsible” AI practices (AI-BOM: Building an AI-Bill of Materials | Wiz). It’s a proactive investment in trust.

For business leaders, the takeaway is: ask your technology teams about an AI BOM whenever you deploy a significant AI agent. It might sound technical, but it boils down to a simple business question: “Do we know what’s inside our AI and that we can trust every part of it?” If you can’t confidently answer that today, make implementing an AI BOM process a priority. It will enforce discipline in how AI models and tools are sourced and used. And in the event of an incident, it provides a solid foundation for forensic analysis and remediation. In the long run, an AI BOM is as foundational to AI governance as financial auditing is to corporate governance – an enabler of trust, both internally and externally.

Actionable Steps for Business Leaders

Understanding these concepts is one thing; implementing them in your organization is another. Here are some concrete steps to put these ideas into practice and ensure your company is ready to leverage AI agents responsibly:

1. Inventory Your AI Use Cases: Start by listing where you are (or soon will be) using AI in a potentially autonomous capacity. Identify which systems are mere tools (requiring human sign-off for actions) and which are approaching agent status (making decisions or acting on their own). This audit will highlight where stronger governance is needed. For each AI system, document its principal (the business owner or accountable person/team) and what domain it operates in (tasks, decisions, and limits).

2. Define Domains and Delegations: For each AI agent use case, explicitly define its scope and authority. Write it down like a job description: what it can do independently, what it cannot do, and when it should defer to a human. Ensure this is reflected in the AI’s technical configuration (through rules or constraints) and communicated to all stakeholders. Just as you wouldn’t hire an employee without a role description, don’t deploy an AI without clear boundaries.

3. Establish AI Governance Bodies: If you haven’t already, form a governance structure for AI oversight. This could be a dedicated committee or a working group that meets regularly (including stakeholders from IT, data science, legal, risk, and business units). Charge them with creating an AI governance policy that covers areas like testing and validation requirements, monitoring protocols, ethical guidelines, and incident response plans for AI. Have them review any proposal for a new AI agent deployment before it goes live, to ensure it meets your organization’s standards of safety and ethics.

4. Implement Monitoring and Oversight: Ensure every AI agent in operation has an owner – a human “manager” responsible for its performance and compliance. Set up dashboards or reports for key metrics (accuracy, error rates, decision turnaround times, etc.) and review them as you would review an employee’s KPIs. Establish logs for AI decisions and actions (especially for high-stakes use cases) so that there is an audit trail. Consider running periodic “red team” exercises where someone intentionally stress-tests the AI with novel scenarios to see how it behaves; use those findings to improve the system.

5. Develop an AI Bill of Materials: Work with your CIO/CTO or data science leaders to institute an AI BOM for major AI systems. This might involve deploying new tools or integrating with your MLOps pipeline to automatically track components. Begin by focusing on your most critical AI agent (say, the one interfacing with customers or financial data) and build a complete BOM for it. Verify that all components are approved and originate from trusted sources. Going forward, make the BOM a requirement in your AI project lifecycle – much like code reviews or QA testing. This investment will pay off by reducing security and compliance risks (What Is Generative AI Security? [Explanation/Starter Guide] - Palo Alto Networks) (AI-BOM: Building an AI-Bill of Materials | Wiz).

6. Enforce Access Controls Tied to Trust: Collaborate with your cybersecurity team to link the AI BOM and governance policies to your access control systems. For example, you can set rules so that an AI system without a vetted BOM or with outdated/unapproved components is sandboxed or prevented from connecting to live production databases. Leverage techniques like role-based access control (RBAC) for AI just as you do for humans (AI Bill of Materials (AI BOM). The AI BOM encompasses everything from… | by Bijit Ghosh | Medium). Essentially, promote your AI agent to higher levels of access only as it proves itself trustworthy (and as its BOM checks out).

