François Chollet on the limitations of LLMs in reasoning — LessWrong