YC Decoded · World Models · Jul 17, 2026

The Key Thing Human Brains Have That AI Is Trying To Learn

Ankit and Francois map the gap between pattern-matching and genuine understanding — and why closing it may require machines to dream.

Watch

▶

A 74-minute episode of Decoded in which Ankit Jain and a guest work through the mathematics and motivation of world models — why they may be the missing key to sample-efficient AI.

Why this belongs in the latticework

Every latticework needs a model for how learning works. The dominant one — that intelligence scales with data volume — has powered a decade of progress. But Ankit opens this episode by naming the hard constraint that scaling alone doesn't touch: sample efficiency. The question isn't how many parameters fit on a chip; it's how much a system learns per additional data point.

This matters for the latticework because it shifts the framing from accumulation to leverage. A world model is a lever: instead of requiring ten thousand real interactions, a brain (or an AGI) simulates them internally. The transcript makes this vivid with the asteroid example — humanity's planetary-defense systems need zero new observations to plot an intercept trajectory because the physics world model is already precise enough to plan decades ahead.

The map-vs-territory distinction runs beneath the whole episode. The hosts are asking: does current AI have a map of how the world works, or does it only have a lookup table of stimulus-response pairs? Their answer — backed by neuroscience and by a basketball study that should startle anyone who builds learning systems — is that the brain's map is so faithful it can substitute for reality. That is the key thing human brains have that AI is trying to learn.

Models the source amplifies

The episode's richest gift to the latticework is empirical evidence for mental simulation. The basketball study Ankit cites splits subjects into three groups: physical practice, imagined practice, and control. Physical practice yields a 24% improvement; imagined practice yields 23%. The gap is noise. For a latticework reader, this isn't a curiosity — it's a proof that the brain's internal forward model is accurate enough to drive the motor-learning system without ever touching a ball.

Second-order thinking is the cognitive move the hosts keep returning to without naming it. First-order AI predicts the next action. Second-order AI predicts the next state of the world, then derives the action. The difference sounds minor; it isn't. Planning a spacecraft intercept requires simulating dozens of future states before committing to a burn. The same move — simulate forward, then choose — is what allows humans to decide without experience. Francois calls it the crux of the whole problem.

Leverage shows up here too. A world model is not just a better map; it's a multiplier on every future experience. A system that can simulate 1,000 trajectories for the cost of one real interaction has a 1,000× information advantage. The hosts frame AGI not as a larger model but as one that has internalized this leverage — that can imagine its way to skill rather than grind its way there.

Models the source troubles

The episode quietly dismantles the strongest form of the scaling hypothesis — the belief that sufficient data plus sufficient compute will produce human-level general intelligence. The hosts don't dispute that scale has been transformative. What they dispute is its sufficiency. Their argument is structural: evolution didn't make the neocortex bigger just to memorize more patterns. It made it bigger to simulate. If the neocortex's purpose is forward modeling, then a system built only on pattern retrieval is architecturally incomplete, no matter how large it gets.

The episode also contradicts the behaviorist reading of intelligence — the stimulus-response account that dominated psychology for decades and still lurks in how some practitioners think about reward-based AI. The basketball experiment is the cleanest refutation available: if intelligence were purely a matter of action-outcome conditioning, imagined practice couldn't work. The gain only makes sense if the brain has an internal model that the imagination can drive. S-R alone doesn't predict 23%.

There's a subtler contradiction too. Many product teams use complexity bias when designing AI systems — piling in more inputs, more parameters, more training tasks — when the episode argues the missing ingredient isn't complexity but a different kind of structure entirely. A smaller system with a good world model may outlearn a larger system without one. Bigness is not a substitute for architecture.

Models worth adding

The episode introduces a unit that deserves a permanent place in any thinker's vocabulary: intelligence per sample. Defined explicitly as "how much smarter does the system get for each additional data point," it gives practitioners a clean ratio to optimize — one that captures what scaling models tend to obscure. A system doubling in size but flat on intelligence-per-sample is burning resources, not gaining capability. The ratio also makes it obvious why the human brain is extraordinary: we learn new games, new languages, new motor skills from a handful of examples. Our intelligence-per-sample is many orders of magnitude higher than our best models.

The hosts introduce the VAS triplet — Value, Action, State — as the minimum viable ontology for world-model-based reinforcement learning. Rather than predicting a reward directly (the classic RL move), a VAS system predicts the next state, derives a value for that state, and only then chooses an action. The extra step is costly at inference time but pays enormous dividends in sample efficiency: the system can plan over imagined trajectories instead of waiting for real feedback. It's a concrete architectural instantiation of second-order thinking.

The deepest new model arrives in the final minutes: the brain as optimizer, not just model. Ankit articulates a view in which the neocortex doesn't merely house a world model — it is the optimizer, with world models as tools it invokes and discards. This inversion matters: it suggests that AGI won't be achieved by finding the right architecture for a world model, but by finding the right architecture for a system that builds, runs, and replaces world models dynamically. The optimizer is the brain; the models are expendable.