Generative AI: The 2026 Field Guide

The most speculative frontier is the one OpenAI's Sora technical report tried to claim, that scaling video generation models eventually produces something more than a video generator. It produces an implicit model of physics, object permanence, scene composition, and the temporal logic of the real world. Sora's report called this 'video generation models as world simulators,' and the claim has been controversial since the day it was published.

ANALOGY

Imagine the difference between a dollhouse you can look at and a dollhouse you can shrink down and walk inside. The dollhouse you look at is a video model: you can press a button and get a beautiful pre-rendered scene, but you cannot interact with it, cannot change what happens, cannot ask 'what if I opened that door instead.' The dollhouse you can walk inside is a world model: physics applies, objects respond to your actions, the scene persists when you turn around and come back. Today's video generators are firmly in the first category, they can produce footage that looks plausible most of the time, but they fail at exactly the kind of internal consistency you would need for the second category. The open question is whether scaling alone gets us from one to the other, or whether something fundamentally different has to be invented. If the strong scaling story is right, robots train themselves in generated worlds, game engines become obsolete for many use cases, and entire categories of simulation get cheap. If it is wrong, video generation remains a creative tool and never becomes a simulator. The honest answer is we will know within two years.

The strong version of the claim is that a sufficiently large video model has internalized enough about how the physical world works that you can use it as a simulator, for robotics training, for scientific computation, for counterfactual reasoning. The weaker version is that video models exhibit some emergent world-modeling behaviors but are nowhere near reliable enough to be used as actual simulators. The empirical state of evidence in 2026 is somewhere between these, the best video models can produce footage that respects basic physics most of the time, but they fail in characteristic ways (objects passing through each other, gravity reversing, conservation of mass being violated) often enough that calling them 'simulators' is generous.

Fig.diagram

■A world model closes the loop. Instead of rendering a fixed clip it takes an action, predicts the next frame, updates its internal state, and waits for the next action. Video generation becomes an interactive simulator you can steer.

The most interesting work in this direction is happening at Google DeepMind, at Fei-Fei Li's World Labs, and at a handful of smaller research groups. DeepMind's Genie 3 is the current frontier for interactive video environments, it takes a text promptPromptThe text description you provide to a model to specify what you want it to generate. or image and generates a navigable space that the user can explore in real time, with the model maintaining spatial consistency as the viewer moves through it. Genie 3 is the third major iteration of this work (Genie 1 was a 2D environment generator, Genie 2 added richer 3D navigation in late 2024), and it is explicitly positioned by DeepMind as a step toward generative game engines and robotics simulation environments.

World Labs, founded by Fei-Fei Li in 2024, shipped Marble in late 2025, which is the clearest product expression of the world-model idea to date. Marble takes a single image or text prompt and generates a persistent, interactive 3D environment, you can walk around in it, interact with objects, save the state, and come back to it later with the world in the same configuration you left it. The fal/a16z State of Generative Media 2026 report explicitly flagged Marble as the moment world models moved from 'prototype to product,' meaning that for the first time a generative 3D environment was good enough to ship to users rather than just demonstrate in a paper. Marble is currently positioned for creative and game development use cases, but the technology obviously generalizes to robotics training, architectural visualization, and education as the quality improves.

The other players to watch are OpenAI (which has reframed its internal compute spend toward world simulation research as part of the stated rationale for the Sora shutdown), Meta (whose Mango model under Alexandr Wang's Superintelligence Lab is expected to push on world-model capabilities), and the Chinese labs (Tencent's HY-WorldPlay released RL post-training code in March 2026 for building real-time interactive world models on top of HunyuanVideo). This is the most contested research direction in generative imagery right now, and the labs investing in it are betting that world models are the next major category after video generation, not the same category scaled up.

If world models work, really work, in the sense that they become reliable enough to use as simulators, they change a lot. Robotics training stops requiring physical robots and uses simulated worlds instead. Game engines become obsolete for some use cases, replaced by generative world models that produce environments on demand. Architectural visualization, scientific simulation, and education all get new tools. The economic implications are enormous, but the technical question of whether the scaling story actually works is still open. The honest answer in 2026 is that we will find out within the next two years, and the answer will reshape the field one way or the other.

INSIDE SCOOP

There is a structural dynamic worth noting. The labs working on world models are mostly the same labs that are working on artificial general intelligence, DeepMind, OpenAI, Anthropic, Meta, and a small number of others. The reason is that the underlying technical question (does scaling produce understanding?) is the same in both cases, and the techniques transfer between text and video and 3D. World models are not really a separate research direction from AGI research; they are AGI research applied to a specific domain. This means that the labs with the most resources for AGI work are also the ones with the most leverage on world models, and the open-source ecosystem is much weaker here than it is for image and video generation. If world models become important commercially, the field will look much more like LLMs (a few dominant players, very few open weights) than like image generation (a healthy open ecosystem).

World models and interactive generation

Get the week's moves in your inbox.