Ex-DeepMind researcher's stealth startup builds agents that plan for the unexpected

A bare office, hung with humanoids

According to MIT Technology Review, Danijar Hafner's brand-new startup occupies a nearly empty office in San Francisco's SoMa district: no company name on the door, little furniture, and racks of humanoid robots imported from China suspended like marionettes across the open space. Hafner, 31, left Google DeepMind in the fall of 2025 to found the company, which remains in stealth mode.

He is deliberately vague about the venture, but frames it as a continuation of his long-running effort to let AI operate in environments it never encountered during training. The humanoids are the physical embodiment of that work. As the profile points out, a robot sent into a person's home has to cope with a floor plan and furniture it has never seen before.

Acting inside a world model

Hafner's method is model-based reinforcement learning. He builds world models — AI models designed to emulate physical reality — and trains agents inside them. The agent treats the model as a stand-in for the real world, learns how to act there, and then uses those experiences to predict future outcomes, something Hafner describes as dreaming or imagining. That ability is what lets agents, and the robots they are embedded in, navigate unfamiliar situations in real life.

MIT Technology Review notes that the technique differs from standard robotics practice: it allows agents to execute massively complicated tasks without the real-world trial-and-error training that has traditionally been required.

From Atari to Minecraft to physical robots

Hafner grew up in a rural town in northeastern Germany, learned programming from a neighbor, and got hooked on AI through online courses in high school. In 2015, as a second-year engineering undergraduate at the Hasso Plattner Institute in Potsdam, he became a student researcher at Google Brain. He went on to roughly a dozen internships and positions across Google Brain and Google DeepMind in the UK, Canada and the US, working with figures including Geoffrey Hinton and Ashish Vaswani, coauthor of the transformer-defining paper "Attention Is All You Need."

His research record is a progression of agent milestones. PlaNet let agents take actions by planning ahead. Dreamer 2 was the first agent to reach human-level performance on Atari 2600 games using a world model. Dreamer 3 was the first to solve the Minecraft Diamond challenge, mining the game's gems on its own. Dreamer 4 went a step further, learning to mine diamonds from an offline dataset of recorded gameplay videos without ever interacting with the game directly.

He has since carried the approach into physical reality. His DayDreamer project used the Dreamer algorithm to let robots operate themselves in novel environments and react to new experiences — such as being pushed over — without any specific training for those events.

Timothy Lillicrap, a former manager and coauthor at Google, told MIT Technology Review that Hafner stands out even among elite researchers, estimating he sits in the top half of 1 percent, and that he would single-handedly build things that would take entire teams of engineers.

Hafner is not revealing what comes next, only hinting at the scale of his ambition: "I was interested in solving a problem that would change the world."

Why it matters

Agent reliability is a central obstacle to putting AI to work outside the sandbox. Systems that only perform when reality matches their training data break down exactly where autonomy matters most. Hafner's world-model approach offers a different bet: agents that rehearse inside learned simulations of reality and then imagine their way through situations they have never faced. If that transfers from Atari, Minecraft and lab robots to general-purpose humanoids, it could remove the slow, expensive trial-and-error training that has long constrained robotics and shorten the path to machines useful in homes and workplaces. That a researcher of this pedigree has left DeepMind to commercialise the idea signals that planning for the unexpected is increasingly seen as a product direction, not just an academic one.