Context

World Labs co-founder Justin Johnson joins MTS hosts Theo Jaffee and Sofia Puccini the same day his company announces Atlas, a new multimodal "world model" that generates, reconstructs, and simulates 3D environments with precise camera control. Johnson lays out World Labs's core thesis: just as language models became a general-purpose horizontal engine for working with text, world models could become an equivalent horizontal layer for visual and physical intelligence, applicable across entertainment, gaming, construction, and robotics. For a PM, the episode is a useful case study in platform-thesis product strategy: betting on a new category of foundational model rather than a single vertical application, and the concrete tradeoffs that come with committing to that bet before the products built on top of it exist.

The Big Idea

World models, AI systems that understand and simulate physical, three-dimensional space rather than just text, are a distinct and equally horizontal category to language models, and building the underlying model with deep, precise creative control (not just generation quality) is what will let it become a foundational layer other industries build on, rather than a one-off demo.

Johnson frames Atlas explicitly as a base model announcement, not a product launch, reflecting a deliberate platform-first strategy: prove out the general capability before committing to how it gets packaged for any single use case.

Key Insights

Atlas does three fundamentally different jobs in one model

Johnson breaks down Atlas's capabilities into three distinct tasks. Generation creates entirely new environments that never existed, starting from a text or image prompt and letting a user fly a camera through the resulting space. Reconstruction takes real-world input, as few as one photo, up to 100 or more, and rebuilds that actual physical space in 3D, with accuracy scaling directly with how many reference images are provided. Simulation takes a reconstructed space and models how objects or robots would behave within it. Johnson is explicit that these three capabilities serve different industries (VFX and gaming lean on generation and reconstruction, robotics leans on simulation), which is part of why he frames the model as horizontal infrastructure rather than a single-purpose tool.

Spatial grounding keeps long generations from degrading

A common failure mode in AI video generation is that outputs get progressively jumbled or start "hallucinating" their own direction the longer the generation runs. Johnson's fix in Atlas is to ground every reference image in actual 3D space rather than treating it as a flat 2D prompt: a user can stage multiple reference images at specific points in 3D and then precisely steer the camera's path between them, effectively leaving a trail of anchor points the model can follow. He describes this as turning long-horizon generation from "a slot machine" into something a user actively directs, with demonstrated control even at a full minute of generated video, an unusually long horizon for this kind of model.

Rebuilding from scratch beat iterating on the prior product

World Labs's first product, Marble, routed nearly everything (video, 3D scenes, meshes) through a 3D Gaussian splat representation as a central bottleneck. Johnson says the team deliberately broke that dependency in Atlas, rearchitecting so that 2D outputs (video, images) generate directly rather than being rendered down from an intermediate 3D representation, with 3D only computed for applications that specifically need it. This is a concrete example of a team choosing a full architectural rebuild over incremental iteration once a foundational technical constraint (the Gaussian-splat bottleneck) was identified as limiting future scale and unification.

Real-to-sim-to-real could compress robot onboarding to minutes

Johnson describes a workflow where a general-purpose, pre-trained robotics foundation model still needs adaptation to a specific physical environment (a particular factory floor, a particular table setup) before it will perform reliably there. Today, that typically requires collecting many demonstrations of a robot, or a human teleoperating one, physically working in that exact space. Johnson's proposed alternative: take a handful of casual phone photos of the space, use Atlas to reconstruct it in 3D, describe the desired task in natural language, and have an AI agent build a simulation to fine-tune the robot against, potentially compressing environment-specific adaptation down to a span of minutes rather than an extended data-collection process.

Meeting existing workflows beats forcing a full AI-native switch

Johnson pushes back on the idea that explicit 3D representations like Gaussian splats or meshes are becoming obsolete now that models can output pixels directly. His argument: real-time applications on embedded or mobile devices still need something a device can render client-side, and entire industries (gaming, VFX, architecture, design) have existing pipelines built around these explicit 3D formats. Rather than asking those industries to abandon established workflows all at once for a fully AI-native approach, he treats supporting both explicit 3D output and direct 2D/video generation in the same model as a deliberate strategy to meet users where they already are.

Mental Models & Frameworks

World models as the visual-physical analog to language models

Johnson's central strategic framework: language models succeeded by being a general, horizontal engine for one modality (discrete text tokens) that then found applications across countless industries. He argues world models can follow the identical pattern, but for visual and physical understanding, becoming a horizontal layer usable across entertainment, VR, construction, and robotics rather than a narrow, single-purpose tool. This framing directly explains World Labs's product sequencing: build and prove the general-purpose base model first (Atlas), and let specific vertical products get built on top of it later.

Creative control over generation quality as the differentiator

Johnson repeatedly frames World Labs's product philosophy in opposition to what he calls "generative slot machines," tools where you pull a lever (submit a prompt) and hope for a good result without feeling ownership over the outcome. His alternative goal is deep, precise control (3D spatial staging, exact camera paths, not just text descriptions) so a creative user feels like a director actively guiding the output, not a gambler hoping for a good roll. This is presented as a deliberate product design principle, not just a technical nice-to-have, and it shapes concrete architecture choices like treating camera position as a native input modality rather than an afterthought.

