Danijar Hafner, a 31-year-old researcher formerly of Google DeepMind, has launched a new startup in San Francisco. While the venture remains in stealth mode, the company is focused on advancing artificial intelligence agents capable of planning ahead to navigate environments they have not previously encountered.
The startup’s workspace features a collection of imported humanoid robots, which serve as the physical platform for Hafner’s software. His approach centers on model-based reinforcement learning, where AI agents are trained within "world models"—simulations designed to emulate physical reality. By "dreaming" or imagining future outcomes within these models, the agents can adapt to real-world scenarios, such as unfamiliar home layouts or furniture, without requiring extensive trial-and-error training.
Hafner’s career includes significant contributions to AI research, including the development of the PlaNet model and the Dreamer series. His Dreamer 3 agent notably achieved human-level performance in Minecraft, while Dreamer 4 demonstrated the ability to learn tasks solely from offline video data. His DayDreamer project further bridged the gap between virtual training and physical robotics, allowing machines to react to novel experiences like being pushed.
Before founding his startup in the fall of 2025, Hafner held numerous roles at Google Brain and Google DeepMind, working alongside prominent figures such as Geoffrey Hinton and Ashish Vaswani. Timothy Lillicrap, a researcher at Google DeepMind, noted Hafner’s exceptional technical capabilities, stating, "I get to interact with a lot of really smart people in research at Google, and he easily sits in the top half of 1%."
While Hafner has not disclosed specific details regarding his current project, he indicated that his goal is to address challenges that could have a significant global impact. Source: MIT Technology Review
暂无评论。快来发表第一条评论吧。