World Labs Raises Stakes in Spatial Intelligence: Fei-Fei Li Tells a16z “World Models, Not LLMs, Are the Next Frontier”

June 2025 — Palo Alto, CA

In a wide-ranging fireside chat with Andreessen Horowitz partners Martin Casado and Eric Torenberg, Stanford professor and World Labs co-founder Fei-Fei Li declared that large language models (LLMs) are “lossy compression” of reality, while world models—AI systems that perceive, reconstruct, and reason in 3-D space—will unlock the next wave of applications far beyond robotics.

From Lab Sketch to Unicorn in Three Months

Founded in early 2024, World Labs has already closed two funding rounds totalling US $230 million, attracting a heavyweight cap-table—Andreessen Horowitz, Radical Ventures, NEA, Nvidia’s NVentures, AMD Ventures, and Intel Capital. The rapid capital infusion pushed the company’s valuation past the US $1 billion mark in just twelve weeks, making it one of the fastest-forming unicorns in generative-AI history.

Li revealed that the startup’s seed round was subscribed “within days, not weeks,” because institutional investors see spatial intelligence as the missing bridge between today’s chat-based AI and tomorrow’s embodied, environment-aware systems. Andreessen partner Martin Casado called World Labs “a full-stack moonshot,” spanning new data pipelines, rendering engines, and multimodal model architectures.

Why Spatial Intelligence Pre-Dates—and Outranks—Language

Li opened the discussion with a personal anecdote: Years ago, a corneal injury forced her to live temporarily without stereoscopic depth perception. “Navigating a familiar street became terrifying,” she said. “I spoke the same language, but without 3-D cues I could no longer judge a car’s distance. That convinced me spatial understanding sits at intelligence’s ground floor.”

Evolution backs her up: insects, fish, and mammals mastered spatial cognition hundreds of millions of years before symbolic language emerged. Modern LLMs, impressive as they are, still “talk” about the world without seeing or simulating it. “Remove a blindfold and a human reconstructs a room instantly; words alone cannot replicate that reconstruction,” Casado observed.

The Technical Inflection: NeRFs, Volumetrics, and Hybrid Pipelines

World Labs’ founding team reads like a who’s-who of cutting-edge vision research:

• Ben Mildenhall, co-inventor of NeRF (Neural Radiance Fields), which ignited today’s 3-D scene-reconstruction boom.

• Christopher Choy, whose sparse convolution and volumetric research slashed the cost of high-resolution spatial models.

• Justin Johnson, an early innovator in generative adversarial nets (GANs) for style transfer, now adapting those techniques to textured world synthesis.

Li insists the project’s novelty is not just better math but a systems marriage of deep learning with film-grade rendering. Training world models demands petabytes of photorealistic, view-consistent data—something only a deeply integrated graphics + AI pipeline can feed at scale.

Beyond Robots: A Universal Creative Engine

While embodied robotics remains an obvious beneficiary—giving machines the “sense of space” they need to manipulate objects—Li argues the addressable market is vastly wider:

• Architecture & BIM — Rapid “scan-to-model” workflows for renovation, site planning, and code compliance.

• Industrial design — AI that parses a single product photo, infers hidden geometry, and auto-generates CAD variations.

• Film & gaming — Instant asset generation and physically correct scene layouts, slashing VFX budgets.

• Digital multiverses — Parallel virtual worlds tailored for education, therapeutic exposure therapy, or purely artistic exploration.

Casado illustrated: “Snap a phone pic of your living-room coffee table; the model fills in obscured legs, derives material properties, and lets you rescale or recolor it—no Blender skills required.”

LLMs Are the Prologue, Not the Finale

Li stressed she is not anti-LLM: “Language is humanity’s crown jewel. But it compresses reality—like a JPEG discards detail.” World models re-inject physics, lighting, and causality, enabling AI to reason about mass, torque, occlusion, and temporal change—concepts textual training alone cannot ground.

Transitioning from Stanford lab to venture-backed firm was pragmatic: Training multi-billion-parameter spatial networks consumes GPU-years and orchestration complexity beyond academic budgets. “We needed industrial-scale compute and an org capable of uniting graphics engineers, roboticists, and data-center operators under one roof,” she said.

A New Design Paradigm on the Horizon

The conversation veered philosophical when Casado quipped, “We’re walking evolution backward.” Spatial cognition appeared in trilobites 500 million years ago; articulate language is a blink-ago upgrade. So if AI hopes to rival human dexterity and reasoning, it must master the older faculties first.

Li’s closing salvo: “I’ve waited my entire career for hardware, data, and algorithms to make true 3-D world modeling feasible. The real universe isn’t made of words—and neither will the most powerful AIs of the decade ahead.”

With public previews promised “before the end of 2025,” the industry will soon see whether World Labs can turn this vision into the platform that reshapes robotics, design, and the very nature of digital experience.

See Original Video