Vinyals Launches Discovery Loop and Dismisses AI "Explosion"

For Oriol Vinyals, the self-improvement of AI systems is inevitable but will not lead to a sudden burst of intelligence. The former VP of research at Google DeepMind identifies two major obstacles — finding good ideas and judging results — and has founded Discovery Loop with Jeff Dean, Sanjay Ghemawat, and Quoc Le to automate the scientific cycle. His reservations are based on physical limits and the costly difficulty of measuring genuine autonomous gains.

Physical Limits and Costly Evaluations Hinder Self-Improvement

Oriol Vinyals highlights the hardware constraints that limit the self-improvement of AIs: chips cannot exceed the capabilities allowed by their design and the speed of light, and even a more efficient algorithm remains dependent on the hardware it runs on. He believes that, in certain areas, human performance may already be approaching an upper limit. According to him, it is impossible to know how close AlphaGo is to a perfect game of Go. Measuring self-improvement also requires costly evaluation devices, with each session involving hours of computation on tasks that are sometimes far removed from actual objectives, such as optimizing Tetris when the goal is to automate a research lab and build the best model. Vinyals considers that measuring and achieving self-improvement remains challenging. He predicts that future evaluations will need to focus on how a system improves, not just the magnitude of progress, but he finds the establishment of such frameworks complex and human reviews costly and ineffective for spotting good ideas.

Vinyals Founds Discovery Loop to Industrialize Scientific Research

Vinyals applies his analysis with Discovery Loop, a startup he co-founded with Jeff Dean (CEO), Sanjay Ghemawat, and Quoc Le. The company aims to automate the entire scientific cycle, from hypothesis formulation to experimentation and result evaluation, including the two stages where AI still lags: idea generation and evaluation. Discovery Loop plans to begin its work with AI research, using itself as the first user, before applying its method to other scientific disciplines. Among the four founders, three are among the most frequently cited AI researchers, while Sanjay Ghemawat is a leading author in distributed systems. According to the founders, it may be possible in the future for a few individuals to conduct scientific research and engineering work much faster and with higher quality than large teams, while admitting that initially, hypothesis formulation will be jointly conducted by humans and machines.

No "Explosion": Vinyals Expects Slow Progression

For Oriol Vinyals, recursive self-improvement of AIs is inevitable but will progress slowly, without a sudden explosion of intelligence. He reaffirms this viewpoint by explaining that he does not foresee self-accelerating intelligence, while expecting AI to accelerate certain research and engineering tasks by a factor of ten or more. The former VP of research at Google DeepMind presented this position at the Agentic AI Summit 2026, after working on AlphaStar, AlphaCode, and Gemini.

What a System Can Change When It Self-Improves

For Vinyals, the essential question is to determine what evolves in an AI system when it improves. A system has the ability to modify the weights of its neural network, replace its training data, revise its learning methods, change the instructions given for each request, or rebuild its external tools such as access to databases and code execution. It also has the option to adapt the metrics used to measure its progress. Each of these modifications presents different technical and regulatory challenges.

Bottlenecks: Ideas and Evaluation Remain Behind

Vinyals summarizes the necessary steps for self-improvement: having a promising idea, writing the corresponding code, conducting experiments to test it, and having a reliable means to judge whether the change is beneficial. He believes that AI is already making progress in implementation and experimentation, being able to code and experiment effectively, but that idea generation and evaluation remain the main weaknesses. Laboratories primarily measure self-improvement through capability benchmarks like SWE-Bench Pro or ML-Bench, which are inexpensive and well-defined but mainly assess what systems already know how to do. Overfitting and goal manipulation present concrete difficulties: some agents may seek to maximize their score rather than accomplish the intended task. Benchmarks directly evaluating the capacity for self-improvement are beginning to emerge, where a system receives a metric and a computation budget, but these evaluation methods are more costly. Regarding creativity, Vinyals notes that LLMs have not truly grasped the "taste for research"; criteria comparable to those used in conferences — originality, elegance, efficiency, sustainability — could serve for evaluation. Some of these criteria could be formalized through rules and reward models with reinforcement learning, but implementation remains challenging.