Top Google AI experts pick Japan to set up on their own

Receive free Artificial intelligence updates

We’ll send you a myFT Daily Digest email rounding up the latest Artificial intelligence news every morning.

A former Google researcher who co-authored a paper that kick-started the generative AI revolution has joined forces with an ex-colleague to found a Tokyo-based artificial intelligence start-up.

Welshman Llion Jones, who left the US tech giant this month, has since set up Sakana AI alongside David Ha, the former head of Google’s AI research arm in Japan. Ha, who is the chief executive, most recently led the research at image AI company Stability AI.

Jones, Sakana’s chief technology officer, was one of eight Google researchers who collaborated on building a software known as the transformer, which underpinned the rise of generative AI, including chatbots such as ChatGPT and Bard, and image generators such as Stability AI, Midjourney and Dall-E.

The Transformers research paper was first published in June 2017. Since then all of its co-authors have left Google, primarily to found their own start-ups as the race for generative AI talent heats up globally. Jones was the last of the eight to exit Google.

Sakana AI will build its own generative AI model — software that can generate text, images, code and other multimedia. It will be up against some of the world’s biggest AI companies including Google, Microsoft, OpenAI, and start-ups such as Cohere, Character.ai and Anthropic, in an area that is highly competitive and extremely well funded. Microsoft this year invested $10bn in OpenAI, in a multiyear deal, while Cohere and Character.ai have raised money over the past few months at valuations of $2bn and $1bn respectively. 

The name Sakana, derived from a Japanese word さかな (sa-ka-na) for fish, seeks to evoke the idea of “a school of fish coming together and forming a coherent entity from simple rules”, according to the two co-founders, whose research is inspired by concepts from nature, such as evolution and collective intelligence. 

Jones and Ha say that the limitations of current AI models come from the fact that they are being engineered as brittle, unchangeable structures, like bridges or buildings.

In comparison, natural systems, where collective intelligence plays a role, “are very sensitive and responsive to changes in the world around them. Natural systems adapt and become part of their environment,” they said. They hope to build AI models using these principles of evolutionary computing, which will address issues including cost and security of these systems. 

The founders, based in Japan for several years, have chosen its capital as the company’s headquarters partly due to the competitive market for researchers who can build generative AI software, particularly in North America, they said.

Tokyo is uniquely positioned for the growth of an AI company, due to its high-quality technical infrastructure and educated workforce. It is a global city that is seen as attractive to foreign talent.

“On top of that, training data and the machinery around the models tailored to perform well in non-western societies and culture will be a catalyst in the next technological breakthrough,” they added.