I've been wondering this too. if it's as capable as they think, i would expect more discoveries resulting from previously unexplored intersections of different specialist fields, given the vastly greater breadth of the knowledge corpus
People need to think extensively for making such connections. We have only now reached the stage where LLMs do that for extended periods and even then they are mostly used for giving answers to specific questions.
Agents are coming. That changes things.
I think what this gets wrong is that humans have a lot of preloaded cognitive abilities that llm dont come with. Part of the difficulty with LLMs is you are both trying to introduce the cognitive abilities and the knowledge at thr same time
As fake as it seems
— & it does seem fake —
I start to worry about some crazy overhang developing
What happens when we figure out, "oh, the llm is missing a few tricks to enable it to dynamically explore its own cognition" (or whatever the trick turns out to be) & the thing rips down the road
I think there is already an overhang. I’m not sure how popular of an opinion it is, but I think there’s quite a lot of compute available on earth, if you count devices that aren’t typically “dedicated” to AI, that if a system can harness it they’ve got enough to bootstrap themselves to ASI
There is definitely a huge overhang, see how deepseek R1 was able to be trained for way less compute than a lot of previous models. There's also a huge amount of research that has been 'ignored' or seriously pursed, like bitnet, mamba, etc. I think there are still HUGE overhangs to be found
Claude is fine and the new model seems good but yeah, this is like a slide you would show an investor trying to convince them of future growth rather than something realistic
it annoys me that so many people just repeat what LLM company executives say about future performance totally uncritically
they might be right! they might not! in a world where humans still have to think about things it's important to do that
