The thesis, the observations, and the tension at the center of the work.

The frontendwhat we mean— click of learning and professional work has always been the constraint — not human intelligence, not the availability of knowledge, not effort. For most of history the best frontends we had — the teacher, the lecture, the textbook, the specialized software package — were fixed, one-to-many, non-adaptive, and jargon-gated. Some people fit through the filter and we called them talented. Most did not, and we either blamed them or forgot them. Large language models together with modern interactive interfaces are the first serious plural, adaptive, multimodal frontends the species has ever had access to.

LeDesign was built to ship those frontends commercially, in fields where proprietary siloswhat we mean— click have been pricing people out of their own tools. LeResearch is the other half: the research that asks whether the reframing holds under evidence, across domains where no product roadmap would fund the work — and refuses to treat the inherited framewhat we mean— click as settled just because the frame is what everyone already agrees on.

Sapolsky's work on behavior, biology, and environment lands on one conclusion we cannot step past: human capacity is not fixed at birth. It is shaped continuously by the environment it encounters — and that environment includes every frontend through which knowledge is offered. What we call smartness is downstream of who happened to fit the particular filter a given society built at a given time, using a given paradigm, for a given distribution of people.

Post-industrial schooling built one linear frontend. A teacher at the front, a classroom in rows, a single pace, a single vocabulary, a single language, a single canonical text. Students who fit the filter were called able. Students who did not were called unmotivated, slow, or worse — and most importantly, came to believe the framing about themselves. The filter defined who was smart. The smart ones were permitted to define the filter. The people outside that loop were not less able; they were differently served by a frontend designed for a mean that never existed, enforced as if it were physics.

This lands as a moral claim, not just a factual one. If capacity is environmental, structural failure to provide the environment is a moral failure rather than an individual one. If the frontend is part of the environment — which it has always been — then the design of the frontend is a public-health question, a public-mind question, and only incidentally a UX question. Nobody gets to wash their hands of it by saying “we are just building software.”

Eight glasses of water a day. Two thousand calories. Three meals on a schedule set by industrial factory shifts. Eight hours of sleep. The nine-to-five workday. The classroom itself. Each of these was a specific decision by specific people for specific reasons, most of which had nothing to do with what the human body or mind actually needed. And yet they are absorbed, operationally, as truths — as infrastructure, the same way gravity is infrastructure.

Very few of them survive contact with the evidence. They are not uniformly wrong; some are approximately right for some people some of the time. The point is not the specific number. The point is that a contingent frame is being accepted as a natural law by people who would, in other contexts, be the first to insist on seeing the evidence.

Section 2 named the outcome — that contingent decisions calcify into infrastructure. This section names the mechanism: the same biological and cognitive normalization that makes humans adaptable to almost any environment also makes change at low gradient invisible to us, and reserves our awareness for change steep enough to fire many sensors at once.

The species-level argument is straightforward. The same body and brain that allowed humans to inhabit the Arctic and the equator, the savanna and the city, did so by continuously recalibrating what counts as normal. A fixed reference for normal would have killed the lineage. A drifting reference, retuned to whatever is most recently present, kept it alive. The cost of that adaptation is not metaphorical: things that move slowly enough do not register as change. They register as the world.

The implication for inherited systems — the imagined orders, in the anthropological sense, that humans started building once group sizes exceeded what direct social cognition could hold (roughly Dunbar's number) — is that slow drift gets absorbed into the frame without ever entering deliberation. A nine-to-five workday, a classroom of thirty students in rows, eight glasses of water a day: each of these started as a gradient steep enough to be visible (a labor strike, a Prussian school decree, a magazine column), and each ended as something nobody remembers having decided. The shock made them legible. The normalization made them invisible.

The inverse case — the paradigm shift, in Kuhn's sense — is what happens when a change exceeds the normalization range fast enough that many sensors fire simultaneously. Body, society, ecosystem, market, institutional rule-set: all reading out of adjacent normality at once, with no available frame to absorb the input. That dissonance is what “feels like a paradigm shift.” It is also when public attention, regulation, journalism, and resistance briefly catch up to the change. Then normalization resumes, the frame absorbs the new state, and the cycle resets.

The political question — who steers the gradient — is not symmetric. The same actor who would lose a fast public argument can usually win a slow private one, because nobody is looking at the slope. A great deal of what shapes contemporary life — recommender systems sorting hiring, lending, news, dating; the steady extension of monitoring at work; the steady contraction of what counts as professional judgment — moved on a slope shallow enough to never trigger broad debate. The shock-and-normalize cycle is not just a description of how change happens. It is, increasingly, an instrument: a way of allocating which changes get democratic friction and which do not.

