All statements he made about world models did not already turn out to be wrong. He may be right in the long term.

LLMs work well for search problems with the help of a strong verifier.

As it turns out, every problem is ultimately a search problem. But not every domain has an easily available verifier.