Huberman LabPoppy Crum on Huberman Lab: How AI Reshapes Learning
Crum explains cortical plasticity reshapes with every tool you use; AI sensor feedback and active encoding accelerate skill gain without losing cognitive depth.
CHAPTERS
- 4:20 – 12:00
Neuroplasticity, Technology, And Why Your Brain Is Still Very Changeable
Crum outlines her view that human brains are far more plastic than most people assume and that every interaction with technology reshapes neural maps. She uses the somatosensory homunculus and examples like thumbs and ankles to show how practice, environment, and tools reallocate finite neural resources.
- •Neuroplasticity is ongoing and heavily influenced by technology and environment statistics.
- •The classic Penfield homunculus is outdated; modern life would enlarge thumbs and driving-related regions.
- •Expertise creates more neurons devoted to a function, with finer specificity and resolution.
- •Technologies like robots also change our brains when we collaborate with them.
- •We should be consciously asking how current and future tools are architecting our brains.
- 12:00 – 25:00
Cities, Soundscapes, Perfect Pitch, And Sensory Imprints
The conversation explores how noise profiles of cities sculpt hearing and sensitivity, and how specific sounds become deeply encoded. Crum then explains absolute pitch, its variability across historical tunings, and how her own pitch perception shapes her experience of environments.
- •City noise profiles can predict where someone lives and change hearing thresholds.
- •Exposure to particular animal calls or environmental sounds heightens sensitivity to those frequencies.
- •Crum has absolute pitch and experiences sounds categorically, like colors.
- •A demonstrates how A440 vs A415 (Baroque pitch) are social standards, not absolutes.
- •Perfect or absolute pitch is not purely beneficial; it can make flexibility in music harder.
- 25:00 – 41:40
Smartphones, Texting, And Lossy Compression In Human Communication
Huberman and Crum analyze how smartphones and texting have created new composite brain activities, blending thumb motor control, internal and imagined voices, and rapid back-and-forth communication. Crum reframes texting acronyms and shorthand as perceptual compression, preserving or even enriching experience despite ‘loss’ of explicit information.
- •Texting couples fine thumb motor control with language and emotional processing in new ways.
- •You now ‘hear’ both your own and others’ voices in your head while texting.
- •Younger generations may literally anchor safety and social connection to phone charge status.
- •Acronyms and shorthand (LOL, etc.) are lossy compression algorithms that trigger rich internal states.
- •Whether these tools are good or bad depends largely on how they shape priors and neural connections.
- 41:40 – 1:05:00
Priors, Context, And How Technology Can Make Us Smarter Or Dumber
Crum describes humans as products of sensory systems, experiential priors, and expectations—illustrated by her daughter yelling “Minions” at a NASA Earth image. She introduces Bayesian brain ideas and argues tech is powerful when it enriches data and insight, but harmful when it replaces core cognitive work.
- •Our perception is shaped by sensory limitations, experiential priors, and immediate context.
- •Bayesian processing underlies our ability to navigate probabilistic situations in daily life.
- •Technology can refine situational intelligence or erode it, depending on use.
- •Using tools to get faster output without learning leads to long-term cognitive dependence.
- •Design goal: use tech to expose patterns and data we can’t see, not to avoid thinking.
- 1:05:00 – 1:18:20
Gaming, Neuroplasticity, And Building Performance-Enhancing Tools
Drawing from her Stanford course on neuroplasticity and video gaming, Crum explains how games reshape low-level visual functions and higher-order probabilistic inference. She shows how these principles can be used to build training systems for athletes with real-time sensor feedback that increases neural resolution around key performance metrics.
- •Gamers have superior contrast sensitivity and faster probabilistic decisions compared to non-gamers.
- •Training non-gamers with ~40 hours of action games permanently improves these abilities.
- •Video game benefits transfer to real-world situational awareness and reaction time.
- •Crum’s course designs closed-loop training tools (e.g., auditory feedback based on calf sensors for soccer players).
- •More granular, real-time feedback helps the brain differentiate subtle performance variations and learn faster.
