MIT Technology Review published a profile on 8 September 2026 of Danijar Hafner — a 31-year-old researcher who left Google DeepMind in autumn 2025 and now runs a stealth startup in San Francisco. Hafner is behind the Dreamer series, agents that learn inside a world model rather than in the real environment.
Key takeaways
- Hafner left Google DeepMind in autumn 2025, the startup operates in San Francisco's SoMa district
- PlaNet: an agent that plans actions inside a learned model of its environment
- DreamerV2: the first world model to reach human-level performance on Atari 2600
- DreamerV3: the first autonomous solution to the Minecraft Diamond challenge
- Dreamer 4: diamond mining learned from recorded gameplay alone
A world model instead of trial and error
Classic reinforcement learning needs millions of trials in the environment. Model-based RL reverses that order.
The learning loop inside a world model
Real data improves the simulation, not the policy directly. That lets the agent predict the outcome of a move it has never made. Formally, the world model learns the distribution over the next state of the environment.
Symbol meaning
- …
- state of the environment at step t, as a latent representation
- …
- action taken by the agent at step t
- …
- learned transition model with parameters θ
Four generations of Dreamer
| Model | Breakthrough | What it proved |
|---|---|---|
| PlaNet | planning inside a learned model | an agent can plan, not merely react |
| DreamerV2 | human level on Atari 2600 | world models match model-free methods |
| DreamerV3 | Minecraft Diamond unaided | it handles long sequences under sparse reward |
| Dreamer 4 | diamond from recordings alone | training with no environment interaction at all |
The hardest of these thresholds fell to DreamerV3, because Minecraft Diamond combines a long chain of dependent steps with a sparse reward. A separate project, DayDreamer, carried these algorithms onto physical robots.
Who is behind it
Danijar Hafner grew up in rural north-eastern Germany, the child of classical musicians. A neighbour taught him to program. He landed a researcher role at Google Brain in 2015, in his second undergraduate year at the Hasso Plattner Institute in Potsdam. He then moved through a dozen positions at Google Brain and Google DeepMind across the UK, Canada and the US, working with Geoffrey Hinton and Ashish Vaswani among others.
He easily sits in the top half of 1%.
Timothy Lillicrap, researcher at Google DeepMind and co-author of the DreamerV3 paper.
Why it matters
World models are the one serious alternative today to scaling an LLM, if the goal is an agent acting in a physical setting. The difference is practical: a language model predicts the next token, a world model predicts the next state of the environment.
Dreamer 4 showed this can be trained from video alone, which removes robotics' most expensive ingredient — collecting data on real hardware. Humanoid robots imported from China sit in the startup's office, so the direction is legible.
What's next
- The startup remains in stealth — no name, no disclosed investors and no announced product
- Humanoid robots imported from China standing in the office point to world-model testing on physical hardware, but the company has confirmed no deployment
- MIT Technology Review discloses neither team size nor a date for the new group's first publication
Sources
- MIT Technology Review — This AI entrepreneur is developing agents that can plan ahead for the unexpected
- arXiv — Mastering Diverse Domains through World Models
