I'm surprised that OpenAI Comms would approve this: GPT-5 "foreshadows a recursive self-improvement loop" In OpenAI's Preparedness Framework, recursive self-improvement is a Critical risk (if at a certain rate), which would call to "halt further development"

To be clear, it sounds like Sebastien isn't describing an especially fast loop. He's also talking about foreshadowing, not being here today per se I was still surprised OpenAI would use this term about its AI though. Then I realized it's also used in "The Gentle Singularity"

@CJKRaymond For instance, if a technique like this caused a big speedup in model gens (because OAI is no longer bound by limited HQ data), that'd meet the lagging indicator defn Either way, I'm not saying this is a dangerous RSI! I'm just surprised that OAI didn't veto this description