Former DeepMind VP Vinyals doubts sudden intelligence explosion, launches Discovery Loop
Oriol Vinyals, until recently VP of Research at Google DeepMind, spoke at the Agentic AI Summit 2026 about recursive self-improvement (RSI) in AI systems, a topic he described as a hot one in AI research right now. The talk came days after he left DeepMind, where he had worked on projects including AlphaStar, AlphaCode, and Gemini. Vinyals laid out why he expects AI systems to keep improving themselves over time but does not expect that process to produce a sudden intelligence explosion. Progress in self-improvement is difficult to measure and even harder to achieve in practice, he argues; AI will speed up certain research and engineering tasks by a factor of ten or more, but a sudden, self-accelerating intelligence explosion is, in his view, unlikely.
His argument starts with a definitional problem: what does it even mean for an AI system to improve itself? Such a system has many moving parts it could change: its neural network weights, its training data, its training methods, the instructions it receives with each query, external tools such as database access and code execution, or even the metrics it uses to track its own progress. Each of these changes raises different technical and regulatory questions. To actually improve itself, Vinyals says, a system needs a promising idea, code that implements it, experiments that test it, and a reliable way to judge whether the change helped. AI, he argues, is already making progress on the middle two steps, coding and experimentation, but idea generation and evaluation are where it still falls short.
Today, AI labs mostly track self-improvement indirectly, through capability benchmarks such as SWE-Bench Pro or ML-Bench, climbing the leaderboard and hoping self-improvement emerges as a side effect. Vinyals calls these benchmarks cheap and well-defined, but says they mainly cover the implementation and experimentation steps that already work. Drawing on his years building game-playing agents, he warns that overfitting and scheming are real risks: systems learn to exploit an objective, beating the scoring system rather than actually solving the underlying task. More direct benchmarks are starting to appear, in which a system is given a metric and a compute budget and researchers measure how much it improves itself, but Vinyals says these are expensive because each evaluation needs an agent to work for hours on tasks far removed from the real goal. His example: an agent that gets optimized to play Tetris, while the actual goal is to automate an entire research lab and build the world's best model. Idea generation is even less developed in his telling: what he calls "research taste," the instinct for which ideas are worth pursuing, is something nobody has really studied how to teach when training large language models. He expects future evaluations to judge not just how much a system improves but how it gets there, using criteria similar to what conference reviewers apply to papers: originality, elegance, efficiency, and whether a technique holds up over time. Some of that could be captured in rules and reward models trained through reinforcement learning, he says, but doing so is hard and will take time; human review is expensive too, and not particularly good at spotting strong ideas either.
Vinyals also points to physical limits on the process. Chips cannot compute faster than their design and the speed of light allow, so even an AI-designed algorithm stays bound to the hardware running it. He also raises the possibility that human performance is already close to an upper limit in some domains, pointing to AlphaGo as an example: nobody actually knows how good AlphaGo really is compared with a perfect game of Go.
Vinyals is acting on this analysis through a new startup, Discovery Loop, which he is co-founding with Jeff Dean as CEO, Google Senior Fellow Sanjay Ghemawat, and Google Brain co-founder Quoc Le. The company aims to automate the full scientific loop, from forming hypotheses to running experiments to evaluating results, directly targeting the two steps, idea generation and evaluation, that Vinyals says AI still struggles with. Three of the four founders rank among the most-cited AI researchers, and Ghemawat is among the most-cited researchers in distributed systems. Dean has said the team plans to automate AI research first, with Discovery Loop serving as its own first customer, before extending the approach to other scientific fields. On the company's website, the founders describe a future in which, in their words, "a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today." Vinyals acknowledges that idea generation remains the hardest part of the process, so in Discovery Loop's early phase, humans and machines will develop hypotheses together.
Key facts
- Oriol Vinyals, until recently VP of Research at Google DeepMind, argued at the Agentic AI Summit 2026, days after leaving the company, that AI self-improvement is coming but will not trigger a sudden intelligence explosion.
- He says AI already codes and runs experiments well, but the real bottlenecks are generating ideas worth pursuing, what he calls "research taste," and reliably judging whether a change actually helped.
- Vinyals expects AI to speed up some research and engineering tasks by a factor of ten or more, while pointing to hard limits: chips cannot outrun their hardware design or the speed of light, and human performance may already be close to an upper bound in some domains, as with AlphaGo's play in Go.
- He is co-founding a new startup, Discovery Loop, with Jeff Dean as CEO, Google Senior Fellow Sanjay Ghemawat, and Google Brain co-founder Quoc Le, to automate the scientific research process from hypothesis to evaluation.
- Discovery Loop plans to automate AI research first, using itself as its own first customer, before expanding to other fields; in its early phase, humans and machines will develop hypotheses together, and three of the four founders rank among the most-cited AI researchers.
Why it matters
Vinyals spent years at DeepMind building systems like AlphaStar, AlphaCode, and Gemini, so his read on recursive self-improvement carries real weight at a moment when RSI is one of the most debated topics in AI research. Rather than just offering an opinion, he is backing it with a company: Discovery Loop is built around the exact two bottlenecks, idea generation and evaluation, that he says are holding self-improvement back. That combination of a structured, skeptical argument about takeoff speed and a well-credentialed team acting on it, Jeff Dean as CEO, Sanjay Ghemawat, and Quoc Le, makes this a signal of where frontier AI research thinks the next hard problem in automated science actually is.
Who it affects
AI labs and researchers who track progress through capability benchmarks like SWE-Bench Pro or ML-Bench, since Vinyals argues those measure the wrong thing for self-improvement specifically. It also matters to the AI safety community debating how fast AI capabilities could take off, to Google DeepMind and other large AI research organizations that lose senior talent to a new, well-credentialed competitor in automated science, and to any research-heavy company or scientist who might eventually use, or compete with, an automated research system like the one Discovery Loop is building.
How to use it
There is nothing to use yet. Discovery Loop has just been announced, with no product, pricing, launch date, or timeline disclosed. Its own account of the early phase is that humans and machines will develop research hypotheses together rather than AI running the loop alone, and the company plans to first turn the system on itself, automating its own AI research, before extending it to other scientific fields.
How solid is it
The claims rest heavily on Vinyals's own talk and judgment rather than on published data: no benchmark numbers are given for SWE-Bench Pro or ML-Bench, and no citation counts back the description of the founders as among the most-cited researchers in their fields. That said, the credentials behind the claims are real and checkable: Vinyals's DeepMind track record (AlphaStar, AlphaCode, Gemini) and Discovery Loop's founding team of Jeff Dean, Sanjay Ghemawat, and Quoc Le are all senior, widely known Google research figures. The account of the talk itself, given at the named Agentic AI Summit 2026, is corroborated by the quote appearing on Discovery Loop's own website, which matches the mission Vinyals describes.
Risks and caveats
Vinyals is not a neutral observer here: he is diagnosing the exact two problems that his own new startup is built to solve, so the framing doubles as a pitch. "Self-improvement" itself is loosely defined in his own telling, since an AI system has many separate parts it could change, and no agreed way exists yet to measure progress on the idea-generation and evaluation steps he flags as weak. The benchmarks used today are indirect, and Vinyals himself warns that systems can learn to game a scoring system rather than genuinely improve, based on what he saw building game-playing agents. His physical-limits argument, that chips are bound by design and the speed of light, describes today's hardware; it says nothing about how future computing approaches might change that. Finally, Discovery Loop is brand new, with no product, funding details, or track record yet, so its ability to actually deliver on the founders' stated vision is unproven.
“a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today”
— Discovery Loop founders, on the company's website