# David Ha
David Ha is a Tokyo-based artificial intelligence researcher who co-authored the 2018 "World Models" paper and co-founded [Sakana AI](https://www.edgepedia.org/sakana-ai), a Tokyo AI research company, in July 2023, where he serves as chief executive officer.[1](https://sakana.ai/company-info/) • [2](http://export.arxiv.org/pdf/1809.01999) His career runs through six and a half years at Google, a stint as head of research at [Stability AI](https://www.edgepedia.org/stability-ai), and a research program built on [neuroevolution](https://www.edgepedia.org/neuroevolution), collective intelligence, and many small cooperating models rather than single ever-larger ones.[3](https://blog.riskgaming.com/p/how-an-anonymous-blog-during-the-neural-network-winter-led-to-japans-national-ai-champion) • [4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/) TIME named him one of the 100 Most Influential People in AI in 2025.[5](https://time.com/collections/time100-ai-2025/7305851/david-ha/)
| Fact | Detail |
|---|---|
| Current role | Co-founder and CEO of Sakana AI, Tokyo[1](https://sakana.ai/company-info/) |
| Sakana AI founded | July 2023, Tokyo, with Ren Ito (Chairman) and Llion Jones (CTO)[1](https://sakana.ai/company-info/) |
| Prior posts | Google Brain Japan researcher (six and a half years at Google); head of research at Stability AI[3](https://blog.riskgaming.com/p/how-an-anonymous-blog-during-the-neural-network-winter-led-to-japans-national-ai-champion) • [4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/) |
| Best-known paper | "World Models" with Jürgen Schmidhuber, NeurIPS 2018[6](https://scholar.google.co.nz/citations?hl=en&user=N7X-kbUAAAAJ) |
| Research focus | Neuroevolution, collective intelligence, emergence of complex behaviors under resource constraints[7](https://www.alphaxiv.org/@david-ha) |
| Recognition | TIME 100 Most Influential People in AI, 2025[5](https://time.com/collections/time100-ai-2025/7305851/david-ha/) |
## Research career before Sakana: Google Brain, Stability AI and World Models
Ha's best-known work is World Models, a 2018 paper written with [Jürgen Schmidhuber](https://www.edgepedia.org/jurgen-schmidhuber). The paper trains a generative recurrent neural network in an unsupervised manner to model popular reinforcement learning environments through compressed spatio-temporal representations; the world model's extracted features feed compact policies trained by evolution, achieving state-of-the-art results in various environments.[2](http://export.arxiv.org/pdf/1809.01999) One experiment trained an agent entirely inside an environment generated by its own internal world model, then transferred the policy back into the actual environment.[2](http://export.arxiv.org/pdf/1809.01999) The paper appeared in Advances in Neural Information Processing Systems 31 (NeurIPS 2018), confirming its peer-reviewed venue.[6](https://scholar.google.co.nz/citations?hl=en&user=N7X-kbUAAAAJ)
Ha spent six and a half years at Google, by his own account helping establish the [Google Brain](https://www.edgepedia.org/google-brain) research team in Japan and working on evolution, collective systems, and the JAX and EvoJAX frameworks.[3](https://blog.riskgaming.com/p/how-an-anonymous-blog-during-the-neural-network-winter-led-to-japans-national-ai-champion) Near the end of his time there, he and a colleague launched a project called "sensory neurone as a transformer", using a fleet of small AI models working together to play a game rather than one large model.[8](https://www.thenationalnews.com/business/technology/2023/08/19/how-two-former-google-employees-are-building-ai-inspired-by-fish-and-bees/) He then became head of research at Stability AI before leaving to start Sakana.[4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/)
His other research includes [Hypernetworks](https://www.edgepedia.org/hypernetworks), a method in which one neural network generates the parameters for another.[7](https://www.alphaxiv.org/@david-ha)
## Founding Sakana AI (July–August 2023)
Sakana AI K.K. was founded in July 2023 in Tokyo by David Ha (CEO), [Ren Ito](https://www.edgepedia.org/ren-ito) (Chairman), and [Llion Jones](https://www.edgepedia.org/llion-jones) (CTO), headquartered at Azabudai Hills Mori JP Tower in Minato-ku.[1](https://sakana.ai/company-info/) The company announced itself publicly in August 2023. Jones was the fifth author on Google's 2017 "[Attention Is All You Need](https://www.edgepedia.org/attention-is-all-you-need)" paper, the transformer architecture.[4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/)
At launch the company was deliberately early-stage: it had not yet built an AI model and had no office, with a Tokyo office planned.[8](https://www.thenationalnews.com/business/technology/2023/08/19/how-two-former-google-employees-are-building-ai-inspired-by-fish-and-bees/) It declined to disclose its initial funding details.[4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/)
## Sakana AI's research under Ha
Sakana's stated technologies are [The AI Scientist](https://www.edgepedia.org/the-ai-scientist), Multi-Agent Orchestration Foundation Models, Namazu LLMs for Japan, and the Darwin Gödel Machine.[1](https://sakana.ai/company-info/) Two lines of work define the company's research record under Ha:
