What this argues. This is a position piece rather than a survey. I am not claiming machines will stay behind on benchmarks. They will not, and on any task with a score that race is effectively decided. The claim is that "outsmarting" is the wrong comparison, because natural intelligence is not a performance system that happens to run on biological hardware. It is a survival system, and its intelligence is inseparable from what it exists to do. Plenty of people I respect disagree with this.
The oldest embarrassment in the field
Moravec's paradox has been sitting in AI's lap since the 1980s and has never really been dislodged. The things humans experience as difficult, like symbolic logic, chess, and integration by parts, turned out to be comparatively easy to mechanize. The things a two-year-old does without apparent effort, like walking over uneven ground, recognizing that a half-occluded object is still an object, or picking up a cup they have never seen before, turned out to be extraordinarily hard.
The standard explanation is evolutionary time. Sensorimotor competence has had hundreds of millions of years of optimization behind it and formal reasoning has had a few thousand. That is true as far as it goes, but it undersells the point. The difficulty is not that perception and movement are larger computational problems. It is that they have a different shape. They are not problems with answers. They are continuous negotiations with a world that does not hold still, carried out by a body that has something at stake in how they go.
Cognition grew out of staying alive
Every nervous system is first a homeostatic device. Before it was in the business of solving problems it was in the business of keeping internal variables inside survivable bounds: temperature, glucose, oxygen, hydration. Cognition developed as an extension of that regulatory loop rather than as a separate faculty added on top of it.
This is why Damasio's work on somatic markers cuts so deep. Patients with damage to the ventromedial prefrontal cortex often retain intact IQ, intact logic, and intact language, and become catastrophically bad at deciding anything at all. They can enumerate options all afternoon. What they have lost is the bodily valuation signal that makes one option matter more than another. Their reasoning is fine. Their reasoning is also useless, because reasoning was never the mechanism that ranked outcomes.
A system with no homeostasis has no stakes. Give it an objective and it will optimize that objective beautifully, but the objective arrived from outside. It was not generated by any condition the system is in, and nothing about the system's continued existence depends on the result. The architecture of natural intelligence, from attention through memory consolidation through emotion, is organized around the fact that something is always at risk.
Situated in a world rather than trained on descriptions of one
The enactivist tradition, and specifically Varela, Thompson and Rosch's argument that cognition is enacted through a body's coupling with its environment rather than computed over internal representations of it, has aged well. An organism does not perceive a world and then work out what to do about it. It perceives a world already carved into possibilities for action.
Gibson called these affordances. A chair affords sitting to an adult and climbing to a toddler. Neither of them computes this. It is simply what a chair looks like to a body of that size with those goals, because the perceiving system and the acting system are the same system.
This is also why the frame problem, which was Dreyfus's long-running objection to symbolic AI, never really disappeared. An embodied creature does not face a combinatorial explosion of possibly-relevant facts when its situation changes, because relevance is not computed. Relevance is what a body with needs finds salient. Systems that learn from descriptions of the world inherit an enormous amount about how the world gets described. What they do not inherit is a standpoint from which some parts of it matter more than others.
Priors that were already paid for
A newborn is not a blank slate running gradient descent. Infants arrive with structured expectations about objects, agents and number, and these show up in looking-time studies far earlier than they could plausibly be learned from data. Gopnik's work on children as active experimenters describes what happens next. They do not passively fit whatever data arrives. They intervene, run small experiments, and go out of their way to produce exactly the observations that would discriminate between competing hypotheses.
Two features of that are hard to reach by scaling:
- The priors are cheap to use and were paid for over evolutionary time. A child needs a handful of examples where a model needs millions, because the child is not starting from nothing.
- The curiosity is endogenous. Nobody assigns a toddler an exploration bonus. The drive to find out originates inside a system that will still be alive tomorrow and stands to benefit from knowing.
Gigerenzer's ecological rationality makes a complementary point about the supposed irrationality of human heuristics. Measured against formal optimality, human shortcuts look like defects. Measured against the actual conditions of decision-making, which involve incomplete information, time pressure and real consequences, they are frequently better calibrated than the optimal procedure, because they are fitted to the environment that produced them rather than to an idealized version of it.
The asymmetry that does not close
Predictive processing, where the brain is understood as minimizing prediction error with perception and action as two routes to the same goal, makes the difference precise. A biological system can reduce surprise in two ways. It can update its model, or it can change the world so that the prediction comes true. Both routes are available at all times, because it has a body in the same world its predictions are about.
The second route does not transfer. Learned systems minimize error by updating, and that is the only lever they have on the loop. It is a different relationship to being wrong.
What follows and what does not
None of this predicts a capability ceiling. Machines will continue improving at essentially everything measurable, and I expect to be wrong about specific timelines in the embarrassing direction. None of it is mystical either. I am not appealing to consciousness, souls or Gödel. Everything above is mechanism.
The claim is narrower and I think more durable. "Smarter" is a comparison that requires a shared purpose, and there is not one. Natural intelligence is what a system does when it has to keep itself alive, generates its own goals as a consequence, and finds relevance by having something at stake. Artificial intelligence is what a system does when it optimizes objectives it was handed, with nothing at stake, in a world it is not in. These are different kinds of thing, and one can exceed the other at every task on a list without ever occupying the position the other occupies.
There is a practical edge to this, which is why I care about it while building robots. If the difference is one of degree, safety is a capability problem and the answer is to keep improving the model until it is reliable enough. If the difference is one of kind, then no amount of capability produces a system with skin in the game, and the stakes have to be supplied externally through monitors, constraints, and humans who remain accountable for the outcome. That is an engineering commitment, and it is the one I would rather build on.