Simulation: The Next Frontier for AI

Research bridges the possible and the impossible. In the age of prediction, we made it possible to train general-purpose models to classify text and images with high accuracy. In the age of reasoning, we are making it possible for models to solve challenging math, coding, and other complex, multi-step problems.

But what about messy, real-world questions whose answers depend on the emergent result of many humans interacting over time? For example:

• How would productivity and culture change for our organization if we allow remote work?
• How should we redesign the third-grade math curriculum for millions of students?
• How would clinical culture shift if physicians were evaluated on team outcomes?

These questions require something more: they require a fine-grained understanding of what will happen in any given situation. In other words, they require us to simulate.

At Simile, we believe that we are about to enter the age of simulation, the next frontier of AI. Simulation is about understanding ourselves and our environments so well that one can play forward any imaginable “what if?” scenario.

Research challenges

The Simile team has pioneered the field of AI simulation. We have developed multi-agent simulations with emergent behavior, simulated online social interactions, and created agents that accurately represent real individuals. To unlock the full potential of simulation, we must tackle the following challenges:

1. We must develop high-fidelity models of people and their environments. Today’s language models do not capture the nuances of human behavior. We need novel data collection strategies that capture this latent knowledge and to train foundation models that extrapolate to novel situations.
2. We must perform large-scale simulation efficiently. How do we simulate eight billion people over a year? We must develop multi-scale models that allow us to simulate both the macro- and micro-level dynamics of entire populations over time.
3. Our simulations must instill trust. Our models must produce calibrated probability estimates over the distribution of possible outcomes. Here, the simulation itself serves as an interpretable artifact tethered to a concrete reality.

What simulation unlocks

To start, a simulator allows us to predict the future: given the current state, simulate what will happen next. But a simulator can do much more than this. Simulators are causal models of the world. In terms of Pearl’s causal hierarchy, we can evaluate interventions: if we were to make a certain decision, what would happen? Or more ambitiously, we can answer counterfactual questions: if we had implemented a certain decision, what would have happened? The ability to answer such questions enables not just better decision-making, but also a deeper understanding of ourselves and the world.

The age of simulation is beginning. A predictive model can generate the optimal action, but it cannot explain why. Reasoning models can tell stories, but these are not necessarily grounded in reality. Simulation provides a complete auditable trace for the world’s most complex questions. Simulating what will happen is harder than predicting what to do. But we believe it is the true path to robust superintelligence. Simulation sits on the boundary between the possible and the impossible. At Simile, we are working to answer the exciting research questions required to cross that boundary. Come join us on this journey.