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

概览

This episode features Andrew Huberman’s conversation with Dr. Terry Sejnowski, a computational neuroscientist at the Salk Institute, about how algorithms can help explain brain function. The discussion centers on an intermediate “algorithmic level” between neurons and behavior, where learning rules, prediction, reward, and error correction connect biology to cognition.

A major thread is learning: the basal ganglia, dopamine, reward prediction, punishment, procedural practice, sleep spindles, and memory consolidation. Sejnowski argues that cognitive learning and procedural learning work together, and that active practice, testing, sleep, and exercise are essential for durable learning.

The conversation then expands into AI, large language models, medicine, psychiatry, Parkinson’s disease, energy, mitochondria, cognitive velocity, consciousness, dreams, and self-attention. Rather than treating AI as a replacement for human expertise, Sejnowski frames it as a tool that can partner with humans, generate hypotheses, analyze evidence, and perhaps help model future outcomes.

分段落总结

[00:00] Episode Framing and Guest Introduction

[事实] Huberman introduces Dr. Terry Sejnowski as a professor at the Salk Institute who directs the Computational Neurobiology Laboratory. [事实] The episode is framed around computational neuroscience, AI, algorithms, dopamine, motivation, learning, exercise, Parkinson’s disease, Alzheimer’s disease, and mitochondrial function. [事实] Huberman says the goal is to make computational neuroscience and AI accessible rather than intimidating.

[05:18] Why Algorithms Matter for Understanding the Brain

[事实] Sejnowski describes brain study across many spatial levels, from molecules and synapses to neurons, circuits, cortical areas, and the whole nervous system. [事实] He contrasts bottom-up reductionism with top-down behavioral approaches and says neither has fully answered the big questions. [事实] He proposes an intermediate “algorithmic level” as a useful way to connect biological implementation with behavior.

[10:19] Basal Ganglia, Practice, and Goal-Directed Action

[事实] Sejnowski gives the basal ganglia as an example of a brain system that learns sequences of actions to achieve goals. [事实] He says the basal ganglia help actions become better through repeated practice, such as learning a tennis serve. [事实] He extends this idea beyond movement, saying expertise in fields like medicine, finance, or neuroscience also requires repeated practice. [事实] The basal ganglia interact with motor regions and with prefrontal cortex, linking action learning with thinking.

[12:50] Go/No-Go Control and Adolescent Development

[事实] Huberman asks whether basal ganglia “go” and “no-go” functions apply to thoughts as well as actions. [事实] Sejnowski says the prefrontal cortex loop with the basal ganglia is among the last systems to mature in early adulthood. [事实] He says adolescents may lack fully developed no-go control for planning and actions, which can affect decisions that are not in their best interest. [推测] The discussion implies that inhibitory control over thoughts and actions depends on developmental maturation of frontal-basal ganglia loops.

[13:37] Reward Prediction, Dopamine, and Value Functions

[事实] Sejnowski says the brain uses a simple algorithm to learn action sequences: predicting the next reward. [事实] When an action produces more or less reward than expected, the brain updates synapses and builds a value function. [事实] He connects this learning rule to dopamine and to reinforcement learning in AI. [事实] He says AlphaGo used the same broad class of algorithm.

[16:47] Reward, Punishment, and Social Learning

[事实] Sejnowski says the value function is updated every time a person acts and accumulates over time. [事实] He says small rewards can shape learning gradually through trial and error. [事实] He says negative punishment can produce one-trial learning and gives traumatic experience and PTSD as examples. [事实] He says a large part of prefrontal cortex is devoted to social interactions and learning cultural expectations.

[19:13] Procedural Learning Versus Cognitive Learning

[事实] Sejnowski contrasts procedural learning with cognitive, step-by-step thinking. [事实] He says procedural learning is automatic and more efficient once learned. [事实] Huberman gives examples such as piloting, scuba diving, and practicing tasks that cannot be mastered only from books. [事实] Sejnowski says classroom instruction and homework map onto cognitive and procedural systems working together.

