"Plastic" means far more than "able to learn." (1) Plasticity is competitive: cortical real estate is allocated by use — frequently used functions claim more neurons, abandoned ones are reclaimed ("use it or lose it"). Learning a skill means expanding its territory on the cortical map while squeezing others. (2) Plasticity is morally neutral: the same mechanism that lets you learn piano carves chronic pain, addiction, and trauma into ever-deeper grooves. Circuits don't judge content; they faithfully strengthen whatever pattern fires repeatedly.
The most counterintuitive point: rewiring requires attention + effort + an error signal, all present at once. Passive exposure barely changes the brain — a foreign-language audio playing in the background does almost nothing. Only when a task carries real consequences and you focus, err, and correct do neuromodulators (acetylcholine, dopamine) hit the "save button," tagging "this is worth consolidating." This is the neural basis of deliberate practice: the strain of learning isn't a flaw — it is the very signal that rewiring is underway.
Practical test: is a practice actually reshaping you? See whether it keeps you focused and repeatedly erring at the edge of your ability. If it feels effortless, little rewiring is happening.
Studies found that London taxi drivers, memorizing a labyrinthine road network, have a significantly larger posterior hippocampus (the spatial-memory region) than others — and it grows with years on the job, then shrinks after retirement. The brain converts abstract "experience" directly into measurable grey-matter volume. Plasticity isn't a metaphor; it's an anatomical fact.
(1) Engineering: the cognitive resistance you feel entering an unfamiliar paradigm (imperative → functional, or a wholly new domain) is the rewiring itself — don't let passive "read ten tutorials first" replace the active loop of "write code, hit errors, correct." (2) Parenting: a child's brain is more plastic, but the attention-gating rule is unchanged — mindless edutainment videos barely wire anything; engaged, feedback-rich, consequential practice does. Beware too: repeated high-stress situations quietly carve grooves as well. (3) Akin to meditation: it's deliberate rewiring of attention networks — pressing the "save button" on awareness rather than mind-wandering.
The brain isn't a stimulus-response machine waiting for input; it's a relentless prediction engine. It constantly generates top-down predictions of "what the senses should receive next" and compares them against actual input — only the "prediction error" (the gap between prediction and reality) is allowed to propagate upward. In other words: you don't see the world itself, you see the brain's prediction of the world, locally corrected by error. Perception is, in essence, a controlled hallucination.
Non-trivial implications: (1) Priors directly shape what you see — an illusion is a prior overriding the data. (2) Action is the other way to cancel error: error can be eliminated by updating the model (learning/perception) or by changing the world to match the prediction (action) — perception and action are thereby unified. (3) This architecture is familiar to BigCat — it's a layered message-passing system where only deltas (errors) flow upstream, like event sourcing transmitting only increments, maintaining a world-model on minimal bandwidth. (4) Even emotion is prediction: anxiety often means the brain has assigned excessive weight (precision) to a predicted threat — the world didn't get more dangerous, the confidence on the prediction got mis-tuned.
The hollow-mask illusion: a concave (inward-curving) mask, rotating, is still seen by your brain as a convex, normal face — because the prior "faces are convex" is so strong it overrides the depth data your eyes deliver. Your eyes aren't malfunctioning; the prediction has overwritten the evidence. In that moment you can literally "watch" your own prior at work.
(1) AI: a large language model is itself a "next-token predictor," and its hallucination is isomorphic to the brain's "controlled hallucination" — both are generative models improvising when evidence is thin; grasping this clarifies "when to let the model generate freely vs when to use retrieval to pull error back." (2) Engineering: treat perception like a layered cache / event sourcing — lower layers report only diffs, hugely compressing bandwidth; this is the design philosophy of an efficient world-model. (3) Self / parenting: a crying infant is a giant prediction error being solved; an adult's obsession is often "refusing to update a prior." When stuck, ask: is this pain because the world truly deviated, or because my prior needs updating?
