Fei-Fei Li

persona · active · confidence: medium · reviewed 2026-09-21 · tier: quarterly

Who she is

Affiliation: Co-founder & CEO, World Labs (spatial-intelligence / world-model lab; out of stealth 2024-09-13 with $230M at a valuation "over $1 billion" from Andreessen Horowitz, NEA and Radical Ventures — TechCrunch, 2024-09-13; a further $1B round on 2026-02-18 including $200M from Autodesk, with AMD, Emerson Collective, Fidelity and Nvidia — the company "declined to say whether the latest round boosted its valuation"; press reports a month earlier put the target at $5B — TechCrunch, 2026-02-18, SiliconANGLE, 2026-01-23). Co-founders: Justin Johnson, Ben Mildenhall, Christoph Lassner. Role check run 2026-09-21: the v1 file (reviewed 2026-06-16) did not mention World Labs at all — a two-year-old CEO role with $1.23B raised was absent.
Stanford: inaugural Sequoia Professor of Computer Science; founding co-director of the Stanford Institute for Human-Centered AI (HAI), 2019 → 2026-05. On 2026-05-04 Stanford merged HAI with Stanford Data Science under James Landay as "Denning Director"; "Co-founder Fei-Fei Li takes on a new university-wide role as Special Advisor on AI" to President Jonathan Levin, co-chairs the institute's advisory council with John Hennessy, and will "retain her title as Stanford HAI's founding director and senior fellow" (Stanford HAI, 2026-05-04; interview, same date). That change landed six weeks before the v1 "review" and was missed. Her Stanford Profiles page still lists "Co-Director, Stanford Institute of Human-Centered AI (HAI) (2019 - Present)" alongside "Senior Fellow at HAI" — Stanford's own pages disagree; the dated HAI announcement is trusted here (Stanford Profiles). Partial academic leave "until December 2025" is documented (TechCrunch, 2024-08-14, quoting her profile); an extension beyond that appears in search summaries only and could not be read off a primary this cycle — flagged, not asserted.
Other hats (Stanford Profiles, fetched 2026-09-21, dates as listed there): Special Advisor to the UN Secretary-General (2023–; appointed to the UN Scientific Advisory Board, Stanford HAI, 2023-08-03); Co-Founder and Chairperson of the Board, AI4ALL (2015–); board, Computer Vision Foundation (2019–); member, Council on Foreign Relations (2020–). Scientific Partner, Radical Ventures (Radical team page, © 2026) — a fund that is also an investor in her company. Co-lead of Governor Newsom's Joint California Policy Working Group on AI Frontier Models (appointed 2024-09-29, the day SB 1047 was vetoed). Past: VP and Chief Scientist of AI/ML, Google Cloud (2017–2018, profile); director, Stanford AI Lab (2013–2018, profile); independent director, Twitter (2020-05 → removed 2022-10-27 on Musk's takeover — Wikipedia, secondary).
Recognition: creator of ImageNet (14,197,122 images, 21,841 synsets — image-net.org); 2025 Queen Elizabeth Prize for Engineering, one of seven "pioneers of modern machine learning" (with Bengio, Dally, Hinton, Hopfield, Huang, LeCun), presented by King Charles III at St James's Palace on National Engineering Day, 2025-11-05 (QEPrize; QEPrize year review, 2025-12-19; Stanford HAI, 2025-11-07); TIME 2025 Person of the Year as one of eight "Architects of AI" with Zuckerberg, Su, Musk, Huang, Altman, Hassabis and Amodei (CBS, 2025-12-11). Memoir The Worlds I See (2023).
One-line position: AI must be built around human needs, not just technical benchmarks — and the people building it must include voices beyond Silicon Valley's default demographic if the technology is to serve broad human wellbeing. Since 2024 that principle has a $1.23B technical corollary: language models are "wordsmiths in the dark," and the next frontier is spatial intelligence — machines that see, reconstruct and act in three-dimensional space ("From Words to Worlds," 2025-11-10).

Discipline & technical bet

Computer-vision scientist whose signature result was a dataset, not a model — ImageNet proved that scale of labeled data could move a whole field (2009–2012). She now runs a closed, venture-funded lab making one concrete wager in four parts. (1) Spatial intelligence is a capability distinct from language: today's models "can't navigate mazes, recognize shortcuts, or predict basic physics" and remain "eloquent but inexperienced, knowledgeable but ungrounded" (Substack, 2025-11-10). (2) The path is generative world models that reconstruct, generate and simulate 3D scenes from pixels — Marble (2025-11-12; "a first-in-class generative multimodal world model" taking "text, images, video, or coarse 3D layouts" and exporting "Gaussian splats, meshes, or videos" — World Labs) and Atlas (2026-09-01; "an omni model that we pretrained from scratch to natively operate on text, images, video, and 3D," with three functions, "Camera-Controlled Generation," "Spatial Reconstruction," and "Space-Time Simulation," output "up to 1 minute of video at 1440p" — World Labs). (3) The unit of learning: co-founders describe Atlas as "really new view prediction," and "new view prediction is to spatial intelligence what next-token prediction was to large language models" (a16z podcast with Martin Casado, 2026-09-04) — the explicit claim that this is the GPT-moment recipe for a second modality. (4) Robotics' binding constraint is data, not algorithms — "the biggest problem right now in robotics is actually data" (same episode) — solved by generating training worlds: the SceniX acquisition brought a "real-to-sim-to-real (R2S2R) engine that turns one physical task into many controllable, reusable worlds" (World Labs, 2026-07-28).