[22:26] Education, Practice, and Learning How to Learn

[事实] Sejnowski says the brain has two major learning systems: cortical cognitive learning and subcortical procedural learning involving basal ganglia. [事实] He criticizes attempts to reduce practice in schools, arguing that practice is necessary for learning. [事实] He and Barbara Oakley created a free online course called “Learning How to Learn.” [事实] He says the course has been taken by about 4 million people in 200 countries, across ages 10 to 90.

[24:08] Testing, Active Learning, and Generalization

[事实] Huberman says testing is not only evaluation but also helps identify errors that support learning. [事实] Sejnowski says the brain does not memorize like a computer and requires active engagement. [事实] He says solving problems independently engages trial-and-error procedural learning. [事实] He links this to AI, saying large language models are not merely parroting data but must generalize to new cases.

[26:40] Distributed Brain Activity and Human Recordings

[事实] Huberman asks where learned action functions are stored in the brain. [事实] Sejnowski says neuroscience does not yet know the answer. [事实] He says new optical recording methods show that tasks engage many brain areas at once, not only isolated modules. [事实] He describes collaborating with Mass General Hospital to study people with epilepsy who have invasive cortical recordings over weeks.

[30:15] Traveling Waves, Sleep Spindles, and Memory Consolidation

[事实] Sejnowski says studies in humans revealed circular traveling waves during sleep. [事实] He says these occur during non-REM transition states and are related to sleep spindles lasting about one or two seconds. [事实] He says sleep spindles are important for consolidating daytime experiences into long-term memory. [事实] He says the hippocampus replays experiences and helps integrate new information into cortex without overwriting existing knowledge.

[35:00] Sleep, Exercise, Ambien, and Memory Tradeoffs

[事实] Huberman says sufficient sleep is important for enough sleep spindles and memory consolidation. [事实] Sejnowski says sleep is not the brain turning off but the brain entering different states with important functions. [事实] He describes Sarah Mednick’s experiment in which zolpidem, also known as Ambien, doubled sleep spindles and improved recall of material learned before taking the drug. [事实] He says the downside is that experiences after taking the drug can be lost from memory. [推测] The example supports the broader point that pharmacological enhancement often involves tradeoffs.

[40:03] Pharmacology, Nootropics, and Biological Tradeoffs

[事实] Huberman and Sejnowski discuss the idea that changing one brain or body system can create costs elsewhere. [事实] Huberman mentions steroids, growth hormone, modafinil, stimulants, sedatives, and sleep drugs as examples where benefits may come with costs. [事实] Sejnowski says drugs can unbalance systems that evolved to balance many processes. [事实] Huberman states that behaviors will generally prevail as tools.

[42:14] Psychedelics, Connectivity, and Synaptic Pruning

[事实] Huberman asks whether increased brain-wide connectivity from psychedelics such as psilocybin is beneficial. [事实] Sejnowski says infant brains form many synapses in the first two years and then prune them. [事实] He says synapses are energetically expensive, so pruning reduces energy use and preserves useful connections. [事实] He says aging tends to involve loss of connectivity, while old memories remain strong because they were laid down earlier.

[46:08] Adult Learning and the MOOC Audience

[事实] Sejnowski says adults can still learn, though not always as quickly as younger people. [事实] He says the largest demographic for “Learning How to Learn” is people aged 25 to 35. [事实] He says many learners are already college educated and use the course to improve learning while working or managing adult responsibilities. [事实] The course uses short videos, quizzes, tests, forums, and teaching assistants.

[50:02] Technology, Fundamentals, and Intuition

[事实] Huberman asks how people should decide which old skills still need to be learned when calculators, Google Maps, and AI can perform many tasks. [事实] Sejnowski says technologies provide tools but do not remove the need for foundational understanding. [事实] He gives the example of students producing wildly wrong calculator answers because they lack number intuition. [事实] He says tools can make work faster and more accurate, but procedural foundations help preserve intuition.

