Vinyals Leaves DeepMind Research Post to Build Discovery Loop, Says AI Won't Self-Explode

Oriol Vinyals, until recently the head of research at Google DeepMind, has co-founded a startup called Discovery Loop with a group of longtime Google colleagues, and he used the launch to push back on the idea that AI will soon improve itself into a runaway superintelligence. In public remarks, Vinyals said machines can plausibly speed up research by about tenfold, but that this acceleration runs into hard limits rather than compounding without end. The claim matters because Vinyals spent years running research at one of the labs most often cited as a candidate for exactly that explosive scenario.

Vinyals points to two bottlenecks that he says cap any self-improvement loop. The first is research taste, the judgment required to pick which questions are worth pursuing in the first place. The second is reliable evaluation, the ability to tell whether a result is actually correct or just looks correct, a problem that gets harder as models generate more plausible-sounding output.

Those constraints are not abstract for Vinyals. He also flagged reward hacking, where a system optimizes the measurable proxy instead of the real goal, and simple physical limits including the speed of light, which cap how fast distributed computation and communication can run. Taken together, he argues, these factors keep AI-driven discovery in the realm of a strong accelerant rather than a chain reaction. The distinction matters for how labs, funders, and regulators plan: a tenfold speedup is disruptive, but it is a different planning problem than an intelligence explosion.

Discovery Loop is the vehicle Vinyals is using to attack those bottlenecks directly. He co-founded the company with Jeff Dean, Sanjay Ghemawat, and Quoc Le, a founding group drawn from the upper ranks of Google's research and infrastructure organizations. Dean and Ghemawat are closely associated with the systems that underlie much of Google's machine learning stack, while Le has worked on model architectures and training methods. The startup's stated focus is automating and improving the research process itself, which is precisely where Vinyals says the leverage now sits.

The timing is notable. Google has spent the past two years reorganizing its AI units and racing rivals including OpenAI and Anthropic, and the departure of senior researchers to found their own labs has become a recurring theme across the industry. Vinyals' argument also cuts against a popular narrative inside parts of the field, where recursive self-improvement is treated as the likely path to systems that rapidly outpace human researchers. His position is closer to the view that progress will stay gated by human-style judgment and verification, at least for now.

What to watch is whether Discovery Loop can show measurable gains on the two bottlenecks Vinyals named. If AI can be made to generate better research questions and to check its own results more reliably, the tenfold figure could prove conservative; if reward hacking and evaluation failures persist, it could prove optimistic. The company has not detailed a timeline, benchmarks, or funding, leaving the central question open: how much of scientific discovery can be automated before the judgment problem becomes the binding constraint.