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

Andrew Huberman and Terry Sejnowski explore how the brain works at the 'algorithmic' level between molecules and behavior, centering on the reinforcement-learning value function that governs all motivation via dopamine. They contrast cognitive vs. procedural learning and argue both are essential, warning against schools dropping practice-based learning. The conversation covers practical tools for better learning, sleep spindles and memory consolidation, exercise and mitochondria as the key to lifelong energy, and the surprising parallels between large language models and the human brain. They also dig into ketamine, schizophrenia, the glutamate hypothesis, Parkinson's, free will, consciousness, dreams, and how AI can act as an 'idea pump' to forage future scientific discoveries.

• Sejnowski says neuroscience now knows the actual brain algorithm for learning sequences of actions: simply predict the next reward (temporal difference learning) - the same algorithm that powered DeepMind's AlphaGo.
  00:13:37
• He criticizes California schools for trying to eliminate procedural (practice-based) learning, calling it 'ridiculous' and 'crazy.'
  00:22:31
• Ambien (zolpidem) doubles sleep spindles, and a study found it lets you remember twice as much of what you learned before taking it - but it wipes out memories formed after.
  00:37:02
• A China study found Alzheimer's onset was earliest in people with no education and latest in the most educated, suggesting cognitive reserve delays the disease.
  01:11:21
• Ketamine fights depression by inducing 'a touch of schizophrenia' - over-exciting an under-excited depressed cortex back into balance.
  01:35:12
• Locked-in Parkinson's patients given L-DOPA started talking again; though they barely moved, they cognitively believed they were moving at super velocity - a 'set point' problem.
  01:51:48
• Sejnowski's colleague Rusty Gage uses large language models as an 'idea pump,' feeding them all his experiments and the literature to generate new experiment ideas.
  01:38:50

• There is more input coming into the visual cortex from the motor system than from the eye itself.
  00:27:43
• The brain often needs only one example to generalize - like learning the whole process of going to a restaurant after a single visit.
  00:25:36
• Sleep spindles (1-2 second traveling waves circling the cortex) consolidate the day's experiences into long-term memory without overwriting existing knowledge.
  00:31:19
• The peak demographic benefiting from the 'Learning How to Learn' course is ages 25-35, not students - and the course has been taken by 4 million people in 200 countries.
  00:46:23
• A writer found ChatGPT far less mentally fatiguing once she treated it politely like a human, tapping into the brain's effortless social circuits.
  01:01:58
• Exercise is described as 'the best drug you could ever take' - it replenishes mitochondrial energy and benefits every organ system, including the brain.
  01:09:45
• In a dermatology study, expert doctors and AI each scored ~90% on skin lesions, but doctors using AI jumped to 98% by combining different strengths.
  01:42:27
• AI trained on hurricane data can predict Florida landfall far more accurately than supercomputer simulations - and does it on a laptop in 10 minutes.
  01:45:31
• The biggest difference between an LLM and a human brain: a quiet human keeps generating internal thoughts, while an LLM goes blank with no self-generated activity.
  02:09:37
• LLMs show 'in-context learning' - getting better through a dialogue despite having no plasticity or new training - a mystery neuroscientists also can't explain for humans.
  02:11:43