Women Building the AI Era, Part 2- Fei-Fei Li

In 2009, a Stanford researcher named Fei-Fei Li presented a dataset called ImageNet at a poster session almost nobody stopped to look at. It held 14 million labeled photographs, built over two years by paying people on Amazon Mechanical Turk to sort images by hand.

Three years later, two graduate students trained a neural network on that same dataset using a pair of gaming graphics cards. The system, later named AlexNet, cut the error rate on the year's image-recognition contest by close to half against the next best entry. That result is widely credited as the spark that set off the deep learning era, and much of what the AI industry has built since.

Li spent the years after that as one of the field's most cited researchers and its most visible advocate for building AI carefully. Now she runs a company that has raised $1.2 billion on a simple wager: the dataset problem she solved for flat images still hasn't been solved for the three-dimensional world, and whoever solves it next controls the next layer of AI infrastructure.

The Woman Who Taught Machines to See

Li didn't stop at ImageNet. She spent over a decade at Stanford building the university's Human-Centered AI Institute, did a stint as Google Cloud's chief AI scientist, and became one of the field's loudest voices arguing that AI safety and AI progress aren't in conflict. Her nickname in the industry, "the godmother of AI," isn't hype so much as a rough description of her role: she built the substrate a huge share of modern computer vision was trained on.

In January 2024, she co-founded World Labs with three collaborators, each carrying a specific piece of the technical puzzle. Justin Johnson, a former student of Li's, brought years of work on 3D generative modeling from Meta's AI research lab. Christoph Lassner had built an early sphere-based rendering technique that helped set the stage for a 3D graphics method called Gaussian Splatting. Ben Mildenhall came in for his expertise in the same kind of rendering and spatial reasoning. The company's premise was that large language models, for all their fluency with text, have almost no grasp of physical space, and that gap is worth building a company around.

Did You Know?

The same Nvidia GPUs that trained the 2012 neural network on Li's ImageNet dataset became, over the following decade, the hardware backbone the entire AI industry now runs on. That origin story is detailed in the twenty-year bet that turned a graphics card company into AI's essential infrastructure. Nvidia's venture arm is now also an investor in Li's new company.

Betting on the Next Modality

World Labs came out of stealth in September 2024 with $230 million and a $1 billion valuation, before it had shipped anything. That alone would have been a striking debut. Instead, it kept raising: by early 2026 it had closed roughly $1.2 billion across two rounds, with a valuation reported near $5 billion, backed by Nvidia, AMD, Autodesk, Fidelity, and a roster of venture firms led by Andreessen Horowitz.

The Autodesk check is the tell here. Autodesk sells the software that architects, game studios, and film-effects houses use to build 3D worlds by hand, one of the most labor-intensive parts of those industries. A model that generates a navigable 3D environment from a single photo or a paragraph of text isn't just a research curiosity to a company like that. It's a bet on which company builds the tool that replaces years of manual modeling work.

The round's structure carries its own signal too. Andreessen Horowitz didn't just write a check: general partner Martin Casado has taken an active role helping shape World Labs' product and research direction, an unusually hands-on arrangement for a lead investor. Early backers also included Ashton Kutcher's Sound Ventures, alongside Adobe, Databricks, and Nvidia's venture arm, a spread that runs from Hollywood money to the chipmakers that would supply the compute either way.

The Product: A World in a Sentence

World Labs shipped its first commercial product, Marble, in November 2025, reaching general availability in February 2026 alongside the larger funding round. Feed it a photo, a video clip, or a text description, and it generates a persistent, editable three-dimensional space, viewable in a browser or a VR headset like Vision Pro or Quest 3.

It's a strange thing to explain to someone who hasn't tried it. The easiest comparison is a movie set built overnight from a single sketch: real enough to walk through, editable enough to move the furniture, generated in minutes instead of months. Marble's early customers span game studios prototyping levels, architects previsualizing buildings, and robotics teams that need synthetic 3D environments, cheaper and faster to generate than filming or building the real thing, to train machines that operate in the physical world.

"World model" is the industry's term for an AI system that doesn't just describe a place in words, but builds a working sense of how that space behaves, enough to predict what's around a corner or how a room looks from another angle. Researchers have chased that idea since at least the 2010s; Marble is one of the first versions an ordinary customer can actually use.

A Race With More Than One Runner

World Labs isn't alone in this bet, which is part of why the money is moving so fast. Google DeepMind has its own world model, Genie, now on its third version and built for real-time interactive 3D scenes. Runway, better known for AI video generation, released a competing world model aimed at filmmakers rather than robotics or games. Yann LeCun, Meta's former chief AI scientist, left in 2026 to start a company chasing the same idea from a different technical angle.

None of these labs agree yet on what a world model is even for, whether it's a filmmaking tool, a robotics training ground, or a step toward something closer to general intelligence. That disagreement is exactly why investors are spreading bets across several teams instead of waiting to see who wins first.

Knowlegic Perspective

World Labs is a useful window into how AI capital moves once a modality gets crowded. Text-based AI has dozens of well-funded competitors, and by 2026 the returns on funding one more of them looked increasingly uncertain. Spatial intelligence is emptier ground, with fewer credible teams and a founder whose track record includes literally building the dataset that made the last paradigm shift possible.

That's also why the investor list matters as much as the product does. Nvidia, AMD, and Autodesk aren't index-fund money looking for the next hot ticker. Each of them sells something that becomes more valuable if 3D-generative AI succeeds: chips that render it, and software that competes with, or gets replaced by, it. Watching who writes a strategic check is often a faster way to see where an industry thinks the puzzle to be solved next actually is than reading the pitch deck.

It also explains why World Labs isn't the only well-funded team chasing this. When Google, Runway, and a fresh company from one of the field's most senior researchers are all racing toward the same rough idea at once, that's usually a sign capital thinks the category itself, not any single company, is where the value is about to land.

Li's bet is really a bet about where the next several billion dollars of AI infrastructure spending goes next, and for now, the industry's most strategic money is following her.

Sources & References

• Fei-Fei Li's World Labs comes out of stealth with $230M in funding, TechCrunch (2024)

• World Labs lands $1B, with $200M from Autodesk, to bring world models into 3D workflows, TechCrunch (2026)

• Fei-Fei Li's World Labs in funding talks at $5 billion valuation, Bloomberg (2026)

• Fei-Fei Li's World Labs speeds up the world model race with Marble, its first commercial product, TechCrunch (2025)

• How world models became AI's next frontier, The Deep View (2026)