Introducing Sakana: The Future of Foundation Models, Inspired by Adaptive Systems

Written by
Jon Rezneck

We are excited to announce Geodesic’s participation in the seed financing round for Sakana AI and to back David Ha, Llion Jones, and Ren Ito.  

Based in Tokyo, Sakana AI is focused on creating a new kind of foundation model based on nature-inspired intelligence, harnessing swarm collective intelligence and evolutionary computing principles through biomimicry in its AI model production. The founders believe their model will not only provide results equal to larger models but also be more economical, secure, and adaptable.  

Their thesis revolves around adaptive systems built with small “swarms” of autonomous agents that cooperate to accomplish a desired objective or produce a requested output (like a school of fish, which is what Sakana means in Japanese). Also worth noting, Sakana’s models will initially be trained on Japanese characters, representing a departure from traditional Western Large Language Models trained solely on Latin-based characters.  

David and Llion believe that the immense complexity of our world requires a new kind of foundation model built using the principles of complex adaptive systems. Today modern neural network systems are built similarly to how we build bridges and buildings, i.e. thoughtfully designed but not adaptive. Natural systems adapt and become part of their environment, a core component of the inspiration behind Sakana.  

We view the launch of Sakana AI in Japan as a strong validation of Japan, its talent, and its ecosystem.  We look forward to working with David, Llion, and Ren!