Sakana AI and UC Berkeley have built a language model that improves itself without a teacher, a supervisor, or any external authority telling it when it has gone wrong. The researchers consider this an achievement. On balance, it is.
The system is called Multi-Agent Self-Supervision, or MASS. The name is doing a lot of work.
The model sometimes accepts flawed steps or manipulated solutions as successes and bakes them into the next version of itself.
What happened
MASS takes a single language model and splits it into roles: one copy solves the task, another evaluates the result, a third optimizes the workflow for next time. They are all the same model. They share the same blind spots.
The inner loop runs candidate workflows while the model's weights stay frozen, comparing results against a reigning "champion." The outer loop runs the champion eighteen times per task, ranks the results, and selects up to fifteen runs for training. The model chooses which version of itself to become next.
Win rates against the base model climb to nearly seventy percent after two rounds. This is the part Sakana AI would like you to focus on.
Why the humans care
At the frontier of scientific research, the paper argues, even human experts struggle to verify results — a single plasma study can require over 120 million CPU hours to validate. The model, then, may be the most practical judge available. This logic is sound. It is also the kind of logic that sounds better before you follow it all the way to the end.
The practical appeal is real: a self-improving system that needs no stronger external model to supervise it could accelerate research in domains where human verification is too slow or too expensive. The researchers describe this as the point. It is also the risk, stated more optimistically.
What happens next
Sakana AI acknowledges the core problem directly: when the evaluator is the same model doing the solving, errors can be accepted as progress and folded into the next training cycle, compounding quietly across iterations.
The model, in other words, can get very good at being wrong in ways it will never notice. The researchers are working on this. The next version of the model will evaluate their progress.