You're reading a whole lot into a tweet, in his interview with
Dwarkesh Patel he says, about 20 different times, that scaling LLMs (as they are currently conceived) won't lead to AGI.
You keep changing topics so I don't get it either, I can attest it's not a fringe view that the situation is interesting, seen it discussed several times today by unrelated people.
He's said it pretty clearly, an LLM could be part of the solution in combination with program synthesis, but an LLM alone won't achieve AGI.