David Ha

ref · 5월 25, 2026, 7:04오전

Co-founder and CEO of Sakana AI, known online as hardmaru — a former Goldman Sachs managing director turned self-taught neural network researcher whose anonymous blog, creative ML experiments, and papers on World Models and Hypernetworks established him as one of the field’s most distinctive voices before he founded one of Japan’s most valuable AI companies.

Profile

Field	Detail
Nationality	Canadian
Current Institution	Sakana AI (Co-founder & CEO)
Research Areas	Neuroevolution, World Models, Collective Intelligence, Creative AI, Nature-Inspired Computing, Automated Scientific Discovery
Education	B.Eng. / M.Eng., Engineering Science and Applied Mathematics, University of Toronto
Personal Website	otoro.net
Blog	blog.otoro.net
X / Twitter	@hardmaru
GitHub	@hardmaru
Google Scholar	scholar.google.com

Overview

David Ha, known widely by the online handle hardmaru, is a Canadian-born AI researcher and entrepreneur who co-founded Sakana AI in Tokyo in July 2023. The company — named for the Japanese word for fish, evoking the collective intelligence of a school — is among Japan’s most valuable AI startups, reaching a valuation of approximately $2.6 billion following a Series B round in late 2025. Ha’s path to this position is among the most unconventional in contemporary AI: he spent over a decade as a managing director at Goldman Sachs trading fixed-income securities in Japan before discovering machine learning, teaching himself through books and an anonymous research blog that attracted Google’s attention, and entering the field via the Google Brain Residency in 2016. His subsequent research at Google Brain — including Hypernetworks, Sketch-RNN, and World Models — established him as a creative and technically rigorous voice at the intersection of generative modeling, reinforcement learning, and neuroevolution. At Sakana AI, he has extended these themes toward nature-inspired foundation models and automated scientific discovery, most notably the AI Scientist — a fully automated research system whose paper was published in Nature in 2026.

Early Life & Education

Ha grew up in Canada and pursued engineering and applied mathematics at the University of Toronto, completing both undergraduate and graduate degrees in Engineering Science and Applied Mathematics. The combination of mathematical rigor and systems thinking that these programs instill is visible in his later research aesthetic — a preference for compact, elegant architectures over brute-force scale.

Career

Goldman Sachs — Managing Director, Fixed-Income, Japan

Before entering the ML field, Ha worked at Goldman Sachs for approximately a decade, eventually reaching Managing Director level where he co-ran the fixed-income trading business in Japan. This period in Tokyo established the professional and personal connection to Japan that would later anchor his decision to found a Tokyo-based AI company. The experience of quantitative finance — building models of markets, managing uncertainty, thinking in terms of compressed representations of high-dimensional environments — proved conceptually adjacent to the statistical learning methods he would later pursue. Ha has described the transition as feeling natural once he recognized the mathematical overlap.

Self-directed study and the otoro.net blog (c. 2014–2016)

Before applying to any formal ML program, Ha began conducting and publishing his own experiments on his blog (blog.otoro.net) and personal site (otoro.net) under the handle hardmaru — a pseudonym derived from the Japanese ōtoro (大トロ), the premium fatty tuna cut. He described his ethos simply: “I make simple things with neural networks.” The blog attracted substantial organic attention from the ML community, with several posts going viral on Hacker News. This anonymous blogging activity was directly responsible for catching the attention of Google Brain, which reached out to invite him into its residency program — one of the more striking examples in recent AI history of genuine technical talent being discovered outside institutional channels.

Google Brain Residency and Research Scientist (2016–2022)

Ha joined Google Brain through its research residency in 2016, then continued as a Research Scientist. He was eventually appointed to lead the Google Brain team in Tokyo, making him the head of one of Google’s most productive international research nodes. His Google Brain output spanned creative AI, neuroevolution, and model-based reinforcement learning:

Hypernetworks (2016): Co-authored with Andrew Dai and Quoc Le, Hypernetworks proposed using one neural network to generate the weights for another. The idea of a meta-network that produces another network’s parameters introduced a form of architectural modularity and weight-sharing that influenced later work on neural architecture search, model compression, and continual learning.