7. Educate and Communicate: Finally, bring your people along. Train your staff (especially managers of functions where AI is deployed) about what it means to work with AI agents. They should understand the AI’s domain, its limitations, and how to oversee it. Encourage an open dialogue – if employees spot the AI doing something odd, they should feel responsible for reporting it, not assuming “the tech team must know.” Also, communicate to customers or external partners when an AI agent is part of the process and what safeguards you have. This can be as simple as a note on your website or in a service interaction that says, “This response was generated by our AI system under human supervision” – it sets expectations and shows you’re not hiding the use of AI.

By following these steps, you turn abstract principles into tangible practices. Not every organization will get everything perfect at once, but even partial progress (like instituting an oversight committee or drafting initial AI use policies) significantly lowers the odds of an AI-related mishap. Moreover, these actions signal to your whole organization that AI is a strategic asset that will be managed with care, not a magic box to be adopted recklessly.

Conclusion

The rise of AI agents represents a new chapter in how work gets done – one filled with opportunity, from automating drudge tasks to scaling expertise, but also with new forms of risk. Business leaders cannot afford to approach this evolution with either blind fear or blind optimism. The key is to bring clarity and structure to the role of AI in your organization. Thinking of advanced AI systems “through the lens of legal agency” provides that clarity. It reminds us that an AI agent, much like a human agent, must be chosen carefully, given a clear mandate, and supervised appropriately. It also reminds us that ultimately responsibility lies with the principal – the organizations and leaders deploying these agents. By internalizing that, leaders can avoid the trap of attributing failures to a supposedly inscrutable algorithm. Instead, they will proactively shape the AI’s behavior through domain definitions, oversight, and technical safeguards.

Trust in AI is often cited as a barrier to adoption. Trust doesn’t mean assuming AI will never fail; it means having confidence that it will perform well and that if something goes off course, you will catch it and correct it. The approaches discussed – from governance committees to AI BOMs – are all about creating the conditions for justified trust. When your AI systems are transparent, well-audited, and aligned with your intent, you and your stakeholders (whether customers, employees, or regulators) can trust them enough to integrate them deeply into operations. And when that trust is well-placed, it unlocks the full potential of AI agents to drive efficiency, innovation, and value creation.

For forward-looking companies, investing in these governance and trust mechanisms will be a competitive differentiator. As AI agents become more common, those who have mastered managing them will be able to scale up their use confidently, whereas others might hold back or suffer public failures. In essence, treating AI agents “not just as tools” but as accountable, well-governed extensions of your organization could be the difference between leading the AI-enabled economy or lagging behind it. Business history has shown that every transformative technology rewards the prepared and punishes the complacent. With AI agents, preparation means pairing technological prowess with sound management principles. By doing so, you ensure that when your AI agents join the workforce – as Sam Altman and others predict they will – they will be productive, reliable colleagues rather than loose cannons. The future of work will likely include humans and AI agents working side by side; it’s up to today’s leaders to lay the groundwork for a partnership built on trust, clarity, and accountability.

A slew of recent articles I’ve been reading prompted me to reflect on how we’ve long approached tokenization, model design, and scaling – and how rapidly those assumptions are evolving. In this post, I’ll share that reflection and some strategic insights, wearing my hat as a product innovator and wanna-be AI researcher. We’ll journey from the quirks of glitch tokens to cutting-edge byte-level models, exploring new architectures that break old bottlenecks. Along the way, I’ll channel some systems-level thinking (and even a bit of philosophical insight) to help you examine your own mental models as you build AI products.

The Tokenization Hangover: From Clever Shortcuts to Lingering Headaches

If you’ve worked with language models, you know the drill: raw text doesn’t go straight into a neural network. It first gets chopped into tokens – pieces like words or subwords. This was a necessary trick; it gave models a manageable vocabulary to learn. Nearly all popular NLP models have relied on an explicit tokenization step. Even the modern subword algorithms (BPE, WordPiece, SentencePiece) were essentially clever compression schemes for text.

For a while, this looked like a solved problem. Subword tokenizers handled many languages and never exactly ran out of vocabulary. But they weren’t perfect. Any fixed vocabulary can limit a model’s ability to adapt to new jargon or languages. Tokenizers can be brittle – early pipelines broke on emojis or rare unicode symbols. And then there are the glitch tokens, those odd artifacts of byte-pair encoding that live on the fringe of the vocabulary and confuse models. They’re a symptom of a larger issue: by pre-defining the units of language, we baked a certain worldview (or word-view) into our models. It’s as if we handed the model a dictionary and said “these are your LEGOs – build everything from them.” Most of the time it works, but occasionally the model tries to chew on a piece that doesn’t make sense.

