Google AI Leaders Exit to Launch Discovery Loop

Four of Google’s most influential artificial intelligence researchers are departing the search giant after decades of service to launch a new startup called Discovery Loop. The founding team includes Jeff Dean, who spent nearly 27 years at Google as the cofounder of Google Brain, chief scientist of Google DeepMind, and technical colead on the Gemini model. Joining Dean, who will serve as CEO, are his long-time collaborator Sanjay Ghemawat, DeepMind VP of research and Gemini colead Oriol Vinyals, and Google Brain cofounder Quoc Le, the mind behind the AutoML-Zero project.

Operating as a public benefit corporation, Discovery Loop aims to build systems where AI acts as an autonomous researcher. The startup plans to automate the scientific method by running thousands of iterative loops of proposing, implementing, and evaluating experiments. Initially, Discovery Loop will focus on improving its own machine learning algorithms, with Le noting the potential to discover entirely new transformer architectures. Eventually, the team plans to generalize these autonomous loops to accelerate breakthroughs in chip design, biology, drug discovery, and material science.

The venture has secured backing from major Silicon Valley investors, including Khosla Ventures and Radical Ventures, whose managing partner Jordan Jacobs will join the board. Despite the massive talent loss, Google is participating as a founding investor. Alphabet CEO Sundar Pichai attempted to retain the engineers, but Google ultimately agreed to take an equity stake, provide cloud compute resources for the startup's first year, and collaborate on a machine learning research framework.

For AI practitioners and researchers, this migration represents a massive shift in where foundational AI research is conducted. If Discovery Loop succeeds in deeply automating the research process, small, agile teams of developers could soon leverage these autonomous loops to out-invent massive corporate research laboratories. By shifting AI from a tool used by human researchers to an autonomous researcher in its own right, the startup could fundamentally change how new model architectures and scientific discoveries are generated.