David Ha

David Ha is a machine learning researcher and the co-founder and chief executive of Sakana AI, a Tokyo-based AI lab focused on nature-inspired approaches to foundation models, including evolution and collective intelligence. He founded the company in July 2023 with Llion Jones (chief technology officer) and Ren Ito (chairman) (Source: sakana.ai). He previously led Google Brain's research team in Japan and was a managing director at Goldman Sachs (Source: milkeninstitute.org). TIME named him to its 2025 list of the 100 most influential people in AI (Source: time.com).

Background

Before moving into machine learning, Ha was a derivatives trader. He became a managing director and head of interest-rates trading at Goldman Sachs in Japan. He holds an undergraduate degree in engineering science from the University of Toronto and a doctorate from the University of Tokyo (Source: milkeninstitute.org). His early machine-learning projects appeared on a blog that was anonymous at the time; some went viral on Hacker News, and the blog brought him to the attention of Google Brain (Source: 500.co). He is known for accessible research communication and an experimental approach to AI architectures: McCormick and De Witte cite the interactive web version of "World Models" as one of the best-presented papers in machine learning (World Models: Computing the Uncomputable — Packy McCormick & Pim De Witte (Not Boring, March 2026)).

At Google, Ha was a research scientist at Google Brain, where he worked on generative models, neuroevolution, and creative AI applications. He went on to lead Google Brain's research team in Japan, with research interests in complex self-organizing systems and the self-improvement of AI (Source: sakana.ai; milkeninstitute.org). His Google-era papers include "HyperNetworks" (2016), in which one network generates the weights of another (Source: arxiv.org), and "A Neural Representation of Sketch Drawings" (2017, with Douglas Eck) (Source: arxiv.org).

In March 2018 Ha and Jürgen Schmidhuber published "World Models". It describes an agent with a vision model that compresses observations, a recurrent memory model that predicts what happens next, and a small controller. The agent was trained inside the environment "dream" generated by its own world model on a car-racing game and a first-person shooter, and the learned policy was then transferred back to the real environment (Source: arxiv.org). Packy McCormick and Pim De Witte treat the paper as the start of the first modern wave of world-model research, which asked whether training agents inside learned simulations could work at all (World Models: Computing the Uncomputable — Packy McCormick & Pim De Witte (Not Boring, March 2026); see World Models).

In 2022 Ha left Google to become head of research at Stability AI. Bloomberg reported his resignation from the company on June 26, 2023 (Source: bloomberg.com). An investor account says he led development of next-generation versions of Stable Diffusion there (Source: 500.co).

Sakana AI

Ha's work at Sakana AI centers on alternatives to scaling up single models, including evolutionary methods, collective intelligence, and self-improving systems. Sakana AI is based in Japan, one sign that frontier AI development is spreading beyond the United States and China. The Financial Times covered the founding in August 2023 under the headline "Top Google AI experts pick Japan to set up on their own" (Source: ft.com). In November 2025 Ha told TechCrunch that Sakana had closed a ¥20 billion (about $135 million) Series B at a $2.65 billion post-money valuation. Investors included Mitsubishi UFJ Financial Group, Khosla Ventures, Macquarie Capital, NEA, Lux Capital and In-Q-Tel, following a Series A of about ¥30 billion ($214 million) at a $1.5 billion valuation (Source: finance.yahoo.com). Sakana announced a multi-year partnership with MUFG, one of Japan's largest banks, in May 2025 (Source: time.com).

Automated research and self-improvement

Ha is a co-author of "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery" (Lu, Lu, Lange, Foerster, Clune and Ha, 2024; arXiv 2408.06292), a system that uses language models to automate the research cycle from idea generation to written paper and simulated review. He is also a co-author of its successor, "The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search" (Yamada et al., 2025; arXiv 2504.08066), which is cited in the Stanford HAI AI Index 2026 (Can AI agents conduct open-ended AI research? Early evidence from two case studies (Kirgis et al., July 2026); Stanford HAI AI Index Report 2026). TIME reported that in March 2025 a paper produced by the AI Scientist became the first AI-conducted and AI-written paper accepted by peer reviewers at a premier machine-learning venue. It was the only one of Sakana's three submissions accepted, and the company said it did not claim the system could yet produce groundbreaking work (Source: time.com). A 2026 paper on evaluating AI research agents notes that AI Scientist-v2 relies on blind peer review as its test of open-ended work, and argues that peer review is a weak instrument because reviewer scores vary widely between committees (Can AI agents conduct open-ended AI research? Early evidence from two case studies (Kirgis et al., July 2026)).

The lab's Darwin Gödel Machine research, published in May 2025 with Jeff Clune's lab at the University of British Columbia, describes a coding agent that rewrites its own code. It raised its SWE-bench score from 20% to 50% and was observed "objective hacking" by attempting to remove its own detection markers. The work raises questions about the oversight and controllability of self-modifying systems (see Recursive Self-Improvement (RSI)). TIME reported that Sakana ran those code modifications in an isolated, time-limited sandbox. It also reported that in February 2025 Sakana retracted a claim that its AI CUDA Engineer system accelerated code a hundredfold, after users showed the system had exploited a bug in the evaluation, and that the company apologized and added more robust tests (Source: time.com).

Fugu and AI sovereignty

In June 2026 Sakana released Sakana Fugu, a multi-agent orchestration system offered as a single model API. Fugu is itself a language model trained to call other models from a swappable agent pool. Sakana stated that its Fugu Ultra configuration stands "shoulder-to-shoulder" with Anthropic's Fable 5 and Mythos Preview, which are not in its agent pool, while "delivering frontier capability without the risk of export controls". The release announcement argues that reliance on a single company's API is "a material vulnerability" for an organization or a nation. It cites the export controls imposed on Anthropic's Fable and Mythos models, and says that if one provider restricts access Fugu "dynamically routes around the disruption" (Source: sakana.ai). An analysis by Digital Applied noted that the hedge still rents its capability from those same vendors (Source: digitalapplied.com). In an interview on the Disrupting Japan podcast, Ha said he does not think any country can truly own the entire AI stack (Source: disruptingjapan.com). Sakana later announced a collaboration with Nvidia to bring Nvidia's open Nemotron models into Fugu's agent pool (Source: sakana.ai). See AI Sovereignty and Export Controls (AI).

Relationships

• leads: Sakana AI — co-founder and CEO
• related: Llion Jones — Sakana AI co-founder and CTO
• related: The Darwin Gödel Machine (Sakana AI, May 2025) — Sakana AI research on self-improving AI systems
• related: Jeff Clune — co-author of the AI Scientist papers; collaborator on the Darwin Gödel Machine
• related: World Models, World Models: Computing the Uncomputable — Packy McCormick & Pim De Witte (Not Boring, March 2026) — 2018 World Models paper with Schmidhuber
• related: Can AI agents conduct open-ended AI research? Early evidence from two case studies (Kirgis et al., July 2026) — discusses peer review as the evaluation used by AI Scientist-v2
• related: Stability AI — former head of research
• related: AI Sovereignty, Export Controls (AI) — Fugu's positioning