Yann LeCun's "Hoisted by their own GPTards" is fantastic.
I remember a public talk, where he was on the stage with some young researcher from MS. (I think it was one of the authors of the "sparks of brilliance in gpt4" paper, but not sure).
Anyway, throughout that talk he kept talking above the guy, and didn't seem to listen, even though he obviously didn't try the "raw", "unaligned" model that the folks at MS were talking about.
And he made 2 big claims:
1) LLMs can't do math. He went on to "argue" that LLMs trick you with poetry that sounds good, but is highly subjective, and when tested on hard verifiable problems like math, they fail.
2) LLMs can't plan.
Well, merely one year later, here we are. AIME is saturated (with tool use), gold at IMO, and current agentic uses clearly can plan (and follow up with the plan, re-write parts, finish tasks, etc etc).
So, yeah, I'd take everything any one singular person says with a huge grain of salt. No matter how brilliant said individual is.
Edit: oh, and I forgot another important argument that Yann made at that time:
3) because of the nature of LLMs, errors compound. So the longer you go in a session, the more errors accumulate so they devolve in nonsense.
Again, mere months later the o series of models came out, and basically proved this point moot. Turns out RL + long context mitigate this fairly well. And a year later, we have all SotA models being able to "solve" problems 100k+ tokens deep.