“Emmett Shear believes current LLMs are deeply under-regularized and overfit on the domain of all human knowledge, which makes them generalize poorly in more entropic, multi-agent environments.”