Oriol Vinyals launches Discovery Loop to automate research
Former DeepMind VP Oriol Vinyals has co-founded Discovery Loop to automate scientific research, arguing that AI self-improvement will be a gradual process rather than a sudden explosion.
Oriol Vinyals, the former vice president of research at Google DeepMind, has teamed up with prominent computer scientists to launch Discovery Loop, a startup aimed at automating the scientific research process. Speaking at the Agentic AI Summit 2026 shortly after his departure from Google, Vinyals outlined his vision for the company alongside co-founders Jeff Dean, who serves as CEO, Sanjay Ghemawat, and Quoc Le. The venture plans to use its own technology to automate AI research first before expanding into other scientific disciplines.
The launch coincides with Vinyals' skeptical take on recursive self-improvement, a concept where AI models iteratively upgrade themselves. While Vinyals, who previously worked on landmark DeepMind projects like AlphaStar, AlphaCode, and Gemini, believes AI can accelerate engineering tasks by a factor of ten or more, he dismisses the idea of a sudden intelligence explosion. He argues that self-improvement is exceptionally difficult to execute and measure in practice due to physical hardware limits, the speed of light, and fundamental cognitive bottlenecks.
According to Vinyals, self-improvement requires generating ideas, writing code, running experiments, and evaluating results. While current AI systems excel at coding and experimenting, they struggle with generating novel ideas and evaluating outcomes. Existing benchmarks like SWE-Bench Pro and ML-Bench only measure implementation and experimentation. Vinyals notes that teaching AI what he calls "research taste"—the instinct for identifying which hypotheses are worth pursuing—remains an unsolved challenge. He warns that without this, systems often optimize for the wrong metrics, such as mastering Tetris instead of solving broader scientific problems.
Discovery Loop aims to solve these bottlenecks by automating the entire scientific cycle from end to end. Because generating ideas remains the most difficult hurdle, the startup will initially have humans and machines collaborate to form hypotheses. The ultimate goal is to enable small teams to conduct high-quality research much faster than massive organizations of human scientists can today.
This is our own summary of reporting by The Decoder