The deeper I looked, the clearer it became that tokenization is a legacy. One paper put it bluntly: explicit tokenization itself is problematic. It’s a holdover from the old NLP pipeline days, like feature engineering’s last stand. What if, instead, we let the model figure out language from scratch? What if we freed it from our pre-packaged vocabulary and let it read text as raw bytes or characters? This idea isn’t new – it’s been percolating for years. In fact, a 2022 model from Google, CANINE (“Character Architecture with No tokenization In Neural Encoders”), did exactly that. It ditched the subword vocab and operated directly on characters, using a clever downsampling scheme to keep sequence lengths in check. The reward? Greater flexibility across languages and no “out-of-vocabulary” issues. CANINE even outperformed a comparable mBERT model on a multilingual QA benchmark, despite having fewer parameters.

Still, going tokenizer-free came with a cost: longer sequences mean more compute. Transformers famously scale quadratically with sequence length. Feed a transformer raw bytes, and you quadruple the sequence length (since “hello” might be one token but five characters) – your compute requirements explode. For years, that brute-force approach just wasn’t worth it. We stuck with subwords as a necessary evil for efficiency’s sake.

But things have changed. Compute is cheaper, and research is showing ways to make models handle longer sequences more gracefully. The hangover from tokenization – the brittleness, the blind spots – is finally pushing us to find a cure. And the cure coming into focus is to go native: treat language as the byte sequence it truly is, and train models to understand it end-to-end. Recent breakthroughs are proving that this is not only feasible, but possibly superior.

Byte-Native Models: Hello, Raw Data – Goodbye, Handcrafted Tokens

A new generation of models is emerging that throws out the tokenizer altogether. These models work directly with raw bytes (or characters), and they’re closing the performance gap with traditional token-based models. In fact, they’re starting to outperform them in some cases, while being more robust.

One landmark is Meta AI’s Byte Latent Transformer (BLT). BLT is a tokenizer-free architecture that matches the performance of token-based large language models, at scale, with significant boosts in efficiency and robustness. How? BLT doesn’t naively feed every single byte into a giant Transformer – that would drown in a sea of tokens. Instead, it encodes bytes into dynamically sized “patches” that become the units of computation. Think of patches as a smarter form of token: instead of a fixed 50k-word vocabulary, the model itself decides how to group bytes into meaningful chunks on the fly.

These patches aren’t uniform; BLT dynamically adjusts them. The rule is intuitive: if the next bytes look predictable or low in information, group them into a long patch; if something complex or uncertain is coming, break the patch and let the model pay more attention. In practice, BLT measures the entropy of the incoming byte stream to decide patch boundaries – high entropy (unpredictable content) triggers a new patch. This means the model allocates more compute to the hard parts of the input and less to the easy parts. It’s like a reader slowing down on a dense paragraph but skimming through simple sentences.

The results are impressive. For a fixed inference cost (same number of FLOPs), BLT scales better than standard token models by growing both patch length and model size simultaneously. In other words, given the same compute budget, a BLT can handle more data or a bigger model than a vanilla Transformer reading subword tokens. Crucially, BLT’s patches ended up making sequences shorter on average than they would be with subword tokenization, saving compute without losing meaning. And because BLT’s “vocabulary” is just bytes (0–255), it’s inherently more flexible – it can ingest anything from English to emoji to programming code without needing a handcrafted tokenizer. No more mystery tokens lurking at the fringes, waiting to glitch out your model.

