This is along the same lines I imagined it as well. And yes modern LLMs are nothing like what I’ve described. We’d need kilowatts of power per every single LLM conversation if we were to gradient descend them per conversation…. Funny enough human brain runs at like 10 watts. So I imagine there’s some huge paradigm shift from what LLMs do to what human brain does. You don’t get 1000 time efficiency boost from just putting all the world’s ram into it.