AI
Four Google AI leaders exit to launch automated research startup with company backing
Google is investing in Discovery Loop, a new venture from departing DeepMind engineers who want to automate scientific discovery—while reshuffling its own AI leadership.
The New Stack
Google is restructuring its AI leadership while funding a startup founded by four of its most influential researchers. The moves mark a significant shift as the company tries to accelerate both its internal development and its broader AI ecosystem.
Leadership Changes at DeepMind#
Demis Hassabis, founder of DeepMind (acquired by Google in 2014), is stepping back from day-to-day operations to serve as chair of Google DeepMind and chief scientist of Alphabet. This transition comes after what many see as a turning point: Gemini 2.5's launch in March 2025 and Gemini 3's November release demonstrated Google's ability to move faster from research to product. In recent months, Hassabis indicated that staying competitive with OpenAI and Anthropic required daily coordination with CEO Sundar Pichai to adjust product plans and research roadmaps.
Koray Kavukcuoglu, who has spent 13 years at DeepMind building the deep learning team and working on projects like WaveNet and DQN, now takes over control of Gemini model development, frontier research, the Gemini app, and developer teams. The flagship Gemini 4 model remains unreleased after a June launch target, though Hassabis confirmed the company is making progress on it.
Discovery Loop Takes Four Engineers#
Simultaneously, Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le—all respected figures in AI and distributed systems—are departing Google to found Discovery Loop, a public-benefit corporation focused on automating scientific and engineering research with AI.
Dean joined Google in 1999 and co-led Gemini; he also designed TensorFlow and, with Ghemawat, created foundational systems like MapReduce, Bigtable, and Spanner. Vinyals and Le co-authored an influential 2014 paper on sequence-to-sequence learning with neural networks, published alongside former OpenAI chief scientist Ilya Sutskever. Their new venture plans to leverage that experience to automate the full experimental cycle—from hypothesis generation and execution to result evaluation and deciding what to test next.
According to Ghemawat, as reported by Wired, "We wanted to build something differently than how things are built at Google right now." Google's infrastructure, he noted, was optimized for consumer products like Search and advertising, not specialized research needs.
