Organizing model copies into teams feels promising for self-improvement. Blog: pub.sakana.ai/mass/ Paper: arxiv.org/abs/2610.12176 Work by Hyunin Lee, Jinglue Xu, Jeffrey Seely, Yujin Tang, and UC Berkeley collaborators (Somayeh Sojoudi, Matei Zaharia, Donghyun Lee)!
Recursive Self-Improvement through Multi-Agent Self-Supervision arxiv.org
AI Weekly's analysis
→
- A new arxiv paper proposes MASS, which alternates evolutionary workflow optimization with supervised fine-tuning on self-generated multi-agent trajectories.
- With Qwen3.6-27B over two MASS cycles, the authors report 1.2-1.6x higher performance per output tokens on four open-ended public benchmarks.
- Multi-agent traces look more training-efficient: a student trained on them beats a single-agent student trained on 1.4x more tokens.
Read full analysis →
Incredible work by Robert Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn, and the team! GitHub: github.com/SakanaAI/AI-... Blog: pub.sakana.ai/agentic-scie... arXiv: arxiv.org/abs/2610.18933
GitHub - SakanaAI/AI-Scientist-v3: The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery github.com
Incredible work by Robert Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn, and the team! GitHub: github.com/SakanaAI/AI-... Blog: pub.sakana.ai/agentic-scie... arXiv: arxiv.org/abs/2610.18933
The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery arxiv.org
AI Weekly's analysis
→
- The pipeline decomposes computational research into eight stages, each powered by coding agents with 'agentic self-verification' and rubric checks.
- A development version of the system submitted a paper that was accepted for publication at ALIFE 2026 in co-scientist mode.
- The authors' ablations argue reliable scientific autonomy depends on the cross-agent orchestration layer, not only on base model capability.
Read full analysis →
↻
hardmaru reposted
I am incredibly proud of our Tokyo team for shipping this. By orchestrating the world’s models, we are delivering the resilient blueprint required for AI sovereignty. Read our full vision and results here: sakana.ai/fugu-release 🐡
Sakana AI sakana.ai
View on Bluesky ·
♥ 12
↻ 0
↩ 0
·
6 from the directory shared this ·
119d ago
“World Models in Natural and Artificial Intelligence” brings together pioneers including Douglas Hofstadter, Michael Levin, Josh Tenenbaum, Samuel Gershman, and Melanie Mitchell to ask: What if the next leap in AI requires not just more data, but systems that model themselves?
royalsocietypublishing.org
↻
hardmaru reposted
Sakana AI
@sakanaai.bsky.social
Fugu-Ultra is now live on OpenRouter! ⚡ We share a core vision with the OpenRouter team: the future of AI isn’t a single monolithic model, but the collective intelligence of the world’s best models working together. Try it: openrouter.ai/sakana/fugu-... 🐡
Fugu Ultra - API Pricing & Providers openrouter.ai
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
↻
hardmaru reposted
By orchestrating the world's models, we are building the resilient infrastructure required for AI sovereignty. Try: sakana.ai/fugu Blog: sakana.ai/fugu-max-rel... 🐡
Sakana Fugu — Multi-agent System as A Model sakana.ai
AI Weekly's analysis
→
- Fugu routes tasks through a dynamic multi-agent pipeline exposed as a single OpenAI-compatible API, removing orchestration setup from users.
- The system draws on two ICLR 2026 papers: TRINITY assigns Thinker/Worker/Verifier roles; Conductor uses reinforcement learning to design coordination strategies.
- Fugu Ultra scored 73.7 on SWE Bench Pro and 93.2 on LiveCodeBench; base Fugu reached 95.5 on GPQA-D, per Sakana's own benchmark reporting.
Read full analysis →
Organizing model copies into teams feels promising for self-improvement. Blog: pub.sakana.ai/mass/ Paper: arxiv.org/abs/2610.12176 Work by Hyunin Lee, Jinglue Xu, Jeffrey Seely, Yujin Tang, and UC Berkeley collaborators (Somayeh Sojoudi, Matei Zaharia, Donghyun Lee)!
Multi-Agent Self-Supervision (MASS) pub.sakana.ai
AI Weekly's analysis
→
- In two rounds of training, the Round 2 model beat the base model on test tasks 69.9% of the time and beat its own Round 1 version 60.8%.
- Multi-agent trajectories hit a 68.3% win rate while consuming 28.8M supervised tokens against 42M for the single-agent approach, a 31% saving.
- On public benchmarks MASS reports 1.60× score-per-token on MLR-Bench, 1.35× on DSBench, 1.20× on ScienceAgentBench and 1.28× on AstaBench.
Read full analysis →
By orchestrating the world's models, we are building the resilient infrastructure required for AI sovereignty. Try: sakana.ai/fugu Blog: sakana.ai/fugu-max-rel... 🐡
Sakana AI sakana.ai
AI Weekly's analysis
→
- Fugu Max prices at $2 per million input tokens and $6 per million output tokens, which Sakana claims runs 40-60% below Sonnet 5, GPT 5.6 Terra, and Kimi K3.
- The system routes each task across a swappable pool of open-weight and specialized models, with NVIDIA's Nemotron folded in via an August 2026 collaboration.
- Fugu Ultra v2 scores 48.3 on Chartography against Opus 5's 27.3, and does so without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool.
Read full analysis →
