Sakana AI's MASS lets one model optimize its own multi-agent workflows and then train on the results
- MASS is a homogeneous recursive self-improvement method where one shared model acts as executor, optimizer, and evaluator, and it improves score per output token over the base model on research benchmarks.
- The inner loop searches for better team workflows with the model weights fixed, using the model's own judgments, while the outer loop fine-tunes the executor on runs of the best workflow it found.
- Only a self-evaluation signal guides the loop on non-verifiable research tasks, so the authors check progress with outside judges, GPT-5.5 and Claude Opus 4.8, whose verdicts are withheld from MASS.
- The paper is arXiv 2610.12176 and the code is in the SakanaAI/mass GitHub repository, with Hyunin Lee of UC Berkeley as lead author and co-authors from Sakana AI and UC Berkeley.
Hacker News opinions
Sounds like we're still on track for ai-2027.com, for better or for worse.
Are we reading the same website? The China predictions are fantastical. They aren't falling behind, and distilling is enough, so they don't need to steal weights.
Everything from mid-2026 on is wrong. China released Kimi K3 at parity with Claude Fable, the CDZ thing obviously didn't happen, and the S&P 500 is up 13.9% YTD, which is not out of the ordinary.
I'm still putting it at 2029 for real-world work acceleration.
Pretty cool. It has a fine-tuning step, so it's not just a harness optimizer. I think the hard part will be keeping the training on rails and not Goodharting into an irrelevant metric.
In the strictest sense I don't think this is RSI yet. Most of the recent math breakthroughs are interesting applications of stuff already in the corpus, not something new. Self-improvement, sure, but I'm not sure it meets the recursive part.
The code they released doesn't match the paper, and the results are much narrower than the wording like "Odyssean moment" suggests.
This is not a fad, it's a consequence of the Bitter Lesson. RSI is the only game in town now.