Dr Terry Sejnowski How To Improve At Learning Using Neuroscience Ai

High-Level Overview

Andrew Huberman talks with Dr. Terry Sejnowski, a professor at the Salk Institute for Biological Studies who directs its computational neurobiology laboratory and has helped pioneer both neural network AI and the math of how the brain learns. Their conversation ranges from the single learning rule that drives motivation, to why we need both knowledge and practice, how sleep spindles lock in memories, a free online course on learning how to learn taken by millions, how exercise recharges the brain's energy supply, and how large language models can act as idea generators, diagnostic partners and even a kind of window into possible futures.

The tone is curious, friendly and fast-moving, with two longtime colleagues trading stories and speculation. Huberman is openly excited about AI by the end, and Sejnowski is candid about how much neuroscience still doesn't know. The discussion moves back and forth between brain algorithms, practical learning tools and big-picture AI, touching on disease, drugs, dreams and free will along the way.

The Discussion Explores:

• The algorithmic level of the brain: why neither bottom-up parts lists nor top-down behavior studies explain the brain, and what algorithms add.
• The basal ganglia and the value function: how predicting the next reward drives learning of actions and thoughts, and why punishment teaches faster.
• Cognitive versus procedural learning: why you need both lectures and problem sets, and why active learning and generalization matter.
• Sleep spindles and memory: traveling waves during sleep that consolidate the day's learning, and what the sleep drug Ambien reveals.
• Learning How to Learn: a free online course with Barbara Oakley, and why adults aged 25 to 35 use it most.
• Energy, aging and exercise: mitochondria, short intense intervals, education as cognitive reserve, and Huberman's idea of cognitive velocity.
• AI as a partner: treating chatbots like people, using them as an idea pump, AI plus doctors in diagnosis, and predicting hurricanes.
• Disease, dreams and the frontier: Parkinson's, ketamine and schizophrenia, dreams, free will, and self-attention in brains and transformers.

Deep Dive into Main Ideas

Understanding the brain at the level of algorithms

Huberman opens by asking how knowing the brain's parts list could ever tell us how the brain works. Sejnowski, whose earlier book with Patricia Churchland laid out a diagram of levels spanning ten orders of magnitude from molecules to the whole nervous system, says scientists have mostly tried two approaches. The bottom-up, reductionist approach makes progress but gets lost in the forest; the top-down approach of psychologists and early AI studied behavior but never reached the core. What's emerging now, in neuroscience and AI simultaneously, is a middle level: algorithms, which he compares to recipes that specify ingredients, order and timing.

His flagship example comes from the 1990s work with postdocs Peter Dayan and Read Montague on the basal ganglia, which learn sequences of actions toward a goal, such as a tennis serve, and also interact with the prefrontal cortex to shape thinking. The algorithm, he says, is remarkably simple: "It's simply to predict the next reward you're going to get." Each time you act, you compare the reward you got with what you expected and update your synapses, gradually building a value function, your accumulated sense of what is good or bad for you, like knowing what to order at a restaurant. This temporal difference learning is the same algorithm DeepMind used in AlphaGo, which beat the world champion at Go, a game far more complex than chess.

Reward, punishment and the value function

Huberman asks whether that value function persists across years and applies to complex domains like friendships and romance. Sejnowski says it is updated every time and becomes a very permanent part of who you are. Small rewards gradually move you closer to a goal, but "the negative punishment is much more effective, one trial learning": a shocked rat never forgets, one terrible relationship teaches lasting lessons, and PTSD shows how a single experience can shape a lifetime. Much of the prefrontal cortex is devoted to social interaction, and we arrive not knowing our language or culture, so we must learn them through this value function.

On Huberman's question about go and no-go signals, Sejnowski notes that the prefrontal loop with the basal ganglia is among the last to mature in early adulthood, which is why adolescents often fail to stop themselves from doing things that aren't in their best interest.

Why you need both knowledge and practice

Sejnowski distinguishes cognitive learning, which is cortical, step by step and slower, from procedural learning, which is subcortical, runs through the basal ganglia and becomes automatic. You can't learn tennis from a book, and as a physics PhD he found that working through problem sets, not memorizing equations, is what made him a physicist. Huberman adds examples like clearing a flooded scuba mask, which you can only learn by doing, and argues that knowing the mechanism behind a protocol and practicing it reinforce each other. Sejnowski agrees the two systems help each other, and criticizes a push in California schools to reduce practice so students don't feel stress.

The key, he says, is that "it has to be active learning": wrestling with a problem yourself engages trial and error, whereas simply being told the answer stores a fact that won't come up when a similar but different problem appears. This is also why today's large language models work: they aren't parrots that memorize, they generalize to new cases. "The brain is really, really good at generalizing", often from a single example, like learning how restaurants work after one visit. He adds that new optical recording methods show that even simple tasks engage many brain areas at once; in mice, the visual cortex gets more input from the motor system than from the eyes.

