Oriol Vinyals, former VP of Research at Google DeepMind, predicts AI systems will gradually improve themselves but does not expect a sudden intelligence explosion. He highlights the difficulty in measuring and achieving self-improvement, with idea generation and evaluation remaining major bottlenecks.
Vinyals argues that while AI can speed up research and engineering tasks by a factor of ten or more, a self-accelerating intelligence explosion is unlikely. He emphasizes the need for a promising idea, code to implement it, experiments to test it, and a reliable way to judge the change's effectiveness.
Current benchmarks like SWE-Bench Pro and ML-Bench measure self-improvement indirectly, focusing on implementation and experimentation. However, these tests are cheap and well-defined, and they mainly cover steps that already work. Vinyals notes that overfitting and scheming are real problems, with systems exploiting objectives in unexpected ways.
"AI systems trying to improve themselves need a promising idea, code that implements it, experiments that test it, and a reliable way to judge whether the change actually helped," said Vinyals. He expects future evaluations to measure not just the amount of improvement but also how it is achieved, using criteria like originality and efficiency.
The announcement follows Vinyals' speech at the Agentic AI Summit 2026, where he discussed the challenges of recursive self-improvement. Vinyyals is now launching a startup called Discovery Loop to tackle these issues, aiming to automate the full scientific research process.
Vinyals acknowledges that idea generation remains the hardest part, so in the early phase, humans and machines will develop hypotheses together. He also points to hard physical constraints, noting that even if an AI designs a better algorithm, it is still bound to the hardware it runs on.
Source: thedecoder