Two rival bets frame hers. Google DeepMind's Genie 3 (2025-08-05) generates worlds "you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p" and is framed as "a key stepping stone on the path to AGI" (DeepMind) — interactive-first, AGI-framed; Marble is persistent-and-downloadable-first, AGI-agnostic. LeCun's AMI Labs raised $1.03B (2026-03) for non-generative JEPA world models — same diagnosis of LLMs, opposite architecture (predict in an abstract space, not pixels). Her lab is the only one of the three whose commercial product shipped first and whose research paper hasn't.

Key claims (Says)

Notable predictions — with falsifiable checks

Revealed behavior (Does)

Feels

Hope that spatial intelligence is her second ImageNet — being early on the thing the field then organizes around; she says it in her own frame, "our dreams of truly intelligent machines will not be complete without spatial intelligence" (manifesto, via TechCrunch 2025-11-12). Underneath: a builder's need to be the responsible adult and the Maven memory — the 2018 leaked emails ("avoid at ALL COSTS any mention or implication of AI"; weaponised AI "is red meat to the media to find all ways to damage Google" — Silicon Republic, 2018-05-31) are the one public moment when human-centered language and corporate messaging collided in her own inbox. "Because I believe in humanity... all of us have a responsibility in ushering AI to a better state" (TIME, 2025-12-09) is the thing she wants to be true.

Hears

Co-founders Johnson, Mildenhall and Lassner; a16z's Martin Casado (lead investor and her podcast host); Radical Ventures, where she is a partner; Stanford's leadership triangle (Levin, Hennessy, Landay); policy co-leads Cuéllar and Chayes; SceniX's Yunzhu Li on robotics; the UN Secretary-General's board. An input diet that skews to investors, Stanford and Sacramento — little contact with labor economists or the external critics (Gebru, Crawford) who share her stated values.

Sees

What one dataset did to a field — so she reads robotics through ImageNet: "the biggest problem right now in robotics is actually data" (a16z, 2026-09-04), and buys a data-generation company. As a working-group co-lead she has read every frontier lab's safety practices from the regulator's side while running a frontier-adjacent lab herself — a vantage no other bench member holds. And Stanford's talent pipeline, which she has drawn on for two decades.

Incentive map

Re-derived from scratch this cycle; v1 carried none. Who pays her: World Labs (CEO and founder equity in a company reported at $5B; investors a16z, NEA, Radical Ventures, NVIDIA, AMD, Autodesk, Emerson Collective, Fidelity); Stanford (Sequoia Professor; Special Advisor on AI to the President); Radical Ventures (Scientific Partner — she is a partner at a fund that holds a position in her own company, so her public thesis is also her fund's marketing); Stanford HAI's industry programme, whose Founding Member is Google and whose Strategic tier (Accenture, American Express, AXA, Doosan, EY, IBM, Infosys, Itaú Unibanco, PepsiCo, PwC, SAP, Wells Fargo) buys "faculty-specific research projects aligned with member's priorities" with a $550K "wallet" (Stanford HAI industry page, fetched 2026-09-21) — the tension HAI itself acknowledged in 2019: "While we value and appreciate the support of donors and partners, HAI maintains complete intellectual and operational independence" (Etchemendy, 2019-10-10). What she's selling: spatial intelligence as the next frontier; Marble subscriptions and World API credits; Atlas early access; and a "trust but verify" policy model that favors transparency over liability — the regime a16z-backed "little tech" wanted. What she can't say: "LLM agents will subsume spatial reasoning"; "robotics is decades out"; "Marble is a 3D asset generator, not a world model"; "California's regime harmed startups"; anything hostile to a16z's regulatory line; "AGI is near" (it would hand the frontier framing to the labs she shares the TIME cover with). Access dependence: Newsom appointee, UN advisor — positions renewed by being useful, not by being right. Rule-25 discount: high on spatial-intelligence timelines and "next frontier" claims; moderate-to-high on AI-policy positions (the a16z relationship predates the op-ed); moderate on labor claims (no direct stake); low on ImageNet-era historical claims. No --investing twin exists to diff.

Theories aligned with

What she's reacting against

Where she overlaps / splits (with Rich)

Track record

Empirical vs normative

Weak spots / open questions

Rich's take

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

Sources

migrated v1→v2 2026-09-21 · backup: _archives/fei-fei-li.html.bak-20260921 · converts-from: people/fei-fei-li.md · AI & Society domain