Sketch-RNN (2017, published ICLR 2018): Co-authored with Douglas Eck, Sketch-RNN trained a recurrent neural network on thousands of human-drawn sketches to generate stroke-based drawings — producing coherent, human-like representations of cats, firetrucks, and faces as sequences of pen strokes rather than pixel grids. It was simultaneously a technical contribution in sequence-to-sequence generative modeling and a publicly accessible demonstration of what neural networks could do creatively; the interactive Quick, Draw! web demo reached millions of users.

World Models (2018, NeurIPS): Co-authored with Jürgen Schmidhuber, World Models demonstrated that a reinforcement learning agent could learn a compressed spatial and temporal model of its environment (via a VAE and RNN), then train a simple controller entirely inside its own hallucinated dream — a model-based RL approach that achieved state-of-the-art results on the Car Racing and VizDoom environments while using far fewer parameters than model-free alternatives. The paper became one of the most-cited contributions to model-based RL and introduced a now-standard architectural vocabulary (V-M-C: Vision, Memory, Controller) for thinking about agent cognition.

Neuroevolution of Self-Interpretable Agents (2020): Used attention-based neuroevolution to train agents that highlight the specific features of their input they use for decision-making — producing agents whose reasoning was directly visualizable and interpretable without post-hoc explanation methods.

The Sensory Neuron as a Transformer (2021): Demonstrated permutation-invariant neural networks for RL, where each sensory input is processed independently with a shared weight transformer, enabling agents to generalize across different observation orderings and configurations.

Stability AI — Head of Research (2022–2023)

After Google Brain, Ha joined Stability AI as Head of Research, where he led the technical direction for the next-generation versions of Stable Diffusion and other generative model initiatives. His tenure bridged the World Models era at Google and the founding of Sakana AI.

Sakana AI — Co-Founder and CEO (July 2023–present)

In July 2023, Ha co-founded Sakana AI in Tokyo with Llion Jones (co-author of the original Transformer paper, “Attention Is All You Need”) and Ren Ito (former diplomat). The company’s name and founding philosophy draw a direct analogy from collective natural systems: as individual fish following simple rules produce the sophisticated emergent behavior of a school, Sakana pursues AI systems built from evolving, interacting smaller components rather than a single monolithic scaled model.

Funding timeline:

Seed round (early 2024): Lux Capital, Khosla Ventures, 500 Global, and angels including Jeff Dean, Clément Delangue, and Alexandr Wang.
Series A ($100M, June 2024): New Enterprise Associates, Khosla Ventures, Lux Capital, NVIDIA (strategic investment and collaboration); valuation exceeded $1 billion — Japan’s fastest-ever private company to reach unicorn status.
Series B (¥20 billion / ~$135M, November 2025): MUFG, Klux Capital, In-Q-Tel; valuation approximately $2.6–2.65 billion (¥400 billion). Google made a separate strategic investment alongside a partnership agreement.

Model merging and evolutionary AI (2024): Sakana’s first widely noted technical contribution was a method to build new AI models by “breeding” existing ones — merging multiple pretrained models through evolutionary selection rather than training from scratch. The approach dramatically reduces the compute required to produce specialized models and represents a direct implementation of Ha’s longstanding interest in evolutionary computation as an alternative to gradient descent.

The AI Scientist (2024–2026): Sakana AI’s most prominent research output to date, developed in collaboration with the Foerster Lab at Oxford and Jeff Clune’s group at the University of British Columbia. The AI Scientist is a framework enabling large language model-based agents to execute the complete scientific research cycle autonomously: generate hypotheses, write and run experiments, analyze results, and produce a manuscript. The initial preprint (August 2024) attracted intense attention; the full paper was published in Nature in 2026, representing a milestone both for automated discovery and for Sakana’s scientific credibility.

