World Labs

Fei-Fei Li's journey to founding World Labs spans three decades of AI research. In 2024, after witnessing the transformative impact of large language models (GPT-2, ChatGPT), she realized a critical gap: while language models captured conversational intelligence, they missed spatial understanding—the 3D reasoning that humans use constantly. Li had pioneered ImageNet (2006-2012), the dataset that powered the deep learning revolution and AlexNet's breakthrough in 2012. Now, she saw that spatial intelligence was the missing piece needed to unlock embodied AI, robotics, and immersive human-AI collaboration.

World Labs assembled a small but elite team of 30 researchers and research engineers, many from top AI labs, computer graphics, and computer vision backgrounds. Co-founders Justin Johnson, Christoph Lassner, and Ben Mildenhall brought deep expertise in frontier models. Over a year-plus development cycle, the team tackled an extremely difficult problem: building a generative model that outputs genuinely 3D worlds—not just videos or 2D projections. The breakthrough came when they demonstrated the ability to prompt with text and images, then navigate freely through generated 3D environments. They also developed real-time video generation capable of running on a single H100 GPU.

World Labs launched Marble—their first product—roughly one month before the podcast interview aired (late 2024). Rather than narrow targeting, they opened it to creators, designers, game developers, and researchers. Immediate organic adoption revealed diverse use cases: virtual production studios discovered it could cut VFX production time by 40x, game developers began exporting Marble scenes into game engines, roboticists reached out to generate synthetic training data, and unexpectedly, psychology researchers requested it for exposure therapy and psychiatric research environments.

The product-led approach worked remarkably well. By making Marble freely accessible and discoverable, World Labs uncovered use cases they hadn't anticipated—particularly in clinical psychology. The intentional feature visualization (rendering dots before full texture) became a delightful user experience that researchers noticed and praised. However, the source material doesn't detail any failures or pivots, suggesting either a very focused execution or early-stage operation where challenges hadn't yet emerged publicly.

World Labs is positioned as a frontier model company building spatial intelligence as a horizontal platform. Marble represents the first step, with the team already collaborating with major studios (Sony) and seeing rapid adoption across entertainment, games, simulation, and research. Li frames spatial intelligence as equally important—or more important—than language models, positioning the company to capture value as robotics, VR, and embodied AI mature over the next decade.