Newshot

田渊栋’s startup co-founder: The fastest path to AGI can only be AI building AI | CoRL 2026

AI summary

Recursive co-founder Jeff Clune: “QD, open-endedness, and AI-generated algorithms—I think they are the next generation of great AI.” On November 12, the final day of CoRL 2026, Jeff Clune, a professor at the University of British Columbia and a founding member of Uber AI Labs, was invited to speak at a workshop on Continually Self-Improving. In May this year, 田渊栋, former Meta FAIR research director and lead of ELF OpenGo, founded Recursive Superintelligence with seven other co-founders, and Jeff Clune is one of them.

Translation(model translation, not the source's own text)

Article title: 田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

Body:

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

November 12, the final day of CoRL 2026. The main conference had ended, and 32 Workshops were running back to back. One of them focused on Continually Self-Improving, and Jeff Clune, a professor at the University of British Columbia and a founding member of Uber AI Labs, was invited to give a talk.

In May this year, 田渊栋 (former Research Director at Meta FAIR and lead of ELF OpenGo) co-founded Recursive Superintelligence with seven other co-founders, and Jeff Clune is one of them. The company raised $650 million in its first round at a post-money valuation of $4.65 billion, co-led by Google GV and Greycroft, with both compute giants 英伟达 and AMD betting on it at the same time.

Most teams flying the banner of recursive self-improvement (RSI) are working at the post-training level: self-rewriting code, self-optimizing prompts, self-searching agent workflows. Recursive is the only company in the world doing RSI at the pre-training stage. 田渊栋 recently revealed that their system can already write high-performance GPU Kernels on its own—understand this as the low-level compute programs of a graphics card, the core code that talks directly to the hardware—and on some tasks it writes better than professional engineering teams.

Jeff Clune's research on self-evolution was also aimed from the very beginning at the most upstream link: model training. In 2019, he published a paper that seemed almost like science fiction at the time, proposing that the fastest path to AGI is to build an "AI that can build AI," which he named AI-generating algorithms. The roadmap has three pillars: architectures that evolve themselves, learning algorithms that evolve themselves, and automatically generated training environments. Note the wording of the third pillar, "training environments," not "inference environments" or "deployment environments."

Six years later, this roadmap is being cashed out one item at a time: the Darwin Gödel Machine, which can rewrite its own code, pushed SWE-bench from 20% to 50%; the AI scientist, which can set its own problems, run its own experiments, and write its own papers, had submitted manuscripts pass peer review. Jeff Clune laid out his work in more detail in the talk.

The following is an edited transcript of the on-site talk brought to you by AI Tech Review (Leiphone.com (official account: Leiphone.com)).

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

There is a paradox in life: the harder you fixate on a difficult problem, the more likely you fail; if you ignore the big goal for a while, you can actually go further.

In the history of science and technology, every extremely hard thing has the same property. In the era of cooking with fire, if all you wanted was to cook faster, you would never build a microwave. Before computers existed, if you only funded people who could calculate faster, you would never get computers. Because getting computers required someone to work on electricity and vacuum tubes, and these technologies were not invented with "calculating faster" as their goal at all.

So the only way to solve hard problems is to act like a basic scientist: try what you can do and what you can solve, then switch between goals, picking new goals based on what you think is "possible" and "interesting."

The idea of "switching goals" comes from a 2016 paper. Suppose you work on robots and you want the robot to move forward, but the algorithm starts learning to crawl and to stand on one leg for balance. Don't throw these away as useless little moves. Although they don't noticeably speed up forward locomotion, they may be stepping stones toward it. We want the algorithm to catch this kind of "hey, this is interesting" discovery.

This is the idea of innovation engines. Natural evolution and cultural evolution are both innovation engines, and they share one recipe: an archive holds everything that has ever been invented; take something out, modify and combine it slightly, and if it becomes a better version, put the better one back into the archive.

Today I'll talk about three classes of algorithms, all trying to reproduce this power of innovation: quality diversity (QD) algorithms, open-ended algorithms, and AI-generating algorithms.

QD does not seek a single optimum; instead it asks across the whole space: what fundamentally different good solutions can I actually find for this problem?

The most popular and most representative is MAP-Elites, which I built more than a decade ago. First pick a few variation dimensions you care about (such as height), add a performance metric (such as speed), then describe the robot with a parameter vector, evaluate its metrics, and place it into the corresponding cell of a multi-dimensional archive. Then start the loop: take one out, modify it, re-evaluate it, and if it is the best ever seen in that cell, swap it in, and repeat. What it gives you is not a single point but an entire surface—where it is good, where it is bad, where it can go, all visible.

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

In 2015 we used QD to make soft robots. Classic optimization could only obtain low-performance solutions; deliberately encouraging diversity actually produced less diversity, because the system found high-performance solutions and was unwilling to trade performance for diversity. Switching to QD, with the same compute, the exploration coverage was completely different. You can even "do science" inside these archives: trace the final best solution in a cell back through its ancestors, and you'll find the search took a very winding path.

In 2021 we again used QD to solve the hard exploration problem in reinforcement learning: reach as many states as possible with the highest quality. We used this to play Montezuma's Revenge, and it blew the ceiling off the field, surpassing the human world record, clearing all levels along the way, and beating all the best exploration methods at the time.

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

But no matter how strong QD is, it is still trapped in a single environment. This leads to open-ended algorithms, the thing I have studied my whole life. We want an algorithm that can keep innovating forever—even if it runs for a billion years, you would still want to come back and look, because it is still innovating.

