AI-Feynman as a benchmark for what we should be aiming for — LessWrong