Huberman LabHow to Improve at Learning Using Neuroscience & AI | Dr. Terry Sejnowski
CHAPTERS
- 0:00 – 13:00
Introduction, Guest Background, And The Limits Of ‘Parts List’ Neuroscience
Huberman introduces Dr. Terry Sejnowski, outlining his role in computational neuroscience and AI. They frame the central question of how the brain actually works beyond naming structures, and why both bottom‑up (parts) and top‑down (behavior) approaches have been insufficient on their own. Sejnowski introduces the idea of an intermediate, algorithmic level as the key to linking brain implementation to behavior.
- •Huberman describes Sejnowski’s position at the Salk Institute and his pioneering role in computational neurobiology and AI.
- •The traditional ‘parts list’ view (neurons, synapses, brain regions) fails to explain consciousness and complex behavior.
- •Bottom‑up reductionism and top‑down behaviorism/AI programming both hit limits in explaining real intelligence.
- •Sejnowski highlights the ‘algorithmic level’—recipes or rules executed by circuits—as the crucial missing layer.
- •New methods in neuroscience and AI now allow discovery and testing of such algorithms in real neural circuits.
- 13:00 – 42:00
Basal Ganglia, Value Functions, And Dopamine As A Learning Algorithm
Sejnowski explains how the basal ganglia learn action sequences through reinforcement learning, driven by dopamine-based prediction error signals. This mechanism underlies not only motor skills like tennis but also complex thinking and social behavior. They discuss how rewards, punishments, and one‑trial learning shape a lifelong ‘value function’ that guides decisions in domains from dining to relationships.
- •Basal ganglia learn sequences of actions for goals (e.g., tennis serve, surgery, professional expertise).
- •These circuits form ‘go’ and ‘no‑go’ pathways for both actions and thoughts (in concert with prefrontal cortex).
- •Dopamine encodes reward prediction error—better or worse than expected—which updates synaptic weights.
- •Repeated experience builds a stable value function, encoding what’s good or bad for you across life domains.
- •Negative, high‑magnitude events (shocks, bad relationships, trauma) can produce one‑trial learning and PTSD.
- •The same reinforcement learning algorithm underlies AlphaGo’s success and the brain’s procedural learning.
- 42:00 – 55:00
Procedural Versus Cognitive Learning And Why Practice Still Matters
They contrast procedural learning (automatic skills) with cognitive learning (explicit knowledge) and argue that both are essential. Sejnowski criticizes educational moves to remove practice because it is stressful, noting that drills and problem sets are how the brain automatizes expertise. He and Huberman use examples from sports, physics, scuba diving, and school math to illustrate why practice is irreplaceable.
- •Procedural learning (basal ganglia) is needed for rapid, automatic execution; cognitive learning (cortex) for explicit reasoning.
- •Reading about tennis or flying can’t substitute for real practice; execution under time and stress is different.
- •School structure (lecture then homework) reflects this dual system: explanation then proceduralization via problems.
- •Practice and testing are not just evaluation—they are key mechanisms of learning and error‑driven refinement.
- •Educational trends that remove repetitive practice to avoid ‘stress’ undermine long‑term competence.
- •AI success (LLMs) relies on generalization, not rote memorization; likewise, brains must generalize beyond memorized facts.
- 55:00 – 1:10:00
Learning How To Learn: The MOOC And Practical Study Strategies
Sejnowski describes the free Coursera MOOC ‘Learning How to Learn’ he co‑created with Barbara Oakley. Aimed originally at students, it became massively popular among 25–35‑year‑olds already in the workforce who need to upskill. They discuss how active recall, dealing with procrastination and test anxiety, and understanding how the brain encodes information can make learning more efficient at any age.
- •‘Learning How to Learn’ is a free, global online course taken by ~4 million people, ages 10–90.
- •It emphasizes that brains require active engagement: self‑testing, problem solving, spaced repetition.
- •The course covers procrastination, exam anxiety, and practical techniques for retaining and applying knowledge.
- •Unexpectedly, the largest user group is working adults (25–35) needing to learn new skills without re‑enrolling in degrees.
- •Older learners can compensate for slower plasticity by using more efficient strategies the MOOC teaches.
