Investors poured over $1.6 billion into world model companies in just a few months — a new AI category trained to simulate physical reality rather than process language. Ars Technica spoke with researchers from MIT, Runway, and World Labs, revealing both the potential and the unresolved challenges of this technology.
Key takeaways
- Runway raised $315M in February 2026 for world models (GWM-1)
- World Labs (Fei-Fei Li) and AMI (Yann LeCun) each raised approximately $1B
- Key difference from LLMs: real-time, continuous interaction — not turn-based Q&A
- Primary applications: training robots on synthetic data and 3D asset generation
- Technically: autoregressive diffusion?Autoregressive diffusion: Generating video frame by frame, each new frame conditioned on the previous ones — which makes real-time interaction possible. instead of generating all frames at once
What is a world model
"World model" is an overloaded term — the researchers themselves admit it. In the broadest sense, it describes any AI model that takes an action as input and predicts what will happen in the environment next. Runway defines it as "an AI system that builds an internal representation of an environment and uses it to simulate future events."
The key difference from LLMs lies in the interaction mode. Ben Mildenhall, co-founder of World Labs, explains: language models are turn-based — user types, model responds. A world model should operate synchronously, in real time, where many things are happening at the same time. Fei-Fei Li, World Labs co-founder and a computer vision pioneer, identifies three defining properties: generating worlds with perceptual, geometrical, and physical consistency; multimodality by design; and the ability to output next states based on input actions.
Money flows to physical AI
In just a few months, the sector attracted substantial capital. Runway raised $315M in February 2026, focusing on GWM-1. Google DeepMind unveiled Genie 3 in August 2025 — a model building real-time interactive simulations. World Labs released Marble, a tool for generating immersive 3D environments from text or images. Yann LeCun launched AMI, a company betting on models that interact with the physical world.
| Company | Product | Funding |
|---|---|---|
| Runway | GWM-1 | $315M |
| World Labs | Marble | ~$1B |
| AMI | — | ~$1B |
| Google DeepMind | Genie 3 | — |
Investors are drawn by concrete applications: training and testing robots on synthetic data, generating 3D assets for game development and VFX, and scientific simulation.
The technique: autoregressive diffusion
Most current world models generate video as their primary output. Standard video models generate all frames simultaneously through a denoising process — this achieves high quality but makes real-time interaction impossible.
Runway: autoregressive diffusion
Runway adopted autoregressive diffusion: frames are denoised sequentially, and the user can intervene between batches. Anastasis Germanidis, Runway's CTO, explains: "We denoise one or a few frames at a time. We present it to the user, the user provides an action that influences the next frames." The drawback is enormous compute demand and a long-term memory problem.
World Labs: Gaussian splatting
World Labs chose a different approach: Marble generates environments as collections of Gaussian splats?Gaussian splatting: Representing a 3D scene as a cloud of millions of tiny translucent blobs (Gaussians) instead of a polygon mesh — light and fast to render. — a volumetric format compatible with standard 3D workflows. The output is static and exportable, but has no dynamics or physics.
The bitter lesson on scaling
Both approaches share a core assumption: models should learn world representations themselves, without hardcoded 3D geometry or physics equations. This is Richard Sutton's "bitter lesson" — the observation that attempts to embed human knowledge into AI algorithms have historically lost to scaling compute and learning from raw data.
Germanidis from Runway: "Scale and pixel prediction should be enough for 3D consistency and statefulness to emerge on their own." Sitzmann from MIT is more cautious: video models may provide useful physics representations in latent space?Latent space: The internal, compressed space in which a model encodes data features (here: scene physics) in a form not directly human-readable., but whether this is a necessary or sufficient path to embodied AI remains an open question.
Where world models help robots
The most significant potential application is generating synthetic training data for robots. The data problem for humanoids is the inverse of self-driving cars: autonomous vehicles gather data by driving the same roads they will operate on. Humanoid robots in homes face unlimited variety — and data is scarce.
Germanidis: "You can generate synthetic data faster and more safely than you can collect it in reality." Sitzmann warns that small deviations in contact simulation, friction, or force can break robot control policies.
Why this matters
World models will not replace LLMs anytime soon, but they signal a direction a significant portion of AI research is moving: from language reasoning toward spatial and physical understanding. If these models prove effective at generating training data for robots, they could unlock progress in home robotics currently blocked by data scarcity. The investment figures — over $1.6B in just a few months — signal the industry treats this not as a lab experiment but as a bet on the next phase of AI.
What's next
- Runway plans to unify its three GWM-1 models into a single general model — no timeline given, but intent confirmed by CTO Germanidis
- World Labs is actively working on adding dynamics to the currently static Marble environments, according to Mildenhall this is an active research priority
- Evaluating world models for robot training requires benchmarks comparing policies trained on synthetic data versus real data — Sitzmann identifies this as the critical next step
Sources
- Ars Technica — Simulating everything, sort of: The promise and limits of world models
- Runway — Introducing Runway GWM-1
- World Labs — RTFM: Real-time generative frame-based model
