Discovery Loop: Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals launch an AI discovery venture

A new venture called Discovery Loop has publicly introduced its plan to automate scientific discovery, arguing that the experimental loop — propose an experiment, run it, examine the results, repeat — remains slow and labor-intensive in many domains because it depends on sequential human effort.

A founding team from Google's research ranks

The founding team is Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals, who describe decades of close and impactful collaboration. Collectively, the site states, they represent three of the most-cited researchers in artificial intelligence and two of the most-cited researchers in distributed systems.

Their combined track record spans Google Search, Ads, News and Translate; infrastructure such as the Google File System, MapReduce, BigTable and Spanner; and AI work including TensorFlow, Pathways, TPUs, AlphaChip, AlphaStar, AlphaCode and AlphaFold, as well as word2vec, sequence-to-sequence models, chain-of-thought reasoning, mixture-of-experts architectures and several generations of large language models.

Automating the experimental loop

Discovery Loop says it is building systems that use frontier AI models and large-scale computational infrastructure to propose, run and learn from evaluations without a human sequencing every step. The aim is to run thousands of experiments in parallel, compressing iteration time and increasing both the quantity and the quality of scientific and engineering output.

The company will start with machine learning research and engineering and act as its own first customer: automated ML capabilities will first be used to optimize its own technology stack before it expands to other domains.

From ML to engineering grand challenges

Longer term, Discovery Loop says its approach should apply to any learning loop with measurable outcomes in science and engineering. It names National Academy of Engineering Grand Challenges as eventual targets, including engineering better medicines, advancing health informatics, making solar energy economical, providing access to clean water, securing cyberspace and engineering the tools of scientific discovery.

The mission stated on the site is to build AI systems that automatically solve important problems in machine learning, science and engineering so the benefits of science and technology arrive faster. The team says it is building a lean, in-person organization around that goal.