- •The course is domain‑agnostic: it doesn’t teach content, but how to acquire and consolidate any content.
- 1:10:00 – 1:53:00
Brain‑Wide Connectivity, Sleep Spindles, And Memory Consolidation
The conversation turns to how real tasks engage widespread cortical networks and how new optical methods reveal global interactions. Sejnowski then details sleep spindles and hippocampal replay as mechanisms for safely integrating new memories into cortical knowledge. They discuss zolpidem’s enhancement of sleep spindles and memory, its tradeoffs, and the broader role of exercise and sleep stages in learning.
- •Modern recordings show tasks engage distributed cortical areas, not isolated ‘modules’.
- •Human intracranial recordings (e.g., epilepsy monitoring) reveal traveling waves and sleep dynamics at high resolution.
- •Sleep spindles during light non‑REM sleep are crucial for consolidating hippocampal memories into cortex.
- •Hippocampus replays daytime experiences, triggering spindles that ‘knead’ information into existing networks without overwriting.
- •Zolpidem doubles spindle counts and can double memory retention for pre‑drug learning but impairs formation of memories after ingestion.
- •REM sleep is likely particularly important for motor tuning and emotional/neural remodeling; cycles of REM and non‑REM are both needed.
- •Daytime exercise increases sleep quality and likely spindle expression, indirectly supporting consolidation.
- 1:53:00 – 2:31:00
AI As Social Partner, Idea Pump, And Future Predictor
They discuss how interaction style changes AI output: treating ChatGPT politely and conversationally reduces fatigue and improves usefulness, likely because it taps human social circuits. Sejnowski describes colleagues using LLMs as ‘idea pumps’ to suggest new neuroscience experiments. The pair speculate about using AI to simulate large‑scale clinical trials or future scenarios in fields like schizophrenia treatment and hurricane prediction.
- •A writer found ChatGPT less fatiguing and more productive when she treated it like a human collaborator.
- •Humans have well‑tuned circuits for social dialog; framing AI interaction that way recruits these circuits efficiently.
- •Researchers (e.g., Rusty Gage) feed their experimental data and literature into LLMs to generate novel experimental ideas.
- •AI can analyze large scientific corpora, critique statistical methods, and weight evidence across dozens of papers rapidly.
- •Weather models using AI (trained on simulations and historical data) can predict hurricane landfall more accurately in minutes.
- •Future applications could use AI to rapidly explore hypothetical clinical trials, metabolic interventions, or policy scenarios.
- 2:31:00 – 3:06:00
Mitochondria, Exercise, Aging, And Cognitive ‘Velocity’
Huberman raises the issue of declining energy with age, leading to a discussion of mitochondria as cellular power plants. Sejnowski explains how mitochondrial decline and some medications sap vigor, while exercise boosts mitochondrial function across tissues. They introduce the idea of ‘cognitive velocity’—the felt speed and depth of thinking—and link it to circadian rhythms, temperature, and deliberate interval‑type stress.
- •Mitochondria generate ATP and decline in number and efficiency with age and some drugs, reducing overall vigor.
- •Exercise upregulates mitochondrial biogenesis and function in muscles and brain, improving cognition and resilience.
- •Education level and lifelong cognitive engagement delay onset of Alzheimer’s symptoms—evidence for cognitive reserve.
- •Huberman describes ‘cognitive velocity’ as the sweet spot where reading/thinking is slightly challenging yet efficient.
- •Circadian phases, core body temperature, and transient stress (interval efforts) strongly modulate this cognitive window.
- •Short bouts of intense physical exertion (intervals) parallel mental intervals: controlled, time‑limited stress that drives adaptation.
- 3:06:00 – 3:24:00
Neuromodulators, Ketamine, Schizophrenia, And Depression
The discussion shifts to psychiatric disorders, focusing on ketamine as a window into schizophrenia and depression mechanisms. Recreational ketamine at high/frequent doses can induce transient, full‑blown psychosis by disrupting NMDA‑mediated inhibition and causing cortical overexcitation. At lower, controlled doses, the same mechanism can ‘correct’ underactive circuits in severe depression, illustrating how neuromodulatory balance is central and context‑dependent.
- •Schizophrenia likely begins early in development; late‑adolescent onset reflects long‑running circuit and genetic issues.
