In 30 seconds
Fan-Yun Sun presents a framework for world models and describes Moonlake’s approach to generating controllable simulations. He discusses explicit state, learned rendering, agent training and the cost of making simulations useful.
- 1State and actionDefine what a simulation represents and what actions it permits.
- 2Clinical limitsPatient simulations need validation and cannot replace clinical evidence.
- 3ControllabilityA generated world should be editable for the intended task.
- 4Error accumulationFrame-by-frame errors can accumulate over time.
- 5Explicit stateA simulation scaffold and visual renderer have different roles.
- 6Generated curriculaTraining in generated environments is still a transfer research problem.
- 7EfficiencyCompute cost constrains practical deployment.
Takeaways
- Define what a simulation represents and what actions it permits.
- Patient simulations need validation and cannot replace clinical evidence.
- A generated world should be editable for the intended task.
- Frame-by-frame errors can accumulate over time.
- A simulation scaffold and visual renderer have different roles.
- Training in generated environments is still a transfer research problem.
- Compute cost constrains practical deployment.
Read the edited text by chapter
What is a world model?
States, actions and the history of simulation.
Stanford AI Club introduces Fan-Yun Sun, co-founder and CEO of Moonlake AI, to discuss world models and generated simulations. The talk combines a technical framework, product demonstrations and the founder’s vision for future applications.
Sun describes research on generating environments and training embodied agents within them. The intended goal is to build policies that can eventually be useful outside the simulated setting.
He uses a racing car to explain multimodal reasoning: observations can include visual motion, sound and expected consequences. A useful simulation must represent the aspects relevant to the decision being studied.
Sun uses “world model” broadly for a data-driven simulation that helps predict or decide. He notes that modeling and simulation have a long history, even though today’s terminology is newer.
Games illustrate how interactive environments can be constructed from objects, rules and behaviors. His ambition is to use modern models to make more kinds of such environments easier to create.
Sun discusses Will Wright’s simulation games, where designers create rules and players produce varied outcomes. The example separates the rules of a world from the particular stories that emerge within it.
Traditional game creation involves specifying geometry, physics, interactions and scripted behaviors. Each choice determines what the environment can represent and what users can do in it.
He compares this with the Wright brothers’ wind tunnel: a small physical model isolated selected aerodynamic properties so they could test designs. A model is useful because it captures relevant features, not because it reproduces every aspect of reality.
Sun proposes a framework for evaluating simulations: define the states, define available actions and specify how the system transitions between states. The application determines which details must be represented.
He uses [Genie 3](https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/) as an example of a model that generates interactive visual environments. DeepMind’s own description notes limitations in action space, duration and accurate real-world representation.
A drug-response example illustrates a different choice of state and action: a model could represent patient conditions and treatment changes. Such simulations require rigorous validation and are not a substitute for clinical evidence.
A generative engine
Controllable scenes and rendering.
Sun says Moonlake is interested in scenarios that are expensive or dangerous to stage, such as rare driving events. Simulations can help explore them, but their value depends on how well they represent the real situation.
The company’s proposed generative simulation engine emphasizes controllability: users should be able to alter a generated environment’s layout and properties rather than accepting a fixed video.
Sun describes a first layer that uses models to construct a scene and specify geometry, objects and physics in a simulation engine. He likens the process to a developer choosing components and assigning behavior.
A second learned rendering layer converts an engine’s output toward a more realistic visual appearance. The explicit simulation state and the generated visual layer serve different purposes in this design.
The demonstration includes generated games and interactive scenes. These examples illustrate the system the company is building; they do not independently establish physical accuracy across arbitrary tasks.
Sun contrasts an explicit scene scaffold with repeatedly predicting the next video frame. In an autoregressive video model, errors can feed into subsequent predictions and accumulate over a long interaction.
In Moonlake’s approach, code and the simulation engine maintain aspects such as geometry and state, while a learned renderer supplies appearance. This can reduce some forms of drift, but its accuracy remains dependent on the engine and generated setup.
Training and validation
Synthetic labels and agent curricula.
Explicit state can also provide synthetic labels for training or evaluation. Labels produced by a simulator describe that simulator; they must be checked before assuming they transfer faithfully to the physical world.
Sun imagines an iterative loop in which agents train on generated tasks, receive feedback and encounter harder tasks. Whether such a curriculum produces robust real-world skills is a research question, not a result proved by the talk.
In questions, he says the demonstrated system uses a modified game engine with additional tooling. He notes that the underlying physics implementation can be changed for different domains, subject to the engine’s limitations.
Future use and cost
Applications, limits and efficiency.
His longer-term vision is on-demand simulations for decisions, education, autonomy and other applications. These are proposed future uses, and each would need evidence that the simulated behavior is appropriate for its domain.
For social or educational scenarios in particular, plausible simulated people are not reliable evidence of how real individuals will behave. Such outputs would require careful validation before informing consequential choices.
Sun identifies efficiency as a major technical challenge. A simulation’s cost has to be weighed against the value and frequency of the decision it supports, especially when long runs require substantial computation.
Asked whether the explicit engine and learned renderer will remain separate, he says the architecture may evolve. His current argument is that structured state offers a practical route to controllable, efficient simulations.