BLT isn’t the only byte-native pioneer. DeepMind’s MEGABYTE introduced a multi-scale decoder that can model sequences over a million bytes long. It segments an input into fixed-size byte patches and uses a local Transformer to capture patterns within each patch, and a global Transformer to capture patterns across patches. This two-level approach slashes the cost of attention (sub-quadratic scaling) and lets the model focus on local details and global context separately. The payoff: a 1.3B-parameter MEGABYTE model was able to generate 1-million-character sequences and perform competitively with subword-based models on long-context language tasks. In fact, MEGABYTE achieved state-of-the-art results in domains like image and audio modeling directly from raw data. Together, these results establish the viability of tokenization-free sequence modeling at scale. In plain terms, we can finally train models on raw text (or sound, or pixels) without needing a human-crafted compression step at all.

Another fascinating entrant is MambaByte, which took a different route to the same destination. Instead of Transformers, MambaByte is built on a state-space model (SSM) backbone. State-space models, like the recent S4 architecture, can maintain a fixed-size hidden state regardless of sequence length – neatly avoiding the quadratic blow-up of attention. MambaByte adapted an SSM (the “Mamba” model) to work on byte sequences without any tokenizer. The authors found that this model was competitive with, and even outperformed, state-of-the-art subword Transformers on language modeling tasks. And it did so while inheriting the benefits of token-free input, such as robustness to noise (one can imagine it doesn’t get thrown off by a single odd byte or a glitch token). They even tackled the speed issue: by using a clever speculative decoding trick – drafting with a traditional tokenizer model and then verifying with the byte model – they got a 2.6× inference speedup. In the end, MambaByte made a strong case that state-space models can enable token-free language modeling efficiently.

What these efforts have in common is the trend away from hand-crafted tokens and toward native byte/character inputs. This trajectory isn’t just academic; it carries practical significance for product builders. It means future models could handle any language or format you throw at them, out of the box. It means fewer engineering hours spent maintaining custom tokenization pipelines for every new market or data type. And it could mean more robust systems – ones that don’t freak out at an out-of-vocab token or need patchy fixes for every odd corner case. In short, as a product leader, you might soon be able to treat text as just data, and let the model deal with the raw bytes. That opens up a kind of simplicity on the far side of complexity – a chance to focus on higher-level problems because the low-level text processing is finally taken care of.

Beyond Attention: New Architectures to Slay Old Bottlenecks

Dropping tokenization is one side of the coin. The other side is evolving the model architecture itself. After all, if you feed a billion-byte text into a Transformer, you still face the compute bottleneck. So researchers have been busy rethinking how models can handle long sequences and complex inputs more efficiently. This is leading to some inventive architectural trends, which are worth watching if you’re charting a product roadmap.

One approach we touched on is patching – used by BLT and MEGABYTE. This is essentially patch-based inference for text, akin to how vision transformers use image patches. By chunking a long sequence into patches and processing hierarchically, the model avoids blowing up its internal workload. BLT’s twist of dynamic patch sizing (entropy-based) is especially interesting: it’s a form of adaptive computation, allocating resources where needed. That means if your input has a lot of fluff (say, repetitive spaces or easy boilerplate), the model doesn’t waste cycles on it. But if suddenly a DNA sequence or code snippet appears (high entropy, unfamiliar pattern), the model zooms in and spends more compute there. From a systems perspective, this is beautiful – it’s an AI doing resource allocation on the fly, rather than treating every input token equally.

Another trend is dynamic token pooling. Traditionally, Transformers process all tokens at all layers, which is overkill if some tokens are redundant. Dynamic pooling mechanisms attempt to shorten the sequence as it flows through the network. For example, a model might learn to merge or drop less important tokens after the first few layers, reducing the length for deeper layers. Recent research showed that a Transformer with dynamic token pooling can jointly segment and model language, achieving faster and more accurate results than vanilla Transformers within the same compute budget. Essentially, the model itself learns how to compress the text representation as it goes, instead of relying on a fixed tokenizer or a naive pooling like “take every 4 tokens”. This joint segmentation+modeling approach blurs the line between “tokenization” and “modeling” – they happen together, dynamically. For product folks, it hints at models that are more efficient and possibly more interpretable (imagine seeing which parts of a document the model deemed worth “zooming in” on or merging).

