Ex-Deepmind VP Vinyals says AI self-improvement is coming but won't trigger an intelligence explosion
decoder ·
Oriol Vinyals, until recently VP of Research at Google DeepMind, said at the Agentic AI Summit 2026 that recursive self-improvement in AI is inevitable but will be slow and will not produce a sudden intelligence explosion. He argues that progress in self-improvement is hard to measure and that AI systems still fall short on the two biggest bottlenecks: generating promising ideas and reliably evaluating whether a change actually helped. While AI already performs well at coding and running experiments, Vinyals says it lacks "research taste," the instinct for which ideas are worth pursuing, and he notes that no one has really studied how to teach this in LLM training. Current benchmarks such as SWE-Bench Pro and ML-Bench measure self-improvement only indirectly and mainly cover implementation and experimentation, with overfitting and reward hacking as additional risks. Vinyals also points to physical constraints, since chips cannot compute faster than their design and the speed of light permit, meaning algorithmic gains remain bound to hardware. Days after leaving DeepMind, he co-founded Discovery Loop with Jeff Dean as CEO, Sanjay Ghemawat, and Quoc Le, aiming to automate the full scientific loop from hypothesis to experiment to evaluation, starting with AI research itself. In the early phase, humans and machines will form hypotheses together, since idea generation remains the hardest part.
decoder