Bloomberg Tech

Fei-Fei Li on Why World Models Generate Worlds, Not Words

Fei-Fei Li· Co-founder & CEO of World Labs at World Labs
·~10 min·English·Bloomberg
MultimodalRoboticsAI SafetyAI Company
TL;DR

Fei-Fei Li argues the next AI substrate is spatial intelligence — world models that generate navigable 3D worlds instead of words — and that the harder questions of safety, benchmarking, and governance are human problems to be solved at the ecosystem level, not left to any single lab or to fatalism.

01Core Mental Model

World Models Output Worlds, Not Words

<strong>A language model outputs a stream of symbols, while a world model outputs a 3D world you can move a camera through</strong> — a change in output type, not just quality, that makes it a different kind of model.

The models generate beautiful pixels, very accurate geometric structure of those 3D or 4D world, and which in turn can allow people, creators, designers, robotics developers to have precise control of the cameras or the construction of worlds.

— Fei-Fei Li, Bloomberg Tech
Key Insight
The interesting claim is not that world models look better — it is that they change the output type. An LLM hands you a one-dimensional symbol stream; a world model hands you a navigable space with real geometry. Everything downstream (robots, design, simulation) follows from that one shift.
02The Product

Atlas Fuses Two Problems Vision Solves Apart

<strong>Atlas combines pixel generation with 3D reconstruction in one model</strong>, and Li says it outperforms specialized state-of-the-art models on those tasks.

It also tries to solve a very hard problem in the field of computer vision that combines the generation of beautiful pixels with the reconstruction of intricate 3D structure.

— Fei-Fei Li, Bloomberg Tech
Key Insight
For a decade, photorealistic generation and accurate 3D reconstruction were different research tracks with different specialist models. Folding them into one model is the news: it means the pretty picture and the correct geometry are designed to remain consistent, which is exactly what a robot or a camera operator needs.
03The Application

Where Data Is Scarce, the Model Dreams It

<strong>Specialized robotics domains such as pharmaceutical manufacturing and lab science have limited real data</strong>, so a world model can simulate environments for both training and evaluation.

But using world models to simulate and then turn that simulation into very realistic training ground as well as evaluation ground helps to develop robots.

— Fei-Fei Li, Bloomberg Tech
Key Insight
This reframes the world model as data infrastructure, not a demo. In domains where collecting real robot data is slow and expensive, the model manufactures the scenarios to train on and the scenarios to grade against — the classroom and the exam — which is how a data-starved field can still make a policy robust.
04The Frame on Safety

A Double-Edged Sword, Held by Both Edges

<strong>Li refuses to treat world-model alignment as special: it is the same calculus as every powerful technology — pursue the benevolent advance and keep safety in the forefront, both at once.</strong>

technology is a double-edged sword. We develop it with the intention to be benevolent and helpful to humans, but we have always to put safety and responsibility in the forefront of the development and deployment of technology.

— Fei-Fei Li, Bloomberg Tech
Key Insight
By putting world models in the same bucket as electricity and the steam engine, Li quietly rejects AI exceptionalism. The move has an edge: if the risk is not novel, then neither are the tools for handling it — governance, institutions, and deployment discipline — which is precisely where she steers the rest of the interview.
05The Governance Answer

We Cannot Leave It to Fatalism

<strong>Asked about pacing the frontier, Li rejects fatalism and assigns responsibility to a broad ecosystem</strong> spanning the public sector, open and closed source, academia, government, and civil society.

I believe responsibility is so important. We cannot leave it to fatalism.

— Fei-Fei Li, Bloomberg Tech
Key Insight
Her answer names many actors — public sector, open and closed source, academia, government, civil society — and centers none of them, not even a single frontier lab. Spreading the responsibility this wide is also spreading the power, which is why her ecosystem framing reads as much as a governance stance as a safety one.
06The Missing Yardstick

Benchmarks Are Public Infrastructure

<strong>World models still lack a shared yardstick, and Li argues the maker's internal benchmark can't be the last word — evaluation has to become an ecosystem-level responsibility.</strong>

ImageNet 15 years ago was one of the very first benchmark of AI. I continue to believe importance of benchmark from independent bodies and public sector bodies like academia as well as the shared responsibility of industry of government

— Fei-Fei Li, Bloomberg Tech
Key Insight
Li is citing her own resume as the argument: ImageNet, the benchmark she built in academia, is why the last era of AI was measurable at all. The implicit warning is that if world-model evaluation is left entirely inside the companies that ship them, there is no independent scoreboard — and no one outside can tell whether the safety claims hold.
07The Human Element

The Threat Is Within Ourselves

<strong>Li relocates the existential-risk question from the machine to the people: the danger and the control both live in how humans wield and govern the technology.</strong>

AI it's not about AI. It's about humans. It's about our shared as well as individual dignity, agency, as well as our responsibility.

— Fei-Fei Li, Bloomberg Tech
Key Insight
This is the through-line stated as a worldview: the failure modes she cares about are not a model waking up, but humans mis-wielding a powerful tool. It is a humanist's answer to a doom question — and it is also why every earlier section routed the hard problems to institutions and people rather than to the model itself.