Most companies calling themselves world models are not. That is less a criticism of the companies than a problem with the term. In March, Alexandre LeBrun, CEO of Yann LeCun's AMI Labs, predicted that within six months, every company would call itself a world model to raise funding.1 Dealroom now counts a record $3.2B of 2026 venture funding into self-described world-model companies.2 The label covers video generators, robot simulators, forecasting engines, CRMs, and agent runtimes.
The idea underneath is one of the most important in AI, and the hype is making it harder to see. This post sets out what a world model is, why it matters, and where we are investing. A real world model can answer one question that the others cannot: if we take this action, what happens next?
What is a world model?
The idea is simple. Before you act, try the action on a model of the world and see what it predicts. The most capable robots, self-driving cars, and game-playing AI all work this way.
Penn's CIS 6280, a new graduate course taught by Jiatao Gu, defines world models as "learned representations and predictors of environment dynamics for perception, planning, and decision making."3 In plain terms, a world model captures the current state of something and predicts how that state will change, including what happens when you act on it.
That is two things joined. A representation of the state: what is true about the system right now, including what cannot be observed directly. And a predictor of dynamics: how that state changes over time and in response to actions.
Neither half is enough on its own. A representation without a predictor is a map, not a simulation: it shows where things stand but not where they are going. A predictor that cannot take an action is a forecaster: it can tell you what is likely to happen, but not what will happen if you intervene. The line that matters is the action. "Will churn rise next quarter?" is a forecast. "If we raise prices 5%, what happens to churn?" is a world model question.
A world model does not decide. It computes the consequences of candidate decisions, and something else chooses among them: a robot's controller, a software agent, or a CFO. The decider proposes several actions, the world model plays each one forward, and the decider picks the future it prefers.
Why does that matter? Because whoever is deciding can now plan before acting. Most AI in business today predicts a number, such as a churn score or a demand forecast, and hands it to a person, who then has to work out the consequences of each option in their head or in a spreadsheet. A world model does that work for the decider: it rolls forward fifty pricing paths, a dozen hiring plans, or several orders for a risky migration, and compares the results before anything real is touched.
A world model also compounds. Every action taken produces new data about what it actually did, so the model gets better at the exact decisions it is used for. That matters more as decisions are handed to agents. When an action cannot be undone, whether a robot arm is moving or a pricing agent changing a price, the agent is only as good as its model of what the action will do.
A four-part test
The clearest definition we have found comes from biology. A September 2026 paper in Cell, "World models for biomedicine," from Marinka Zitnik's lab at Harvard, sets out what a world model must do.4 We read it as a four-part test. A world model:
- represents the state of a system,
- accepts a specified action,
- predicts the resulting state, and
- continues simulating from its own prediction.
The fourth requirement is the one most products miss. A model that takes an action and returns one answer is a forecaster with an extra input. A world model can keep going: take its own prediction as the new state, try the next action, and roll the future forward.
Run the test on what is being called a world model today. A video generator that simply plays out a scene shows a plausible future, but there is no way to act inside it, so it fails the second requirement. A CRM with a language model reading over a company's records holds state but predicts nothing about actions. A tabular model that answers one prediction per question carries no state from one question to the next. None is a world model yet.
That research raises a warning that holds across all domains. Querying a simulator for consequences does not guarantee it reveals true cause and effect. Deep architectures learn from records, where historical correlation frequently masks actual causation. When an enterprise historically adjusted prices upward only during periods of surging demand, any architecture trained strictly on those archives would infer that higher rates stimulate revenue.
The algorithm captures the sequence, not the underlying driver. This makes the fundamental evaluation question straightforward: is the system learning from active interventions in which decisions were deliberately manipulated and outcomes measured, or merely from passive observations?
What is actually new?
The idea is older than deep learning. Penn's CIS 6280 opens by tracing its lineage. In 1943, the psychologist Kenneth Craik argued that the mind tests actions on a "small-scale model" of reality before acting.5 In 1960, Rudolf Kalman's filter, later used to guide Apollo, estimated a spacecraft's position from noisy measurements.6 AI researchers then taught agents to learn their own models: Richard Sutton in 1990, and David Ha and Jürgen Schmidhuber in 2018, whose agent learned to play a video game inside a model it built from the screen.7 Each step worked, but only where the model could be written by hand or the world was as small as a game.
Businesses have been modeled for decades, too: marketing-mix models, demand forecasts, every CFO's spreadsheet. But in all of these, an analyst chose the variables and wrote the relationships. What is new is that the representation can now be learned from messy, real-world data. Deep learning removed that step for pixels, and the same capability is now arriving for tables, event logs, human video, and software traces.
Hand-coded models miss the situations that matter. An analyst's demand model has a variable for price and one for season, but none for the competitor that launched last month. A learned model can pick up signals like that on its own, once it has enough data.
Data is also why the money has gone where it has. Some of the best-funded companies so far did not start as world models. They started with a narrower product, such as marketing attribution or forecasting, collected data through it, and only later became world models.
When you need one
A thermostat does not need a world model. It just reacts. When the task is fixed, and feedback is fast and cheap, reacting is enough; a world model is overhead.
A world model earns its keep when the goal is new, when the action cannot be undone, or when feedback arrives after the damage is done. A self-driving car cannot learn what happens when it runs a red light by running a thousand of them. Most consequential business and political decisions are similar: the construction of a new facility, a major acquisition, or an interest rate decision. You cannot run the experiment twice, and the result shows up quarters later.
