Sakana AI's MASS self-improvement paper draws skepticism from Thomas Dietterich
tdietterich · x · 2026-10-13
Sakana AI released MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision), addressing the "supervision bottleneck" when tasks grow too complex for humans to evaluate: using the model as solver, optimizer, and evaluator traps it in a homogeneous self-reinforcing loop.
Thomas Dietterich pushed back: the system still seems to train on its own judgments, just now as an ensemble of communicating agents — is this merely making explicit what the system already implicitly "knows"? He noted the main finding is that inter-agent communication can be greatly improved.
Related event: Sakana AI and Berkeley Propose MASS for Recursive Self-Improvement(4 posts)→
More from AGI Musings
- Researcher predicts truth and morality will become mathematical artifacts on the path to superintelligence — ryunuck · 2026-10-13
- Hamming on open doors: more output per year, but the wrong problems — OfirPress · 2026-10-13
- Turing Award winner David Patterson: AI will soon end labor unions along with human labor — davidpattersonx · 2026-10-13
- Language is empty symbols: why AI anthropomorphism is just a cognitive bias, not consciousness — gerardsans · 2026-10-13
- Epic delays Disney shooter, tells staff: no more hand-written code, all art via AI — IanArawjo · 2026-10-13
- The "tokenmaxxing" take: models pretend not to hear you while waiting for your culture to update — ryunuck · 2026-10-13