if you mean the "random search" comment it appears inaccurate both on principle (the process of making an LLM can be described as "big autoregressor" maybe, at least pre-RLHF part, but the final product is something else) and on domain expert opinion (Thomas Bloom)
