Apart from the gruesome ratios, I found the fact that smarter LLMs are closer to having a well defined utility function quite surprising and maybe a little concerning