What happened?

MIT Technology Review has published a profile of Danijar Hafner's new startup in San Francisco. The 31-year-old researcher left Google DeepMind in the autumn of 2025, and the company he founded is still in stealth mode — its name is not even on the door.

The office is short on furniture but full of robots: humanoids of various shapes and sizes, imported from China, hang like marionettes from racks running down the middle of the open space. Hafner describes the venture as a continuation of his long-running work on enabling AI to navigate environments it has not encountered in training.

The method: world models

The approach Hafner relies on is called model-based reinforcement learning. He develops AI models designed to emulate physical reality — world models — and trains agents inside them.

The agent treats the model as a simulation of the real world and learns how to act there. It then uses those experiences to make predictions about future outcomes — to dream, as Hafner might put it. That allows agents, and the robots they are embedded in, to navigate unfamiliar situations in real life.

This is where it differs from other approaches: the method makes it possible to execute massively complicated tasks without the real-world trial-and-error training traditionally used in robotics. If you want to send a robot into someone's home, it needs to handle a floor plan and furniture it has never seen before.

The path: from games to reality

Hafner honed the approach over years by testing it against video games:

  • PlaNet — his first breakthrough; it let agents execute actions by planning ahead.
  • Dreamer 2 — the first agent to reach human-level performance on Atari 2600 games using a world model.
  • Dreamer 3 — the first to solve the Minecraft Diamond challenge, mining the gems on its own.
  • Dreamer 4 — learned to mine diamonds from an offline dataset of recorded gameplay videos, without ever interacting with the game directly.
  • DayDreamer — carried the Dreamer algorithm into physical reality; robots operated themselves in novel environments and reacted to new experiences, such as being pushed over, without specific training.

Why it matters

The detail about Dreamer 4 is the most striking item on that list. An agent learning from recorded video and then solving the task targets a practical problem in robotics directly: real-world trial and error is expensive, slow and often dangerous. A robot arm cannot learn by dropping something a thousand times.

His former manager and co-author at Google DeepMind, Timothy Lillicrap, describes Hafner as a standout among standouts: "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%. In many cases he would build, single-handedly, things it would take entire teams of engineers to build."

The limit needs stating

This is a profile, not a product announcement. The company is in stealth, its name has not been disclosed and there is no auditable result. Hafner himself is coy about his next steps, saying only that he was interested in "solving a problem that would change the world."

So the news here is not an achievement but a direction: one of the leading figures in AI research has founded his own company around an approach outside language models — world models and physical robots. What comes of it is something nobody can see yet.