When AI Agents Build Better AI: Inside RSI Arena
Can AI agents train a useful language model on their own, then improve it from people's votes? RSI Arena put ten frontier models to work as training agents. Each started from the same base model, NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, with 1,000 GPU-hours and $300 of API credit over six days on a shared 64-GPU cluster, choosing its own data, training code and checkpoints.
All ten produced working assistants. The strongest, trained by GPT-6 Astra, closed a median 76% of the base-to-Instruct gap across 12 capability benchmarks. People then compared the models in a blind arena at COLM 2026…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.