{"slug": "fei-fei-li-s-world-labs-launches-atlas-to-build-walkable-ai-worlds", "title": "Fei-Fei Li's World Labs Launches Atlas to Build Walkable AI Worlds", "summary": "Fei-Fei Li's World Labs released Atlas on September 1, 2026, a multimodal world model that generates up to one minute of 1440p video from one or more images with pixel-perfect camera control, aiming to create persistent, walkable 3D worlds for video, design, robotics training, and future versions of its Marble product. The company claims Atlas is 'a first of its kind multimodal world model trained from scratch' and can reconstruct real spaces from sparse inputs like cellphone footage for robotics simulation.", "body_md": "*Fei-Fei Li's World Labs released Atlas on September 1, 2026, and the real claim is bigger than another AI video demo: it wants to make 3D worlds you can revisit and use.*\n\nYou give Atlas one photograph, or a handful of them, and World Labs says the model can turn that input into a scene you can move through with controlled camera paths. That's the bet. The company isn't trying to sell you another flat clip generator with a nicer prompt box. It wants Atlas to become the spatial layer beneath video, design work, robotics training and future versions of Marble, its existing world-generation product.\n\nFei-Fei Li, who co-founded World Labs in 2024, announced the launch on X as 'a first of its kind multimodal world model trained from scratch.' The company blog gives the more technical version: Atlas is an omni model pretrained from scratch to work across text, images, video and 3D, using a multimodal autoregressive diffusion transformer. Plainly, World Labs is saying the model treats camera position and 3D space as native inputs, not as vague instructions hidden inside a text prompt.\n\n## What Atlas Actually Does\n\nThe most useful number is simple. World Labs says Atlas can generate images and videos from one or more images with pixel-perfect camera control, up to one minute of video at 1440p. In one demo, the company says it hand-designed a camera path through a scene from a small number of reference images and had Atlas generate the full route. If that holds up outside the launch page, you get something closer to a controllable set than a one-off video.\n\nStart there. A Sora-style model can give you a clip, and a good one can make that clip look expensive. Atlas is trying to preserve the place behind the clip. Come back to the same spot, and the promise is that the geometry and lighting are still there. World Labs calls that spatial intelligence, a phrase Li has used for years to describe the missing piece in AI systems that can talk fluently but still struggle with the physical world.\n\n[Singapore Still Can't Fix Its Developer Shortage Even With Vibe Coding Tools](https://startupfortune.com/singapore-still-cant-fix-its-developer-shortage-with-vibe-coding-tools/)\n\nSingapore's founders now build with Cursor, Lovable and Replit, but ManpowerGroup's 2026 survey found AI development and AI literacy are the country's two hardest skills to hire for. The tools multiplied faster than the engineers who can be trusted to run what they build. - [how to fix Singapore developer shortage crisis](https://startupfortune.com/singapore-still-cant-fix-its-developer-shortage-with-vibe-coding-tools/) - [AI coding tools versus hiring real engineers](https://startupfortune.com/singapore-still-cant-fix-its-developer-shortage-with-vibe-coding-tools/)\n\nThis isn't World Labs' first swing. The company made Marble generally available on November 12, 2025, with a free tier and paid plans now running up to $95 a month. Marble turns text, images, videos, panoramas or coarse 3D layouts into editable 3D environments. Then World Labs launched the World API on January 21, 2026, so developers could generate navigable worlds directly from software. Atlas sits above those efforts as the new base model, and the company says it will power future World Labs products.\n\nRobotics is the harder prize. World Labs says Atlas can reconstruct real spaces from sparse inputs, including ordinary phone footage, then generate the RGB and depth observations a simulated robot would see as it moves. In its robotics examples, the company says two large environments were captured with cellphone video and 24 frames each were used for reconstruction. That's not a small gap from hand-built simulation assets, where collecting and cleaning scenes alone can slow the whole training loop.\n\n## The Race Around It\n\nAtlas doesn't arrive alone. Google DeepMind's Genie 3, announced in August 2025 and now used in Project Genie for U.S. Google AI Ultra subscribers, generates interactive worlds at 20 to 24 frames per second and 720p, according to DeepMind's own materials. Nvidia's Cosmos world foundation models are already aimed straight at robotics and autonomous vehicles, and Nvidia said last year that Cosmos models had been downloaded more than 2 million times.