The brain as an AI
A dinner-table memory game sends Juan into a reflection on how the mind works like a model. Trying to recall a math teacher's name, his brain ran a filter, not her, not her, she's the math one, then reconstructed the name by analogy to a teacher whose name it did have, feminized. Memory, he thinks, compacts information down to a minimal thread you can grab and pull to rebuild the whole; some memories condense and survive, others are deleted outright, like a classmate his partner insisted existed but that Juan's mind had erased so completely it took a Facebook photo to believe. Then the AI parallel, and where it breaks: neural nets have their dimensions predefined and train to fill them, but the brain seems to create a new dimension each time it needs a new concept, bounded only by the hardware. His conclusion: you can't know everything, human or machine, both hit a limit and need specialization; we only seem able to learn anything because we never live long enough to fill the space, where a model can read the whole internet in a day.
// trace: where this idea came from
- ↳ video diary @ 3:47 (el cerebro hizo un filtro, no es esta, no es esta, reconstruyó el nombre por analogía)
- ↳ video diary @ 5:53 (la memoria condensa hasta un hilito que se agarra y reconstruye)
- ↳ video diary @ 12:13 (la red tiene dimensiones predefinidas; el cerebro crea una dimensión nueva por concepto)
- ↳ Entry 221-3: He who grasps too much (el mismo límite que fuerza la especialización, aquí en humanos y máquinas)
A dinner with old friends turned into a memory game, and the game turned into a theory of mind. Juan found he could recall almost every teacher’s full name, knowledge he didn’t know he had, and watched his own brain retrieve one it had misplaced. His mother described a good math teacher; nothing came. Then his brain ran a filter, not this one, not that one, she’s the math teacher, and finally reconstructed the name by analogy to a teacher whose name it did have, the same name in feminine form ▸ 3:47. Memory, he decides, compacts information until only a minimal thread remains; grab the thread and the whole thing reconstructs ▸ 5:53. Some memories condense and wait under dust; others are deleted outright, like the classmate his partner swore existed and Juan’s mind had erased so cleanly it took a Facebook photo before he’d believe it ▸ 8:26.
el cerebro no llena su límite; solo se le acaba el tiempo →
Then the parallel, and where he thinks it breaks. Embeddings and neural nets already work a little like this, each dimension of a high-dimensional space encoding some feature. But a network has its dimensions predefined and trains to fill them, whereas the brain seems to create a new dimension whenever it needs a new concept, growing the space as it goes, bounded only by the hardware in your skull ▸ 12:13. The conclusion lands on specialization: you can’t be expert in everything, human or machine, both saturate and both end up needing to specialize ▸ 13:36. We only seem able to learn anything at all because we never live long enough to fill the space, a model can read the entire internet in a day, and we don’t get a hundred years ▸ 14:18. The human capacity, then, is mostly unused…
// continued in
no entry has continued this idea yet: the arc is still open