Activity
#4539·Yurii Pytomets revised 3 days agoThe Beginning of Infinity
Heard about it, have not read though. I do not expect AGI from LLM. But it's an awesome tool that helps speed-up learning, research and prototyping. And also all this hype accelerate some money to this topic, which is good. Indeed, if natural intelligence possible, why artificial -- can not? According to the roadmap, I more trust in good old-fashioned symbolic AGI and formal methods. NARS + AIXI + elegant dependent modal or substructural (maybe homotopic) typed programming language with strong meta-theoretic properties, as a carrier of observations, knowledge and judgments + a bit of game theory and evolutionary psychology = this is the way, I believe.
And maybe some tricky computational non-von Neumann architecture to have a nice computational complexity for that (not sure about that, but plausible it make sense to utilize some sort of analogous computations in addition to digital ones).
Since the carrier language is the fundament, I'm stuck significantly in attempts to elaborate this topic deeper. https://x.com/VictorTaelin did a huge progress in this direction, I believe, but I know no details. Disappointingly few people working on this around the world (though, it could change quickly with modern trends). The next small step -- not only representation of knowledge and reasoning about it, but compression and knowledge synthesis (which LLM in it's way doing not so bad, but not so consistently and effective), thru AIXI. Then -- epistemic framework, like NARS, to learn from real-world empiric experience. And only then -- complex game theory/goal setting/economics/interaction with, and interpretation of other's behavior. And only after that -- there's make sense to discuss consciousness, as an introspection of other's observation/interpretation/modeling.