llm is not agi by ewan

there’s a common trend to centre everything around ai (gen ai, specifically), and for most, it’s about llms. the end goal rn for foundation model labs, is agi.

well, i’m no ai researcher, but it shouldn’t take one to know that this isn’t the path to agi. llms are just probabilistic token predictors that cannot live independently of external data. this is a contrast to rl (reinforcement learning) systems that can exist in a system independently and come up with novel data.

i’m not sure why rl is not at the core rn, subjugated to processes such as rlhf, but i can tell that llms are not the future of agi. i’d previously thought rl was possible at scale whereby a model will be trained on data that’ll be introduced in an experience like manner (as with children), but i’m not a researcher and this might be a really dumb take.

maybe, what we’re seeing is a smoke screen as true agi won’t be handed out for commercial purpose (as seen with most frontier tech), maybe not. but i belive llm is not agi.