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

Danijar Hafner is a 31-year-old AI entrepreneur based in San Francisco’s SoMa district. He grew up in a rural town in northeastern Germany, where his parents were classical musicians. Hafner developed an early interest in AI after learning programming from a neighbor and taking online courses during high school. He studied engineering at the Hasso Plattner Institute in Potsdam, Germany. In 2015, he joined Google Brain as a student researcher and later worked at Google DeepMind in the UK, Canada, and the US. Hafner collaborated with AI pioneers such as Geoffrey Hinton, often called one of the godfathers of AI, and Ashish Vaswani, coauthor of the influential research paper Attention Is All You Need, which introduced the transformer technology used in modern large language models. His former manager at Google, Timothy Lillicrap, described Hafner as exceptionally skilled, noting he is in the top half of 1% of researchers at the company.

Hafner’s new startup focuses on developing AI agents capable of planning ahead for unexpected situations, a critical challenge for robots operating in human environments like homes. Unlike traditional robotics, which relies on extensive real-world trial-and-error training, Hafner uses model-based reinforcement learning. This approach involves creating world models—AI systems designed to simulate physical reality—and training agents within these models. The agents learn to act by imagining future outcomes, enabling them to navigate unfamiliar environments without prior exposure. Hafner’s robots, imported from China, are designed to handle scenarios they were not specifically trained for, such as adapting to new floor plans or furniture arrangements. This method aims to bridge the gap between virtual training and real-world application, making robots more adaptable and practical for everyday use.

Hafner has achieved several milestones in AI development through his world model approach. In 2019, he created PlaNet, an agent that could plan actions by predicting future outcomes within a simulated environment. His Dreamer series further advanced these capabilities: Dreamer 2 became the first agent to achieve human-level performance in Atari 2600 games using only a world model, while Dreamer 3 solved the Minecraft Diamond challenge by mining in-game gems autonomously. Dreamer 4 marked another breakthrough by learning to mine diamonds from pre-recorded gameplay videos, without direct interaction with the game. These achievements demonstrated the potential of Hafner’s methods to train AI agents in virtual environments and transfer their skills to real-world tasks, reducing the need for extensive physical testing.

Hafner’s research has recently expanded beyond video games into the physical world with his DayDreamer project. Using the Dreamer algorithm, robots equipped with this technology can operate in novel environments and adapt to new experiences, such as recovering from being pushed over, without specific prior training. This shift from virtual to physical applications is a significant step toward deploying AI-driven robots in real-world settings like homes, offices, or healthcare facilities. Hafner’s new startup, founded in the fall of 2025 after leaving Google DeepMind, aims to further develop these capabilities. While details about the startup remain under wraps, Hafner hints at ambitions to address a problem with global impact, suggesting a focus on scalable and transformative applications for AI agents.