DrFeiFei says spatial intelligence is AI's next frontier

Today’s AI models can write brilliant essays and generate hyperrealistic video clips, but ask them to judge the distance between a coffee cup and the edge of a table, and they are basically guessing at random.

AI pioneer Dr. Fei-Fei Li wants to change that. In a new Substack essay, the Stanford professor argued that large language models are essentially wordsmiths in the dark, eloquent but completely ungrounded in physical reality. According to Li, the next defining frontier of AI isn't bigger text models, but spatial intelligence, systems that can perceive, reason about, and act within the 3D physical world.

Backing up that manifesto, her startup World Labs just launched its first commercial product, a tool named Marble that turns simple prompts into downloadable 3D environments.

Wordsmiths in the dark

Li pointed out that long before humans learned to communicate with words, intelligence evolved through the simple loop of perception and action.

Babies grasp space, gravity, and scale through touch long before speaking. Human history is packed with physical breakthroughs, from measuring the circumference of the Earth through shadows to physically building 3D wire models of the DNA double helix. None of these cognitive leaps could have happened through text alone.

Yet current multimodal models fall flat in the physical realm. Li noted that state-of-the-art models barely perform better than chance when estimating distance, relative size, or orientation. They fail at navigating mazes, struggle with mental rotation, and produce AI videos that lose physical coherence after just a few seconds.

To bridge this gap, AI needs what Li calls world models, generative systems capable of understanding persistent, physics-consistent 3D spaces.

Generating 3D spaces for game engines

World Labs, which Li co-founded in 2024 with Justin Johnson, Christoph Lassner, and Ben Mildenhall, is moving fast. After raising $230 million last fall, the company rolled out Marble this week.

Marble allows users to generate interactive 3D worlds from text, image, or video prompts. The generated environments can then be downloaded and plugged directly into game engines like Unreal Engine and Unity.

Pricing comes in four subscription tiers. The Free plan offers up to 4 world generations per month. The Standard plan costs $20 a month for 12 generations. The Pro plan runs $35 a month for 25 generations and commercial usage rights. The Max plan hits $95 a month for 75 generations.

Early hands-on tests show users creating open-air castles with waterfalls and hobbit-style houses. While visually impressive, physical limitations remain obvious. You can only walk a few steps into the generated scene before hitting a literal invisible wall in the 3D mesh.

Why it matters

For 3D artists, indie game developers, and VFX scouts, Marble offers an immediate, tangible shortcut for rapid ideation and prototyping. Instead of spending days modeling a rough 3D greybox environment, teams can prompt a space, export it, and start tweaking inside Unreal Engine.

Across the broader industry, the battleground is clearly moving beyond chatbots and flat pixel generators. With tech heavyweights like Google and Tencent also chasing world models, spatial reasoning is rapidly becoming the next consensus bottleneck.

We think World Labs made a clever strategic trade-off here. Rather than getting bogged down in the endless hardware grind of general-purpose robotics right away, the team found a viable commercial beachhead in 3D creative workflows, bringing in software revenue while training their spatial models.

Still, the gap between a visual 3D asset generator and a true physics-aware world engine remains massive. The fact that users still hit invisible walls after taking three steps shows that Marble is currently an impressive spatial scene generator, not yet the simulated physical reality Li envisions.