Gemini & TPUs engineer Patrick C attended a dinner in New York and discussed some interesting topics. Feels like I’ve underestimated the practical applicability of the RSI concept itself. Right now, most of the performance optimizations by major inference providers in the U.S. (including fireworks) are discovered through proxies, and performance engineers are now spending most of their time managing and cycling through RSI. Time to seriously rethink and apply RSI in my own workflow.