\n\nThat makes Nvidia the awkward comparison. Reuters reported in February 2026 that World Labs raised $1 billion from investors including Nvidia, AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company and Sea, with Autodesk putting in $200 million. Forge private-market data later listed World Labs at a $5.29 billion post-money valuation. Nvidia is a backer today, but it also owns much of the hardware and software stack that physical AI companies already touch. That could change the relationship fast.\n\nFor now, the two look useful to each other. Nvidia sells the compute and simulation tooling. World Labs sells the idea that one model can generate and reconstruct a place, then simulate it with enough consistency for builders to use. If you're building in robotics, VFX, gaming or design, that distinction matters. You don't just need a pretty frame. You need a world that survives the next camera move.\n\n## The Receipts Are Thin\n\nThe strongest part of the Atlas launch is also the part you should be most careful with. World Labs says third-party human raters preferred Atlas over rival video models on camera-path adherence, including 81% against Gemini Omni Flash and 93% against FLUX 3. The caveat is obvious. Those are World Labs' tests, published by World Labs, and the comparison gave Atlas native camera inputs while rival models received camera directions through text.\n\nFrankly, that caveat should stay in the story. Atlas may still be better, and the demos are hard to ignore, but a native camera interface is a structural advantage before the first frame is judged. World Labs also says Atlas outperformed open-source reconstruction models including Pi3X, VGGT-Omega 1B, Depth Anything 3 and MapAnything. Meta's VGGT-Omega project page, though, posted an August 18, 2026 notice warning that an ancestor checkpoint of the released 1B model may have benchmark contamination. You shouldn't read that leaderboard as settled science.\n\n[MIT Says AI Can Now Complete Almost Any Undergraduate Assignment](https://startupfortune.com/mit-says-ai-can-now-complete-almost-any-undergraduate-assignment/)\n\nMIT's Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training released a report on August 25 finding that AI can now credibly complete almost any undergraduate assignment, from essays to proofs to code. The committee is urging departments to fast-track curriculum changes and instructors to shift toward oral exams and portfolios... - [ai can now complete undergraduate assignments at mit](https://startupfortune.com/mit-says-ai-can-now-complete-almost-any-undergraduate-assignment/) - [how generative ai is changing college coursework standards](https://startupfortune.com/mit-says-ai-can-now-complete-almost-any-undergraduate-assignment/)\n\nThere's also a commercial blank space around the launch. Atlas entered early access with select partners, but World Labs did not name those partners, publish a research paper, list pricing, provide a model card or give a general-availability date. Its current API documentation still centers on Marble endpoints. The receipts still matter.\n\nAtlas is current, ambitious and worth watching because it turns World Labs' spatial-intelligence pitch into a product people can begin to test. It is not yet proof that Li's company owns the category. Over the next year, the real test is whether developers can use Atlas outside carefully chosen demos and whether robots trained through its simulated views improve in the real world. There's also whether DeepMind or Nvidia makes the same idea cheaper and easier to adopt first.\n\n**Also read:** [Broadcom's $60 Billion AI Backlog Faces Its Biggest Test Today](https://startupfortune.com/broadcoms-60-billion-ai-backlog-faces-its-biggest-test-today/) • [Meta Permanently Disables Cameras on Thousands of Tampered Smart Glasses](https://startupfortune.com/meta-permanently-disables-cameras-on-thousands-of-tampered-smart-glasses/) • [McKinsey Survey Finds 32% of Firms Now Building Software Instead of Buying It](https://startupfortune.com/mckinsey-survey-finds-32-of-firms-now-building-software-instead-of-buying-it/)", "url": "https://wpnews.pro/news/fei-fei-li-s-world-labs-launches-atlas-to-build-walkable-ai-worlds", "canonical_source": "https://startupfortune.com/fei-fei-lis-world-labs-launches-atlas-to-build-walkable-ai-worlds/", "published_at": "2026-09-02 08:56:46+00:00", "updated_at": "2026-09-02 09:22:10.837996+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "ai-products", "ai-research"], "entities": ["World Labs", "Fei-Fei Li", "Atlas", "Marble", "Google DeepMind", "Genie 3", "Project Genie", "Google AI Ultra"], "alternates": {"html": "https://wpnews.pro/news/fei-fei-li-s-world-labs-launches-atlas-to-build-walkable-ai-worlds", "markdown": "https://wpnews.pro/news/fei-fei-li-s-world-labs-launches-atlas-to-build-walkable-ai-worlds.md", "text": "https://wpnews.pro/news/fei-fei-li-s-world-labs-launches-atlas-to-build-walkable-ai-worlds.txt", "jsonld": "https://wpnews.pro/news/fei-fei-li-s-world-labs-launches-atlas-to-build-walkable-ai-worlds.jsonld"}}