- 1:18:20 – 1:31:40
AI As Coach: DIY Computer Vision For Skills And Health
Crum describes using AI and no-code tools to build apps that analyze swimming strokes, gait, and other behaviors from smartphone video. She differentiates democratizing elite-level analytics for everyone from replacing coaches, emphasizing that parents and individuals can use these insights to reinforce good coaching and training.
- •Anyone can use AI platforms (Perplexity Labs, Replit, etc.) to create domain-specific analyzers.
- •Simple examples: analyzing arm entry angle, roll, velocity consistency in freestyle swimming.
- •The value is in richer, more frequent analytics than human observation alone can provide.
- •Similar strategies can be applied to running, workplace performance, and other skills.
- •AI developers should focus on skill amplification and insight, not human replacement.
- 1:31:40 – 1:45:00
Digital Twins: From Reef Tanks And Flights To Human Bodies
Crum introduces digital twins as data-based representations of physical systems used to gain ongoing insight and improve control. She gives intuitive examples (air traffic controllers, airline pricing, reef aquariums) and explains that a digital twin of your health or performance is about capturing relevant data layers, not replicating your entire body.
- •Digital twins are digital representations of key parameters of physical systems used for monitoring and decision-making.
- •Examples include air traffic displays, dynamic airline pricing models, and a sensor-rich reef tank dashboard.
- •The goal is not a full-body clone, but the right interoperable datasets to detect issues and optimize conditions.
- •Integrating biometric, environmental, and behavioral data is where digital twins become powerful.
- •Future health and performance twins will inform proactive adjustments rather than reactive fixes.
- 1:45:00 – 2:00:00
AI, Cognitive Load, And The MIT Study On Writing With LLMs
They unpack an MIT study comparing students writing with pure brainpower, with search engines, or with LLMs. Crum uses cognitive load theory to explain why LLM-written work produces less long-term learning: germane load (schema-building effort) drops dramatically. She warns that professionals overusing LLMs for core tasks risk a shallower, less generalizable understanding over time.
- •Cognitive load has three components: intrinsic (difficulty), extraneous (presentation/context), and germane (schema-building).
- •Germane load is where real learning happens; LLMs often remove it.
- •EEG data showed far less neural engagement when students used LLMs to write.
- •People with higher initial competence tend to use AI to deepen understanding, not bypass it.
- •Doctors and lawyers relying heavily on LLMs may match outputs now but lack deeper pattern recognition and extrapolation later.
- 2:00:00 – 2:11:40
Sleep States Are Well-Mapped. Waking States Aren’t—Yet.
Huberman notes that sleep states (REM, slow-wave) are well understood and easily optimized with tech, but waking cognitive states are poorly defined, making optimization harder. Crum argues that richer, integrated data streams—from body, local environment, and broader context—are needed first, after which AI can help discover and support distinct waking state profiles.
- •We can label and optimize sleep stages, but have coarse concepts of waking states (e.g., ‘focus,’ ‘relax’).
- •To classify and optimize waking states, we must integrate internal physiology, local environmental data, and external context.
- •HVAC systems, cars, and rooms can be turned into adaptive interventions, not just static environments.
- •Future systems will aim at shifting state (e.g., into deep work) based on personalized markers and goals.
- •AI will be well-suited to identify latent state categories once enough multimodal data is collected.
- 2:11:40 – 2:28:20
Non-Contact Sensing And The Coming Age Of Ambient Intelligence
Crum describes how CO₂ sensors, microphones, thermal cameras, and eye tracking in devices or glasses can infer emotional states, stress, and engagement without wearables. She gives vivid examples such as CO₂ rises in cinema audiences during suspenseful scenes and pupil-size tracking via smart glasses, arguing that our environments will become ‘aware’ and responsive.
- •CO₂ patterns in rooms correlate with collective suspense and affect (e.g., in screenings of Free Solo or Hunger Games).
- •Pupil size reflects arousal and cognitive load; future glasses and devices will track it continuously.
- •Eye-tracking and ambient light data can be combined to normalize for light and isolate arousal.
- •Consumer sensors already rival or exceed many medical devices, but regulation lags behind capabilities.
- •Ambient intelligence can proactively adjust conditions (light, temp, sound) or warn of risk (e.g., resident fatigue).
