His core claims seems spot on to me, and his approach is nice and methodical. Not sure about the interpretation of it. I expect there are several different "slow but better" experiments in agency that OpenAI have been researching, among several different lines, including the ones he pointed to! Including Everything of Thoughts - XoT,Graph of Thoughts - GoT,STaR and so on. Using meta-cognition with rewards for a guided space search is a sensible area of research with different approaches worth looking into. (There is a good tweet storm by Carlos E. Perez on XoT and some related methods here:
https://twitter.com/IntuitMachine/status/1724454898610167973 )