a16z Podcast

World Labs on new view prediction, a new base-model primitive

Fei-Fei Li, Justin Johnson & Ben Mildenhall· Co-founders at World Labs
·~44 min·English·a16z
MultimodalRoboticsTrainingAI Company
TL;DR

World Labs' Atlas model reframes 3D as new view prediction, a base-model primitive on par with next token prediction, unifying reconstruction and generation so a handful of photos can rebuild and re-film an entire scene.

01Core Mental Model

A Third Base-Model Primitive

World Labs frames its new model Atlas around new view prediction, putting it in line with the primitives that defined other model families: next token prediction for language models and next frame prediction for video models.

You know, LLMs are built on next token prediction. We've seen video models as being built on next frame prediction. Atlas is really new view prediction.

— Justin Johnson, a16z Podcast
Key Insight
Naming a primitive is a claim about what a model family is for. Framing spatial intelligence as new view prediction tells you Atlas is meant to answer one question over and over: what does this scene look like from a camera placed here, at this moment?
02Architecture

Reconstruction and Generation, Fused

Atlas does something computer vision kept in separate research tracks for half a century: it both reconstructs the parts of a scene the cameras actually saw and generates the parts they never captured, inside a single model.

It's the first time we have a unification of pixel generation and pixel reconstruction.

— Fei-Fei Li, a16z Podcast
Key Insight
The two halves are not just bundled for convenience; they need each other. Any pixel no camera saw is a hole a pure reconstructor cannot fill, so the model has to imagine it, which is a generative act built on the geometry it already recovered.
03Why It Is Different

Grounded, Not a Slot Machine

Because every input image carries a 3D camera pose, Atlas reconstructs a scene with precise, repeatable control, instead of the retry-until-it-looks-right guessing that text-prompted video models rely on.

if we had four views of this room, one at each corner, you can put those into the model and then get an exact replication of everything you see in this room and it's not going to guess what's in the other corner

— Ben Mildenhall, a16z Podcast
Key Insight
This is the difference between a creative toy and a production tool. When each frame is anchored to a known viewpoint, the output is stable enough to build on, which is what lets professionals treat a generated world as a persistent stage rather than a lucky screenshot.
04The Capability

Sparse Beats Dense

Traditional 3D capture needs hundreds of overlapping photos to cover every crack and crevice of a room; Atlas rebuilds a comparable scene from as few as three, a 50 to 100 times reduction in the inputs a person has to gather.

With Atlas, we can do this with just three cameras. No studio capture, no green screen, no expensive calibration.

— Justin Johnson, a16z Podcast
Key Insight
A big enough quantity change becomes a quality change. When capture drops from a tedious hour to three phone shots, footage no one ever treated as reconstructible, old clips, internet video, becomes raw material for 3D, which widens the market far more than a prettier render would.
05The Bridge to LLMs

Reconstruction Is Long-Context Generation

World Labs treats reconstruction as generation run with a very long context window, so pouring in more views grounds more of the world, the same lesson language models learned when they pushed context length from thousands of tokens to millions.

reconstruction is just like generation with a really long context and you put a lot of stuff in it

— Justin Johnson, a16z Podcast
Key Insight
This reframing is why Atlas jumps where the earlier Marble model stalled. Marble could hold only a couple of images at once; casting the task as long-context generation lets Atlas absorb dozens of views into one coherent scene instead of stitching many small guesses together.
06The Bet

We're at the Beginning

The founders say Atlas is not near the ceiling of its architecture but at the start of its scaling curve, held back today by training compute rather than by data or a missing breakthrough.

we're basically at the beginning and we're basically limited by compute at this point

— Justin Johnson, a16z Podcast
Key Insight
Calling compute the bottleneck is a confidence signal, not a complaint. It says the team believes the recipe already works and only needs more of the same, which is the position every scaling-era lab wants to be in and the reason the released model was sized to a release deadline, not to a limit.
07Robotics

The Simulator Becomes the Planner

Robotics is bottlenecked by data, so World Labs uses Atlas as a neural simulator to turn a few real captures into many varied training scenes, and then makes a deeper claim: a model that can predict how the world responds to an action can also plan the action.

understanding how the world's going to respond to an action is highly related to imagining what kind of action I need to take to make the world respond in a particular way

— Justin Johnson, a16z Podcast
Key Insight
This collapses two robotics stages into one. If the same world model that generates training scenes also carries an internal sense of cause and effect, then the simulator you built to gather data quietly becomes the thing that decides what to do, which is a much larger prize than faster data collection.
08The Thesis

New View Prediction Is 'AI-Complete'

World Labs argues new view prediction is AI-complete, as fundamental to intelligence as next token prediction, and points to evolution for intuition: animals that move evolved eyes, because motion is what constantly presents a new viewpoint to predict.

Because when you move, you see a new viewpoint.

— Fei-Fei Li, a16z Podcast
Key Insight
AI-completeness is a bet that one narrow task, pushed to full generality, forces you to solve everything. By placing new view prediction beside next token prediction, World Labs is staking out spatial intelligence as a road to general intelligence, not a niche tool for 3D content.