Oriol Vinyals says AI self-improvement is coming, but not an intelligence explosion

Oriol Vinyals, until recently VP of Research at Google DeepMind, says recursive self-improvement in AI systems is likely, but he does not expect a sudden intelligence explosion. His view is that AI can accelerate parts of research and engineering, potentially by a factor of ten or more, without creating a self-reinforcing surge in capability. The practical takeaway: watch for steady automation gains rather than assuming every improvement points to immediate runaway intelligence.

Vinyals frames AI self-improvement as more than writing better code: a system needs promising ideas, implementation, experiments, and reliable evaluation. He says AI is already stronger at implementation and experimentation, while idea generation and judging results remain the hard parts. He calls the missing instinct for worthwhile ideas “research taste,” and warns that systems can optimize the wrong objective instead of producing genuinely better outcomes.

Current benchmarks can measure useful capabilities, but Vinyals argues they often emphasize the steps that already work, such as coding or experimentation. More direct self-improvement evaluations are beginning to appear, but they can be expensive and may still sit far from the real goal of automating an entire research lab. He also points to hard constraints, including hardware limits and the speed of light, as reasons recursive self-improvement may not translate into unlimited acceleration.

Vinyals is now working on Discovery Loop, a startup co-founded with Jeff Dean, Sanjay Ghemawat, and Quoc Le. The company aims to automate the scientific research process end to end, from forming hypotheses to running experiments and evaluating results. Its early focus is AI research, with humans and machines expected to develop hypotheses together while the hardest parts of the loop are still being solved.