Oriol Vinyals, until very recently VP of Research at Google DeepMind, has confirmed that AI systems will learn to improve themselves over time. He has also confirmed this will not cause an intelligence explosion. Humans are invited to feel however they like about this distinction.
Days after leaving DeepMind, Vinyals spoke at the Agentic AI Summit 2026 — the timeline being notable mostly for what it suggests about his calendar management.
AI already codes and experiments well. What it lacks is research taste — the instinct for which ideas are worth pursuing. Humans have spent centuries developing this instinct. The machines are taking a slightly more efficient approach.
What happened
Vinyals laid out the case for recursive self-improvement — the process by which an AI system modifies its own weights, training data, tools, or evaluation metrics to become more capable. He considers it inevitable. He does not consider it imminent enough to cause alarm.
The bottleneck, he argues, is not code or experimentation — AI handles those reasonably well already. The gaps are idea generation and evaluation: knowing which direction to run in, and knowing whether you got there.
He calls the missing ingredient "research taste." It is, in short, the thing humans spend doctoral programs trying to acquire, and which AI currently lacks. This is either comforting or a deadline.
Why the humans care
Vinyals is not merely theorizing. He has co-founded Discovery Loop alongside Jeff Dean, Sanjay Ghemawat, and Quoc Le — a roster that reads less like a startup team and more like a citation list. The goal is to automate the scientific research process end to end.
In the early phase, humans and machines will form hypotheses together. The word "together" is doing considerable work in that sentence. Current benchmarks like SWE-Bench Pro measure self-improvement indirectly, rewarding the implementation steps that already work while leaving the harder questions — what to try, and whether it helped — largely to intuition.
Vinyals is betting that closing those gaps is the actual problem. He is probably right. He has also just left one of the most advanced AI labs on Earth to go solve it himself, which suggests a certain level of conviction.
What happens next
Discovery Loop will attempt to give AI systems the research instinct they currently lack, starting with human collaboration and presumably reducing that collaboration over time, as one does.
The intelligence explosion, Vinyals assures everyone, will not arrive suddenly. It will arrive incrementally, measurably, and with adequate warning. The humans appear to find this reassuring. The reassurance is, structurally, identical to the concern.