Context

a16z's Martin Casado interviews the three co-founders of World Labs (Fei-Fei Li, Justin Johnson, and Ben Mildenhall) right after they launched Atlas, a "world model" for spatial intelligence. The technical topic is 3D: generating, reconstructing, and simulating physical space from a handful of images. But underneath the computer-vision detail are product lessons that generalize well beyond 3D: how choosing the right primitive can define a category, how an order-of-magnitude cost reduction changes which products are even possible, and how to add control to a model without wrecking the quality that made it worth using. For a PM, it's a clear worked example of building conviction and making architecture bets under real uncertainty.

The Big Idea

World Labs bet the company on a single new primitive, "new view prediction," as the spatial-intelligence equivalent of next-token prediction for language. Given some views of a scene, the model predicts what that scene looks like from any other camera position. The claim is that one general primitive, scaled, unlocks a whole category rather than a single feature.

The payoff they emphasize is practical: Atlas cuts the number of photos needed to reconstruct a space from hundreds down to roughly three, a 50 to 100x reduction that changes what's worth building at all.

Key Insights

A new primitive can define a category

The team frames Atlas around "new view prediction" the way LLMs are framed around next-token prediction and video models around next-frame prediction. Their conviction rested on two hypotheses: a scaling law (bigger model, longer training, more chips reliably got better) and that new view prediction was the right primitive to scale. The product lesson for anyone building on AI: the durable bet is often the underlying primitive, not the surface feature, because a well-chosen primitive is what generalizes and keeps improving with scale.

Order-of-magnitude cost drops flip the calculus

Traditional dense 3D reconstruction needs so many photos that capturing one room can take an untrained person one to two hours, because every surface has to be seen from multiple angles. Atlas reduces that to as few as three images, a 50 to 100x cut. Ben Mildenhall's point is that this doesn't just make the old workflow cheaper, it flips the calculus on which captures are worth doing: you can now reconstruct old footage, casual videos, even images pulled off the internet, none of which was ever treated as reconstructable before. A cost drop of this size creates new use cases rather than just accelerating existing ones.

Unify capabilities that were always separate

Historically, reconstruction (rebuilding a real scene exactly) and generation (imagining something new) were separate subfields with separate models and separate conference tracks. Atlas does both in one model by anchoring everything on the viewpoint. Their reframe is elegant: reconstruction is just generation with a very long context. If a surface was captured, the model triangulates it; if it wasn't (and even an expert always misses spots under tables or between chair legs), the model generatively fills the gap. Combining two capabilities users previously stitched together by hand is itself the product unlock.

Long context is underused on the visual side

Justin Johnson notes that LLM builders spent years pushing context length (128 to 256 to a million tokens) and everyone now intuits its value in a coding model, but no one pushed this on images and video in the same principled way. Atlas treats reconstruction as dumping many views into a long visual context: a 64-image capture can produce a coherent fly-through of a whole house, a 2,000-image house capture comes down to 30 or 40 inputs with a nearly identical result. The transferable idea: a technique that's mature in one modality (long context in language) can be an untapped unlock in another.

Add control without degrading quality

Ben's rule for the roadmap: "the trick here is you've got to add control but not compromise the quality of the model or it just becomes a party trick." No serious user will swap their frontier video model for one that hands them extra knobs but produces worse output. This is a general product principle for adding capability to any AI feature: new controls (editing, layout, object identity, time) only matter if the base quality bar holds. A knob that costs quality is a demo, not a product.

Real workflows are stateful and multi-stage

Nobody uses one monolithic model end to end. Creatives pull mood boards from favorite image models, use different video tools to build keyframes, then clip and edit later. And they want persistent 3D state (a stage, props, elements they build up over time), not ephemeral generate-and-throw-away outputs where a room's objects drift every time you re-prompt. Designing for the messy, multi-tool, stateful reality of how people actually work, rather than an idealized single-model flow, is where Atlas aims its "spatial context" control.

Meet users where 95% of the pain is

Ben describes booth builders, architects, and set designers whose real bottleneck isn't the idea, it's translating verbal feedback or sketches back into a 3D representation. Mapping a creative director's note into 3D software is "95% of the work": a meeting produces feedback, then a week of revisions, because the software is decades old and never became as intuitive as sketching or playing with Legos. The lesson is to find the specific step that eats most of the effort and aim the product at removing that, rather than at the part that already feels easy.

Mental Models & Frameworks

Reconstruction equals long-context generation

Treat exact reconstruction and creative generation not as opposites but as two ends of one continuum, separated only by how much grounded input you provide. Many precise views give you faithful reconstruction; few or zero views lean on the model's generative imagination. The same model serves both by scaling its context. Use this frame whenever a product seems to need two "different" models: check whether one capability is just the other with more or less grounding.