The intellectual lineage of this section sits across several traditions we have not yet developed in their own right — Castoriadis on the imaginary institution of society, Anderson on imagined communities, Searle on institutional facts, Berger & Luckmann on the social construction of reality, Bourdieu on doxa and habitus, Pauly on shifting baseline syndrome, Schmachtenberger on the metacrisis. Each is named in §12 with a one-line note on how we intend to take it up. This section will be revised as those threads develop.

AI has broken the silos open. It is feasible for one person or a small team to produce the adaptive, multimodal, plural-by-construction frontends that an entire industry of proprietary software gatekeepers previously charged rent for. That observation is correct as far as it goes.

The companion observation we find less comfortable: the word AI has become analytically unusable in public discourse, precisely because everyone — the people selling it and the people rejecting it — benefits from keeping the category undifferentiated. On the hype side, AI is a unitary force with a runaway future, deserving massive capital on the theory that any piece of it might turn out to be the piece. On the refusal side, AI is a unitary threat that consumes water, replaces jobs, and steals public knowledge, deserving rejection as a category because decomposing it is work. Both positions avoid having to name what they actually mean, and the ambiguity serves both.

When someone says “we don't do AI,” the sentence has no referent until a specific row is named. The refusal to decompose is itself the governance pathology — it lets a blanket rejection pass as a considered position, and it lets a blanket enthusiasm pass as the same. Neither side has to do the work of distinguishing a mathematical method from the business model that happens to be monetizing it this year.

This is why the most common critique of AI — that it consumes water, that it is extractive, that it replaces labor — is mostly, in substance, a critique of bubble economics, not of the underlying mathematics. Compute intensity and ecological cost are overwhelmingly concentrated in training frontier foundation models at scale, because scale is what capital markets are pricing, not because the applications require it. Inference on a small, purpose-built, openly-licensed model on commodity hardware is measured in watts, not megawatt-hours. A physics-informed graph neural network for aquifer modeling is AI in the same taxonomic sense as a frontier LLM and has essentially none of the objectionable properties. The refusal to decompose collapses that distinction on purpose, because naming it would concede that work can be done without the business model attached.

LeResearch's existence presumes the distinction is real and operable. Our work happens in small teams, on open hardware at commodity price-points, with models small enough to run on a laptop, trained on openly-published data, under licenses that require outputs to be redistributable. This is AI in the taxonomic sense. It is not the thing the refusal is rejecting. But the dominant discourse has no vocabulary for the distinction, so work like ours sits in a semantic blind spot — present but unclassifiable — until someone builds the vocabulary that separates a mathematical method from the economic model currently monetizing it. Part of what LeResearch exists to do is build that vocabulary.

§4 refused to let AI be a single thing. The same refusal applies to labor. The conventional question — “what jobs will AI replace?” — assumes the current job structure was natural, optimal, or inevitable, when it was none of these. The category of job in its modern sense (wage labor as identity, hours as contract, employer as primary social affiliation) is roughly two hundred years old. Before it: household economies, craft, subsistence, slavery, serfdom, indentured service, common land. The post-WWII professional middle class — the population that experiences “AI replacing my work” most acutely — is younger than penicillin.

Who got which job inside that arrangement was never a measurement of natural capacity. It was a function of race, gender, citizenship, language, schooling access, and whose children were allowed into which credential pipeline. The knowledge worker category that contemporary AI discourse most often centers — paralegal, junior copywriter, customer-service operator, illustrator, translator, programmer — is a particular layer of a particular society at a particular moment. AI does not disrupt the assignment. It operates through it.

When the joint decomposition is applied — which AI, which labor, at which gradient (in the sense of §3) — the picture stops being one story.

These are three different gradients and three different political responses. Calling all of them “AI and jobs” is, structurally, the same move as calling everything from a thermostat to ChatGPT AI: a refusal to decompose that is not neutral. It is the move that lets the shock layer absorb the public attention while the slow layer reorganizes labor without it.

The earlier observation — that doom and hype are the loud sensors that keep the rest invisible — is the same mechanism viewed from the political side. The discourse is not failing to discuss AI and labor. It is discussing the part of AI and labor that fires the sensor, while the part that moves on the slow gradient embeds.

The work the framework therefore wants to do is not “predict which jobs disappear.” That is the wrong question, asked from inside the silo. The work is: name the assignment that produced the current job structure, name the gradient on which any given AI is reorganizing it, name the actor who benefits from which sensor firing, and refuse the version of the conversation that treats either AI or labor as one thing.

