The Godmother Just Told Us Where Law Is Going.
Most Lawyers Weren't Listening.

I'm sitting in a conference hall in San Francisco. Fei-Fei Li, the woman who built ImageNet and fundamentally rewired how machines see the world, is explaining why the entire AI industry is about to pivot away from language. And all I can think is: she's describing the future of litigation.

Let me back up. Dr. Fei-Fei Li is not a legal technologist. She's a Stanford computer science professor, a co-director of the Stanford Institute for Human-Centered AI, and the founding CEO of World Labs, a spatial intelligence company that just raised $1 billion. Nvidia, AMD, Autodesk, and Andreessen Horowitz wrote the checks. That's not venture optimism. That's infrastructure capital betting on a paradigm shift.

Language Models Are Wordsmiths in the Dark

Li opened with a metaphor drawn from evolutionary neuroscience. For half a billion years, the most primitive organisms existed in what she called "profound darkness." Not literal darkness. A complete absence of spatial awareness. They couldn't perceive the world around them. They couldn't sense their own relationship to it.

Then vision evolved. And everything changed. Speciation. Arms races. Nervous system complexity. The entire trajectory of biological intelligence pivoted on a single capability: understanding space.

Her argument: today's large language models are brilliant, but they're operating in that same profound darkness. They process tokens. They generate text. But they have no spatial understanding of the world. No sense of geometry, physics, cause and effect in three-dimensional space.

Now translate that to legal practice. What are we doing with AI in law right now? We're summarizing documents. Drafting motions. Searching case law. All of it is flat. Two-dimensional. Text in, text out. We're building increasingly sophisticated wordsmiths. And they're operating in the dark about the actual world where legal conflicts happen.

What a "World Model" Means for Litigation

Li defines a world model as an intelligent system that can understand space, reason about geometry and physics, predict the next state of the world, and generate three-dimensional environments. Her company, World Labs, is building exactly this with their Marble platform, which creates persistent 3D worlds from text and image prompts.

Right now, Marble generates environments about the size of a hotel lobby. A year and a half ago, Li admitted, they were generating rooms she called "closets." But the trajectory is clear, and the implications for legal practice are staggering if you know where to look.

Every one of these practice areas depends on the factfinder understanding space. And right now, we rely on static photographs, hand-drawn diagrams, and expert witnesses billing $400 an hour to explain what happened where.

Spatial intelligence doesn't replace legal reasoning. It gives legal reasoning a body.

When AI can model the physical world where disputes actually occur, the evidentiary gap between "what happened" and "what the jury understands happened" collapses. That changes trial strategy. It changes settlement posture. It changes everything downstream of fact development.

The Data Problem Is the Same Problem We Have

The most candid moment in Li's talk came when she discussed training data. Language models feast on the internet. Trillions of tokens scraped from billions of pages. But there's no equivalent internet of spatial data. No massive, structured repository of three-dimensional environments annotated with physics and dynamics.

She described this as the hardest constraint on progress. Not compute. Not models. Data.

Sound familiar?

This is the exact constraint facing legal AI. We don't have a shortage of language models that can draft a motion for summary judgment. We have a shortage of structured, jurisdiction-specific, practice-area-specific data that makes those drafts actually useful in Lorain County versus Cuyahoga County versus Erie County.

World Labs is solving their data problem through a hybrid approach: mixing real-world data with synthetic data generated by their own models. Marble's output becomes training data for the next generation of Marble. A data flywheel.

State Prediction: The Concept Lawyers Should Memorize

Li introduced a technical concept that deserves its own section in every trial advocacy textbook: state prediction.

A "state" is a snapshot of the world at a given moment. A world model's job is to predict the next state. Li compared it to a tennis player reading the trajectory of a ball mid-flight. Not reacting. Predicting. Making decisions based on where reality is going, not where it's been.

The language model analogy is next-token prediction. But in spatial intelligence, a "state" is a richer, more complex unit. It includes geometry, physics, and dynamics. It can span milliseconds or minutes.

Now think about legal strategy through that lens.

The Anthropic Question She Dodged

Ed Ludlow, covering the interview for Bloomberg, asked Li directly about the Anthropic-Pentagon litigation. Her response was diplomatic: she hadn't followed the specific case. She pivoted to a general statement about wanting World Labs' technology to be used broadly across industries and sectors.

Fair enough. She's a CEO with a $1 billion raise to protect.

But the question matters. Because the tension between AI developers and government end-users isn't going away. It's intensifying. The FASCSA dispute between Anthropic and the Pentagon raises constitutional questions about compelled commercial relationships that will shape how every AI company operates for the next decade.

And as spatial intelligence matures, those questions get harder, not easier. A language model that writes emails is one thing. A world model that can simulate physical environments with predictive accuracy is something the defense and intelligence communities will want access to. The governance frameworks we build now determine whether that access comes through negotiation or compulsion.

Why She Left Stanford (And Why It Matters to Solo Practitioners)

One of the most revealing moments came when Li explained why she left academia to found World Labs. Not because academia doesn't matter. Because the resource requirements of frontier AI development have exceeded what any university can provide.

Compute. Data. Focused teams of expensive talent. These three inputs are table stakes for building frontier models, and they require revenue-driven capital structures to sustain.

Li was careful to defend academia's role. Blue-sky research. Curiosity-driven inquiry. Training the next generation. But the building, she argued, has to happen in industry.

This is the same tension playing out in legal technology. The innovation isn't going to come from bar associations or law school clinics. It's going to come from practitioners who understand both the technology and the practice, who are willing to bet on building the infrastructure themselves.

That's not arrogance. That's pattern recognition. Li saw it in computer vision in 2024. The same pattern is visible in legal AI in 2026.

The "Godmother" Question

Ludlow raised the "Godmother of AI" moniker. Li's response was one of the most human moments in the entire talk. She admitted she cringed when she first heard it. It's not her personality.

But she accepted it because she understands what it represents: a generation of researchers — particularly women — who built the foundational infrastructure that every frontier model now depends on. ImageNet wasn't just a dataset. It was the paradigm shift that made deep learning possible at scale.

What This Means for Your Practice Tomorrow

Let me bring this back to earth.

If you're a domestic relations attorney in Ohio, you're not going to be building world models next week. But you should be building the data infrastructure that positions you to leverage them when they arrive. That means structured client records. Standardized financial data pipelines. Consistent case outcome tracking across judges, counties, and case types.

Because when spatial intelligence meets legal practice, the firms that already have clean, structured, jurisdiction-specific data will be able to deploy it. The firms that are still running on yellow legal pads and gut instinct will be playing catch-up for years.

The Godmother Told Us Where AI Is Going

Fei-Fei Li didn't mention a single legal use case on that stage. She didn't have to. The architecture she described — systems that understand space, predict states, and generate three-dimensional models of reality — is the exact architecture that litigation has needed since the first trial attorney tried to explain a car accident to a jury with a whiteboard marker.

The Godmother told us where AI is going. The question is whether the legal profession is paying attention.

Find Me at HumanX

I am at HumanX all week on the Startup Track with SilverTung. If you care about this conversation, I want to meet you.

Because this is not going away. And the practitioners who understand both the technology and the law will define the next era.

Related Reading