Intelligence saturates — the body does not

Intelligence saturates — the body does not

· 10 min read

Auf Deutsch lesen →

Listen to this article
Speed

AI narration · automatically generated

Subscribe as a podcast
RSS
Personal note

Stored only locally in your browser — nothing is sent.

Two reading levels: the running text is the through-line — for everyone, no prior knowledge needed. The clearly marked In-depth boxes are for specialists and can be skipped without losing the thread.

An uploaded fly that isn’t one

In March 2026 the Californian start-up Eon Systems published a post with a title straight out of science fiction: “We’ve uploaded a fruit fly.” Founder Michael Andregg meant it literally. His team had taken the wiring diagram of the fly’s brain mapped by the FlyWire project — the connectivity of some 140,000 nerve cells — turned it into a simple computational model, and used it to steer a physically simulated fly body. By the company’s own account, the model reproduces the real fly’s behaviour to 91 percent.

That sounds like crossing a threshold — the old dream of uploading a mind, first rehearsed on a fly. Except it isn’t. Kording, whose own lab is listed among Eon’s advisers, would not use the word “uploading” here, and the reason is precision. A wiring diagram mostly says that two nerve cells are connected — not reliably how strongly, and certainly not what glial cells, neurotransmitters and the constant reshaping of connections add. The worm C. elegans has been fully mapped for almost four decades, and its behaviour still cannot be reliably derived from the diagram. The simulated fly’s behaviour, too, was not derived from the map but to a large extent trained in.

What is missing is not the last percentage point of compute. What is missing is the body. Look at a fly under the microscope, Kording says, and it is stunning how good that body is — “a machine optimised over millions of years to be a fly.” A brain without the body it grew to steer is a script without a stage. The fly thus poses the question this essay is about: how far does intelligence carry on its own — without the body and the world on which it formed?

Two roads lead to the body

There are two very different answers, which should be kept apart even though both end at the body. One is economic: even a bodiless, arbitrarily clever AI hits a limit, because intelligence alone moves nothing. The other is cognitive-scientific: intelligence as we know it is itself a product of the body. The first road is about the usefulness of intelligence, the second about its origin. Let us take them in turn.

The first road: why the usefulness of intelligence saturates

Remarkably, the sharpest argument against AI hubris comes not from a data centre but from economics. Konrad Kording — Penn Integrates Knowledge Professor at the University of Pennsylvania, a movement researcher by training — published a paper with the economist Ioana Marinescu at Brookings in November 2025 that splits the world into two mutually dependent sectors: a physical one — bodies, machines, matter, everything that actually acts on the world — and one of intelligence. Because each needs the other, a sentence follows that cuts against the singularity chatter: “Given complementarity between the two sectors, the marginal returns to intelligence saturate, no matter how fast AI scales.”

The picture is simple. Give someone the task of digging over a garden, and little intelligence — they dig aimlessly. Give them more — they get more efficient. Give them infinite intelligence — they find the perfect path and are still not infinitely fast, because they have to move a body and physics holds. The added value of intelligence runs up against a ceiling set by the physical sector. The paper keeps the economic sharpening — that wages might first rise and then fall under automation, as people are pushed from the screen back into physical work — deliberately cautious: the result of a baseline simulation, explicitly not inevitable. What holds is the core: of the two quantities that must work together, only one scales at the speed of computing. The other stays bound to steel, supply chains and the inertia of matter.

Note what this road does not need: any claim about what intelligence is. It holds for a bodiless super-AI just as for a human. It is an argument about the leverage of intelligence, not about its nature.

The second road: where intelligence comes from

The second road is trickier and more interesting. It claims not merely that intelligence meets its limit at the world, but that it gained its shape from the world. Care is needed here. “Intelligence comes from the body” does not mean “no body, no intelligence” — that hard version would be settled in two sentences.

First objection: language models. They achieve enormous things, entirely without a body. But they form not on the world but on its description — trained on a mountain of text produced by embodied humans. Their meaning is borrowed. They are precisely the example of an intelligence that floats free because it lacks the friction of the body. Tellingly, the vanguard of AI research points there itself: Yann LeCun holds purely textual systems to be insufficient and is working on machines that build a world model from observation and action.

Second objection: locked-in patients. People whose bodies are completely paralysed go on thinking — seemingly proof that cognition works without movement. But this mind was built over decades through a moving body; its concepts, its expectations, its assumptions about the world were acquired sensorimotorically. That the output later falls silent does not erase this inheritance.