The historical contingency of the job as a category — central to the argument above — sits in conversation with Graeber (Bullshit Jobs, Debt) and Zuboff (Surveillance Capitalism) among others, listed in §12.

AI has broken the silos open. The companion observation we find less comfortable: AI also compresses the silos' outputs into one confident-sounding voice and replays them as though they were the record of human knowledge — rather than the record of what got written, by whom, weighted toward the paradigm that decided what deserved to be written. The representation biases do not disappear when you route them through a language model. They calcify. Every query returns a weighted average of the same inherited frame.

A recent small example: an image model was asked to generate a community. Most of the figures in the frame were Black or brown. The one in the center, taller than the others and dressed in a suit, was white. Nobody told the model to do that. The training distribution did — because the training distribution is a statistical record of whose bodies, whose clothes, whose postures were historically placed at the centers of frames in the images the model was trained on. No malice required. The frame IS the malice, carried forward.

And the compression layer is commercially governed. Production models are silently versioned — RLHF updates happen, fine-tunes ship without announcement, A/B tests run on real questions in real production contexts. The answer returned today is not the answer that would have been returned last quarter, and the people using the model typically do not know. “Truth” in the public domain becomes, in practice, downstream of whichever provider has the subsidy runway plus the distribution channel. The model that wins is not necessarily the model that is most accurate. It is the one whose business lasts longest in the AI-investment cycle we are currently inside.

If a society routes its default truth-formation through a small number of privately-governed, silently-versioned models, the capacity for collective error-correction — which has always depended on dissonance, patience, diverse error, and the structural visibility of disagreement — gets thinner. Not zero. Thinner. The early sign is not obvious collapse. The early sign is that disagreement starts to feel eccentric rather than ordinary, and that more people say “the AI says” the way a previous generation said “science says,” as if that were the end of the conversation.

Section 1 named how people who were filtered out by the linear frontend come to believe they cannot understand the work. The mirror observation is less comfortable and less often stated: people who were not filtered out — who hold governance roles, institutional salaries, and the decision rights that flow from them — often refuse to analyze the same work, under different cover. Where the excluded internalize “this is not for me,” the privileged perform “we respect what the people believe.” The public posture is humility. The operational effect is the same: the category is not decomposed, the specific instance is not examined, and the governance choice is made on brand rather than substance.

This is a live pathology in the environmental, advocacy, and community-nonprofit world — and it is not politically symmetric with the first. When an excluded person says “I don't understand AI,” they are reporting on the filter that failed them. When a nonprofit director paid several multiples of the median household income in their service area says “the people don't like AI,” they are invoking a public they have not consulted on a specific question to justify declining to analyze the specific question. The first is a symptom. The second is a decision — one of a set of decisions that cumulatively reproduce the environment the first is a symptom of.

Both failure modes produce the same outcome. They keep the dominant framework — whatever it is, technical or otherwise — from being examined in its particulars. The first happens because the frontend filtered someone out. The second happens because the frontend privileged someone into a position from which examining it would be uncomfortable. Both versions need to be named, because addressing only the first leaves the second intact, and the second is the mechanism by which institutional governance actively reproduces what institutional rhetoric claims to fight.

LeResearch commits to being honest about both versions, including when the second version shows up inside our own organization — which it will, because no structure is immune to it, and the only durable defense is a governance split (board authority vs. executive authority vs. research authority) that makes substantive analysis the default behavior rather than a thing that has to be insisted on under pressure.

For publicly documented cases of this pattern in operation — including the NYC Department of Education ChatGPT ban-and-reversal as a near-textbook execution and the Memphis xAI opposition as a counter-example of substantive (not pathological) refusal — see /philosophy/cases. The framework should be triangulated against the public record, not just against any participant's account.

This is not an argument against AI as a frontend. LeDesign uses language models precisely because they are the first adaptive, multimodal, plural-by-construction frontends humans have ever had access to. The argument is against monoculture in the frontend layer. It is against the version of the future where everyone consults the same model, that model is governed by three or four companies, and our collective ability to say “that is wrong, and here is why” atrophies because we stopped exercising it.

This is the tension we find ourselves unable to step past. It is the reason LeResearch exists as a distinct entity — the reason we did not simply expand LeDesign's product work. Research on epistemic hygiene in an AI-mediated knowledge ecology is not a product requirement that resolves in eighteen months. It is slow, it is unglamorous, it does not resolve cleanly, and it touches every substrate we work on — from how a child meets an aquifer for the first time, to how a household assembles nutrition for a week, to how a court evaluates a calibrated groundwater model under equitable apportionment. The work is the ecology itself.

