Jeff Clune
Last reviewed
Jun 3, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 2,058 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 2,058 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jeff Clune is a computer scientist known for research on open-endedness, evolutionary algorithms, deep reinforcement learning, and what he calls "AI-generating algorithms" (AI-GAs), the idea of building systems that learn to produce general AI on their own rather than having humans hand-design every piece [1][2]. He is a professor of computer science at the University of British Columbia and a Canada CIFAR AI Chair at the Vector Institute, and over his career he has held senior research roles at OpenAI, Uber AI Labs, and Google DeepMind [1][3]. In January 2026 he became a co-founder of Recursive (also reported as Recursive Superintelligence), a startup building self-improving AI [1][4].
His name is attached to a string of well-cited results: the 2014 study showing that deep neural networks can be easily fooled, the MAP-Elites quality-diversity algorithm, the Go-Explore exploration method published in Nature, OpenAI's Video PreTraining (VPT) agent that learned to play Minecraft from internet video, and recent work on self-improving agents such as Automated Design of Agentic Systems and the Darwin Godel Machine [5][6][7].
| Item | Detail |
|---|---|
| Education | BA Philosophy, University of Michigan (1999); MA Philosophy (2005) and PhD Computer Science (2010), Michigan State University [2] |
| PhD advisor | Charles Ofria [8] |
| Current roles | Professor, Computer Science, UBC (2024–); Canada CIFAR AI Chair and faculty, Vector Institute (2022–); co-founder, Recursive (2026–) [2] |
| Prior roles | Research Team Manager, OpenAI (2020–2022); Senior Research Advisor, DeepMind (2023–2025); Senior Research Manager and founding member, Uber AI Labs (2017–2019) [2] |
| Research focus | Open-ended and AI-generating algorithms, quality-diversity, neuroevolution, deep RL [1][3] |
| Citations | ~46,800; h-index 64 (Google Scholar, 2026) [5] |
Clune grew up in Michigan and came to computer science through philosophy rather than engineering. He earned an honors BA in philosophy from the University of Michigan in 1999, then an MA in philosophy from Michigan State University in 2005 [2]. He stayed at Michigan State for a PhD in computer science, completed in 2010, working at the BEACON Center for the Study of Evolution in Action under Charles Ofria [2][8]. His dissertation centered on evolving artificial neural networks with generative encodings, a thread that runs through much of his later work.
After his doctorate he moved to Cornell University as a postdoctoral fellow in the Department of Mechanical and Aerospace Engineering, where his advisor was Hod Lipson [2]. The Cornell years produced some of his most-cited early biology-adjacent papers, including a 2013 study on "The evolutionary origins of modularity" in the Proceedings of the Royal Society B, which asked why biological networks tend to be modular and tested the hypothesis in silico [8].
Clune joined the University of Wyoming as an assistant professor of computer science in 2013, later holding the Harris Associate Professorship there [2]. While at Wyoming he also took an industry turn. From 2015 to 2016 he was a senior leader at Geometric Intelligence, a machine learning startup founded by Gary Marcus, Zoubin Ghahramani, Kenneth Stanley, and Douglas Bemis [2][9]. (Contrary to some accounts, Clune did not found Geometric Intelligence; he was a senior member of its research team.) When Uber acquired the company in December 2016, its people formed the core of Uber AI Labs, and Clune became a founding member and senior research manager of the new lab, a role he held from 2017 to 2019 [2][9].
In 2020 he moved to OpenAI as a research team manager, where he led work on open-endedness and multi-agent learning until 2022 [2][10]. He returned to academia at UBC, first as an associate professor from 2021 and then as a full professor from 2024, while also taking on a Canada CIFAR AI Chair and faculty membership at the Vector Institute in Toronto from 2022 [1][2]. Alongside the UBC post he served as a senior research advisor at Google DeepMind from 2023 to 2025, and as a scientific advisor to Yoshua Bengio's LawZero safety nonprofit in 2025 and 2026 [2]. He is a co-author on the 2024 Science paper "Managing extreme AI risks amid rapid progress," signed by Bengio, Geoffrey Hinton, Stuart Russell, and others [8].
The organizing idea behind much of Clune's research is the AI-GA, set out in his 2019 arXiv paper "AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence" [6]. The argument starts from a pattern he sees across machine learning: hand-designed pipelines tend, eventually, to be beaten by learned ones. He extends that logic to the project of building general AI itself. Rather than hand-engineer an AGI, why not design an algorithm that learns to produce it, the way Darwinian evolution produced human intelligence on Earth?
He frames the approach around three pillars: meta-learning architectures, meta-learning the learning algorithms themselves, and automatically generating the environments and curricula that agents train in [6]. The third pillar, generating an endless stream of increasingly hard challenges, connects directly to his interest in open-endedness, the property that lets natural evolution and human culture keep innovating without ever converging on a final answer.
