The high-tech landscape of San Francisco’s South of Market (SoMa) district is no stranger to stealth-mode startups operating out of austere, minimalist office spaces. Yet, even by the eclectic standards of Silicon Valley, the new headquarters of Danijar Hafner’s latest venture presents an arresting tableau. On any given day, the expansive, sparsely furnished warehouse floor stands largely vacant, devoid of corporate signage, rows of desks, or the customary trappings of a heavily funded artificial intelligence firm. Instead, the central fixtures of the room are suspended from steel racks running down the spine of the wide-open space. Hanging like marionettes from these overhead frames are humanoid robots of varying shapes, sizes, and configurations, imported directly from manufacturing hubs in China.
For Hafner, a 31-year-old researcher whose pedigree includes heavy-hitting stints at Google Brain and Google DeepMind, this sparse room represents the physical crucible of a grander ambition. While the startup remains officially unannounced and cloaked in operational secrecy, Hafner’s overarching mission is clear: he is attempting to solve one of the most persistent bottlenecks in modern robotics—teaching artificial intelligence to navigate environments it has never encountered during its training phase. The humanoid robots dangling from the SoMa racks are the physical embodiment of this endeavor. If autonomous machines are ever to move beyond structured factory floors and successfully integrate into unpredictable human spaces like residential homes, hospitals, and offices, they must possess the cognitive flexibility to adapt instantly to novel floor plans, shifting obstacles, and entirely unforeseen physical scenarios.
The Mechanics of Machine Imagination: Model-Based Reinforcement Learning
To bridge the chasm between static training data and the chaotic reality of the physical world, Hafner relies on an advanced paradigm known as model-based reinforcement learning. Rather than forcing a robot to stumble through thousands of hours of physical trial and error—the traditional, computationally expensive, and potentially destructive approach that has long dominated robotics—Hafner constructs sophisticated internal representations of reality called world models.
Within these computational simulations, an AI agent is deployed to learn the underlying physics and dynamics of an environment. The agent treats the world model as an empirical sandbox, running countless internal simulations to learn optimal behaviors and anticipate future outcomes. In practical terms, the algorithm learns to "dream" or "imagine" scenarios before they happen. This predictive capability equips the autonomous agent—and by extension, the physical hardware it controls—with the capacity to react intelligently to unfamiliar environments in real time (IRL).
The efficiency of this approach marks a profound departure from standard machine learning workflows. While conventional deep learning models often require massive, continuous real-world feedback loops to adjust their weights, Hafner’s methodology enables software agents and their mechanical hosts to execute extraordinarily complex tasks without subjecting physical hardware to the wear-and-tear of iterative, real-world mistakes.
A Prodigy’s Journey: From Rural Germany to the Pinnacle of AI Research
Danijar Hafner’s trajectory toward the bleeding edge of artificial intelligence began far away from the venture capital hubs of California. Growing up in a quiet, rural town in northeastern Germany, Hafner was raised by parents who both worked as classical musicians. Despite an artistic household, his own curiosities pulled him toward the logical structures of computing. He taught himself the fundamentals of programming under the guidance of a neighborhood mentor, and by the time he reached high school, he was actively enrolled in online courses exploring the theoretical foundations of artificial intelligence.
"I was always fascinated with how thinking works," Hafner reflects, noting that computer science offered a tangible canvas upon which to model cognitive processes.
That early fascination catalyzed a rapid academic and professional ascent. In 2015, while balancing his second-year undergraduate engineering studies at the Hasso Plattner Institute in Potsdam, Hafner secured a coveted position as a student researcher at Google Brain. That initial foot in the door unlocked a prolific string of more than a dozen internships and research roles across Google’s premier artificial intelligence divisions, including Google Brain and Google DeepMind in the United Kingdom, Canada, and the United States.
During his tenure within these elite labs, Hafner collaborated directly with legendary figures who shaped the modern generative AI landscape. He worked alongside Geoffrey Hinton, widely heralded as one of the founding godfathers of modern deep learning, and Ashish Vaswani, a co-author of the landmark 2017 research paper "Attention Is All You Need," which introduced the transformer architecture powering virtually all contemporary large language models.
Peer Recognition and the Evolution of the Dreamer Architecture
Among the researchers who witnessed Hafner’s technical output firsthand, the consensus is that his capabilities place him in an exceedingly rare tier of computer science talent. Timothy Lillicrap, a prominent researcher at Google DeepMind who previously managed and co-authored papers with Hafner, speaks of his former colleague with unequivocal praise.
"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%," Lillicrap observes. "In many cases he would build, single-handedly, things it would take entire teams of engineers to build."
Hafner methodically built his reputation by scaling his world-model algorithms and testing them against progressively difficult digital challenges, often utilizing popular video games as proxy environments for complex physical spaces. His early academic breakthroughs included PlaNet, a pioneering model that enabled software agents to execute complex behaviors by planning several steps ahead based on learned environment dynamics.
Following PlaNet, Hafner spearheaded the development of the Dreamer series, which systematically pushed the boundaries of reinforcement learning. Dreamer 2 achieved a major milestone by becoming the first agent to reach human-level performance on Atari 2600 games utilizing an internal world model. The subsequent iteration, Dreamer 3, crossed another major computational threshold by solving the notorious Minecraft Diamond challenge, successfully mining in-game gems autonomously without prior human instruction. Pushing the envelope further, Dreamer 4 demonstrated the ability to learn complex tasks—such as mining diamonds—entirely from an offline dataset of recorded gameplay videos, eliminating the need for the agent to interact with the interactive simulation during the training phase.
Having established the robustness of his algorithms in digital domains, Hafner turned his focus toward the physical world. His experimental DayDreamer project successfully applied the core Dreamer algorithm to physical robots, allowing autonomous machines to operate unassisted in novel environments and recover from physical disturbances—such as being deliberately shoved over—without requiring bespoke, task-specific retraining.
Transitioning to Entrepreneurship: The Founding of a Stealth Venture
In the fall of 2025, after years of groundbreaking contributions within the institutional walls of Google DeepMind, Hafner made the pivotal decision to step away and establish his own independent enterprise. The move aligns with a broader industry trend of elite AI researchers leaving massive technology conglomerates to form agile, hyper-focused startups capable of commercializing foundational theoretical breakthroughs.
While Hafner remains tight-lipped regarding the specific commercial applications, target markets, or immediate product timelines of his new San Francisco-based venture, the underlying implications of his work are difficult to overstate. The convergence of advanced world models and affordable, articulated humanoid hardware represents a holy grail for the robotics industry. If successful, companies attempting to deploy general-purpose robots into unstructured consumer spaces will no longer need to manually program solutions for every conceivable household layout or domestic obstacle. Instead, the robots will possess the foundational cognitive architecture to imagine, adapt, and problem-solve on the fly.
When pressed on the ultimate objective of his new enterprise, Hafner offers only a concise, characteristic hint that underscores his long-standing obsession with artificial cognition.
"I was interested in solving a problem," he says, "that would change the world."
