Google alums go big in Japan with months-old startup Sakana AI

SAKANA AI, a Tokyo-based startup founded by former Google researchers, is capitalising on a surge in interest among Japanese firms seeking to catch up in artificial intelligence.

Japan has given an outsized welcome to the nine-month-old venture, led by David Ha, who helped set up the Google Brain research team in Tokyo, and Llion Jones, one of the authors of the seminal paper that launched the current AI wave.

The 10-person startup has won a government grant of supercomputer time and is exploring projects with heavyweights Nippon Telegraph and Telephone and Sony Group to build out Japan’s AI ecosystem. Sakana is also targeting government agency work as well as subcontracts in Japan’s small but growing defence industry.

Sakana, whose logo sports a red fish swimming away from the rest of the group, sets out to train low-cost generative AI models using small datasets. Its team unveiled three AI models for Japanese speakers last month, part of a demonstration to show how quickly companies can ramp up AI capabilities.

Ha and Jones chose to set up shop in Tokyo partly because of the country’s untapped AI landscape.

Despite pioneering technologies including mobile internet and QR codes, Japan’s cash-only stores and routine use of faxes and corporate seals have earned it a reputation as a digital laggard, trailing behind the US and China in AI.

GET BT IN YOUR INBOX DAILY

Start and end each day with the latest news stories and analyses delivered straight to your inbox.

Sign UpVIEW ALL

“I want to work with the NTTs and the KDDIs and the Sonys, because there’s opportunity. No one’s doing that,” Ha said in an interview. “Doesn’t sound sexy, but it is.” 

Valued at around US$200 million, Sakana’s approach contrasts with those of industry leader OpenAI and others spending billions of dollars training AI on huge datasets.

Sakana says its algorithms can help automate the creation of AI models that process language or images by mimicking natural selection when they combine, test and discard data, reducing the amount of human input required. That saves time and energy, Ha said. 

The amount of electricity or resources needed to train many models is enormous, monopolising a whole data centre for months. “We want to shorten that time to maybe days or weeks with less compute and to have a similar level of performance,” Ha said.

Japan is beginning to draw global attention for its potential as an AI market. Microsoft Corp. is investing US$2.9 billion over the next two years to build out the country’s data centres and cloud computing infrastructure, while OpenAI co-founder Sam Altman is opening a Tokyo office this month.

Microsoft is partnering with SoftBank Corp. on generative AI, while OpenAI is working with Rakuten Group on local AI services.

Sakana also faces competition from Elyza, whose acquisition by the country’s No. 2 telecom operator KDDI Corp. was announced last month. Launched by University of Tokyo researchers, Elyza develops large-scale language models in Japanese for generative AI engines targeting companies and local governments.

Boosted by Ha and Jones’ combined pedigree in the world of AI research, Sakana was one of seven institutions to receive the Economy Ministry’s Generative AI Accelerator Challenge grant this year. The award includes time using Japan’s supercomputing resources to run tests and train models.

The two met during their time at Google, where Jones co-authored the oft-cited paper, “Attention Is All You Need,” and where Ha collaborated on a project that deploys a fleet of small AI models working together, rather than use one large model.

The two teamed up with former Japanese diplomat Ren Ito, who headed online thrift shop Mercari’s push into Europe, to create Sakana. 

The startup, which raised US$30 million from a recent seed round from the likes of Lux Capital, Khosla Ventures and Sony, is fielding hundreds of resumes from around the world, according to Ha.

“Every company in Japan who wants to have their own foundation model can work with us or another company like us to develop their own model,” Ha said. “We think there’s room for Japan to become an innovator.” 

It’s important for Japan — one of Asia’s top-ranked liberal democracies — to play a larger role in developing its own AI, Ha said. Cultural and data bias is inherent in language, affecting AI models, he said. 

“It makes sense for companies to have at least the option of having their own foundation model, even if it’s not like a super high performing one,” said Ha. “They need to own their own kind of fax machine.” BLOOMBERG