LeResearch is how we intend to hold it.

How the thesis becomes day-to-day product and research decisions. These are constraints we accept, not aspirations we aim at.

• ·Not a replacement for teachers, doctors, lawyers, engineers, or scientists. These are frontends with virtues software does not have.
• ·Not apolitical. Water is political. Education is political. Law is political. AI-governance is political. Pretending otherwise is itself a political stance — one that defaults to the status quo.
• ·Not a universal curriculum. The reading levels, representation toggles, depth controls — these are frontends each learner composes. They are not a standard we impose.
• ·Not a static framework. This document will need to be revised. If it stops being revised, we have stopped learning.
• ·Not a closed ecosystem. Open-source, open-data, open-hardware where licensing allows. No “you have to use our full stack” ever.
• ·Not a single-domain research center. Hydrology, food systems, educational frontends, AI epistemics — these are substrate tracks, not the mission. A donor looking for “the water nonprofit” or “the food nonprofit” is in the wrong place.

Across writing, product copy, proposals, documentation:

• Learner-first, not teaching-first. Prefer “the tool makes it possible for someone to…” over “we teach X.” Avoid totalizing constructions (always, every, all).
• Conditional over declarative. “If this is useful…” rather than “this is the right way to…”
• Specific over abstract. A $15K line item defending 150 hours of classroom UX review beats a $40K line for “educational consulting.”
• Humble about what we don't know. Uncertainty is named. Drafts are labeled drafts. Work in progress is labeled work in progress.
• Political when honesty requires it. We do not hide behind neutrality when the facts have stakes. We do not campaign, but we do not obscure either.

This document leans on a set of intellectual traditions developed elsewhere. The list below is the index; the first-pass treatments — three to five paragraphs each, naming the load-bearing argument, the connection to the framework, and the question we have not yet worked out — live on the dedicated open-threads page. Each entry below links to its developed treatment. The order is rough; the groupings are not.

• 1975
  Castoriadis · L'institution imaginaire de la société

  The 'imaginary institution' as the substrate beneath any social order; the distinction between the instituted and the instituting imagination. Direct ancestor of how we use 'imagined orders' in §3.

  →
• 1983
  Anderson · Imagined Communities

  How print capitalism produced the nation as a felt collective among strangers. The mechanism that scales Dunbar.

  →
• 1995
  Searle · The Construction of Social Reality

  'Institutional facts' as objective consequences of collective intentionality. The analytic-philosophy bridge to the same observation.

  →
• 1966
  Berger & Luckmann · The Social Construction of Reality

  The phenomenology of how habit becomes typification becomes institution. The micro-mechanism of §2 and §3 together.

  →
• 1972–1980
  Bourdieu · doxa and habitus

  The space of the 'taken for granted' that frames what can even be argued about. Direct cousin of §2.

  →
• 2011
  Harari · Sapiens

  The popular synthesis of the 'intersubjective myth' thesis at species scale. Useful for vocabulary, contested for analytic depth — to be discussed critically, not adopted.

  →

• 1995
  Pauly · Anecdotes and the shifting baseline syndrome of fisheries

  The empirical naming of normalization-as-blindness in an environmental setting. Cleanest entry point for §3.

  →
• 1962
  Kuhn · The Structure of Scientific Revolutions

  Already central in §3. To be reread alongside Castoriadis: paradigm shift as an instituting moment.

  →
• 2007
  Klein · The Shock Doctrine

  Companion case for the political instrumentalization of the shock half of the cycle.

  →
• ongoing
  Schmachtenberger · the 'metacrisis' framing

  Connects shock-and-normalize cycles to the present-day political case across multiple substrates.

  →

• 2018
  Graeber · Bullshit Jobs

  The historical contingency of 'the job' as a category. Foundation for the labor-side decomposition in §5.

  →
• 2011
  Graeber · Debt: The First 5,000 Years

  The longer arc — wage labor as a recent form among many older economic relations.

  →
• 2019
  Zuboff · The Age of Surveillance Capitalism

  The slow-gradient reorganization of labor and attention via recommender and behavioral systems. Central to §5's 'recommender systems' bullet.

  →

The treatments on the open-threads page are first-pass readings, not finished pieces. They are commitments — to do the reading, to revise the page, and eventually to split out the threads that mature into their own essays, diagrams, or worked cases. Read the developed treatments →

The tagline — “a small contribution to the silos's fall” — is the one-line version of everything above. It means: we are not the ones tearing down the wall. The wall is coming down. We would rather be in the rubble with the people who were priced out, than on the other side charging rent for the door.

Last revised 2026-07-14. Living document.