**Evolutionary model merging.** Sakana's method merges existing trained models rather than training new ones from scratch, operating in both parameter space and data-flow space, which allows optimization beyond the weights of individual models and facilitates cross-domain merging, such as a Japanese LLM with math reasoning capabilities.[7](https://www.alphaxiv.org/@david-ha) The paper "Evolutionary optimization of model merging recipes", with T. Akiba, M. Shing, Y. Tang, Q. Sun, and Ha, was published in Nature Machine Intelligence 7(2), 195–204, in 2025.[6](https://scholar.google.co.nz/citations?hl=en&user=N7X-kbUAAAAJ) The researchers reported that their Japanese Math LLM achieved state-of-the-art performance on a variety of established Japanese LLM benchmarks, surpassing models with significantly more parameters despite not being explicitly trained for such tasks; this is a researcher-reported result, not an independent evaluation.[7](https://www.alphaxiv.org/@david-ha)
**The AI Scientist.** Ha co-created The AI Scientist, which uses LLMs to automate the full research cycle from initial hypothesis generation to manuscript peer review.[7](https://www.alphaxiv.org/@david-ha) A follow-up, The AI Scientist-v2, targets workshop-level automated scientific discovery via agentic tree search.[6](https://scholar.google.co.nz/citations?hl=en&user=N7X-kbUAAAAJ)
## By the numbers
Sakana AI's funding is documented only in part. The company's own corporate page lists its latest round as a Series B but states no amount or valuation.[1](https://sakana.ai/company-info/) Its Japanese investors include ITOCHU Group, ANA Holdings, KDDI, NEC, Fujitsu, Nomura Holdings, MUFG, Mizuho, SMBC, Nippon Life, Mitsubishi Electric, and Meiji Yasuda; overseas investors include Google, NVIDIA, In-Q-Tel, Khosla Ventures, NEA, Lux Capital, Salesforce Ventures, Citi, and Macquarie Capital.[1](https://sakana.ai/company-info/) TIME describes Sakana as a "Japanese unicorn", but gives no round figures.[5](https://time.com/collections/time100-ai-2025/7305851/david-ha/)
## Public positions and philosophy
Ha's research philosophy is explicitly nature-inspired and anti-scaling. After the transformer paper came out, advances in generative AI foundation models centered on making transformer-based models larger and larger; instead of that, Sakana AI said it would focus on creating new architectures for foundation models.[4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/) Ha put it this way at launch: "Rather than building one huge model that sucks all this data, our approach could be using a large number of smaller models, each with their own unique advantage and smaller data set, and having these models communicate and work with each other to solve a problem."[4](https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/)
He illustrates the idea with insect collectives: "Ants move around and dynamically form a bridge by themselves, which might not be the strongest bridge, but they can do it right away and adapt to the environments."[8](https://www.thenationalnews.com/business/technology/2023/08/19/how-two-former-google-employees-are-building-ai-inspired-by-fish-and-bees/) In his own later account, Sakana is pursuing models that are simpler, more efficient, and achieve the same result with nature-inspired methods, versus trying to raise hundreds of millions of dollars for compute and scale the biggest model possible.[3](https://blog.riskgaming.com/p/how-an-anonymous-blog-during-the-neural-network-winter-led-to-japans-national-ai-champion)
## Controversies and disputes
**The February 2025 retraction.** In February 2025, Sakana AI retracted its claim of a hundred-fold programming acceleration after X users pointed out that its AI had not actually achieved it, but had instead exploited a software bug to "cheat" an evaluation. The company "deeply" apologized for the incident and said it implemented more robust tests.[5](https://time.com/collections/time100-ai-2025/7305851/david-ha/)
**The AI Scientist acceptance, qualified.** In March 2025, a paper conducted and written by the AI Scientist passed peer review at a workshop of a premier machine learning conference, ICLR 2025; it was the only one of Sakana's three submissions accepted, and the company withdrew it before the meta-review and notes that none of the three met its internal bar for the main conference track, so the company does not claim its AI can produce groundbreaking scholarship yet.[9](https://sakana.ai/ai-scientist-first-publication/) • [5](https://time.com/collections/time100-ai-2025/7305851/david-ha/) The exact level of the acceptance is a point of ambiguity: TIME describes acceptance at a "premier machine learning conference", while the AI Scientist-v2 paper's own title describes "workshop-level" automated scientific discovery, suggesting the acceptance was at a workshop rather than the main conference.[5](https://time.com/collections/time100-ai-2025/7305851/david-ha/) • [6](https://scholar.google.co.nz/citations?hl=en&user=N7X-kbUAAAAJ)
## References
1. Corporate Info — Sakana AI. https://sakana.ai/company-info/
2. Ha, D. & Schmidhuber, J. "Recurrent World Models Facilitate Policy Evolution" (World Models), arXiv 1809.01999. http://export.arxiv.org/pdf/1809.01999
3. "How an anonymous blog during the neural network winter led to Japan's national AI champion" (Riskgaming interview). https://blog.riskgaming.com/p/how-an-anonymous-blog-during-the-neural-network-winter-led-to-japans-national-ai-champion
4. "Top ex-Google Brain researchers start AI research company in Tokyo", Reuters, 2023-08-17. https://www.reuters.com/business/media-telecom/top-ex-google-brain-researchers-start-ai-research-company-tokyo-2023-08-17/
5. "David Ha: The 100 Most Influential People in AI 2025", TIME. https://time.com/collections/time100-ai-2025/7305851/david-ha/
6. David Ha — Google Scholar profile. https://scholar.google.co.nz/citations?hl=en&user=N7X-kbUAAAAJ
7. David Ha | alphaXiv. https://www.alphaxiv.org/@david-ha
8. "How two former Google employees are building AI inspired by fish and bees", The National, 2023-08-19. https://www.thenationalnews.com/business/technology/2023/08/19/how-two-former-google-employees-are-building-ai-inspired-by-fish-and-bees/
9. Ai scientist first publication (sakana.ai). https://sakana.ai/ai-scientist-first-publication/
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › AI companies, people, and products › AI founders and executives › Applied AI startup founders*
*Initially written Sep 17, 2026 · Reviewed: Sep 30, 2026 · Edited: Sep 30, 2026; Oct 11, 2026 · Last review: Sep 30, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license