The video takes this paper out of context to argue that LLMs are fatally flawed. The paper merely establishes a complexity bound based on the floating point operations a model is capable of beyond which they begin guessing. AFAICT all that it demonstrates is that LLMs do not defy our theories of computation (e.g. P != NP).

Will LLMs only ever be "tools?" idk. Are LLMs sold as more than tools? Certainly. Does this paper prove LLMs will only ever be tools? No.

Sadly, I don't think the youtuber knows what they are talking about.