Sakana AI is a Japanese artificial intelligence research company based in Tokyo. Its approach draws inspiration from natural systems, particularly evolution and collective behavior, to design models and learning methods. Rather than relying solely on increasing the size of neural networks, it explores ways to combine, adapt and enable cooperation between existing building blocks. Its name, which means “fish” in Japanese, reflects this biological inspiration.
An international team rooted in Japan
Sakana AI was founded in 2023 by David Ha, Llion Jones and Ren Ito. The first two come from AI research: David Ha previously worked at Google and Stability AI, while Llion Jones is one of the coauthors of the paper that introduced the Transformer architecture, which has become central to large language models. Ren Ito brings experience as a diplomat and business executive.
The choice of Tokyo is an important part of its positioning. Sakana AI aims to build an international research laboratory based in Japan, connected to the needs and resources of the country’s business ecosystem. The company has attracted international and Japanese investors; Nvidia notably participated in its funding in 2024. This base allows it to stand out in a sector largely shaped by American and Chinese players.
Combining models and equipping researchers
Its work includes evolutionary model merging methods. The principle is to use algorithms inspired by natural selection to search for combinations of already trained models. This approach aims to obtain new capabilities without always starting over with a full training process. Sakana AI has notably demonstrated it with models adapted to Japanese and systems combining text and image understanding.
The company also explores the automation of scientific work. Introduced in 2024 with academic partners, its project The AI Scientist seeks to link several stages of machine learning research: generating ideas, writing code, conducting experiments and writing a paper. It is a research program, not a guarantee of reliable discovery. The validity of experiments, the originality of results and the quality of evaluations remain points that need to be checked.
What next?
For Sakana AI, the challenge is to turn these experimental avenues into robust and useful tools. Evolutionary approaches could make it easier to adapt models to specific languages, professions or computing constraints. Their value will nevertheless need to be established through reproducible comparisons of performance, costs and reliability.
Scientific automation also opens up possibilities, but requires human supervision and robust control mechanisms. The company’s trajectory will therefore depend as much on its research results as on its ability to translate them into practical applications, without confusing automated output with validated knowledge.