Agents reportedly find most performance optimizations at major inference providers

An attendee at a New York City AI dinner said agents find most performance optimizations at major inference providers. They said performance engineers spend most of their time managing agents and RSI loops, and that most people have agents writing, reviewing and submitting pull requests around the clock. The attendee also said most agreed code review is no longer very productive, and relayed a view that too much reinforcement learning on code and math hurts models' language abilities.