Sakana AI Researcher Interview (March 2025 Media Feature)

A Japanese computer vision magazine, Computer Vision Frontline Spring 2025 from publisher Kyoritsu Shuppan, sat down with three researchers from Sakana AI this March. Takuya Akiba, Yuichi Inoue, and Hayate Kuroki talked about why they signed on, what a normal week looks like, and how they think about doing research that actually matters. Sakana AI shared the opening portion of that conversation with permission from the publisher.

Each of them came from a different place. Kuroki had been studying multi-agent systems and collective intelligence, and he says Sakana AI's nature-inspired approach lined up with wanting to build genuinely smart robots grounded in foundation models. Inoue held a PhD, worked previously at a self-driving startup as its first employee, and describes being pulled in after hearing that CEO David was pursuing research he found genuinely interesting. Akiba was the company's first full-time hire and has been working in deep learning since around 2016; for him, the pull was simpler — he wanted to do something fun with the current wave of foundation models, and he says the caliber of people at Sakana AI is, for him, unmatched in Japan.

All three describe their day-to-day as research-first, closer in feel to a university lab than a typical tech company — light on management structure, heavy on freedom to explore. Kuroki, who calls himself a research engineer, says he typically juggles one main project alongside one or two side projects, with actual implementation work mixed into the research. Inoue mentioned a concrete output from that setup: a new Japanese vision-and-language model built using a model-merging technique that Akiba had developed, which was released publicly.

Asked what actually produces original value in research, none of them pretended there's a formula. Akiba was blunt that if there were an easy answer everyone would already be doing it, and that whatever counts as distinctive today can stop being distinctive the moment others copy it — which is why he treats the question as something the team has to keep revisiting rather than solve once. Kuroki pointed to having specialists nearby across different fields, saying that kind of exposure sparks the ideas that eventually turn into something new. Inoue agreed, adding that people with backgrounds unlike his own — the ones behind projects like the model merging work and an AI scientist effort — bring perspectives he wouldn't land on alone.

Akiba added one more thing he cares about when building a team: adaptability. Interests that feel exciting at first can go stale fast as the environment shifts, so he'd rather have people willing to drop what they liked yesterday and chase whatever seems most interesting right now, even if that means contradicting themselves within a day. The interviewer, Waseda doctoral student Hideki Tsunashima, floated a common assumption afterward — that Sakana AI mainly hires through personal connections. The company pushed back on that directly, noting many of its researchers came in through its standard online application page. Since its 2023 founding, the company has centered on research, and as of early March 2025 it has also stood up a business development division to start turning that research into actual products, with hiring plans expanding beyond the research team.