Sakana AI Announces Pretraining Its Own LLM, Launches 'Sakana Sawara' Project
SakanaAILabs · x · 2026-10-27
At its first symposium in Tokyo (featuring a Jürgen Schmidhuber keynote), Sakana AI outlined a three-pillar strategy and announced it is returning to the model layer:
- Pre-training: the 'Sakana Sawara' (tentative) project to build its own LLM from scratch, using its own AI agents to accelerate experiments; chief scientist Takeru Akiba notes compute remains the bottleneck.
- Post-training: Sakana Namazu, adapting LLMs for Japanese and business use cases.
- Orchestration: Sakana Fugu, top-tier performance via orchestration rather than a single model.
The rationale: open models are closing the gap with frontier ones, model-building costs are falling, and RL enables task-specific specialization — so it's time to build models again. CEO David Ha framed it as key to Japan's 'AI sovereignty.' The company is also working on physical-AI world models and Agentic Scientist research.
Related event: Sakana AI Returns to Model Layer with Japanese-Built LLM Sawara(4 posts)→
More from Companies & People
- AI games platform founder: player retention now rivals top-decile mobile games — TAbrodi · 2026-10-27
- The 'Super IC' model is replacing bloated 10-person marketing departments — VibeMarketer_ · 2026-10-27
- exe.dev meetup packs Sentry HQ, with Beyang and Amp in the house — davidcrawshaw · 2026-10-27
- Suno design engineer goes public, invites users to DM bugs for direct fixes — enggirlfriend · 2026-10-27
- Peter Sellars on managing Nikita Bier: like hiring Liam Neeson from Taken to raise a kid — lennysan · 2026-10-27
- 136 projects in one day at Cerebral Valley's Gemini hackathon — schwentker · 2026-10-27