Completeness is not the goal. A good model keeps only what changes the decision. Conant and Ashby made the point in 1970.8 A pricing model does not need the color of the office carpet. It does need to know which customers are on annual contracts. The best world models are defined by what they leave out.
One idea, three worlds
Most people hear "world model" and think of robots. But the same question comes up anywhere there is a history to learn from, an action to take, and a future that depends on it. We see three such worlds.
The physical world. Matter, motion, and contact. The state includes what a camera cannot see: mass, friction, and what is hidden behind another object. The data is video and simulation. The question: if the arm pushes here, does the stack fall?
The economic world. Companies and markets run on incentives and transactions, just as the physical world runs on physics. The state is hidden here, too. Transactions and dashboards are observations. Real demand, customer intent, and the effect of last quarter's price change have to be inferred. Enterprise world models learn one company's own dynamics. Market world models learn the environment in which it operates. The data consists of the company's own records, which no frontier model has seen, as well as market signals. The question is: if we raise the price by 5%, what happens to churn and margin?
The digital world. The software that agents act in: web pages, applications, code, APIs, and tools. State is what the software currently is, plus what is off-screen. Agents mostly get by on reacting today because software returns feedback in milliseconds, and most actions can be undone. That stops working when the action is a payment, a deploy, or a mission that runs for days. The data consists of agent traces generated by every deployed agent. The question: if the agent runs this migration, what breaks?
Where the attention is, and where we are looking next
Almost all of the capital and commentary so far is physical. This week, AMD agreed to acquire Fei-Fei Li's World Labs for $8.2B in stock, seven months after it raised $1B.9 AMI Labs raised a $1.03B seed round in March.1 We invest there too, most recently in a company in stealth building a world model for robot training.
The economic and digital worlds are earlier, with little capital and almost no one yet able to answer "what happens if" with interventional data. We have backed two companies in the economic world. Wood Wide AI, from CMU's Pradeep Ravikumar and Varsha Raj, is building world models of a company's own structured data, so a business can test a decision, such as a price change or a new hiring plan, before making it.¹⁰ Sooth Labs, founded by Yaser Sheikh, Russ Salakhutdinov and Chuck Hoover, is building a world model of the global environment that can simulate how events are likely to unfold under different actions, so organizations can see the consequences of a decision before they commit to it.11
Both worlds have the same structure as the physical one and a clear source of demand: agents that act on real businesses need to know what an action will do before they take it. That is why the rest of this series focuses on them. At pre-seed and seed, the companies built to the definition from the start are still few, and they are the ones we want to back.
What comes next?
This is the first of four posts. The next two go deep into economic and digital world models: what the world is, where the data comes from, and where early-stage companies can win. The last covers what world models do to applications, and why systems of record may be rebuilt as systems of consequence.
If you are building a model that can answer "what happens if" for a business, a market, or the software agents work in, we want to talk.
So the next time someone tells you they are building a world model, ask the question we started with: What happens if?
Notes
- TechCrunch, "Yann LeCun's AMI Labs raises $1.03B to build world models," March 9, 2026. techcrunch.com
- Dealroom, "World models funding hits a record $3.2B in 2026." dealroom.co
- Jiatao Gu, CIS 6280: World Models, University of Pennsylvania, Fall 2026. cis.upenn.edu/~cis6280
- Ayush Noori, Nic Fishman, Ada Fang, Lukas Fesser, and Marinka Zitnik, "World models for biomedicine," Cell, September 17, 2026. sciencedirect.com
- Kenneth Craik, The Nature of Explanation (Cambridge University Press, 1943).
- R. E. Kalman, "A New Approach to Linear Filtering and Prediction Problems," Journal of Basic Engineering 82, no. 1 (1960): 35–45. On its use in Apollo navigation: Leonard McGee and Stanley Schmidt, "Discovery of the Kalman Filter as a Practical Tool for Aerospace and Industry," NASA Technical Memorandum 86847, 1985. ntrs.nasa.gov
- Richard S. Sutton, "Integrated Architectures for Learning, Planning, and Reacting Based on Approximating Dynamic Programming," Proceedings of the Seventh International Conference on Machine Learning, 1990; David Ha and Jürgen Schmidhuber, "World Models," 2018. arxiv.org/abs/1803.10122
- Roger C. Conant and W. Ross Ashby, "Every Good Regulator of a System Must Be a Model of That System," International Journal of Systems Science 1, no. 2 (1970): 89–97.
- AMD, "AMD to Acquire World Labs to Advance the Future of AI Compute," September 28, 2026 (ir.amd.com); TechCrunch, "AMD will acquire Fei-Fei Li's World Labs for $8.2 billion," September 28, 2026 (techcrunch.com); on the $1B round: TechCrunch, February 18, 2026 (techcrunch.com).
- Wood Wide AI (woodwide.ai); Technical.ly, "Inside Wood Wide AI, the Pittsburgh startup taking on one of LLMs' biggest blind spots," June 15, 2026 (technical.ly).
- Sooth Labs, "World Model for Global Foresight" (soothlabs.com); Ardent Venture Partners, "Why We Invested: Sooth Labs" (ardent.vc); Felicis, "Our Seed Investment in Sooth: Foresight as Infrastructure" (felicis.com).