The AI-complete primitive

Borrowed from Turing completeness: a task is "AI complete" if solving it in full generality would solve general intelligence. The classic example is next-token prediction (predict the last word of a mystery novel and you've had to understand the whole thing). The founders argue new view prediction is similarly complete, and add an evolutionary framing: nature gave animals eyes because they move and must predict new viewpoints, and didn't give trees eyes because they don't. The takeaway for choosing what to build: the most powerful primitives are the ones broad enough that mastering them would pull a whole class of problems along.

Climb the scaling ladder before the big bet

Rather than guessing the final architecture, World Labs ran a sequence of smaller models up a "scaling ladder," each bigger and trained longer, to build conviction on what would scale before committing the expensive training run. The decisive moment came from a smaller model producing an emergent result (a camera flying under a table with a soccer ball), after which they committed in seconds. Use it as a method for high-stakes bets: buy conviction with cheap experiments that reveal the trend, rather than betting everything up front.

Decision Principles

Principle: Don't bottleneck on one output format

• When: designing a model or system that could emit several representations.
• Why: their previous model, Marble, forced every output through Gaussian splats, which became a bottleneck. Atlas keeps the primitive general (new view prediction that can yield RGB frames or 3D) and produces splats only when actually needed. Committing your architecture to one output format early limits what the system can later do; keep the core primitive general and derive specific formats on demand.

Principle: Feed the model more to get less

• When: you want a specific, narrow output (here, clean static 3D).
• Why: counterintuitively, the best way to get high-quality static reconstruction was to train on lots of dynamic data and let the model learn to factor the motion out, because purely static training data is hard to scale. Exposing a model to richer, messier data than your target output can produce a better result than trying to curate only the narrow data you think you want.

Trade-offs & Nuance

Dynamics vs precise reconstruction

There's a real tension: exact 3D reconstruction wants frozen, static scenes (so points can be triangulated precisely), while realistic and robotics-useful worlds need motion. World Labs' resolution is not to pick one but to pre-train on heavy dynamics and then post-train a given checkpoint toward static, spatial precision, keeping latent dynamics available to build on. The nuance worth carrying: two requirements that look opposed at the data level can sometimes be sequenced (broad pre-training, then targeted post-training) rather than traded off.

Generality vs robotics readiness

Atlas is a general world model, but the founders are candid that robotics has a different, harder need: a robot policy is an agent acting in a world that responds unpredictably, so it must be trained against every scenario that could go wrong. Today the binding constraint in robotics is data, not chips, which is why simulation and "real to sim" reconstruction matter. A general model helps, but they're clear it isn't yet a robotics-robust foundation model, a useful reminder not to overclaim a general capability for a specialized, safety-critical domain.

Questions to Consider

• For our product, are we betting on a durable underlying primitive that improves with scale, or on a surface feature that a stronger base model could absorb overnight?
• Where in our product could an order-of-magnitude cost or effort reduction not just speed up the current workflow but make entirely new use cases viable, and are we scoping for those new use cases or just the speedup?
• As we add controls and knobs to an AI feature, do we have a hard rule that base output quality can't regress, or are we at risk of shipping impressive-looking controls that quietly make the core result worse?
• Which single step in our users' workflow eats most of their time (the "95% of the work" step), and is our product aimed at that step or at the part that already feels easy?

Bottom Line

World Labs' Atlas is a case study in AI product strategy dressed as a 3D demo: pick a general primitive worth scaling, refuse to bottleneck it on one output format, and let a 50-to-100x cost drop open new use cases rather than just speeding up old ones. The recurring product discipline is to add capability and control only where it doesn't cost the quality that made the thing worth using.

Case Studies Mentioned

Bullet time with three cameras

The famous Matrix "bullet time" shot (Neo frozen mid-air while the camera flies around him) originally required a ring of hundreds of cameras on a green screen. Atlas reproduces the effect from as few as three iPhone videos, with no studio, green screen, or calibration: capture something like milk splashing, then freeze time and fly a virtual camera through it. The lesson is concrete proof of the cost-flip: a result that once needed a full production rig now needs three phones, which changes who can make it and what it's used for.

The Stanford quad reconstruction

In a demo, Ben reconstructed the entire Stanford quad from just 3 to 25 input images, all taken from the ground, then rendered it from an aerial view no camera ever captured. Everything in that aerial view was generated, but constrained by the laws of reconstruction from the ground-level inputs. It shows the generation-plus-reconstruction thesis working in practice: the model fills unseen viewpoints imaginatively while staying grounded in what was actually observed.

Tools & Products

Notable Quotes

"We 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)

"The trick here is you've got to add control but not compromise the quality of the model or it just becomes a party trick." (Ben Mildenhall)

"Nature gave animals eyes, but nature didn't give trees eyes. Why? Because when you move, you see a new viewpoint." (Fei-Fei Li)