3 ms·
> It's weird, but somehow this seems like the least-tractable challenge... Unless we find that "the weights" were reliably derived from observation of deep stru
by ben_w 2y ago
> It's weird, but somehow this seems like the least-tractable challenge... Unless we find that "the weights" were reliably derived from observation of deep structures of adult brain in a way that transcends language and upbringing.
Doesn't seem so to me; while I keep failing to hear any information about precise measurements of in-vivo synaptic weights, if you could do that then in principle one could pop a pre-trained LLM in and it would be at least close enough.
GPT-3.5 is 0.01% of the human brain's complexity so this wouldn't be a huge extra volume either, and that's enough for, what, speaking 30 languages fluently, and having a decent (degree-level but textbook-only) education in every subject?