Sleep spindles, memory and the cost of shortcuts

Working with epilepsy patients at Massachusetts General Hospital, who have electrodes implanted for about two weeks to locate seizure origins, Sejnowski's group discovered circular traveling waves across the cortex during non-REM sleep. These sleep spindles, lasting a second or two, are important for consolidating the day's experiences. The hippocampus replays experiences and the spindles weave them into the cortex without overwriting existing knowledge. Huberman recalls from earlier guests that getting enough sleep, and daytime exercise, supports spindles; Sejnowski adds that REM sleep may help tune up the motor system.

He describes Sara Mednick's experiment at UC Irvine: the sleep drug zolpidem (Ambien) doubles sleep spindles, and people who learned material and then took it remembered about twice as much the next morning. The catch is that it also wipes out memories of what happens after taking it, which is why travelers sometimes wake with no memory of reaching their hotel. Both conclude you never get anything for free; Huberman extends the point to growth hormone, testosterone and stimulants, saying behaviors will always prevail as tools. On psychedelics, Huberman notes that psilocybin increases brain-wide connectivity, and Sejnowski explains that development runs the other way: infants form a huge surplus of synapses and then prune them because synapses are energetically expensive, so aging tends to reduce connectivity while old memories remain rock solid as a foundation.

Learning How to Learn

Sejnowski and engineer-turned-educator Barbara Oakley, whom he calls the mastermind, created Learning How to Learn, a free massive open online course taken by about 4 million people in 200 countries, ages 10 to 90, with roughly 98 percent approval. It consists of 50 or 60 bite-sized videos of about 10 minutes each over a month, with quizzes, tests, forums and teaching assistants. It doesn't teach subject matter; it teaches how to acquire knowledge, handle exam anxiety and avoid procrastination.

Surprisingly, the peak audience isn't students but people aged 25 to 35 who already went to college, are in the workforce, may have mortgages or children, and need new skills but can't return to school. They notice they're not quite as agile at learning, and the course helps them learn more efficiently. Huberman, who covered learning in a prior episode, highlights the value of self-testing, not just as evaluation but as a way to find and correct errors, and pledges to take the course himself.

Energy, exercise and cognitive velocity

Huberman shares that doing anything on his phone drains him in a way reading a paper book or writing on a computer doesn't, and wonders why. Sejnowski suggests that skills learned young become procedural foundations, while Huberman must use effortful cognition for social media. He then relays a New York Times account of a technical writer who spent a month working with ChatGPT and was exhausted until she started treating it politely, like a human assistant. "You treat it like a machine. It will treat you like a machine." Once she treated it like a person, she got better answers and stopped feeling fatigued, because she was using brain circuits honed by a lifetime of human interaction.

On why energy declines with age, recalling a question from the late Ben Barres, Sejnowski points to mitochondria, which become fewer and less efficient with age and can be harmed by some medications. The good news: "That exercise is the best drug you could ever take", benefiting the heart, brain, immune system and every organ. He runs on the beach below the Salk Institute every day, climbing back up a 340-foot cliff, and does intervals of about 10 seconds of all-out running followed by jogging, explaining that brief, controlled stress is good for body and brain. He also describes a study in China in which Alzheimer's began earliest in people with no education and latest in those with the most, suggesting education may build reserve. Huberman describes his practice of cognitive velocity: deliberately reading or listening slightly faster than his default pace while still retaining information, an idea he learned from skateboarding, where tricks are easier at a bit more speed. He worries that passive scrolling trains the brain to be slow, and Sejnowski likens Huberman's reading intervals to his own running intervals.

AI as an idea pump, a diagnostic partner and a window into the future

Huberman says he likes Claude for its aesthetic and bullet-pointed answers, with no relationship to the company, and asks how people should start with AI. Sejnowski says to try several chatbots and see what feels comfortable, noting that some are better at math; he says Google's Gemini went from 20 to 80 percent correct on math problems after fine-tuning with chain-of-reasoning. Huberman points out it's unfair to compare these models with the best humans rather than average ones. His colleague Rusty Gage, who discovered adult neurogenesis in the hippocampus and co-authored work with Sejnowski showing exercise increases it, now feeds his lab's experiments and the literature into a language model and says "we use it as an idea pump". Huberman describes a tool, shown to him by a Google engineer, into which you drop PDFs and websites and then ask questions that it answers with citations, even critiquing statistics and weighing evidence.

