How to Improve at Learning Using Neuroscience & AI | Dr. Terry Sejnowski

Huberman opens by conceding a limit that most popular neuroscience glosses over: knowing the inventory of the brain does not tell you how it works. Cells, neurons, and their specific wiring explain seeing, hearing, and thinking only in the sense of naming components; they leave the central question untouched. His framing question is deliberately concrete — what is this piece of meat in our heads actually trying to accomplish between waking up groggy and the first cup of coffee, or the first trip to the bathroom? The implied claim is that a genuine explanation has to survive at the level of ordinary behavior, not just at the level of anatomy.

Sejnowski answers with the structure of his book with Pat Churchland, Computational Brain, and its diagram of levels of investigation spanning roughly ten orders of magnitude in spatial scale: molecules, synapses, neurons, circuits, cortical areas, and the whole central nervous system. Consciousness has to be located somewhere in that stack, and he argues that neither of the two standard strategies has succeeded. The bottom-up, reductionist program produces real progress on connectivity and development but, in his phrase, "quickly you get lost in the forest." The top-down program — behaviorist laws of behavior, or early AI programs written to replicate intelligent behavior — is described as successful in places but ultimately unsatisfying.

The proposal is an intermediate level: the algorithmic level, sitting between the implementation of a mechanism and the behavior of the whole system. His analogy is a recipe. Baking a cake requires ingredients, an order of assembly, and timing; get those wrong and "it's just a mess." The analogy matters because it locates explanation in a specific kind of object — a procedure with a structure — rather than in either a parts catalog or a behavioral law. He claims this is where progress is now being made simultaneously in neuroscience and AI, which he calls remarkable for this moment in history.

Notably, the disagreement in this opening stretch is not between the two speakers but between two research cultures, and Sejnowski explicitly widens the target from neuroscientists to scientists generally. That is a small rhetorical move with a substantive point: the failure of bottom-up and top-down approaches is not a failure of biology's tools but of the level of description at which both traditions chose to work.

The concrete case is work from the 1990s, when Peter Dayan and Reed Montague were postdocs in Sejnowski's lab, on the basal ganglia — the structure below the cortex responsible for learning sequences of actions aimed at a goal. Tennis serves and thinking alike are framed as practiced sequences: the basal ganglia take over from the cortex and produce actions that get better and better. Crucially, the loop is not only with the motor cortex at the back but with the prefrontal cortex, "which is the thinking part." That anatomical detail is what licenses extending the account from movement to professional judgment — the financier, the doctor, the neuroscientist all improve by finding out what works.

Huberman then presses on a distinction the standard summary often omits: the basal ganglia organize both go and no-go behaviors. Learning an expert golf or tennis swing requires performing some actions and suppressing others. If the same circuitry generates thoughts, he asks, does it also suppress them? His example is a surgeon who must think about what to think about and also about what not to think about, because a frustrating morning with a child could otherwise intrude on a delicate procedure.

Sejnowski accepts the framing and adds a developmental caveat: the prefrontal loop with the basal ganglia is among the last to mature in early adulthood, so in adolescents "it's not the no-go part for planning an action" that is fully online. The result is that the brake often fails to engage against choices that are not in the person's interest. This is a claim about a specific circuit maturing late, not a general statement that adolescents cannot plan.

The algorithm itself is presented as the simplest imaginable: predict the next reward. After each attempt, compare the reward received with the reward expected, and use the discrepancy to update synapses — synaptic plasticity — so the next attempt has a better chance. Over a lifetime this accumulates into a value function: knowledge about what is good and bad for you, built from meals in many restaurants, among other things. The identification with AI is explicit: "This is the same algorithm that was used by AlphaGo," the DeepMind program that beat the world Go champion.

Huberman supplies the difficulty ranking that makes the AlphaGo comparison pointed, calling Go far more complex than chess, and Sejnowski endorses it with a scaling analogy: "Go is to chess what chess is to something like checkers." The reason given is structural rather than merely quantitative — multiple simultaneous local battles, and an order of placement that constrains what becomes possible later. The relevance to the value-function argument is that a single predictive rule can operate in a space where exhaustive lookahead is hopeless.

