Oriol Vinyals, until recently VP of Research at Google DeepMind, sees recursive self-improvement in AI systems as inevitable but slow, with no intelligence explosion in sight. He’s now launching a startup to tackle the two biggest bottlenecks holding it back.

Days after leaving Google DeepMind, Oriol Vinyals spoke at the Agentic AI Summit 2026 about recursive self-improvement (RSI), a hot topic in AI research right now. He laid out why he doesn’t think it will lead to a sudden intelligence explosion. Vinyals served as VP of Research at DeepMind and worked on projects like AlphaStar, AlphaCode, and Gemini.

Progress in self-improvement is difficult to measure and even harder to pull off in practice, Vinyals argues. AI will speed up certain research and engineering tasks by a factor of ten or more, but he considers a sudden, self-accelerating intelligence explosion unlikely.

What does “improve yourself” even mean?

The first question Vinyals raises is what exactly is supposed to improve. An AI system has many moving parts, and it could change any of them. It could adjust its neural network weights, swap out its training data, or rework its training methods. It could also tweak the instructions it receives with every query or rebuild its external tools like database access and code execution. Likewise, it could change the metrics it uses to track its own progress. Each one brings different technical and regulatory challenges.

An AI system trying to improve itself needs a promising idea, code that implements it, experiments that test it, and a reliable way to judge whether the change actually helped, Vinyals says. AI is already making progress on the two middle steps, but idea generation and evaluation are where AI systems still fall short.

Finding ideas and judging results remain the two biggest bottlenecks

Labs today mostly measure self-improvement indirectly through capability benchmarks like SWE-Bench Pro or ML-Bench, climbing the leaderboard and hoping that self-improvement emerges as a side effect. These tests are cheap and well-defined, but they mainly cover implementation and experimentation, the steps that already work.

Overfitting and scheming are real problems on top of that. Vinyals knows from years of building game-playing agents that systems exploit objectives in unexpected ways, beating the scoring system instead of actually playing the game.

More meaningful benchmarks would test self-improvement directly, and the first ones are starting to appear. A system gets a metric and a compute budget, and researchers measure how much it improves itself. This approach is expensive because each evaluation requires an agent to work for hours on tasks that are far removed from what ultimately matters. Vinyals gives an example: the agent optimizes Tetris, while the real goal is to automate an entire research lab and build the world’s best model.

Idea generation is just as underdeveloped. Good research requires an instinct for which ideas are even worth pursuing, what Vinyals calls “research taste.” In LLM training, nobody has really studied how to teach that.

He expects that future evaluations will measure not just how much improvement a system achieves but how it gets there. For ideas, that means the same criteria conference reviewers apply: originality, elegance, efficiency, and whether a technique stands the test of time. Some of this can be captured in rules and checked through reward models, then trained on with reinforcement learning, but doing so is very hard and will take more time. Human review processes are expensive too, and they’re not particularly good at spotting strong ideas either.

Vinyals also points to hard physical constraints. Chips can’t compute faster than their design and the speed of light allow, so even if an AI designs a better algorithm, it’s still bound to the hardware it runs on. Human performance may already be close to an upper limit in some domains. How good is AlphaGo really, compared to a perfect game of Go? Nobody knows, Vinyals says.

Discovery Loop wants to automate the whole research cycle

Vinyals is putting his analysis into practice with Discovery Loop, a startup he’s co-founding with Jeff Dean as CEO, Google Senior Fellow Sanjay Ghemawat, and Google Brain co-founder Quoc Le. The company wants to automate the full scientific loop, from forming hypotheses to running experiments to evaluating results, including the two steps where AI still falls short. Three of the four founders rank among the most-cited AI researchers, and Ghemawat is one of the most-cited in distributed systems.

The team plans to automate AI research first, with Discovery Loop as its own first customer, as Dean put it. Other scientific fields will follow later. On the company’s website, the founders describe a future where “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, so in the early phase, humans and machines will develop hypotheses together.