[53:26] Generational Learning and Social Media

[事实] Huberman asks whether a child using AI to write a song should be considered a songwriter or musician. [事实] Sejnowski frames the issue as generational, saying younger people build foundations around new technologies earlier. [事实] He says people who learn phone-based communication young can become more procedurally fluent with it. [事实] Huberman says social media feels energetically draining to him, and Sejnowski suggests this may be because it was not part of his early procedural foundation.

[61:06] Interacting With AI Like a Human

[事实] Sejnowski describes a technical writer who found ChatGPT exhausting when treating it like a machine. [事实] He says she got better responses and felt less fatigued when she interacted politely, as if with a person. [事实] Sejnowski says ChatGPT mirrors the way it is treated. [事实] He argues that treating AI conversationally may draw on existing human social circuits and reduce cognitive effort.

[63:35] Online Interaction, Templates, and Energy

[事实] Huberman connects social media fatigue to whether online interactions match childhood-developed templates for communication. [事实] He says some online exchanges feel rewarding, while others feel grating because they do not allow full explanation or mutual understanding. [事实] The conversation shifts toward biological energy, vigor, aging, and mitochondria. [推测] The speakers suggest that mismatch between learned social routines and digital interaction formats may increase cognitive cost.

[69:00] Mitochondria, Aging, and Exercise

[事实] Sejnowski says mitochondria supply ATP, which cells use to operate their machinery. [事实] He says aging is associated with fewer and less efficient mitochondria. [事实] He says some drugs taken for illness can harm mitochondria. [事实] He says exercise can replenish energy and describes exercise as beneficial for every organ system, including heart, brain, and immune system.

[70:24] Exercise, Cognitive Reserve, and Alzheimer’s

[事实] Sejnowski says he runs on the beach near the Salk Institute and climbs back up the cliff steps. [事实] He describes exercise as building reserves of energy for later life. [事实] He cites a study in China where Alzheimer’s onset was earlier among people with little education and later among those with more education. [推测] He suggests education may build cognitive reserve that delays Alzheimer’s symptoms.

[72:20] Cognitive Velocity and Productive Stress

[事实] Huberman describes “cognitive velocity” as the pace at which he can read, listen, and retain information. [事实] He says he sometimes increases reading or audio speed slightly to reach a more engaged learning state. [事实] Sejnowski compares this to interval running and says transient, controllable stress can be good for brain and body. [事实] Sejnowski says pushing into an extra gear helps muscles learn what output is required.

[77:23] How to Start Using AI

[事实] Huberman asks whether young and older people should learn to use AI. [事实] Sejnowski says young people are already adopting AI and that AI is a tool people need to learn how to use. [事实] Huberman says he uses Claude AI and likes its style and formatting. [事实] Sejnowski says different chatbots may suit different people and some are better at specific tasks, such as math.

[79:34] Reasoning, AI Accuracy, and Human Comparison

[事实] Sejnowski says Google’s Gemini improved on math problems after fine-tuning with chain-of-reasoning methods. [事实] Huberman notes that people often compare AI systems to exceptional humans rather than average humans. [事实] He points out that ChatGPT has passed exams such as the LSAT and MCAT. [推测] The speakers treat AI performance as imperfect but already useful when judged against practical human baselines.

[82:22] Neural Networks, Hinton, and AI’s Alternative Path

[事实] Sejnowski recalls collaborating with Jeff Hinton on the Boltzmann machine in the 1980s. [事实] He says AI was then dominated by symbolic rules and hand-written programs. [事实] He and Hinton believed nature had solved problems such as vision, so studying biological algorithms could guide AI. [事实] Sejnowski says some learning algorithms are conserved across species, including flies and humans.

[87:06] AI as a Way to Explore Possible Futures

[事实] Huberman suggests AI could forage through knowledge at large scale and model future scenarios without human limits such as sleep. [事实] Sejnowski agrees the idea is promising but speculative. [事实] Huberman gives schizophrenia treatment as an example where AI might examine dopamine, glutamate, metabolism, ketogenic diet, and hypothetical clinical trials. [推测] The discussion frames AI as a tool for generating testable future-facing hypotheses rather than automatically determining truth.

