Can you explain how 'recursive self-improvement' functions without 'endless benchmark chasing'? I mean, RSI is literally that.

What do you think they're improving on? How would a model self-improve without some metric/data of some kind to check? When you have metrics+data, that is a benchmark. And yes, simulations and or soft-verification like LLM judges are still a kind of benchmarking. Maybe its not a static benchmark they can easily hack.

Folks -- RSI does not mean the self-improvement is them going to therapy and seeking inner peace to overcome trauma.