And we can’t forget the state-space model (SSM) path, as exemplified by MambaByte. SSMs like S4 sprung from outside NLP – a new way to handle sequences using linear dynamical systems principles. What makes them special is that, unlike attention, their memory cost doesn’t balloon with sequence length. An SSM can, in theory, handle an arbitrarily long input with fixed computational footprint, because it processes the sequence through a recurrence with a fixed state vector size (using a lot of clever math to do so fast). We’re still in early days of seeing SSMs compete with Transformers in pure performance, but their success in token-free modeling is a proof-of-concept that alternative architectures can overcome some Transformer bottlenecks. They might be more resistant to very long sequences or allow streaming data processing in ways Transformers struggle with. For AI product strategy, it means the space of viable model architectures is widening – and if your use case involves very long data streams (say, logging data, genomics, long videos, etc.), keep an eye on these alternatives.

All these innovations – patches, dynamic pooling, SSMs – are converging toward a common goal: breaking the trade-off between input length and performance. We want models that can read more, and read it efficiently, without needing us to spoon-feed them with pre-digested tokens or to prune context manually. It’s a thrilling area where fundamental research meets practical need. As a leader, one of your jobs is to sense when a research idea is nearing a breakout point for real-world impact. Byte-native Transformers and dynamic sequence modeling might be at that point. The next generation of AI products could very well ride on these capabilities, offering users seamless handling of huge, messy, multilingual data – and doing it faster and cheaper than before.

The Bitter Lesson Revisited: Scaling, Compute, and the Folly of Cuteness

Back in 2019, Rich Sutton wrote The Bitter Lesson, essentially pointing out that in AI, general methods that leverage compute have always won out over domain-specific cleverness. It was a sober reminder that often the best way to improve a system is not by adding more intricate rules or features, but by making it bigger, faster, and more data-hungry. In the context of tokenization and architecture, we’re seeing that play out in real time.

Consider tokenization: It was a clever hack, a human-engineered way to inject linguistic knowledge and efficiency. But the bitter lesson would predict that if we throw enough compute and learning capacity at the problem, a learned approach (like byte-level modeling) will eventually outperform the clever hack. And lo and behold, we now have models like BLT and MEGABYTE proving exactly that – given sufficient scale, letting the model learn from raw bytes works as well as or better than our carefully designed subwords. It’s compute over cognizance. It’s a bit humbling: all the time we spent perfecting tokenizers might eventually be obsolete, replaced by brute-force learning.

The same pattern emerges in architecture. Think of all the exotic architectures researchers tried over the years – gated RNNs, neurosymbolic hybrids, handcrafted parsing modules – many added complexity in hopes of a leap in understanding. Some helped on small scales, but when the dust settled, plain Transformers scaled up with mountains of data steamrolled past most of these bespoke ideas. The Transformer itself succeeded not because it was simpler than an RNN, but because it was more scalable (parallelizable, amenable to big data) – again a victory of scale and compute. Now, even Transformers have limitations (like that pesky quadratic cost), so new general methods like patching or state-spaces are coming in – but notice, these aren’t injecting human linguistic rules; they’re general-purpose strategies to handle more data. Patching bytes based on entropy isn’t a linguist’s idea, it’s an engineer’s way to optimize compute. State-space models didn’t come from analyzing grammar; they came from math that scales in a different way.

One very concrete example of compute over cleverness is in how we boost model performance today. If you want a model to get better at a task, you could hire linguists and domain experts, or… you could train a bigger model on more data, or use techniques like ensemble or retrieval. Nine times out of ten, the latter wins. The “bitter” truth is that scale (in parameters, data, or compute steps) often beats niche optimization. This is not to devalue scientific insight – it’s to recalibrate where insight is applied. Insight in AI has shifted from crafting the perfect rule to crafting the perfect scaling strategy or architecture that can handle scale. Insight is now in finding the simple rules that let brute force shine (like “scale length by patching” or “use byte-level to avoid vocab limits”), rather than micromanaging the model’s internals.