- •Ketamine is an NMDA receptor antagonist used as an anesthetic and party drug (Special K).
- •Back‑to‑back recreational ketamine use can produce psychotic episodes indistinguishable from schizophrenic breaks, which then resolve—implicating glutamate/NMDA and inhibition in psychosis.
- •Chronic schizophrenia research has shifted from pure dopamine hypotheses toward glutamate/excitation–inhibition imbalance models.
- •Low, intermittent ketamine dosing can relieve treatment‑resistant depression by boosting hypoactive cortical circuits, but must be carefully titrated.
- •AI models, had they existed in the 1990s with present capacities, might have linked ketamine, NMDA, and psychosis far earlier.
- 3:24:00 – 4:34:00
Parkinson’s Disease, Dopamine Loss, And Altered Internal ‘Set Points’
Sejnowski explains how Parkinson’s disease, driven by degeneration of dopamine neurons, impairs procedural learning and movement. Patients become extremely slowed or locked in, yet often perceive themselves as moving normally, revealing shifted internal references for speed and effort. L‑DOPA therapy can be dramatically restorative but also illustrates the difficulty of replacing neuromodulators precisely.
- •Dopamine neurons in the brainstem, vital for basal ganglia learning, are vulnerable to toxins like pesticides.
- •Severe dopamine loss induces profound bradykinesia or ‘locked‑in’ states, where patients barely move.
- •Subjectively, some Parkinson’s patients think they are moving reasonably or even fast despite observable slowness—indicating shifted set points.
- •As dopaminergic tone declines, movement thresholds shift; eventually not moving can feel ‘normal’.
- •L‑DOPA (dopamine precursor) revealed dopamine’s role dramatically: comatose or nearly frozen patients regained speech and movement.
- •Symptoms and treatment responses highlight how motivation, movement, and perceived effort are algorithmically linked via dopamine.
- 4:34:00 – 5:10:00
Mind Wandering, Dreams, And Internal Generative Activity
They discuss the functional value of mind wandering and sleep for insight. Sejnowski notes that many people solve problems after sleeping on them, reflecting overnight reprocessing. He distinguishes REM from slow‑wave dreams and mentions that certain drugs and states (e.g., cannabis withdrawal) affect dream patterns, though a complete theory of dreaming is still lacking.
- •Unstructured walks or runs without media provide ideal conditions for mind wandering and creative insight.
- •Sejnowski often gets his best ideas during such exercise, then now uses notes/voice memos to capture them.
- •‘Learning How to Learn’ explicitly advises stepping away from hard problems; diffuse thinking helps solutions emerge.
- •Thinking about a problem before sleep often leads to waking with clarity, likely via overnight hippocampal–cortical processing.
- •REM dreams are vivid, rapidly changing, and associated with acetylcholine in sensory cortex but low prefrontal engagement.
- •Slow‑wave sleep dreams are often repetitive and emotionally intense; they may relate to different aspects of emotional/memory processing.
- •Cannabis use suppresses REM; cessation produces REM rebound with vivid dreams, underscoring sleep’s sensitivity to substances.
- 5:10:00
AI Limitations, Self‑Generated Thought, And Future Brain–AI Research
In closing, Sejnowski points out that unlike humans, current LLMs do not yet have continuous self‑generated thought when not prompted. He describes his NIH Pioneer proposal to study temporal context and self‑attention in both brains and transformers, hypothesizing that basal ganglia and traveling cortical waves could implement something analogous. The episode ends with mutual appreciation and a call for scientists to communicate clearly with the public.
- •Humans generate ongoing internal activity and thought even in silence; current LLMs go ‘blank’ between queries.
- •LLMs show surprising in‑context learning (improving over a conversation without weight updates), mirroring unknown human mechanisms of long‑term working memory.
- •Transformers’ self‑attention mechanisms capture temporal context—linking distant words/concepts in a sequence.
- •Sejnowski’s proposed work aims to uncover how the brain implements self‑attention and temporal context, possibly via basal ganglia loops and cortical traveling waves.
- •He likens the current stage of AI to the Wright brothers’ first flight: off the ground, but control and long‑range direction are early and evolving.
- •Science is inherently social; AI should augment, not replace, collaborative human inquiry and clear public communication.
