[Special Feature: The Future of AI in Japan]: Can Tokyo be a global hub for AI development? Interview with Ren Ito, co-founder of fast-rising Sakana AI

Sakana AI is an AI startup launched in 2023 and headquartered in Minato City, Tokyo Prefecture. The startup is well-known as the fastest company in Japan to become a unicorn. (A unicorn is an unlisted company valued over $1 billion.) It has achieved groundbreaking developments. Some examples include "Evolutionary Model Merge," which uses evolutionary techniques to combine open-source AI models, and "The AI Scientist," which automates machine-learning research from idea generation to paper writing.

This unique startup was founded by three individuals. David Ha, Chief Executive Officer (CEO), switched from a career in finance to become a world-class AI researcher. Llion Jones, Chief Technology Officer (CTO), previously worked for Google. He is one of the eight authors of the cornerstone research paper on large-scale language models (LLMs). Ren Ito, Chief Operating Officer (COO), formerly worked at the Ministry of Foreign Affairs. He has also served as an executive at other startups.

How can we follow up Sakana AI? Ito has been conducting daily interviews and keynote speeches. We asked him for some straightforward suggestions on securing talented individuals for AI development in Tokyo.

Not "overseas and Japan," but "U.S. West Coast and others"

— Sakana AI has been going for just a few months, yet you have already produced some extremely high-impact results, including Evolutionary Model Merge. It appears like researchers from other countries have played a major role. How can we get top-notch researchers to come to Japan?

When the company was launched, overseas investors asked, "Can you get people here in Japan?" They were often worried. However, in reality, we have obtained excellent researchers from abroad to come to Japan. I believe Japan has retained its image as a technological superpower. There's a sense that "if you make good products, they will be used."

With the U.S. at the cutting edge of AI research, I'm constantly preoccupied about what we should be doing here in Japan. Perhaps the most important thing is not to think in terms of "overseas and Japan," but in terms of "U.S. West Coast and others." Although talent is overwhelmingly concentrated in Silicon Valley, there's a certain number of researchers who'd like to work elsewhere. Looking at it from a global perspective, you can see options taking shape outside of Silicon Valley. Google DeepMind is based in London. Mistral and other AI companies are beginning to cluster in Paris, and Toronto is home to Cohere.

Capability and use cases are key

I wonder if Tokyo can become one of these clusters. It's not only the attractiveness of the city or its livability that matters, as often mentioned by people, but also whether there are interesting AI companies that you'd want to work for matters. Therefore, we don't want to focus only on development specific to Japanese language and problems. It's important to present interesting problems that can be expanded outside of Japan. These problems must be different from those being tackled by other AI companies. The main idea is to create "a company with world-class technology that just happens to be in Japan." And I believe we can do that.

There are brilliant AI researchers in Japan as well, and Sakana AI has been able to hire them; however, there's still room to raise the bar for this group. People often talk about the need to create "homegrown AI" from the perspective of security and the like. However, the most important thing is the ability to develop AI. To build capability in Japan, we should see world-class companies willing to do business in Japan. Furthermore, the people we have are also an important draw. Our CTO, Llion Jones, is in Japan, and the second and third Llion Joneses are coming. I think it's about creating that kind of chain.

What we expect from Japan's AI policy is to create use cases (application examples) anyway. An analysis of government support packages from other countries shows that it's more effective for governments themselves to introduce AI. This is rather than working through subsidies, holding events, or tweaking the tax system. A good example would be the Netherlands, where they're using AI for visa screening. This approach will contribute to the development of AI. Most importantly, it will improve the literacy of society as a whole by assuring people that AI can be used in government. I believe this will lead to a society that uses AI, which will ultimately make it highly efficient.

Creating a vision different from the mainstream, but not too different

— Sakana AI achieved an amazing track record really quickly. I suspect that the secret behind this lies partly with your top-notch researchers, but there's your management too.

Of course, we do project management, but the key is "vision management." Vision management means grasping the current trends and providing people with a direction that is "different but not too different."

For example, companies have conventionally improved performance of AI models by inputting more data and computing resources. But these advances began to slow down around the middle of this year. The trend has shifted toward allowing models to interact with each other to improve their reasoning (logical inference) capabilities. We're also estimate that 2025 will finally usher in the era of "agents" that use AI to control individual infrastructure models.

From the beginning, we've questioned the tendency to think "the bigger, the better" when it comes to AI models. Further, we've begun researching agents and reasoning earlier than anyone else. It means we are earlier, not by "three years," but by "six months." The idea is you grasp what everyone else is doing and then start doing something different a little earlier.

Balanced role-sharing among the three co-founders

Naturally, there's no guarantee that one approach will work, so we think in terms of portfolios. At Sakana AI, as co-founders, we are constantly thinking, "What field will come next?" From there, we bring in overseas people who can realize the vision we've come up with. If we think there are 10 areas that could be expanded, we bring in one expert from each of the 10 areas to work on them. The idea is to get a breakthrough on one of them.

— So, you're saying that we need leaders who can create a vision and have the ability to draw people.

The idea is to create a "weird vision" that's different from OpenAI and so forth in the U.S., but, at the same time, not too weird to get people interested in it. Additionally, I think that the ability to draw people arises from showing them a success story that proves "Sakana AI's vision is on the mark."

Another thing is that "vision for development" is different from "vision for use cases." The first is to find a completely new way to create AI models. The second is to solve the world's problems. It's a common mistake to try to force the two together. If we don't find good use cases, disappointment in the AI industry as a whole will spread. In this sense, while use cases are extremely important, examining only use cases may nip Nobel Prize-level R&D in the bud. We, co-founders, divide our roles and maintain a balance between them.

Open source is extremely important

— Since December, you've been one of the seven members of the AI Strategy Council set up by the Tokyo Metropolitan Government.

My background is in the Ministry of Foreign Affairs, and I still have a desire to "make Japan a technology superpower again." People overseas still hold this image of Japan as a country devoted to the Walkman and Nintendo, with advanced technology. That's why I think, "Let's dream the dream of technology" once again.

In addition, Sakana AI is an open-source company. Open source is a world of give and take. Evolutionary Model Merge and The AI Scientist have features that were all created by others. So, we're more than willing to "give" what we make to the public right away. It's extremely important for the community to think, "There's a positive trade-off to letting these guys use open-source content."

In the same way, the city of Tokyo is the external environment—the ecosystem we live in. We hope to contribute to this ecosystem through avenues like the AI Strategy Council.