- innovation is text-based? kevin - RL (alphadev (sorting, hyashing)), GAI (lay of the land, strnegths and weakeness), RL x GAI (alphaproof) why optimize at the level of assembly? alg generated Q. theori measurement Q. any tips for specifying the reward to problems where correctness and latency is unknown or unknowable? -> should have asked "what would be the bottleneck if we were to model startup as reinforcement learning?" (send mail to dan) A. domain knowledge, focus on signals (evaluation) agents can get discover new efficient algorithms to solve impactful problems, evaluation is fast, solving is hard (is entrepreneurship np-hard?) when funsearch is useful: smooth scoring surface, prior is useful