Huberman tests the account in two directions. First, durability: if someone returns to the tennis court once a month across years, can they still draw on the same value function despite the intervening time? Second, scope: does the same mechanism operate in domains that are not motor skills, such as friendship, sibling relationships, and romantic partnership, where people get some things right and some things wrong? He also flags a gap in the description so far — punishment. "We don't get an electric shock when we get the serve wrong, but we can be frustrated."

Sejnowski's answer is that the value function is updated on every occasion and accumulates, and that it is permanent: "It doesn't matter. It's a very permanent part of your experience and who you are." He then grants punishment a larger role than reward, attributing the point to behaviorists in the 1950s. Small rewards work because they move you steadily closer to what you seek — a better serve, a new friend. But negative punishment is "much more effective," producing one-trial learning: a rat trained with small food rewards needs many trials, whereas a rat that is shocked does not forget.

Huberman's gloss — one bad relationship teaching certain lessons permanently — is extended by Sejnowski to post-traumatic stress disorder, which he describes as screwing a person up "for the rest of your life." The move is worth noticing analytically: the same one-trial aversive learning that makes punishment efficient is also the mechanism behind a clinical condition. Efficiency and pathology are not opposite findings here; they are the same property viewed at different intensities.

The social dimension is then folded in. A large part of the prefrontal cortex is devoted to social interactions, which is how humans acquire language, cultural values, and the expectations of a society they are born into without knowing which one it will be. Sejnowski notes this was a 20th-century discovery, that it now appears in AI as reinforcement learning, and that it is procedural rather than cognitive: "With procedural learning, it's automatic." Cognitive thinking is described as much less efficient because it must proceed step by step.

Asked for a coffee-making comparison between knowledge-based and procedural learning, Sejnowski instead offers the sharper case: can you imagine learning tennis from a book? Huberman's airline anecdote — a passenger studying diagrams of aircraft flight, and the relief that the man is not the pilot — sets up the point that pilot training is necessarily a mixture of practical and textbook learning. His scuba example is more precise: dive tables and gas exchange can be studied, but removing a mask underwater, replacing it, and clearing it cannot be simulated away; it has to be done in a pool, and done when it is needed.

Sejnowski generalizes: procedural fluency is essential for anything that must be executed quickly and expertly "so you don't have to think." School already contains this structure, with classroom instruction followed by homework problems, and he draws on his own training as a PhD physicist — memorizing equations is not the same as understanding how to use them, and the problem sets were the core of becoming a physicist.

Huberman then reframes his usual protocol talk as an instance of the same integration. Knowing that morning sunlight reaches the suprachiasmatic nucleus via retinal ganglion cells, and also actually getting the light, are not separate achievements; linking explicit knowledge to a practice reinforces both. Sejnowski agrees without reservation: "You need both." His formulation is that there are two major learning systems, a cortical cognitive one and a subcortical procedural one centered on the basal ganglia, and that "the two go hand in hand."

This is where an actual disagreement surfaces, and it is about institutions rather than mechanism. Sejnowski says that in California, at least, schools are trying to get rid of the procedural component, on the grounds that practice stresses students and that they should not be made to feel they are having difficulty. Huberman reacts bluntly — "That's ridiculous" — and covers his eyes on camera while offering the swimming analogy: a textbook on swimming, then the ocean, with no actual swimming in between, and survival somehow expected. Sejnowski's verdict is "It's crazy."

The evidence offered for the alternative is the MOOC Sejnowski built with Barbara Oakley, Learning How to Learn. It is free, with no paywall, aimed at students but taken by four million people in 200 countries between the ages of 10 and 90, and he reports fan letters almost daily. The scale is notable because it undercuts the assumption that the demand for instruction in learning technique is confined to formal schooling.

Huberman brings in the testing effect: self-testing is not merely evaluation but a way of identifying errors so they can be compensated for. Sejnowski's confirmation sharpens it into a claim about how the brain stores anything at all. The brain "doesn't memorize things like a computer"; it requires active engagement. Solving a problem yourself is where trial-and-error learning happens, in the procedural system. Being told the right answer stores a fact somewhere, but that fact will not automatically surface when a similar-but-not-identical problem appears.