- 2:28:20 – 2:40:00
From Step Counts To Meaningful Metrics: Gamification Done Right
The discussion turns to how simple metrics like sleep scores and step counts altered behavior, and why richer, domain-specific metrics can be even more effective. Crum dislikes the buzzword ‘gamification’ but strongly supports creating engaging feedback loops that train the right circuits rather than just drive superficial behavior change.
- •Sleep scores and 10,000-step goals show that simple quantification can shift habits at scale.
- •The next layer is tailoring metrics to actual performance goals (e.g., focus bouts, acceleration profiles).
- •Computer vision apps can make these nuanced metrics accessible with just a phone camera.
- •Effective ‘gamification’ should create satisfying, skill-enhancing challenges, not just points.
- •People need feedback that relates directly to neural and behavioral adaptations, not arbitrary numbers.
- 2:40:00 – 2:53:20
Ambient AI For Kids, Clinics, And Everyday Safety
Crum imagines digital twins and ambient AI applied to infants, patients, and everyday life to detect issues early and personalize care. They discuss how AI can pick up health signatures from coughs, cries, speech, and subtle vocal modulations, raising the potential for far earlier interventions in diseases like Alzheimer’s, diabetes, or heart disease.
- •AI can identify disease signatures in voice years before clinical symptoms emerge.
- •Patterns in coughs, baby cries, and breathing contain diagnostic information human ears miss.
- •Subtle speech features can indicate neural degeneration, psychosis risk, or metabolic disease.
- •Future baby monitors could go beyond sound and video to health analytics and early alerts.
- •Large barriers remain in regulatory approval and clinical integration despite technical readiness.
- 2:53:20 – 3:11:40
Absolute Pitch, Owls, And Discovering Neuroplasticity Firsthand
Crum recounts how her absolute pitch, and the trouble it caused when playing Baroque music at different tunings, led her to Eric Knudsen’s owl research on map plasticity. Mimicking the owls’ dual maps, she developed a secondary pitch map at A415, a direct personal experience of large-scale cortical remapping that pulled her firmly into neuroscience.
- •Absolute pitch made it hard for her to adjust to historical tunings like A415.
- •Knudsen’s barn owl work showed that sensory maps can be remapped under altered input plus behavioral demand.
- •By repeatedly playing at A415 in high-stakes musical contexts, she developed a second absolute pitch map.
- •This convinced her that human maps are far more flexible than textbooks suggest.
- •Her core obsession became how tech and experience co-drive neuroplasticity across senses and skills.
- 3:11:40 – 3:26:40
Deterministic Behaviors In Nature: Bats, Moths, Spiders, And Monkeys
To illustrate how tightly tuned sensory-motor loops can be, Crum describes moths evading bats via deterministic escape behaviors, spiders tuning their webs to bat frequencies, and marmosets whose eye movements reveal which vocalizations they’re hearing. These examples show how specific stimuli drive rapid, hard-wired responses—principles she wants to harness in human tech and training.
- •Moths use simple neuron thresholds to trigger random flight, then dropping to the ground as bats approach.
- •Some spiders tune webs to resonate with predator echolocation; Crum demonstrates this by singing to orb spiders.
- •Crickets’ bimodal neurons produce opposite behaviors (approach vs. flee) at different frequencies.
- •Marmoset saccades and pupil responses map directly onto different call types (contact, alarm, aggression).
- •Human tech can be designed to trigger beneficial deterministic responses (e.g., posture change, state shift) from specific cues.
- 3:26:40
Closing Reflections: Self-Directed Plasticity In The Age Of AI
They close by returning to the core theme: we can deliberately direct our plasticity using technology, or be passively reshaped by it. Crum emphasizes designing AI and environments that enhance situational intelligence, empathy, and individual flourishing, while Huberman stresses the urgency of being intentional as AI adoption accelerates.
- •Self-directed neuroplasticity allows us to choose which maps to emphasize and strengthen.
- •AI will change brains regardless; the question is whether it enhances or undercuts human capabilities.
- •Thoughtful integration of sensing, feedback, and personalized goals can make tech deeply benevolent.
- •We need awareness of unintended changes (e.g., from smartphones) and a willingness to correct course.
- •Crum remains optimistic that if we design wisely, AI can broaden human empathy and performance rather than diminish them.