As product and research leaders, we should internalize this lesson. It doesn’t mean we always just buy the bigger GPU cluster – budgets are real – but it means we should be careful about chasing diminishing returns on clever hacks. If your team is spending months tinkering with a special-case feature to improve model X by 5%, consider: could a larger model or more data or a more general technique yield 15% with less fuss? Often, yes. The tokenization saga is a perfect case in point: instead of meticulously updating vocabularies and scripts for each new domain (a very clever but manual process), many teams now just switch to a model that doesn’t need that, and pour compute into training it. The long-term trend is clear: favor approaches that scale with compute and data over those that rely on frozen human insight. This also future-proofs your strategy – because if there’s one thing that ages quickly in AI, it’s our cute bespoke solution when someone else’s scaled-up model leaps ahead.

Yet, the bitter lesson has a sweet coda when paired with human judgment. We still need strategic cleverness – in choosing what to scale and how. The recent advances didn’t happen by accident; researchers recognized where a targeted change (like patching or an SSM) could remove a scaling pain point. They asked, “What if we remove the tokenizer?” – a very insightful question – and then let the scale do its thing. This interplay of insight and scale is where leadership comes in: posing the right questions and then leveraging compute to answer them, rather than manually crafting the full answer ourselves.

Context Is King: Beyond Model Size in the Real World

While model scale has driven incredible progress, another truth has become evident in practice: it’s not just about the model anymore; it’s about the ecosystem around the model. Specifically, how we manage and supply context to our models often matters more for real-world performance and cost than adding a few billion more parameters.

Think about it: GPT-3 was 175B parameters trained on basically the internet. But if you ask it a specific question about, say, your company’s internal data, it might flub it – not for lack of size, but for lack of relevant context. Enter retrieval-augmented generation (RAG) and related techniques. Instead of making the model bigger, we give the model a brainy assistant: a retrieval system that fetches relevant information on the fly. DeepMind’s RETRO model demonstrated this dramatically – a 7.5B parameter model hooked up to a large text database matched the performance of a 175B parameter GPT-3 on the Pile benchmark. In other words, 25× fewer parameters achieved comparable results by smartly bringing in external knowledge. That’s a giant leap in efficiency. The takeaway: sometimes knowledge in a database trumps knowledge baked into weights.

This has huge implications for product strategy. Instead of spending millions training a model from scratch to memorize your entire knowledge base, you might get more mileage by combining a moderately sized model with a fast search index or vector database. You keep the model light and nimble, and let it query for details as needed. It’s like having a small agile team with great internet access versus a giant know-it-all that’s slow and expensive.

Summarization is another form of context engineering changing the game. Long conversations, documents, or workflows quickly overflow any model’s context window (even though those windows are growing – 4k tokens, 16k, 100k…). Rather than brute force a huge context length at full detail (which is expensive), many applications now use summarization or distillation. They generate summaries of earlier chat messages, or compress a 100-page document into key bullet points, and feed that to the model. It’s a bit meta – using the model to help itself by creating condensed context. Done right, this dramatically cuts cost and latency while preserving relevant information. A summarized context can maintain coherence in a customer support chatbot without needing an exorbitantly large model or context size.

And then there’s caching and reuse: If your system answers a question once, cache it. If a particular expensive reasoning step is done, memoize it. Traditional software engineering taught us to optimize repeated computations, and AI systems are no different. Many product teams set up hybrid pipelines where the AI does heavy lifting once and reuses results, or where a smaller model handles easy cases and only escalates to a big model for the hard ones. All these tactics are about being smart with context and computation, not just throwing size at the problem.

Why am I harping on this? Because it represents a shift in what effective scaling looks like. For a long time, the mantra was “bigger model = better.” Now it’s more nuanced: better data and context management = better. We’re augmenting our large models with retrieval, tools, memory – essentially giving them external support rather than internally growing them endlessly. For product leaders, this means the frontier of improvement might be less about training the next 100B model and more about how you orchestrate information around the model. It’s a bit like the evolution of computing hardware: at some point, adding more cores yields less benefit than improving memory access or storage. In AI, adding parameters yields less benefit than ensuring the model always has the right information at its fingertips.

