According to Sakana AI's account of its October 26, 2026, symposium, CEO David Ha described three pillars of the company’s AI sovereignty work: using AI agents to help develop the next generation of pre-trained models, adapting large language models to Japanese and business settings through Sakana Namazu, and using Sakana Fugu to pursue advanced performance without relying on a single model. Those are the company’s approaches and aims, not reported performance results.
Sakana AI said chief scientist Takuya Akiba also discussed Sakana Sawara, the tentative name for a pre-training project. Sakana AI said computing capacity remains a bottleneck for pre-training. It described an approach that would use AI alongside a small team to make development decisions about where to spend limited computing resources. The announcement introduced a project, not a finished model.
The company said it had previously emphasized innovation at the AI-agent layer but now believes it is time to return to building models. In a separate post quoting the symposium announcement, @ChiekoTazuke said Sakana AI also plans to develop world models for physical AI and research an Agentic Scientist approach to speeding up that development. Those plans were not part of the company’s list of three pillars.
In her account of the event, @Stefania_druga placed talks on Sawara, The AI Scientist and world models for physical AI at Hitotsubashi Hall. She also said she shared physical-AI demos, including robot arms and AI-assisted lab work. The talks and demos do not establish that Sawara is available.
Comments