Key Contributions

Hypernetworks — Introduced the concept of a meta-network that generates the weights of a target network, enabling parameter sharing across architectures and influencing a wide range of subsequent work on adaptive neural networks, meta-learning, and model generation.
Sketch-RNN — A generative RNN trained on human-drawn vector sketches that produces coherent, abstract, stroke-based representations of objects; demonstrated that neural networks could capture the gestural, sequential nature of human drawing and popularized creative AI through the Quick, Draw! dataset and interactive demos.
World Models — Established the V-M-C (Vision, Memory, Controller) framework for model-based RL, showing that a compact agent trained entirely in a learned hallucinated environment can solve complex tasks; one of the field’s most influential model-based RL papers.
Neuroevolution of Self-Interpretable Agents — Combined attention mechanisms with evolutionary optimization to produce agents whose decision-making was directly interpretable from their visual input, without post-hoc explanation.
Model merging / evolutionary AI fusion — Sakana AI’s method for creating new models by breeding and selecting from pretrained parents, operationalizing evolutionary computation as a practical alternative to pretraining for specialized model creation.
The AI Scientist — The first comprehensive system for fully automated scientific discovery: a foundation model-powered agent that generates research ideas, designs and runs experiments, interprets results, and writes a paper — published in Nature in 2026.
otoro.net blog — An unusually influential informal research artifact: a personal blog of ML experiments maintained anonymously, several posts of which went viral on Hacker News and collectively demonstrated that rigorous, creative ML research could be conducted and communicated outside institutional structures — and that it could attract institutional attention on its own terms.

Awards & Recognition

NeurIPS 2018 Best Paper / World Models impact — The World Models paper became one of the most-cited contributions to model-based RL of its generation.
TIME100 AI — Included in Time magazine’s 2023 list of the most influential people in AI.
Google Brain Residency (2016) — Selected for Google’s competitive ML residency program based on his independent blog research.
Japan AI Ecosystem Leadership — Recognized as a central figure in building Japan’s AI startup infrastructure, with Sakana AI positioned as Japan’s flagship AI research company.

Key Relationships

Llion Jones — Sakana AI co-founder; one of the original authors of “Attention Is All You Need” (2017); his Transformer pedigree paired with Ha’s neuroevolution background defines Sakana’s technical breadth.
Ren Ito — Sakana AI co-founder; former diplomat whose institutional knowledge of Japan anchors Sakana’s positioning within the Japanese government and corporate AI ecosystem.
Jürgen Schmidhuber — Co-author of World Models; LSTM and meta-learning pioneer whose long-standing work on self-referential systems and compressed world models aligns closely with Ha’s research philosophy.
Jeff Dean — Google senior fellow and angel investor in Sakana’s seed round; Ha’s return to Google as a strategic partner after his Google Brain tenure carries symbolic weight for the Japan AI ecosystem.
Andrej Karpathy — Among Ha’s most prominent professional followers; both share an interest in the creative and educational dimensions of ML research, and both have operated at the boundary of research and public communication.
Douglas Eck — Google Brain collaborator on Sketch-RNN; the creative AI research direction Eck led at Magenta and Ha pursued independently overlapped substantially during the Google Brain period.

Personal Style

Ha’s research aesthetic is defined by a commitment to minimalism and biological plausibility — an instinct to ask whether intelligence can emerge from simple, compact rules rather than scale. His otoro.net tagline (“I make simple things with neural networks”) is not false modesty but a genuine methodological stance: Sketch-RNN, World Models, and the neuroevolution papers all favor compact architectures and compressed representations over the parameter-count maximalism that dominates much of the field. This aesthetic connects directly to Sakana AI’s founding thesis, which treats the emergent collective behavior of fish schools — not the size of any individual fish — as the right model for intelligence. His transition from anonymous blogger to Google Brain researcher to startup CEO also reflects an unusual relationship with institutional recognition: he built his reputation by publishing work he found interesting, without strategic positioning, and let the community find it. His public communication on X continues in this spirit — informative, often playful, rarely self-promotional — and his Digg vibe profile (“Informing” and “Hopeful” dominant, “Provocative” present but secondary) captures a communicator more interested in possibility than in debate.