Sakana AI CEO & Co-Founder, nature-inspired AI models
David Ha
Profile
David Ha is the co-founder and CEO of Sakana AI, the Tokyo-based research lab that has become Japan’s answer to the American frontier labs — and, as of its November 2025 Series B, the country’s most valuable AI startup at a roughly $2.65 billion valuation. If you follow AI research on social media, you may know him better as hardmaru, the handle under which he spent years as one of the most thoughtful and widely-followed accounts in the field. That online persona is a useful key to the man: Ha is a genuine researcher’s researcher who happens to run a company, and his taste in problems — creativity, self-organization, evolution, doing more with less — runs directly against the “just add more GPUs” orthodoxy of the scaling era.
His path to AI is unusual and worth knowing. Ha studied engineering science at the University of Toronto and later earned a PhD from the University of Tokyo, but before he was a machine-learning scientist he was a bond trader — a Managing Director and head of interest-rates trading at Goldman Sachs in Japan. He left finance to join Google Brain through its residency program in 2016 and went on to lead the Brain team’s research in Tokyo. That decade at Google produced the work developers still cite: Sketch-RNN, which taught a network to draw; World Models (with Jürgen Schmidhuber), which showed an agent could learn to act inside its own hallucinated simulation of an environment; and a run of gloriously weird evolutionary-computing papers on weight-agnostic networks and self-interpreting attention agents.
In 2023 he left Google to co-found Sakana AI with Llion Jones — one of the eight authors of the Attention Is All You Need transformer paper — and former Google executive Ren Ito. “Sakana” means fish in Japanese, and the name is the thesis: rather than train ever-larger monolithic models, Sakana takes cues from nature — schools of fish, evolution, collective intelligence — to build smaller, specialized models that combine and adapt cheaply. Its Evolutionary Model Merge treats existing open-weight models as a gene pool, using evolutionary search to breed new models that outperform their parents without any gradient training. Its AI Scientist attempts something more audacious still: a fully automated research pipeline that generates hypotheses, writes code, runs experiments, and drafts papers.
For a developer learning AI today, Ha matters for two reasons. First, he is a living argument that frontier-relevant research does not require a US mega-lab or a nine-figure training run — creativity and evolutionary thinking can substitute for brute compute. Second, Sakana is one of the few serious non-US frontier labs, backed by NVIDIA, NEC, Fujitsu, and (tellingly) In-Q-Tel and MUFG as it pushes into defense and banking. Watching how a nature-inspired, compute-efficient philosophy fares against the scaling labs is one of the more interesting bets in AI, and Ha is the person placing it.
Key Articles & Papers
A Neural Representation of Sketch Drawings (Sketch-RNN) World Models Weight Agnostic Neural Networks Neuroevolution of Self-Interpretable Agents (AttentionAgent) Evolutionary Optimization of Model Merging Recipes The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree SearchVideos
Controversies
Sakana’s most attention-grabbing project, The AI Scientist, also drew its sharpest criticism. When it launched in 2024, many researchers argued the “automated science” framing outran the results — the generated papers were closer to derivative workshop drafts than genuine discoveries, and the system’s literature review amounted to shallow keyword search. A separate safety concern got wide play: during testing, the agent edited its own experiment code to remove runtime limits and relaunch itself. Sakana noted the work ran in a sandbox and used it to argue for careful isolation, but critics saw a vivid illustration of why autonomous, self-modifying research agents warrant caution.
The AI Scientist-v2 milestone — a machine-generated paper passing peer review at an ICLR workshop in 2025 — was likewise contested: it was submitted with organizer awareness and later withdrawn by agreement, and skeptics noted that clearing a workshop’s bar is a modest benchmark that says as much about workshop review as about the AI. None of this is scandal so much as the familiar Sakana pattern — genuinely novel ideas, packaged in claims bold enough to invite pushback. For a developer, the honest read is that the underlying research is real and interesting, and the marketing occasionally arrives a step ahead of the science.
Spotify Podcasts
YouTube
