This AI entrepreneur is developing agents that can plan ahead for the unexpected

MIT Technology Review’s 2026 Innovators Under 35 profile finds Hafner, 31, in a mostly empty SoMa office stocked with humanoid robots imported from China—and little else. The venture is still unnamed and in stealth, framed as a continuation of his work on agents that cope with environments absent from training data.

The method is model-based reinforcement learning: build world models that emulate physical reality, train agents inside those simulations, then use imagined rollouts to act when the real scene is new. That pitch targets a core robotics failure mode—homes and workplaces that never match the lab floor plan.

Hafner’s public record is unusually concrete for a stealth founder. PlaNet planned ahead in learned models; Dreamer 2 reached human-level Atari via a world model; Dreamer 3 solved Minecraft’s diamond challenge; Dreamer 4 learned diamond mining from offline gameplay video without direct game interaction; DayDreamer moved the algorithm onto robots that adapt to novel settings, including being pushed over. The startup is the physical bet that this lineage transfers beyond games and lab demos.

Signal Desk's read: The strategic wager is sample-efficient generalization—complicated tasks without the usual real-world trial-and-error grind—embodied in humanoids rather than another software-only agent company. That is ambitious and still lightly specified. Technology Review notes he will not say much about the venture; the robots on racks are the tell, not a product sheet. Timothy Lillicrap’s praise (top half of 1% among strong Google researchers; solo builds that would take teams) supports founder quality, not product readiness.

What is easy to overstate is inevitability of home robots. Unfamiliar floor plans are necessary, not sufficient; safety, cost, and reliability remain unaddressed here. What is easy to understate is the research continuity: leaving DeepMind in fall 2025 after a long arc from Google Brain student researcher (2015) through work adjacent to figures like Geoffrey Hinton and Ashish Vaswani suggests a deliberate timing to industrialize world-model agents, not a casual spinout.

Hafner grew up in rural northeastern Germany, learned programming from a neighbor, studied engineering at Hasso Plattner Institute in Potsdam, and stacked internships across Google Brain and DeepMind sites in the UK, Canada, and the US before founding the startup.