Sakana AI, founded by former Google researchers, introduces AI models developed through a unique method inspired by evolution and collective intelligence.
The startup is led by David Ha and Llion Jones, both former Google researchers known for their contributions to deep learning architectures.
The company has introduced Evolutionary Model Merge, a method that uses evolutionary techniques to combine different open-source models efficiently, creating new models with user-specified capabilities.
The method automatically discovers novel ways to merge models from diverse domains, such as non-English language and math, or vision and language, surpassing human experts’ capabilities.
Successful models from each generation are identified as “parents” for subsequent generations, enhancing model evolution.
Sakana AI announced it had successfully evolved a Japanese Large Language Model (LLM) capable of math reasoning and a Japanese Vision-Language Model (VLM), achieving state-of-the-art results in various benchmarks.
The Tokyo startup has released three Japanese language models, with two models made open-source: EvoLLM-JP, EvoVLM-JP, and EvoSDXL-JP, available on platforms like Hugging Face and GitHub.
Sakana AI said that it anticipates the rise of evolutionary AI, where combining existing models through evolution will lead to the discovery of models with unexpected abilities.
So far, Sakana AI has raised $30m in seed financing, signalling confidence in its potential among investors.
Sakana AI aims to establish Tokyo as a prominent centre for AI innovation, following the footsteps of OpenAI in San Francisco and DeepMind in London.