So the claim has to be measured precisely. For acting in the world — movement, perception, motor control — the body is constitutive: this intelligence is the control of a body in a world. For highly abstract concepts — democracy, prime number — the thesis grows weaker, and that should be conceded openly. The title means the first: intelligence, insofar as it acts in the world, has the body not as a footnote but as a condition.

What AI calls a “world model” — and motor control has long known

This is where it gets interesting for movement science. For what the AI labs are now proclaiming as their biggest breakthrough — “world models”, with which a machine imagines what its action will bring about — is at heart what this field has studied for decades as an internal model: the ability to predict the next state and to learn from the prediction error. AI is rebuilding, with the compute of entire data centres, what a brain does in passing when grasping, walking, balancing. Rodney Brooks put the idea sharply back in 1990: “The world is its own best model” — a system coupled directly to the world need not represent it internally; it queries it.

One should not oversell the closeness: it is the same computational motif, found independently, not proof that biology works only this way. But the convergence is remarkable. The most productive AI architecture of the present lands on the principle that neuroscience found in movement.

In-depth · for specialists What is meant is the forward model in the sense of Wolpert & Kawato (1998): from state and efference copy, predict the next state; learn from the prediction error. Kording & Wolpert (2004, Nature) showed that this estimate combines a learned prior with uncertain feedback in a Bayes-optimal way — the motor system as a probabilistic estimator, measured, not metaphorical. Ha & Schmidhuber’s “World Models” (2018) implement formally the same thing (a recurrent predictor module in latent space), LeCun’s JEPA programme (2022) likewise — prediction in representation space rather than on pixels. It is a formal homology, not an identity: the AI models run almost throughout in simulation, without the noisy, delayed real body their models depend on. And that morphology itself carries part of the control (Pfeifer & Bongard 2006) remains, in silicon, so far out of reach — whether one calls it “computation” or not (contested: Müller & Hoffmann 2017).

The real provocation: supplier or overtaken?

From this follows a question the field should ask itself before others answer it. For decades movement science was the only discipline that seriously measured internal models of the body in the world. Now AI is industrialising exactly that — with data, compute and a pace no lab ever had.

Two futures are conceivable. In one, biomechanics becomes the supplier and touchstone of embodied AI: its models, its measurement data, its test procedures become the reference the machines must be measured against. In the other, the engineering races ahead and the field becomes a spectator to its own core question — the way linguistics had to watch language models “solve” language without needing its theories. Which of the two comes to pass is not settled. But it is being decided right now, and more in data centres than at conferences.

In-depth · for specialists The difference from linguistics may lie in the data regime. Language sat there as a giant corpus, all but free; embodied interaction data of the quality of motion capture, EMG or force measurement does not. Where clean physical measurement stays expensive and scarce, the discipline that gathers and interprets it keeps a bargaining chip — provided it treats its models as a benchmark for the machines, not merely as a description of the human.

Where the thesis must stay honest

An essay that hides its strongest opponents is worthless before experts. The hardest counter-position comes from the camp of amodal cognition: Bradford Mahon and Alfonso Caramazza showed in 2008 that the co-activation of sensorimotor areas during thought does not yet prove embodiment — the same data are compatible with an abstract, symbolic model of concepts. Co-activation is not constitution. For highly abstract concepts that is a fair objection, and the second road should not talk it down. But the ground of this essay is movement and perception — and precisely there even the critics do not dispute the sensorimotor involvement; they only dispute whether it constitutes or merely accompanies. Whoever knows the limits of their own thesis has the better thesis.

What remains

Two roads, one destination. Economically, intelligence meets the world because on its own it moves nothing; cognitively, it already carries the world within it because it formed on it. Either way the body is not the footnote of intelligence but its counterweight. The uploaded fly demonstrates it: a perfect wiring diagram — and still not a fly, because a fly is a body in a world, not its image.

And for the discipline that measures the body, there is both opportunity and warning in this. The next hard problem of AI — a machine that models the physical world well enough to act in it — is exactly its own. The frontier of artificial intelligence has not left the body behind. It is returning to it. The only question is who will hold the cards then.

If this essay gave you something to think about, feel free to share it — and when the next AI miracle is announced, it is worth asking whether it models the world or only its description.

Sources (selected):