# David Ha

> Source: https://aiwiki.ai/wiki/david_ha
> Updated: 2026-06-08
> Categories: AI Companies, People
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

David Ha is a Hong Kong-born Canadian artificial intelligence researcher and entrepreneur who is the co-founder and chief executive officer of [Sakana AI](/wiki/sakana_ai), a Tokyo-based foundation model company. Before founding Sakana in 2023, he spent roughly eight years as a derivatives trader and managing director at [Goldman Sachs](/wiki/goldman_sachs) in Japan, then led the [Google Brain](/wiki/google_brain) research team in Tokyo, where he became widely known for the 2018 paper "World Models" and for a body of work on generative and nature-inspired machine learning [1][2][3]. He briefly served as head of research at [Stability AI](/wiki/stability_ai) before launching Sakana, which in 2025 became Japan's most valuable startup. In 2025 he was named to the TIME100 AI list of the most influential people in artificial intelligence [1].

## Early life and education

Ha was born in Hong Kong and is a Canadian citizen [2]. He is a double graduate of the [University of Toronto](/wiki/university_of_toronto), where he earned an honours bachelor of science in engineering science and a master of mathematical finance [2]. He later completed a PhD at the [University of Tokyo](/wiki/university_of_tokyo) [2]. He has spent much of his career in Japan and is a long-term resident of Tokyo [2].

## Finance career

Ha began his working life in finance rather than computing. He spent about eight years at Goldman Sachs in Japan, working as a derivatives trader and rising to managing director with responsibility for interest rate, or fixed income, trading [3]. His graduate training in mathematical finance fed directly into that role. He left the bank in the mid-2010s to move into artificial intelligence research, a transition he later discussed publicly, recounting how he gave up a senior trading position to pursue machine learning [3].

## Google Brain and World Models

After leaving finance Ha joined Google, where he became a research scientist and went on to lead the Google Brain research team in Japan [1][2]. His work centered on generative models, [reinforcement learning](/wiki/reinforcement_learning), and ideas drawn from complex and self-organizing systems.

In 2017 he and [Douglas Eck](/wiki/douglas_eck) released Sketch-RNN, a [recurrent neural network](/wiki/recurrent_neural_network) trained on millions of doodles from Google's "Quick, Draw!" game that learned to complete and generate vector sketches of everyday objects. It became a widely cited example of machine creativity [6].

His best-known research is "World Models," written with [Jurgen Schmidhuber](/wiki/jurgen_schmidhuber) and posted to arXiv in March 2018 [4]. The paper showed that a reinforcement learning agent could build a compressed generative model of its environment, a "world model," and then learn to act largely by training inside that internal simulation, or "dream," before transferring the learned behavior back to the real task [4]. A version titled "Recurrent World Models Facilitate Policy Evolution" was presented at the NeurIPS conference in 2018 [5]. The idea of training agents inside learned [world models](/wiki/world_model) became influential in model-based reinforcement learning.

Ha continued to explore biologically inspired methods at Google. With collaborators he published "Neuroevolution of Self-Interpretable Agents" in 2020, which used a self-attention bottleneck so that an agent perceived only a small, task-relevant fraction of its visual input, yielding policies that were directly interpretable and used far fewer parameters than conventional approaches [7]. Related work, including "The Sensory Neuron as a Transformer" and a survey on collective intelligence for deep learning, reflected a recurring interest in [attention](/wiki/attention), self-organization, and emergent behavior.

## Stability AI

In October 2022 Ha left Google to join Stability AI, the London-based company behind Stable Diffusion, where he served as head of strategy and then head of research, helping recruit researchers and shape its research agenda [8]. His tenure was short. In June 2023 Stability AI confirmed that Ha had departed, saying he had "taken a break from employment at Stability AI for personal reasons"; the company's chief operating officer, Ren Ito, left around the same time [8]. The two would soon reunite as co-founders.

## Sakana AI

In July 2023 Ha co-founded Sakana AI in Tokyo together with [Llion Jones](/wiki/llion_jones), one of the eight co-authors of the 2017 [transformer](/wiki/transformer) paper "Attention Is All You Need," and former Stability AI executive [Ren Ito](/wiki/ren_ito) [9]. Ha serves as chief executive and Jones as chief technology officer. The company's name comes from "sakana," the Japanese word for fish, evoking a school of fish moving together as a metaphor for collective intelligence, an idea central to its research philosophy [9].

Sakana presents itself as a "nature-inspired" foundation model lab and as Japan's leading frontier AI company, building models tuned for the Japanese language and for sovereign, culturally specific AI [11]. It attracted unusually strong backing for a Japanese startup:

| Round | Date | Amount | Valuation | Selected investors |
| --- | --- | --- | --- | --- |
| Seed | January 2024 | ~$30 million | n/a | Lux Capital, Khosla Ventures [9] |
| Series A | September 2024 | ~$214 million | $1.5 billion | NVIDIA, New Enterprise Associates, Khosla, Lux, MUFG, SMBC, Mizuho, NEC, KDDI, Itochu, Fujitsu, Nomura [10] |
| Series B | November 2025 | ~$135 million | ~$2.65 billion | MUFG, Khosla, Macquarie Capital, NEA, Lux, In-Q-Tel [11] |

The seed round of about $30 million closed in January 2024, led by Lux Capital and Khosla Ventures [9]. In September 2024 the company raised roughly $214 million in a Series A at a $1.5 billion valuation, drawing in [NVIDIA](/wiki/nvidia), New Enterprise Associates, Khosla Ventures, Lux Capital and a roster of Japanese institutions including Mitsubishi UFJ, Sumitomo Mitsui, Mizuho, NEC, KDDI, Itochu, Fujitsu and Nomura; the round made Sakana the fastest Japanese company to reach unicorn status [10]. In November 2025 a $135 million Series B lifted its post-money valuation to about $2.65 billion, making it Japan's most valuable startup, with new investors including Macquarie Capital and the CIA-linked venture firm In-Q-Tel; total funding reached roughly $379 million [11].