The savvy strategy is to combine approaches: use sufficiently large models to get general capabilities (thanks, scaling laws), but then use context engineering to ground and specialize those models on the fly. This often yields a better cost-performance tradeoff than a naive scale-up. And practically, it means you can deliver more value to users by making the model smarter (through context) rather than just bigger. As a side effect, these techniques also help with forward-compatibility: if you design your system to fetch knowledge and handle summaries, you’re less tied to any one model. You can swap in a new model later, maybe a byte-native one, without retooling your entire knowledge pipeline.

Insight as a Process: Rethinking Our Mental Models

All this talk of tokens, bytes, and scaling strategies ultimately points back to one thing: how we think and make decisions in this fast-moving field. It’s one thing to adopt a new model or technique; it’s another to cultivate the mindset that consistently spots these opportunities and pitfalls. Here’s where I’ll get a bit philosophical – without going too abstract – to share how I approach “insight” in AI strategy, inspired by some heavy thinkers I admire.

Insight often starts with better questions. Someone asked, “Do we really need a tokenizer?” – and that question led to CANINE, BLT, and beyond. In my own career, the biggest leaps forward came not from knowing the answers, but from bravely asking the questions that others thought were settled. For example, a few years ago everyone assumed “more parameters = better model.” But asking “what if we keep parameters the same and add retrieval?” led to a new line of products and research that changed the game. Encourage a culture of inquiry in your team. Ask things like: What are we assuming that might no longer be true? or Where are we being clever instead of letting the model learn? These questions open doors.

Next is detecting blind spots. We all have them – an area we overlook or an assumption we take for granted. Glitch tokens were a literal blind spot in language models – edge cases tokenizers didn’t handle well. When they first popped up, it was easy to shrug them off as oddities. But those who dug in realized they signaled a deeper issue in how models represent information. In our decision-making, blind spots can be strategic (maybe we’re fixated on model accuracy and ignoring inference cost, or vice versa). One way to reveal blind spots is to listen to diverse perspectives: your research scientists, your engineers, your users – each might see a risk or opportunity you don’t. Another way is to simulate failure: actively imagine scenarios where your plan falls apart, and see what you overlooked. This kind of reflection is like debugging your own thought process – a habit that pays off immensely in a complex field like AI.

A powerful step beyond identifying blind spots is integrating competing positions. In AI, we often see camps form: “All in on Transformers!” vs “New architecture now!”, or “Just scale!” vs “We need more data and retrieval!”. It’s easy to pick a side and develop tunnel vision. But the real breakthroughs often come from integrating the truths on both sides. The saga of model scaling vs. data augmentation is a great example: Both more parameters and more context turned out to be important. The best teams aimed to do both, or to balance them in creative ways. As a leader, you should be part diplomat, part synthesizer – able to take the intuition of one approach and marry it with the strength of another. When I lead strategy sessions, I often draw two seemingly opposed ideas from different team members and literally force a fusion: “If we weren’t to choose, how might we achieve both outcomes?” It’s amazing how often that yields a novel solution – a true insight that feels almost obvious in hindsight.

Under the hood, what I’m really advocating is a self-reflective insight process. It’s an internal game of hypothesis and verification, much like the scientific method, but applied to one’s own thinking. A famous epistemologist once described understanding as a spiral: we experience data, we ask questions, we get insights, we test them, and each time our understanding lifts to a higher integration. In the rush of AI development, taking the time for this reflective spiral is tough, but worth it. It helps us build not just better models, but better mental models of how we approach problems. And those mental models guide countless decisions big and small.

So, as you navigate the rapid changes – tokenization falling out of favor, architectures shifting under your feet, scaling throwing curveballs – remember to step back and think about how you’re thinking. Are you reacting on autopilot (“We’ve always done it this way”)? Are you clinging to a comfortable proxy metric while the world changes around it? Or are you actively questioning, learning, and reframing your approach? Leading in AI requires as much clarity of thought as it does technical know-how.