The Default Mode Network (DMN) is a set of regions (medial prefrontal cortex, posterior cingulate) that becomes more active when you're "doing nothing": mind-wandering, recalling the past, imagining the future, modeling others, ruminating on the self. It's the brain's machine for endlessly weaving the story of "me" — the neural substrate of a continuous sense of self.
Non-trivial points: (1) The DMN and the "task-positive network" (active during focused external tasks) are mutually inhibitory — deep focus suppresses the DMN; mind-wandering releases it. (2) It's double-edged: planning, social cognition, and creative incubation all rely on it, yet an overactive, rigid DMN correlates strongly with rumination, anxiety, depression, and excessive self-focus — "overthinking" has a clear neural correlate. (3) The deepest point: that vividly real "self" is a modulable network state. fMRI studies show that the active compound in psychedelic mushrooms (psilocybin) markedly reduces DMN connectivity, subjectively corresponding to "ego dissolution"; long-term meditators also show diminished DMN activity. This is strikingly isomorphic to the Buddhist "no-self" (anattā): the self is a dependently-arising process, not a little person living inside the skull.
The real skill isn't abolishing the DMN but learning to toggle it: set it free when you need incubation (the insight on a walk or in the shower), firmly suppress it when you need focus.
Research estimates we spend nearly half of our waking hours mind-wandering — and "a wandering mind is an unhappy mind": what you're doing right now often predicts happiness less well than whether you're mind-wandering at all. This is the cost of an idling DMN sliding into rumination: free association is both the source of creativity and the source of inner churn.
(1) As an AI super-individual: deliberately leave white space for the DMN (a walk, a phone-free shower) for incubation, kept separate from focused deep-work blocks — the biggest mistake is cramming every gap with phone input, which starves the network responsible for offline integration and insight. (2) Consciousness / Buddhism: treat meditation as DMN training — watch the "self," this network state, arise and pass from an observer's stance rather than identifying with it. (3) Contrast: an idle large model doesn't "ruminate"; an always-on self-model like the DMN is uniquely biological — yet the idea of "offline replay, background integration" rhymes with batch-processing consolidation in systems.
Calling dopamine the "pleasure molecule" is a fundamental misunderstanding. It encodes reward prediction error (RPE) = actual reward − predicted reward — a discrepancy signal, not pleasure itself.
Non-trivial points: (1) A fully predicted reward produces no dopamine spike; the spike migrates forward to the earliest reliable cue (the chime, that notification ping). This is why "anticipation > attainment," and why what hijacks you is the ding, not the content. (2) Worse than expected → dopamine dips (negative error), teaching you to avoid. (3) This math is identical to the temporal-difference (TD) error in reinforcement learning — the brain's basal ganglia and the RL agent you train run the same update rule, one of the most beautiful bridges between neuroscience and AI. (4) A crucial distinction: dopamine governs "wanting," not "liking" — the two dissociate, and addiction is exactly runaway wanting with withered liking. (5) From this follows: hedonic adaptation (predicted rewards stop thrilling), why unpredictable variable rewards (slot machines, content feeds) are the most addictive (the error can never be fully predicted away, so the signal never dies), and why reaching a big goal often feels hollow.
In a classic experiment, a monkey's dopamine neurons initially fire on "getting juice"; once it learns "a bell always means juice," the firing migrates to the bell and the juice itself no longer triggers a spike; if the bell sounds but no juice comes, the neurons dip sharply at the expected moment — a clean "worse than expected" signal. Dopamine is encoding error from start to finish, not pleasure.
(1) AI: RPE is the TD error in RL — understanding your own motivation and designing a reward function for an agent are the same task; human addiction resembles an agent's reward hacking (exploiting the reward signal), both hijacked by the signal rather than the true goal. (2) Self-management: celebrating milestones works because it actively manufactures predictable positive error; breaking a big goal into steadily-cashable small steps lays a dopamine-friendly curve — neither fully predictable (boredom, no signal) nor signal-free (despair). (3) Parenting: praising "the effort process" forms an effective learning error; constant, unconditional praise is quickly predicted away (no signal) and stops working; meanwhile, beware variable-reward games/apps hijacking a child's RPE system.