One of Clune's most cited contributions came in 2014, with Anh Nguyen and Jason Yosinski, in "Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images," presented at CVPR 2015 [11]. The team used evolutionary algorithms to generate images that look like meaningless static or abstract patterns to a person but that state-of-the-art image classifiers labeled, with over 99% confidence, as familiar objects like a school bus or a peacock. The result became a touchstone in discussions of adversarial examples and the brittleness of deep networks, showing that high confidence is not the same as understanding. A related line of work, "Understanding Neural Networks Through Deep Visualization" (2015), produced tools for visualizing what individual neurons in a network respond to [5].
Clune is one of the central figures in quality-diversity optimization, a family of algorithms that try to find a large collection of high-performing solutions that differ from each other, instead of a single best solution. With Jean-Baptiste Mouret he introduced MAP-Elites in the 2015 paper "Illuminating search spaces by mapping elites," which keeps an archive of the best solution found for each cell of a behavioral grid [12]. The method built on the novelty search idea developed by Kenneth Stanley and Joel Lehman, which rewards behaving differently rather than behaving well.
The same year, the team applied these ideas to robotics in "Robots that can adapt like animals," a Nature cover article with Antoine Cully, Mouret, and Danesh Tarapore [8]. A legged robot used a precomputed MAP-Elites archive of diverse walking gaits to recover quickly from damage, such as a broken leg, by searching that archive for a gait that still worked, behaving a bit like an injured animal limping its way to a strategy that keeps it moving.
Several of Clune's best-known systems sit at the intersection of open-endedness and deep RL. POET, the Paired Open-Ended Trailblazer (2019), with Rui Wang, Joel Lehman, and Kenneth Stanley, generates a growing population of obstacle-course environments at the same time as it trains agents to solve them, letting solutions transfer between environments as stepping stones; it won a best paper award at GECCO 2019 [13].
Go-Explore tackled the long-standing "hard exploration" problem in reinforcement learning, exemplified by the Atari game Montezuma's Revenge, where rewards are rare and an agent can wander for ages without stumbling on one. The approach, with Adrien Ecoffet, Joost Huizinga, Joel Lehman, and Stanley, remembers promising states it has reached, returns to them deliberately, and only then explores further. The full version was published in Nature in 2021 as "First return, then explore," and it set records on Montezuma's Revenge and was the first to score above zero on Pitfall, beating expert human players on both [7].
At OpenAI, Clune contributed to Video PreTraining (VPT), published in 2022 [14]. The team trained a small set of contractors to label about 2,000 hours of Minecraft video, used that to build a model that infers which actions produced a video, then applied it to roughly 70,000 hours of unlabeled internet footage to pretrain an agent. After fine-tuning, VPT became the first AI reported to craft a diamond pickaxe in Minecraft, a task that takes a skilled human player more than 20 minutes and tens of thousands of actions [14].
Since the rise of large language models, Clune's group has pushed open-endedness toward systems that design and improve themselves. Automated Design of Agentic Systems (ADAS), with Shengran Hu and Cong Lu and presented at ICLR 2025, introduced "Meta Agent Search," in which a meta-agent writes code for new agents, tests them, and accumulates an archive of discoveries, inventing agent designs that outperformed strong hand-built baselines on coding, math, and science tasks [15].
The Darwin Godel Machine (DGM), with Jenny Zhang, Shengran Hu, Cong Lu, and Robert Lange, extends this to agents that rewrite their own code [16]. Named after Jurgen Schmidhuber's theoretical Godel machine and Darwinian evolution, the DGM maintains an expanding lineage of agent variants and empirically self-improves rather than relying on the original's formal proofs. In experiments run with sandboxing and human oversight, it raised its score on the SWE-bench coding benchmark from 20.0% to 50.0% and on Polyglot from 14.2% to 30.7% [16]. The work was a collaboration between UBC, the Vector Institute, and Sakana AI. His group also helped develop The AI Scientist, an end-to-end system for automating machine learning research; a 2026 Nature paper, "Towards end-to-end automation of AI research," describes that line of work [2][8].
In January 2026 Clune co-founded Recursive, a startup reported under the name Recursive Superintelligence that aims to build AI systems that improve themselves in an open-ended loop, automating the frontier-AI development pipeline of evaluation, data selection, training, and research direction [1][4]. The company is led by chief executive Richard Socher, with a founding group that also includes Tim Rocktäschel, Alexey Dosovitskiy, Josh Tobin, Caiming Xiong, Yuandong Tian, and Tim Shi [4][17]. It came out of stealth on May 13, 2026, with $650 million in funding at a reported $4.65 billion valuation, in a round co-led by GV (Google's venture arm) with backing from Greycroft, Nvidia, and AMD Ventures [4][17]. Reporting on the launch singled out Clune for his AI-GAs research and the Darwin Godel Machine, which he co-wrote with several people who became Recursive employees [4][17].
Clune received a National Science Foundation CAREER Award in 2015 and a Presidential Early Career Award for Scientists and Engineers (PECASE) in 2019, the highest honor the United States government gives to early-career researchers [2]. His paper with Mouret and Lipson on the evolutionary origins of modularity was named among the top cited papers of 2013 in its journal, and POET won a best paper award at GECCO 2019 [8][13]. Several of his collaborations have appeared as cover articles in Nature and PNAS, including the camera-trap work that used deep learning to identify and count wild animals in the Serengeti, published in PNAS in 2018 [8].