Trade-offs & Nuance

Explicit 3D versus direct pixel generation is not fully resolved

Johnson describes two live technical paths for how world models ultimately deliver content: an explicit 3D route (generating a Gaussian splat or 3D scene that then gets imported into a game engine or rendering pipeline) and a direct route (the model outputs finished 2D frames without any explicit 3D representation in between). He says Atlas is built to support both, rather than betting the whole product on one path, because different use cases currently favor different approaches: embedded and mobile devices benefit from lightweight explicit 3D formats they can render locally, while VFX work where the final output is just a 2D video may not need an explicit 3D intermediate at all.

Physics accuracy has real limits tied to training data

Asked directly whether Atlas could handle extreme physics scenarios like a black hole, Johnson gives a candid, ungated answer: behavior that is roughly Newtonian, the physics regime the model's pretraining data covers, is likely fixable with a relatively quick supervised fine-tune. Non-Newtonian regimes, like black hole physics or quantum-mechanical scale phenomena, would likely require rethinking both the model architecture and the training data, not just additional fine-tuning. This is a useful, non-hyped example of a founder being specific about where a capability plausibly extends versus where it genuinely does not, rather than implying the model can do everything.

Practical Application

Ground long-form AI generation in explicit spatial anchors

If building a product that generates long-horizon content (video, simulated environments, multi-step outputs), consider whether anchoring the generation to explicit, controllable reference points, the way Atlas stages reference images in 3D space and lets a user define an exact camera path, could prevent the quality degradation that often appears in long single-shot generations.

Architect for both a legacy-compatible format and a native output

When your AI product could serve users with entrenched existing workflows, Johnson's approach with Atlas (supporting explicit 3D output like Gaussian splats and meshes alongside direct 2D/video generation) suggests deliberately supporting a workflow-compatible format alongside a more native one, rather than forcing every user through a single output format that may not fit downstream pipelines like game engines or VFX tools.

Treat environment-specific adaptation as a distinct, minimizable step

For any product involving a general-purpose model that needs to adapt to a specific physical or contextual environment (robotics is the clearest case here, but the pattern generalizes), separate the general capability from the environment-specific fine-tuning step, and look for ways to shrink the cost of that last-mile adaptation. Johnson's real-to-sim-to-real concept, using a handful of casual photos to build a fine-tuning simulation, is a concrete example of treating that adaptation step as an explicit design target rather than an afterthought.

Be precise about a model's real capability boundaries in your own messaging

When describing what your AI product can and cannot do, follow Johnson's example of giving a direct, specific answer about likely failure modes (black hole physics, quantum-scale phenomena) rather than implying unlimited capability. This kind of candor is more useful to a technical audience evaluating whether to build on your platform than vague optimism.

Questions to Consider

• If your product generates long-form or multi-step AI output, could anchoring the generation to explicit, user-controlled reference points reduce the drift or quality degradation Justin Johnson describes in ungrounded long-horizon generation?
• Does your AI product force users through a single output format, or does it, like Atlas, support both a format compatible with users' existing workflows and a more native format for new use cases?
• If your platform needs to adapt to many different specific environments or contexts, have you identified and tried to minimize the cost of that last-mile adaptation step, the way World Labs's real-to-sim-to-real concept targets robot onboarding time?
• When describing your own product's capabilities to technical users or partners, are you being as specific about where it currently falls short as Justin Johnson was about Atlas and non-Newtonian physics, or are you leaning on vaguer, more optimistic language?

Bottom Line

Justin Johnson's account of building Atlas shows a platform bet in action: rather than iterating on an existing product, World Labs rebuilt its entire model architecture around the thesis that world models are a horizontal category as foundational as language models, prioritizing precise creative control and dual support for legacy and native output formats over chasing generation quality alone.

Case Studies Mentioned

Bullet-time shots from three iPhones

One of the most surprising results the World Labs team found while testing Atlas: using just three ordinary iPhones on tripods to capture a synchronized moment (their headline example is a strawberry dropping into a bowl of oat milk), Atlas could reconstruct the scene well enough to let a user freeze the moment and fly the camera around it in a full bullet-time effect, a capability Johnson says worked far better than the team expected from such a minimal three-camera setup.

Marble in VR as a transformative early experience

Johnson notes World Labs has had its earlier product, Marble, working in VR for almost a year, describing the experience of putting on a headset and being immersed in a fully AI-generated world you created yourself as genuinely transformative, and encourages anyone who hasn't tried it to do so, framing it as a preview of the kind of experience Atlas-powered products aim to deliver at greater scale and fidelity.

People to Follow

Justin Johnson

Co-founder of World Labs, a startup building spatial intelligence and world-model products. Johnson leads the technical vision behind Atlas, World Labs's multimodal world model announced in this episode, and frames the company's strategy around world models becoming a horizontal, general-purpose layer for visual and physical AI across entertainment, robotics, and other industries.

Notable Quotes

"We never wanted to be building these generative slot machines where you just pull the thing and hope you get out a good generation." (Justin Johnson)

"We live our lives in this physical built space all around us, and we need models to help us with these things as well." (Justin Johnson)