Post

they had you at the name, didn't they Tim

They came up with a self-improving RL framework where the model generates both the solution as well as RL harness/environment since they’re jointly optimizing the two together, they can discover better next steps to make the problem slightly harder in all the right ways

big labs have grown huge and complex org charts of engineers creating RL environments. This new self-evolving harness might be a necessary piece to getting open weights models to compete it’s finally bumping RL back in line with The Bitter Lesson