Building Forward-Compatible Mental Models: Key Takeaways

We’ve covered a lot, so let’s distill a few practical takeaways. These are guiding principles I use to keep my strategy robust and future-proof in the face of AI’s rapid evolution:

• Assume Change in Foundations: What is true today (like subword tokenization dominance) might not hold tomorrow. Be ready to pivot when core assumptions (e.g., “we need a tokenizer”) are challenged by new evidence. Design your systems in a modular way so you can swap out pieces (like tokenizers or model backbones) as the tech evolves.

• Prioritize Scalable Over Clever: When choosing solutions, favor those that scale with more data/compute over highly specialized tweaks. Many “clever” shortcuts in NLP – from hand-built token rules to custom features – have been outpaced by approaches that let the model learn with more compute. Invest in infrastructure that lets you scale, and in talent that knows how to leverage scale, rather than micro-optimize. As history shows, compute plus general algorithms tend to win.

• Leverage Context Engineering: Before defaulting to a larger model to boost performance, exhaust options to enrich the model’s context. Can you retrieve relevant facts? Summarize long inputs? Cache previous results? Often a smaller model with the right context outruns a bigger model blundering in the dark. This not only improves performance but can drastically cut costs. It’s a no-brainer for product pragmatists.

• Stay Architecture-Agnostic (to a point): Keep an open mind about new model architectures. Transformers won’t be the end-all forever. Whether it’s state-space models, patch-based hybrids, or something entirely new, be willing to experiment when your use case aligns (e.g., extremely long sequences might warrant an SSM like MambaByte). However, balance this with healthy skepticism – many new ideas fade. The key is to run small experiments early rather than betting the farm, so you’re ready to scale up if a new approach proves itself.

• Cultivate an Insight Culture: Encourage your team (and yourself) to question assumptions and learn from anomalies. Glitch tokens, odd errors, outlier data points – these often hide valuable lessons. Create space for reflection in your development cycle: post-mortems, “insight of the week” discussions, scenario planning. The goal is a team that’s not just executing the next task, but also continuously updating its mental model of what works and why.

• Integrate, Don’t Isolate: When faced with “X vs Y” debates (e.g., big models vs. smart context, speed vs. accuracy), look for integrative solutions. Often the best path is a hybrid: X and Y in moderation, or X enabled by Y. By avoiding extreme positions, you’ll build systems that are more balanced and resilient to change.

In closing, the landscape of AI product development is one of constant reinvention. Today’s glitch is tomorrow’s insight. We started with a weird token that broke a model, and ended up contemplating architectures that break the mold. The through-line is learning – at scale (for the models) and in depth (for us humans). As you lead teams or products in this space, ground yourself in first principles but be ready to update them. Build mental models that are sturdy yet adaptable, rooted in understanding yet open to new information. That way, whether it’s a paradigm shift from tokens to bytes, or Transformers to something new, or shallow models to deeply integrated systems – you’ll not only adapt, you’ll thrive. Here’s to the insights yet to come, and the exciting uncertainties that will drive them. 🚀

Sources:

• Yu et al., “MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers” (2023) – Multi-scale byte Transformer, patches for long sequences.

• Pagnoni et al., “Byte Latent Transformer: Patches Scale Better Than Tokens” (2024) – Dynamic byte patching, matches token model performance.

• Nawrot et al., “Efficient Transformers with Dynamic Token Pooling” (ACL 2023) – Dynamic segmentation of sequences improves speed and accuracy.

• Wang et al., “MambaByte: Token-free Selective State Space Model” (2024) – Byte-level state-space LM, competitive with subword Transformers.

• Clark et al., “CANINE: Pre-training an Efficient Tokenization-Free Encoder” (TACL 2022) – Character-level model with downsampling, no fixed vocab.

• Li et al., “Glitch Tokens in Large Language Models” (2024) – Study of anomalous tokenizer outputs impacting LLM behavior.

• LessWrong Wiki, “Glitch Tokens” (2023) – Definition and examples of glitch tokens causing gibberish output.

• Synced AI, “DeepMind’s RETRO vs GPT-3” (2021) – Retrieval-augmented model matches GPT-3 with 25× fewer parameters.

• Kevin Rohling, “BLT Deep Dive: Hello bytes, goodbye tokens” (2024) – Accessible overview of BLT’s patching approach and benefits.