Blog · Field notes · Oct 21, 2026 · 5 min
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.
In eighteen months, world models went from short clips to places you can walk through. A text prompt now becomes a navigable scene; a few photographs become an editable 3D world; a video model can plan a robot's next move. The progress is real and it is visual. What the field has shown so far is space, for minutes. What it has not yet shown is time measured in years, populations of lives that change each other, and predictions scored against a slice of reality the model never saw. Those three are where the next gains lie.
Google DeepMind's Genie 3, announced on 5 August 2025, generates interactive environments from text at 720p and 24 frames a second, navigable in real time. DeepMind reports that its worlds stay consistent for a few minutes, with visual memory reaching back about a minute, and that weather or new objects can be prompted into a scene as it runs.
World Labs made Marble generally available on 12 November 2025. It builds persistent 3D worlds from text, a single image, several images or a short video of a real place, or from a coarse layout of simple shapes, and exports them as Gaussian splats, triangle meshes or rendered video. The world is an asset that can be edited, expanded and composed with others.
NVIDIA launched Cosmos at CES on 6 January 2025 as a platform of world foundation models for robots and vehicles, with a curation pipeline that NVIDIA says can process and label 20 million hours of video in 14 days on Blackwell hardware. Its stated purposes are synthetic data, scenario search and testing policies in simulated situations. In October 2025 it followed with Cosmos Predict 2.5, which folds text-, image- and video-to-world generation into a single model.
Meta's V-JEPA 2, released in June 2025, takes a different route: it predicts in a learned representation rather than in pixels. Pre-trained on more than a million hours of internet video and then on under 62 hours of unlabelled robot video, it was deployed zero-shot on robot arms in two labs and picked and placed objects towards image goals, with no data collected in those labs.
Four teams, four definitions of a world model: a playable scene, an exportable place, a generator of training data, a predictor for planning. Each is a genuine step.
The clearest measurement of the gap comes from Physics-IQ, a benchmark from Google DeepMind and INSAIT published at WACV 2026. It asks video models to continue real recorded scenes that require an understanding of fluids, optics, solid mechanics, magnetism and heat. Testing Sora, Runway, Pika, Lumiere, Stable Video Diffusion and VideoPoet, the authors found that physical understanding was "severely limited, and unrelated to visual realism". Some cases were already solved, which suggests certain principles can be learned from watching. But the property a viewer rates most highly, how real the frames look, told them almost nothing about whether the model knew what happens next.
That finding generalises. A world model is useful in proportion to how well its predictions hold when something changes, and that can only be measured by comparing it with what actually happened.
The first is time. Genie 3's own list of limitations includes interaction that lasts "a few minutes, not extended hours". The questions that matter most for places with people in them run for years: how a district ages, how a cohort's choices compound, how a shock in one decade echoes in the next. Consistency over minutes and consistency over decades are different engineering problems, because over decades the state that matters is mostly invisible: who knows whom, who owes what, who is still alive.
The second is lives. DeepMind lists "interaction and simulation of other agents" among Genie 3's open challenges. The current generation of world models is built around one viewpoint moving through a scene. A society is the opposite case: millions of viewpoints, each with a history, each changing the options of the others. Some effects only exist at that level. They cannot be seen in any single life and they cannot be rendered from a single camera.
The third is held-out reality. Today's world models are mostly evaluated on how convincing their output looks, on downstream task success, or on benchmarks built for the purpose. Physics-IQ is a welcome exception. What the field has rarely done is fix a prediction in advance, then score it once against measurements from a period, place or population the model never touched. Without that step, a world model's account of the future remains a plausible story.
As AI systems take on decisions that affect where people live, how cities run and how machines move among us, the useful question about a world model is how far its predictions can be trusted, and over what range. A model that renders a street perfectly but cannot say how the street's residents will change over ten years answers a different question from the one being asked. The two lines of work meet in the middle: rich places, filled with lives that behave like real ones, checked against the world before anyone relies on them.
Living Worlds is a simulation engine for persistent populations. It creates people, places and environments, runs them through simulated time and records every consequential change, and each model is registered and scored on real data it has not seen.
On time and scale: one world of a million lives runs for ten years on UK 2019 death and birth rates and replays exactly, with an open checker that rebuilds every event without the engine. That study checks replay, and says plainly that its partnership rates are placeholders, so it makes no claim to realism. On lives that change each other: after a decade with more boys born than girls, the men of those years are 3.4 points less likely ever to have had a partner thirty years later, in all 20 runs at a million lives, an effect that is zero when each life is computed alone. Its direction comes from people having to find each other; its size follows placeholder rates.
On held-out reality: a population lived month by month from 1950 matched how many children Americans had had within 0.17 on average, against 0.36 for a frontier model answering from what it has read (Met). Women's life expectancy forecast twenty years ahead came within 0.92 years on figures nobody had read (Met, women only), while an earlier forecast across 155 countries missed its target of 1.0 year at 2.06 for women and 2.42 for men (Not met). Both results are on the research page, and the miss is part of the record.
Sources
Blog
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.
Learned weather models are now operational, faster and often more skilful than physics on ordinary days, and weaker on record-breaking ones. The habits that made them trustworthy, daily scoring against what happened and honesty about extremes, are what simulation of lives and economies needs next.
Living Worlds runs more than a billion life-months a second on one desktop GPU and on one desktop processor, with every birth, death, partnership and separation identical across engines; a million-life world that anyone can replay; and an effect of who is born that appears only when lives have to find each other.