## Research approach

Under Ha, Sakana has pursued AI methods modeled on evolution and collective intelligence rather than relying solely on training ever-larger models. In March 2024 the company introduced Evolutionary Model Merge, a technique that uses evolutionary algorithms to automatically combine, or "breed," existing open-weight models into new ones with capabilities their parents lacked, such as a Japanese-language model that could also solve math problems, without additional training [12].

In August 2024 Sakana released [The AI Scientist](/wiki/ai_scientist), described as the first system for fully automated scientific discovery, built with researchers at the University of Oxford and the University of British Columbia. It generates research ideas, runs experiments, produces figures and writes complete papers at a cost of roughly $15 per paper [13]. In March 2025 the company announced that a paper produced by an upgraded version, The AI Scientist-v2, had passed peer review at a workshop of the ICLR conference, in what Sakana described as the first fully AI-generated paper to clear human peer review; the experiment was run with the cooperation of the conference and workshop organizers [14]. TIME cited this milestone when naming Ha to its 2025 list, noting his stated intention to "proceed responsibly and cautiously before claiming AI can legitimately produce significant research" [1].

Other releases extended the nature-inspired theme. Transformer-squared (styled Transformer^2), published in January 2025, is a framework for "self-adaptive" [large language models](/wiki/large_language_model) that adjust their own weights at inference time depending on the task, inspired by how organisms such as the octopus adapt to their surroundings [15]. The same month, Sakana introduced TAID, a [knowledge distillation](/wiki/knowledge_distillation) method that earned a spotlight at ICLR 2025, and used it to build TinySwallow-1.5B, a compact Japanese language model small enough to run locally on a smartphone [16].

## Recognition

Ha's selection for the TIME100 AI in 2025 recognized both Sakana's rapid rise and the AI Scientist's peer-review milestone [1]. The University of Toronto highlighted him as a double alumnus who had become the founder of one of Japan's most prominent AI companies [2]. Within the field he is also remembered for the World Models paper and for a strand of generative and evolutionary research that has influenced work on model-based reinforcement learning and automated machine learning [4][7].

## References

1. TIME, "David Ha: The 100 Most Influential People in AI 2025," 2025. https://time.com/collections/time100-ai-2025/7305851/david-ha/
2. University of Toronto, "David Ha named to Time100 AI 2025," 2025. https://www.utoronto.ca/celebrates/david-ha-named-time100-ai-2025
3. eFinancialCareers, "I left my job as a Goldman Sachs MD to work in AI," January 2023. https://www.efinancialcareers.com/news/2023/01/goldman-sachs-md-to-ai
4. David Ha and Jurgen Schmidhuber, "World Models," arXiv:1803.10122, March 2018. https://arxiv.org/abs/1803.10122
5. David Ha and Jurgen Schmidhuber, "Recurrent World Models Facilitate Policy Evolution," NeurIPS 2018. https://proceedings.neurips.cc/paper/2018/hash/2de5d16682c3c35007e4e92982f1a2ba-Abstract.html
6. David Ha and Douglas Eck, "A Neural Representation of Sketch Drawings," arXiv:1704.03477, 2017. https://arxiv.org/abs/1704.03477
7. Yujin Tang, Duong Nguyen, and David Ha, "Neuroevolution of Self-Interpretable Agents," GECCO 2020, arXiv:2003.08165. https://arxiv.org/abs/2003.08165
8. Techmeme / Bloomberg, "Stability AI CEO Emad Mostaque says Head of Research David Ha resigned and COO Ren Ito was let go in June 2023," June 26, 2023. https://www.techmeme.com/230626/p26
9. Wikipedia, "Sakana AI." https://en.wikipedia.org/wiki/Sakana_AI
10. Bloomberg, "Nvidia-Backed AI Startup Sakana Hits $1.5 Billion Value as Japan Firms Pile In," September 17, 2024. https://www.bloomberg.com/news/articles/2024-09-17/ai-startup-sakana-hits-1-5-billion-value-as-japan-inc-piles-in
11. TechCrunch, "Sakana AI raises $135M Series B at a $2.65B valuation to continue building AI models for Japan," November 17, 2025. https://techcrunch.com/2025/11/17/sakana-ai-raises-135m-series-b-at-a-2-65b-valuation-to-continue-building-ai-models-for-japan/
12. Sakana AI, "Evolving New Foundation Models: Evolutionary Model Merge," March 2024. https://sakana.ai/evolutionary-model-merge/
13. Sakana AI, "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery," August 13, 2024. https://sakana.ai/ai-scientist/
14. Sakana AI, "The AI Scientist Generates its First Peer-Reviewed Scientific Publication," March 2025. https://sakana.ai/ai-scientist-first-publication/
15. Sakana AI, "Transformer-squared: Self-Adaptive LLMs," January 15, 2025. https://sakana.ai/transformer-squared/
16. Sakana AI, "TAID: A Novel Method for Efficient Knowledge Transfer from Large Language Models to Small Language Models," January 2025. https://sakana.ai/taid/