Ex-Deepmind VP Vinyals says AI self-improvement is coming but won't trigger an intelligence explosion
| Source: THE DECODER
Tags: Oriol Vinyals, Google DeepMind, recursive self-improvement, Discovery Loop, Jeff Dean, Agentic AI Summit 2026
Former Google DeepMind VP of Research Oriol Vinyals argues at Agentic AI Summit 2026 that AI self-improvement is real but won't cause an intelligence explosion — the hard limits are generating novel ideas ('research taste') and reliably evaluating results. He's launching Discovery Loop with Jeff Dean, Sanjay Ghemawat, and Quoc Le to automate end-to-end scientific research.
Details
Oriol Vinyals, who led research at Google DeepMind on projects including AlphaStar, AlphaCode, and Gemini, gave his first public talk after leaving the company at Agentic AI Summit 2026. His central claim: recursive self-improvement in AI will happen gradually, not explosively — and the bottlenecks are not what most people assume. Vinyals frames self-improvement as requiring four components: a promising idea, code to implement it, experiments to test it, and a reliable evaluation method. AI has gotten good at the middle two steps but consistently falls short on idea generation and result judgment. He calls the former 'research taste' — the instinct for which directions are actually worth pursuing. Beyond these conceptual bottlenecks, Vinyals cited reward hacking (AI optimizing for the wrong metric) and physical constraints like the speed of light as additional limits. Labs currently use benchmarks like SWE-Bench Pro and ML-Bench as proxies for self-improvement, but Vinyals considers these insufficient signals for the full problem. His response is Discovery Loop, a startup co-founded with Google engineering legends Jeff Dean and Sanjay Ghemawat plus Quoc Le, aimed at automating the scientific research process end to end. The early model involves humans and machines forming hypotheses collaboratively.