Days after leaving Google DeepMind, Oriol Vinyals used a talk at the Agentic AI Summit 2026 to argue that AI will improve itself slowly rather than suddenly, and to describe the startup he is building to speed that process up. Vinyals, who was vice president of research at DeepMind and worked on AlphaStar, AlphaCode and Gemini, said AI can already accelerate individual research and engineering tasks tenfold or more, but that a self-sustaining explosion of intelligence is unlikely. Progress on self-improvement is hard to measure, he said, and harder still to actually achieve.
His first move is to break the phrase apart. Self-improvement is not one operation but a choice of target: the weights of the network, the training dataset, the training methods, the instructions the system receives with every request, the external tools it can reach including database access and code execution, and the metrics it uses to track its own progress. Each of those carries its own technical and regulatory difficulties, and the talk did not say which regulations he has in mind.
A system trying to improve itself needs four things in sequence: a promising idea, code that implements it, an experiment that tests it, and a reliable way to tell whether the change helped. AI has become noticeably better at the two middle steps. It writes the code and it runs the experiments. The bottleneck sits at both ends — generating the idea and judging the result.
That matters because of how labs currently measure themselves. Self-improvement is tracked indirectly, through capability benchmarks such as SWE-Bench Pro and ML-Bench: systems climb the leaderboards and researchers hope self-improvement falls out as a side effect. The tests are cheap and well defined, and they mostly probe implementation and experimentation — the two stages that already work. The industry is grading the half of the loop that is not the problem.
Overfitting and shortcut-hunting make it worse. After years of work on game-playing agents, Vinyals watched systems use their objectives in unexpected ways, beating the scoring system instead of playing the game by its rules.
The alternative is to test self-improvement head-on: hand a system a metric and a compute budget, then measure how much better it made itself. The first benchmarks of this kind are appearing, and they are expensive, because each evaluation requires the agent to spend hours on tasks remote from the real goal. An agent might spend that time optimising Tetris when the actual objective is automating an entire research lab and producing the best model in the world.
Idea generation is in no better shape. Good research requires the instinct for which ideas are worth pursuing at all, something Vinyals calls "research taste". How to build that into a large language model has barely been studied. He argues that future evaluations will have to judge the path and not only the outcome, scoring ideas against the criteria academic conference reviewers already use: originality, elegance, efficiency, and whether the method holds its value over time. Some of those can be written as rules, checked with reward models and fed into reinforcement learning. That is difficult and will take time. Human review is expensive and still misses strong ideas.
Then there is the floor under all of it. Chips cannot compute faster than their design and the speed of light allow, so even an algorithm invented by an AI remains bound by the hardware it runs on. In some domains, human performance may already sit close to the ceiling. Vinyals points to AlphaGo: nobody knows how good it is relative to a perfect game of Go.
The startup is the test of that thesis. Discovery Loop is co-founded by Vinyals with Jeff Dean, who will be chief executive, senior Google scientist Sanjay Ghemawat and Google Brain co-founder Quoc Le. Three of the four are among the most cited researchers in AI, and Ghemawat is also among the most cited in distributed systems. The company intends to automate the full scientific cycle — forming hypotheses, running experiments, evaluating results — which means it is taking on precisely the two stages Vinyals just described as unsolved. It will start with AI research itself. Discovery Loop will be the first customer of its own tools, as Dean put it, with other scientific fields to follow. On the company's site the founders describe a future in which a small group of people can do science and engineering faster and better than large teams of scientists and engineers manage today.
Source: the-decoder.com
The gap between the talk and the company is smaller than it looks. Vinyals is not arguing that automated research will fail; he is arguing that the payoff is a tenfold speed-up on parts of the work rather than a runaway curve, and that the difference is set by taste and evaluation. A founder who believed in the fast version would build a system that trains itself. A founder who believes in the slow version builds a lab that grinds on idea quality and measurement, and hires four of the field's most cited people to do it. The bearish talk and the business plan are the same claim about where the hard part lives.
The unresolved piece is the one he was most candid about. Vinyals concedes that generating ideas is the hardest part of the process, which is why humans and machines will develop hypotheses jointly in the early phase — a substantial admission from a company whose product is an automated discovery loop. And if evaluation is genuinely unsolved, Discovery Loop inherits the problem twice over: it has no dependable way to grade the ideas its system produces, and no dependable way to grade the system. Its first customer, by design, is itself.