Jim Fan

Linxi "Jim" Fan graduated from Columbia University in 2016 as valedictorian of the class. That summer he became OpenAI's first intern, working on World of Bits, an early attempt to train agents to operate a web browser. He had also interned with Andrew Ng and Dario Amodei at Baidu's AI lab and with Yoshua Bengio at Mila. He took a PhD at Stanford in Fei-Fei Li's vision lab, finishing in September 2021 with a thesis titled "Training and Deploying Visual Agents at Scale."

At NVIDIA he led work on agents that learn in open-ended virtual worlds. MineDojo, a Minecraft benchmark and a model trained on internet videos and wiki pages about the game, won an Outstanding Paper award at NeurIPS 2022. Voyager (May 2023) had GPT-4 write and store its own code to keep exploring Minecraft without human resets, and Eureka (October 2023) had a language model write reward functions that taught a simulated five-fingered hand to spin a pen.

In 2024 he and the University of Texas roboticist Yuke Zhu started NVIDIA's Generalist Embodied Agent Research (GEAR) lab, and at GTC in March 2024 NVIDIA announced Project GR00T, a foundation model for humanoid robots that Fan co-leads. In March 2025 the company released GR00T N1, which it called "the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills." In a Sequoia AI Ascent talk published in May 2025 he proposed a "Physical Turing Test": coming home to a cleaned apartment and a cooked dinner and being unable to tell whether a person or a machine did it. A year later, at AI Ascent 2026, he argued that robotics would follow the large language model playbook, called for video-based "world action models" to replace vision-language-action models, and told the audience, "Our generation was born too late to explore the Earth and too early to explore the stars, but we are born just in time to solve robotics."