Business schools are built around a simple question: how do people make better decisions? Over three days in late summer, the Center for Applied AI and Chicago Booth hosted WM@Booth 2026, a workshop asking a close parallel to that question. How can machines build an understanding of the world good enough to predict what happens next, and choose well because of it?

World models are one of the fastest-growing areas in AI, and this was the third workshop in the series, following earlier gatherings in New York and Montreal. Booth was a fitting host. Its applied AI faculty group is the first new faculty area the school has added in decades, and the setting pushed the conversation toward messy, changing systems like markets, consumers, and the economy, where the world pushes back on the model. The program brought computer scientists, roboticists, economists, and finance practitioners into one room, and the field's biggest disagreements were aired openly.

Day 1 // What Does it Mean for an AI to Understand the World?

The first day opened with a question that ran through every talk: is predicting well the same as understanding? David Donoho, Anne T. and Robert M. Bass Professor of Humanities and Sciences at Stanford University, set the tone by tying a 1956 theorem from the mathematician David Blackwell to what today's transformers seem to learn internally. Aditya Grover, Assistant Professor of Computer Science at UCLA, made the case for diffusion language models that draft an entire answer in parallel rather than one word at a time.

Pivoting to the field of robotics, the subsequent talks broughtt discussion down to earth and to tangible hardware. Leading with Stanford's Jeannette Bohg, Associate Professor of Computer Science, wondered aloud how test-time compute could ever work when a robot hand needs a command every millisecond. Yilun Du, Assistant Professor at Harvard's Kempner Institute and Department of Computer Science, showed video models imagining futures to plan robot actions, and research scientist Danijar Hafner argued that scale, not clever objectives, is what's missing from physical understanding.

Long called the 'godfather of AI' for his foundational contributions to machine learning, computer scientist Yann LeCun, Jacob T. Schwartz Professor of Computer Science at the Courant Institute of Mathematical Sciences at New York University, closed the keynotes with a provocation: models shouldn't predict pixels at all, but should predict in an abstract space that ignores what can't be predicted. Hafner wasn't buying it, and the closing panel got lively. Underneath the disagreement was a shared sense that AI still struggles with the physical world, and that fast reflexes and slow planning both matter.

Day 2 // When the World Is People and Money

Having established a common definition of world models during the first day, the second asked what happens when the world in question is not so much a self-contained space but more abstract. What if the 'world' centers around human behavior—people, money, decisions. Each talk circled the same uncomfortable, and model-challenging fact: people react to being modeled. How, then, is it possible to work with 'clean' data that is free from influence.

Diyi Yang, Assistant Professor of Computer Science at Stanford, argued that a user model is a world model of a person, and showed a laptop assistant that watches your screen and anticipates what you'll do next. Booth's Ralph Koijen, AQR Capital Management Distinguished Service Professor of Finance and Applied AI and Fama Faculty Fellow,  showed that investment portfolios can teach a transformer which stocks and investors resemble one another, borrowing tricks from language models. Director of Research at Google DeepMind, Aleksandra Faust made the case for cheap, noisy, synthetic training worlds, with a memorable instruction to embrace the noise.

The afternoon panel on finance and markets was among the most candid sessions of the workshop. Panelists Edo Airoldi of Temple University, and Bayan Bruss, VP of AI Research at CapitalOne, described a domain with no reliable simulator, shifting rules, and adversaries who adapt. Other attendees argued that prediction markets already work as a calibrated world model, with prices that behave like probabilities. Along the way, lightning talks covered memory for video models, adapting at test time, and tactile sensing for robots. The takeaway: the hardest world to model is one that changes when individual users do.

View Day 2 Recording

Day 3 // Hands on the Keyboard

Day 3 swapped talks for tinkering. Organizer Randall Balestriero, Assistant Professor of Computer Science at Brown University, opened with a tutorial on training JEPA world models without the usual headache, starting from a surprising result: two autoencoders can reconstruct images equally well and still differ by 20 points on downstream tasks. His fix, regularizing embeddings toward a Gaussian so they can't collapse, comes with a small codebase that trains in minutes on a single GPU.

Then came finance. Steven Bravo showed how to pull market data through an API or straight into Claude, while Booth's Bradford Levy, Assistant Professor of Accounting and Applied AI and Asness Junior Faculty Fellow, explained why financial time series are so stubborn: prices absorb information as soon as traders act on it, so the past predicts the future less than it does for something like electricity demand. Adding a dose of realism about cost concerns, Amir Zadeh, Professor of Management Information Systems at Wright State University, noted that simulating a single humanoid robot in a forest can run into the millions.

After lunch, attendees had two hours to roll up their sleeves and build. Projects included a Minecraft world model, a Breakout world model that could not quite find the ball, and one very hopeful trading experiment. Kawin Ethayarajh, Assistant Professor of Applied AI and Kathryn and Grant Swick Faculty Scholar, closed with a tour of what "world model" has meant since the Stoics, and left three open questions. Are explicit models necessary? Should they live in latent space? How do they keep up as the world changes?

View Day 3 Recording

Looking Ahead

Three days of talks, panels, and late-afternoon tinkering left more open questions than settled ones, and the organizers seemed happy about that. Researchers still disagree on whether world models should live in raw data or abstract spaces, whether they need to be explicit, and how to keep them useful when the world changes in response to them. But the room agreed on the hard part: the most interesting worlds to model are the ones full of people, prices, and consequences. 

WM@Booth Photo Recap