AI Researcher's Stealth Startup Targets Agents That Anticipate Surprises
A new San Francisco venture aims to build AI agents capable of adaptive planning in unpredictable environments.
Danijar Hafner's office in San Francisco's SoMa neighborhood is a study in minimalism—mostly vacant desks, sparse furniture, and a door that bears no company name. The startup, still operating in stealth, has just a handful of people on site, including Hafner himself, who previously gained recognition for research in reinforcement learning and world models.
The core idea behind the venture is to develop AI agents that don't just react to immediate inputs but can think ahead and adjust their plans when unexpected events occur. Hafner's approach draws on his academic work, which focused on teaching machines to simulate possible futures and use those simulations to guide decisions in complex, changing settings.
While details remain scarce, the empty office underscores the early stage of the project. Hafner's track record suggests a technical focus, but the broader goal is clear: pushing AI beyond static responses toward systems that handle real-world unpredictability with foresight and flexibility.