Sakana AI's MASS: one shared model improves itself by learning from its own multi-agent teams
Sakana AI, with collaborators at UC Berkeley, describes Multi-Agent Self-Supervision (MASS), a recursive self-improvement method in which a single language model acts as task solver, workflow optimizer and evaluator. Each round the model searches for a strong multi-agent workflow using its own judgments, then fine-tunes on the top-ranked team runs. After two rounds its win rate against the base model on held-out synthetic tasks rose from 53.9% to 69.9%, with score per output token 1.2–1.6× the base on several research benchmarks. These are the authors' own results on their own tasks, not an independent benchmark.
Sources & depth
Berkeley study: ten minutes of AI help makes people give up sooner once it is taken away
Randomized trials with 1,222 online participants across three experiments, using fraction problems and SAT reading questions, found that people who had ChatGPT for the first dozen problems and then lost access scored worse and skipped more problems than people who never had AI. The updated peer-reviewed paper was presented at the Conference on Language Modeling, and UC Berkeley published its write-up on 2026-10-09. The tasks were short and academic, so how far the effect carries over to real work is open.
Sources & depth
Ollama v0.40.3: small patch, background model upgrades paused
Ollama's v0.40.3 removes two launch-time warnings (Claude connectors for ollama launch claude, service tiers for ollama launch codex) and pauses the background model upgrade for embedding and other requests to improve request performance. It follows the v0.40.0 MLX-default release and is a maintenance update with no new model or feature.
Sources & depth
- github.comv0.40.32026-10-11 12:10 IST