Representation before Generation

Working Framework — 2026

On the atrophy of internal models, the limits of the generative paradigm, and who controls what comes next.

Thesis

As LLMs make generation effortless, they accelerate the atrophy of the representational capacity that makes generation meaningful. The solution is not better tools — it is a new practice for building internal world models before, and independent of, AI engagement. And if that capacity lives inside people rather than platforms, it cannot be bought, buried, or regulated away.

The Logical Chain

I. The Problem

What is actually being lost

The visible problem is dependency. People reach for LLMs to draft, design, decide, and describe. The invisible problem is what that dependency prevents from forming.

Learning to articulate why something feels wrong — not just that it does — requires sitting with incompleteness. It requires failure that isn’t immediately resolved. It requires the particular friction of trying to hold a position under pressure and discovering where it breaks. That process, accumulated over years, is how a person builds what we loosely call taste, or style, or judgment. It is not a talent. It is a structure — a world model, built from real encounter with the world.

LLMs short-circuit the process at every point. You never have to sit with not knowing. The gap between intention and articulation disappears. And so the internal structure never forms.

The Crisis — What is actually being lost

The output looks fine. The roots are gone. Humans who never fail at articulating why something feels wrong never develop the capacity to know when it is.

Representational Poverty — The compounding problem

Each generation less able to direct the systems that follow. They have generation without judgment. The articulation gap widens — nobody can say precisely why something violates the spirit of what came before.

The Energy Waste — Wrong layer, enormous cost

Almost all creative and computational energy is spent at the generation layer. Almost none is spent building the representational model that would make generation precise. This is expensive in every sense.

II. The Thesis

The architectural reframe

Yann LeCun’s argument is architectural. Predicting the next token — or pixel — is not wrong because it is technically difficult. It is wrong because it is working at the wrong level. Generating plausible surface is not the same as understanding the structure underneath. A model that predicts every detail of the future will fail, because the future is not fully predictable at the detail level. What is predictable — what can be learned — is structure.

JEPA (Joint Embedding Predictive Architecture) proposes something different: build an abstract representation of what the world means, work in that latent space, and let generation be downstream of that understanding. Deliberately discard unpredictable detail. Stop wasting capacity on noise. Focus on what is structurally true.

“Trying to predict every pixel is not just expensive — it is actively counterproductive. The model wastes capacity on inherently unpredictable visual details instead of capturing high-level predictable concepts.”

The human parallel is exact. A person who has genuinely developed style is not someone who has seen more references than anyone else. They are someone who has built a world model — through friction, failure, cultural immersion, and judgment under pressure — that allows them to know what something would resist before it exists. That model was not downloaded. It was constructed.

LeCun’s Wager — The architectural claim

Generation follows from representation. Without the model underneath, output is statistically plausible but structurally empty. The wrong layer is being optimised at enormous cost.

The Human Parallel — Style as world model

A person who has genuinely developed style has built a world model. They know what something would resist before it exists. This cannot be prompted into existence. It has to be constructed through encounter with the world.

What JEPA Actually Does — Throw away what you can’t predict

JEPA deliberately discards unpredictable detail and learns abstract structure instead. What seems like a loss of fidelity is precision. This is the structural opposite of how LLMs work.

III. Future Problems

What happens if this is true

If LeCun is right — if smaller, contextual, structurally-grounded models built on real-world experience outcompete brute-force generation — then several things follow, and most of them are not being discussed.

The first is an infrastructure paradox of historic proportions. The capital being deployed to build the current paradigm does not become worthless immediately, but it becomes increasingly misaligned. Companies most exposed are those most committed to the current direction — and those are the largest companies in the world. The response will not be orderly transition. It will be acquisition, consolidation, and regulation in defence of existing position.

The second is a shift in what constitutes the scarce resource. Right now it is compute. In a world of world models, the scarce resource is real-world context — embodied, specific, situated data that can only be collected where the world actually is. Whoever controls those pipelines (robotics, sensors, wearables, smart environments) controls what world models learn. This monopoly is harder to see than a data centre and harder to regulate than a model.

The Infrastructure Paradox — The last cathedral

$100B in data centres and nuclear power optimised for the generative paradigm. If LeCun is right, this capital becomes misaligned. Not gradually — structurally, then suddenly.

The Context Monopoly — The next scarce resource

Real-world context becomes the training data for world models. Whoever controls the pipelines that collect it at scale controls what world models learn. This monopoly is harder to see than compute and harder to fight.

The Efficiency Inversion — When the economics flip

Models that are smaller, contextual, and structurally grounded outcompete large generative models for most real tasks. When this happens — a when, not an if — the economic logic of big AI inverts overnight.

IV. Power & Control

Who controls the transition

The decentralisation argument is correct in its instinct but imprecise in its target. Decentralised compute is a partial defence at best. The real question is where the representational layer lives.

Centralised generation is centralised thought direction — not through censorship, but through dependency. If you cannot generate without the infrastructure, and you have not built the internal model to know what you want, you are entirely captured before any censorship is required. The platform does not need to restrict you. It only needs to be the only place where creation is possible.

The historical pattern of every infrastructure transition is consistent: alternatives that threaten the dominant paradigm get acquired or regulated into irrelevance. The window between a better approach existing and being absorbed is narrow. Open research and distributed practice are structural defences, but the most durable protection is the one that cannot be externalised at all.

Centralised Generation = centralised thought direction

The platform does not need to censor you. It just needs to be the only place where creation is possible. Dependency is the mechanism of control, not restriction.

Buy or Bury — The historical pattern

Every infrastructure transition follows the same logic. Alternatives get acquired or regulated into irrelevance. The window between a better approach existing and being absorbed is narrow.

The Real Protection — Representation inside people

The actual protection is representational capacity that lives inside people — built through practice and genuine encounter with the world — and that no external system fully owns. That is what cannot be bought or regulated away.

V. The Response

What can actually be done

The Bauhaus was not just a school. It was a proof of concept for a philosophy — that the separation of craft from art was destroying both — and every workshop, every assignment was generating evidence for or against the thesis. The school was the experiment.

The same structure applies here. The thesis is that humans need to build world models independent of AI, that this can be taught, and that people who do it produce qualitatively different work. An experiment — small, documented, public — generates the evidence. The experiment is the contribution.

This does not require institutional credentials. It requires a clearly argued thesis, intellectual honesty about what is borrowed from whom, a real experiment that generates real evidence, and the credibility to attract the right collaborators. The people who understand the architecture do not care about creative pedagogy. The people who care about creative pedagogy do not understand the architecture. The translation between them has not been made.

The Framework — Human JEPA

A structured practice for building representational capacity independent of AI. Not a methodology for using tools better — a pedagogy for developing the internal model first. What something cannot be. What it would betray.

The Experiment — Proof of concept

Small cohort. Online. Half the curriculum has zero AI involvement — building the model. Half uses AI as material, not oracle. Measure coherence over time, resistance to trend, ability to articulate why something does not belong. Document everything publicly.

The Contribution — The untranslated gap

Decades of building world models professionally without calling them that. The architecture people and the pedagogy people have not met. Nobody has made this translation. That is the gap. That is the work.

The risk is not being too late or underqualified. The risk is diffusion — holding the AI architecture argument, the creativity crisis, the decentralisation argument, and the pedagogy all at once without a single sharp thesis at the centre.

Version 1.0 — Five layers: Problem · Thesis · Future Problems · Power & Control · The Response