No Priors Ep. 117 | With Co-Director of Stanford's HAI & Founder of World Labs Dr. Fei-Fei Li

Summary

• Fei-Fei Li is betting World Labs on spatial intelligence as AI’s missing foundation layer. She defines it as the ability to understand, reason about, interact with, and generate 3D worlds; without it, “AI would be incomplete.” World Labs says it is the first company it knows of tackling the 3D-generation foundation-model problem, with outputs governed by plausible geometry and physics even when the worlds are fantastical.
• The clearest near-term commercial wedge is AI-assisted 3D creation. Li expects designers, VFX artists, game developers, marketers, and other creators to collaborate with models much as developers use Cursor and Windsurf. Generative spatial models could also address a constraint on metaverse, XR, AR, and VR adoption: beyond improving hardware, “we’re looking for content creation.”
• Robotics will require more than scaling vision or imitation data. Li expects a hybrid of many data forms—including video, simulation and synthetic data, teleoperation, and embodied data—while calling simulation underrated but noting that many experts and robotics companies already work on it. Haptics is “truly underappreciated” for manipulation. She also rejects a humanoid-only future: task and energy economics should produce diverse forms—underwater robots should resemble fish, while airplanes are becoming more robotic.
• The core 3D-model challenge begins with data, while productization remains unresolved. Unlike her NLP and LLM colleagues, Li says World Labs does not necessarily have abundant internet-native material; it needs increasingly sophisticated acquisition, processing, engineering, and synthesis. Productization is another challenge because 3D is interactive rather than passively consumed: “Nobody wakes up and says, ‘I’m just going to sit here and watch 3D.’”
• Li’s career offers a specific precedent for contrarian data bets creating new model categories. Around 2003, she built a 101-category dataset after her adviser proposed 100, then scaled the same conviction into ImageNet’s 15 million labeled images across thousands of categories—despite being told she might not get tenure. AlexNet’s eventual breakthrough validated “that conjecture that no one believed in.”
• Li’s research prescription is simply: “Be fearless.” She locates productive ambition between being “somewhat delusional and crazy” and “rationally bold”; excessive rationality means failing to identify problems big enough, while complete craziness can let many things go wrong. World Labs accordingly recruits across graphics, vision, data, generative AI, infrastructure, optimization, engineering, and product rather than treating spatial intelligence as a homogeneous problem.
• Li’s end-state remains human-centered augmentation, especially where society is short of expertise and care. Healthcare lacks discovery, diagnosis, precision medicine, accessible treatment, chronic-disease support, and better aging—not human relevance. Her governing call is that “AI is a tool to help people,” while preserving love, relationships, prosperity, and justice as human values that machinery should not take away.

Deep dive

Not yet available upstream; scheduled sync will retry.