Sakana Sawara is the provisional name Sakana AI gave its new pretraining effort Oct. 26, broadening a strategy that had focused on combining open-source systems.
The project is one of three parts of the company’s AI strategy. Sakana Namazu covers post-training, while Sakana Fugu handles orchestration, Sakana AI said.
The company said AI agents would help plan pretraining experiments. It gave three reasons for the work: lower model-development costs, open models closing in on leading systems, and reinforcement learning’s potential to tailor models to specific tasks.
Sakana AI said computing resources remain a constraint on pretraining. Before the announcement, it had emphasized evolutionary methods for combining existing open-source models. The Sawara plan adds in-house pretraining from scratch to that approach.