Whoever can crack this, I think, deserves a major prize. As far as I know, there are currently only two open-ended algorithms: natural evolution and cultural evolution. The key is that every problem solved opens new problems, rather than converging toward the score of some benchmark.

A few years ago at Uber AI Labs we built POET, the first time the environment itself was treated as part of the genome, letting the environment archive keep growing while also allowing goal switching. Interestingly, POET invented and solved many very hard and very different problems.

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

Let's step back and look at the ambition of the whole field: AGI, ASI, RSI, call it whatever you like. In a 2019 paper on AI-generating algorithms, I pointed to a trend: the more compute and data there is, the more hand-designed processes and components get replaced by learned ones. Apply this logic to AGI, and the fastest path to AGI is an AI that can build a better AI. I called it AI-generating algorithms then; today people call it recursive self-improvement.

To make this work, at least three pillars must move forward: architecture (Transformer will not be the last), learning algorithms, and automatically generated training environments. Today I will mainly talk about the third pillar.

Truly open-ended algorithms need a huge environment search space, but the vast majority of environments are either impossible, too hard, or too easy. How do you ensure a new environment is "interestingly different" from the past? Now you can just ask a foundation model: I've done these 12 environments, what is the next interesting skill? This lets you bypass the long-standing difficulty of "how to define interesting."

Proposing tasks is not enough; you also need to know how to automatically design a reward for any task and determine whether it has been solved. This is called a general reward function. One path is to feed task videos to a video-language model; another is to have GPT write the reward code on the spot. Uniform sampling and single metrics both fail, but OMNI will gradually learn a smart curriculum.

But we are still trapped in the kitchen. To be more ambitious, we need to enter a "Darwin-complete" search space.

One is code. We built OMNI-EPIC, letting the system generate any environment with code, and so we get these obstacle courses, and it can also generate logic puzzles, programming problems, and robots that clean restaurants—the types can be very different.

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Building AI | CoRL 2026

The second is the neural network itself, a project called Genie, a world model I built with colleagues at DeepMind. You give it text, images, or sketches, and it generates an environment. The first generation was only one second of blurry 2D video; nine months later Genie 2 had 3D worlds; eight months after that, Genie 3 was basically a AAA-game-level renderer, and Waymo uses it to train autonomous driving. Imagine how strong Genie 4, 5, 10 will be?

Next, look at how general this playbook is. First, the automatic design of agent systems: previously people hand-built complex agent workflows; better to directly search over agent systems and let AI build better AI, and the result is that they quickly surpassed all hand-designed scaffolds of the time. Second, we let an archive of high-performing agents keep growing; it gets better, and often it has to get worse first before it gets better. Further ahead is Hyperagents, letting agents modify their own code and also improve their ability to self-improve. The same idea is also used for continual learning: learn to take notes on tasks, and then use them on the spot when a similar task comes up later.

Finally, the AI scientist. A paper we published in Nature a few months ago discussed a question: can AI complete the entire process of writing a paper? The answer is yes.

CoRL 2026 | 田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026

It can generate ideas, design and run experiments, analyze and visualize, write up manuscripts, and peer-review itself. We took some of its papers to peer review, they were accepted, and at the workshop, their scores were higher than 45% of the human submissions from back then. And there is a scaling law: the better the model, the better it does. So the future will not be just some paper, but AI scientists continuously reading their own discoveries, one innovation after another.

This is one of the hottest topics in AI right now: recursive self-improvement. A friend and I founded a company called Recursive, aimed squarely at open-ended AI. In the ten years of image generation, from utterly terrible to nearly perfect, we are probably somewhere in the middle now, and recursive self-improvement is probably just barely at the point of being usable.

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026

Finally, I want to say that QD, open-endedness, and AI-generated algorithms—these things will keep creating stepping stones, switching goals, and learning to innovate, and are very likely the fastest path to superintelligence. So I encourage you to think about how to use these ideas in your own research field; I think they are the next generation of great AI.

▎Q: Your method seems stuck on the characteristics of the environment—the diversity and scale of these features, and the cost of evaluation. How do you think this should be solved?

A: Many people have already used these ideas on robots, including the projects I mentioned earlier. You can also take what we did in our Nature paper, learn a high-quality task library, and then figure out how to use it for transfer. You can train in simulation with a massive, diverse set of tasks to acquire many skills; try them in the real world; then additionally run a small algorithm to train only on the things that transfer well.

The above is a refined on-site summary of the key takeaways. For the full transcript of the talk, you can add the author's contact information below.

The top international robotics conference CoRL 2026 has already taken place. Researchers, students, and readers interested in the latest AI progress are welcome to join our Leiphone.com community group.

Here, you can meet peers, exchange research, share admissions information, scientific collaboration, dinner meetups, submissions, teaming up to attend the conference, and more, and find potential partners.

During the conference, we will also continue to bring on-site reporting updates, interpretations of guests' talks, summaries of frontier directions, and forum highlights.

? Group entry portal: Scan the QR code to join the group or add WeChat szhsn610, with the note: CoRL + unit/school + name + direction.

田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026
田渊栋 Startup Partner: The Fastest Path to AGI Can Only Be AI Creating AI | CoRL 2026

Leiphone.com original article, unauthorized reproduction prohibited. See the reprint notice for details.

  • Facebook AI Research researcher 田渊栋: Applying artificial intelligence to ...
  • Martin Davis's latest interview: Machine learning is a process of convergence, ...
  • The strongest general board-game AI, an interpretation of the AlphaZero reinforcement learning algorithm
  • JD Digits discloses its full federated learning strategy layout for the first time, Bo Liefeng explains the two ...

Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 雷锋网
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