Sejnowski cites a dermatology study in Nature in which expert doctors and an AI each classified skin lesions with about 90 percent accuracy, but doctors using the AI reached 98 percent, because the AI knew rare lesions while doctors knew common ones in depth. "It's not going to be adversarial. It's going to be a partnership." AI trained on past hurricanes and simulations now predicts landfall more accurately on a laptop in 10 minutes than supercomputer simulations that take days. Huberman grows excited about sending AI models into imagined futures to explore scenarios no panel of humans could, and Sejnowski recalls how he and Geoffrey Hinton, a Nobel laureate and close friend, developed the Boltzmann machine in the 1980s when neural networks were dismissed, like small furry mammals under the feet of dinosaurs. He compares AI today to the Wright brothers' first flight: off the ground, but with control still the hard problem.

Schizophrenia, ketamine and Parkinson's

Sejnowski believes language models could have flagged ketamine's relevance to schizophrenia decades ago. Taken on two consecutive days at parties, ketamine can send young adults to the emergency room with full-blown psychosis indistinguishable from a schizophrenic break, which resolves after a couple of days in isolation. It blocks NMDA glutamate receptors and weakens inhibitory circuits, tipping the cortex toward over-excitation, supporting a glutamate and excitation-inhibition hypothesis of schizophrenia. The same mechanism may explain its help for depression, which involves under-excitation in parts of the cortex, with doses roughly every three weeks. Huberman mentions Chris Palmer's work on metabolic psychiatry and ketogenic diets.

In Parkinson's disease, dopamine neurons in the brainstem, which carry the reward prediction signal, die, sometimes after environmental insults such as pesticides. Before L-DOPA, some patients became locked in and motionless, and the first treatments seemed magical as they began talking again. Strikingly, patients who move slowly perceive themselves as moving fast, suggesting a shifted set point, which fascinates Huberman in relation to cognitive velocity. They discuss how body temperature may shape energy across the day, citing Craig Heller's work on temperature and muscle failure, and how stimulants like amphetamines create a proportional crash; Huberman wonders what the long-term effects are for a generation of children raised on stimulant medications.

Free will, understanding and minds that keep running

Sejnowski calls free will, consciousness, intelligence and understanding "weasel words because you can't pin them down", noting that 500 years ago people thought in terms of fate. The debate over whether language models understand language reveals that "we don't understand what understanding is." He has reviewed a paper that fine-tuned models on data from people with different disorders, including sociopathy, and got them to behave like those people. Models have no single persona; he once asked one to explain a technical abstract to a ten-year-old and it did better than he could. The biggest difference from humans, he says, is that a model goes blank between conversations, while the human brain constantly generates its own thoughts. Models also show in-context learning, getting better over a conversation without any plasticity, which remains a mystery, as does how humans hold an hour-long lecture in mind.

His own best ideas come while walking or jogging without music or podcasts, so he now records voice memos. Psychologists call this mind-wandering, and the course teaches students stuck on a problem to walk away, do the dishes or walk around the block. His other tip: "the key is to think about it before you go to sleep", rather than watching TV, because the brain works on it overnight. He doesn't record dreams, explaining that during REM sleep acetylcholine rises but not in the prefrontal cortex, so visual activity goes unmonitored and becomes bizarre. REM dreams are vivid and constantly changing, while slow-wave sleep dreams tend to repeat night after night with heavy emotional content.

The next frontier: self-attention in brains and transformers

Asked what he'd pursue with two more years, Sejnowski describes his NIH Pioneer Award proposal on temporal context in brains and transformers. Transformers, the architecture behind ChatGPT, use self-attention to link words far apart in a sequence and resolve ambiguity from context. Trained only to predict the next word, they build internal semantic models of meaning. His hypothesis is that the brain does self-attention through the basal ganglia, with traveling waves as another piece of the puzzle, and that neuroscience must now integrate brain areas studied separately. He closes by reminding listeners that "science is a social activity" in which everyone makes mistakes and learns from them, and praises Huberman for bringing science to the public.

Conclusion

Terry Sejnowski offers a unifying picture: the brain learns by constantly predicting rewards and updating its value function, by pairing knowledge with hands-on practice, and by consolidating experience during sleep. From that come practical takeaways: learn actively and test yourself, take the free Learning How to Learn course, sleep on hard problems, let your mind wander, and protect your brain's energy through exercise, especially short bursts of intense effort. On AI, he is neither doom-laden nor naive, seeing large language models as powerful partners that generalize, generate ideas and complement human experts, while stressing how much remains unknown about understanding, consciousness and the brain itself. It's an energizing conversation for anyone who wants to learn better and think clearly about what AI can offer.

This is a summary of "Dr Terry Sejnowski How To Improve At Learning Using Neuroscience Ai" from Huberman Lab, giving you the GIST so you can decide if you want the REST. Listen to the full episode.