[90:43] Ketamine, Schizophrenia, Depression, and Glutamate

[事实] Sejnowski says ketamine can temporarily induce symptoms resembling schizophrenic psychosis, including auditory hallucinations and paranoia. [事实] He says ketamine binds NMDA glutamate receptors and weakens inhibitory circuits, increasing cortical excitation. [事实] He describes a glutamate hypothesis of schizophrenia involving imbalance between excitatory and inhibitory cortical systems. [事实] He says ketamine may help depression because depression is associated with lower excitatory activity in some cortical regions, and carefully titrated ketamine may restore balance. [推测] His “touch of schizophrenia” phrasing implies ketamine’s therapeutic window depends on dose and context.

[97:42] LLMs as Research Idea Pumps

[事实] Sejnowski says his colleague Rusty Gage uses large language models as an “idea pump” for research. [事实] The models are given prior experiments and literature access, then asked to suggest new experiments. [事实] Huberman describes an AI tool that can ingest PDFs or URLs, answer questions from them, cite sources, compare rigor, critique statistics, and weigh evidence. [推测] The conversation suggests AI may accelerate literature synthesis and hypothesis generation, especially when paired with expert judgment.

[102:17] AI and Human Experts in Medicine

[事实] Sejnowski describes a dermatology study where expert doctors and AI each identified skin lesions with about 90% accuracy. [事实] He says doctors using AI improved to about 98% accuracy. [事实] He explains that AI may recognize rare cases from large datasets while doctors retain deep experience with common presentations. [事实] He argues AI should be viewed as a partner rather than a replacement for humans.

[104:21] Prediction, Hurricanes, Autism, and Early AI

[事实] Huberman asks whether AI can predict physical events such as traffic or accidents. [事实] Sejnowski says AI has already been used to predict hurricane landfall using prior hurricane data and simulations. [事实] He says trained AI can make such predictions much faster than traditional supercomputer simulations. [事实] Huberman asks whether AI could help clarify autism diagnosis increases, and Sejnowski says it might depend on data and disease complexity. [事实] Sejnowski compares today’s AI stage to the Wright brothers’ first flight: airborne, but still early and limited by control.

[110:05] Parkinson’s Disease and Dopamine Set Points

[事实] Huberman describes Parkinson’s disease as dopamine neuron depletion causing movement difficulty and cognitive or mood dysfunction. [事实] Sejnowski says Parkinson’s affects dopamine cells involved in procedural learning and global neuromodulation. [事实] He says these neurons are vulnerable to environmental insults, especially toxins such as pesticides. [事实] He says L-DOPA produced dramatic effects in people who had become unable to move or speak normally. [事实] He says Parkinson’s patients may feel they are moving quickly even when they are moving slowly, describing this as a set point issue.

[113:56] Brain States, Temperature, and Stimulants

[事实] Huberman says terms like focus, motivation, and flow are not precise scientific measures of waking brain states. [事实] He proposes cognitive velocity as one possible metric of brain state. [事实] Sejnowski says he gets most done in the morning and has another productive period after dinner. [事实] They discuss body temperature, circadian rhythms, enzyme rates, and possible links to cognition. [事实] They discuss amphetamines and prescription stimulants, noting that high activation can be followed by a low-energy crash.

[120:15] Free Will, Consciousness, and Undefined Concepts

[事实] Huberman asks about consciousness and free will. [事实] Sejnowski says words such as free will, consciousness, intelligence, and understanding are hard to pin down because they lack agreed definitions. [事实] He says it is difficult to solve a scientific problem without a definition people agree on. [事实] He says debates about whether large language models “understand” language reveal that humans do not yet have a clear account of understanding.

[124:18] LLM Personas, Diversity, and Social Interpretation

[事实] Huberman asks whether LLMs can be tuned toward different cognitive or behavioral profiles. [事实] Sejnowski says models have been fine-tuned with data from people with different disorders and can behave like those profiles. [事实] They discuss the possibility of tuning models for political leanings, values, creativity, or sensitivity to emotional tone. [事实] Sejnowski says words are only part of communication and that tone, facial expression, and visual information also carry meaning.

