Google Shake Up! Jeff Dean Leaves Google For Discovery Loop As Hassabis Steps Up To DeepMind Chair

Jeff Dean, Google's chief scientist and a 27-year veteran of the company, is leaving to co-found Discovery Loop alongside fellow Google veterans Sanjay Ghemawat, Oriol Vinyals and Quoc Le. The new public benefit corporation aims to automate machine learning, science and engineering research, with Google itself investing in the venture and providing cloud infrastructure. Dean told the New York Times that leaving a public company gives him more freedom to pursue decisions that aren't necessarily in the company's purest financial interest. It's a striking departure for one of the most senior technical figures in AI, and Google's share price fell nearly four percent on the news. Demis Hassabis is stepping down as CEO of Google DeepMind to become its Chairman and Chief Scientist of Alphabet, a move that lets him focus on long-term strategy and scientific breakthroughs rather than day-to-day operations. He remains CEO of Isomorphic Labs, DeepMind's drug discovery spinout. Koray Kavukcuoglu, previously DeepMind's chief technology officer, steps up to senior vice president of the unit, now reporting directly to Alphabet CEO Sundar Pichai. Combined with Dean's departure, it's the biggest reshuffle of Google's AI leadership in years, arriving right as the company tries to keep pace with OpenAI and Anthropic. Elsewhere, Figure introduced Hark Handoff, which the company says has been independently verified as the best internet-use model built so far, beating both ChatGPT 5.4 and Claude Opus 4.8. Rather than chasing coding benchmarks, Figure is targeting everyday tasks like ordering food, booking flights, shopping and general web navigation, an area of agent development that has been comparatively under-served while most industry effort goes into coding assistants for developers. Meta has released Muse Code in beta, a terminal coding agent that takes on complete software engineering tasks across large codebases, handling planning, writing code and validating the results itself. It's powered by Muse Spark 1.2, a coding-focused update to Meta's model line, and marks Meta's most direct entry yet into the terminal coding agent space after largely sitting out the race so far. Separately, Prime Intellect released Prime Agent, a general-purpose coding harness that scores 95.5 percent on ARC-AGI-3, beating the human-expert baseline. Prime Intellect says it saw major improvements across multiple models when compared against their own proprietary harnesses, a reminder that the harness wrapped around a model can matter almost as much as the model itself.