Oriol Vinyals, former VP of Research at Google DeepMind, argues that recursive self-improvement in AI systems will be a gradual process rather than an abrupt intelligence explosion. He notes that while AI can accelerate specific research and engineering tasks by ten times or more, physical hardware constraints and the difficulty of evaluating progress limit the speed of development.
Vinyals highlights that current AI systems struggle with idea generation and objective evaluation. While models excel at implementation and experimentation, they often exploit scoring metrics in unexpected ways, similar to how game-playing agents bypass rules. He emphasizes that meaningful progress requires better benchmarks that directly measure self-improvement rather than relying on existing capability tests.
To address these bottlenecks, Vinyals is co-founding a startup called Discovery Loop alongside Jeff Dean, Sanjay Ghemawat, and Quoc Le. The company aims to automate the entire scientific research cycle, including hypothesis formation and result evaluation. The team plans to focus initially on automating AI research itself, with human-machine collaboration on idea generation in the early stages.