[129:09] Self-Generated Thought and In-Context Learning

[事实] Huberman describes an example where an LLM appeared to pause and look at landscape images during a task. [事实] Sejnowski says a key difference is that a human brain continues generating internal thought in a quiet room, while an LLM goes blank after interaction. [事实] He says neuroscience does not yet know how brains generate continuous activity that supports planning. [事实] He says LLMs show in-context learning despite no synaptic plasticity during inference, and that this remains mysterious.

[132:39] Walking, Mind Wandering, and Creativity

[事实] Sejnowski says he gets his best ideas while walking or jogging, not while sprinting. [事实] He does not listen to music or podcasts during these runs and instead listens to his own thoughts. [事实] Huberman relates this to “wordlessness” and low-input states. [事实] Sejnowski says psychologists study this as mind wandering, which often produces “aha” moments.

[136:03] Letting Problems Incubate

[事实] Sejnowski says “Learning How to Learn” teaches students to stop and do something else when stuck on a concept. [事实] He gives examples such as cleaning dishes or walking around the block before returning with a clearer mind. [事实] He says people often solve problems after sleeping on them. [事实] He advises thinking about a problem before sleep so the brain can work on it overnight.

[138:17] Dreams, REM, and Slow-Wave Sleep

[事实] Sejnowski says he does not record his dreams and says there is no complete theory of why we dream. [事实] He says neuromodulators are generally downregulated during sleep, while acetylcholine rises during REM sleep but not in prefrontal cortex. [事实] He says this may help explain why REM dreams are bizarre and visually vivid. [事实] Huberman says cannabis cessation can produce REM rebound and increased dreaming. [事实] Sejnowski says REM dreams are vivid and changing, while slow-wave sleep dream reports are often repetitive and emotionally heavy.

[143:28] Temporal Context, Transformers, and Self-Attention

[事实] Sejnowski says his NIH Pioneer Award proposal is titled “Temporal Context in Brains and Transformers.” [事实] He explains transformers as deep learning architectures with self-attention. [事实] He says self-attention helps connect distant words and resolve ambiguity through context. [事实] He says predicting the next word can lead transformers to build internal semantic representations. [事实] He hypothesizes that the brain’s version of self-attention may involve the basal ganglia.

[147:26] Putting Brain Systems Together

[事实] Sejnowski says traveling waves may help explain how cortex is organized and how it interacts with basal ganglia. [事实] He says neuroscientists have often studied brain regions independently. [事实] He wants to understand how different brain areas work together computationally. [事实] Huberman closes by thanking Sejnowski for his work on public education, learning, exercise, mitochondria, cognition, and longevity.

[150:00] Science as a Social Activity and Closing

[事实] Sejnowski says he has benefited from students and colleagues and describes science as a social activity. [事实] He says scientists learn from each other, make mistakes, and make progress through those mistakes. [事实] He praises Huberman’s ability to communicate science to the public. [事实] Huberman directs listeners to Sejnowski’s work, the learning portal, and his book through the show notes.

播客点评/总结

This episode is valuable because it connects brain science, AI, learning theory, sleep, exercise, and disease through one recurring idea: brains and machines can be understood partly through the algorithms they use to predict, update, and generalize. The strongest sections are the explanations of value functions, procedural learning, sleep spindles, and the practical importance of active learning.

A major highlight is the repeated bridge between neuroscience and AI. Sejnowski avoids treating AI as either magic or threat; instead, he emphasizes partnership, tool use, expert interpretation, and the possibility that AI can reveal useful analogies for brain function.

The episode is broad and occasionally speculative, especially when discussing AI’s ability to model future scientific outcomes, consciousness, free will, and cognitive velocity. Those parts are engaging but should be read as exploratory unless clearly tied to specific evidence in the transcript.

This episode is best suited for listeners interested in neuroscience, learning, AI, medicine, sleep, and practical cognitive performance. It is less suited for listeners looking for a narrow protocol-only episode, because much of the value comes from conceptual synthesis across fields.