Agentic Entity Simulation

The term “simulation” spans everything from digital twins to world models to AI-generated personalities designed to predict human behavior and decision-making. This week, we want to broadly categorize the types of simulations emerging in the market today, deep dive into the history of agent-based simulations, and attempt to extrapolate how this could evolve. If we can simulate people, what else can we simulate? The ability to create synthetic entities, whether that is a person, company, or regulator, has the potential to unlock a fundamentally new layer of experimentation, forecasting, and decision-making.

Types Of Simulation Technology

The term simulation is extremely broad and encompasses any activity in which synthetic data or scenarios are used to learn about a real-world phenomenon. This can very broadly be put into three categories:

• Digital Twin: This is a 3D representation of a physical object or place that exists inside of a digital setting. Use cases range from a static view of an apartment building to a live 3D replica of an airport, enabling real-time traffic management. Unreal Engine was famously used to replicate the full city of Shanghai in 2020. Battle Road (a Konvoy portfolio company) created a digital twin of the entire globe to help the DoW simulate battles across multiple nations and thousands of entities. While digital twins can be extremely valuable for decision-making, integration can be complex, and the initial upfront cost can be high.
• World Model Simulation: A world model predicts the next state of an environment, the way that a language model predicts the next word. This can be used to simulate future events, understand causality, and interact with physical laws. The goal of these models is to learn the dynamics of the physical world well enough to simulate it and reduce the cost of real-world trial and error. For example, large amounts of human demonstration video can teach a model how to do a task, which can then be mapped onto a robot.
• Agent-Based: This is one of the most recent approaches to simulation, popularized by a research paper from Stanford titled “Generative Agents: Interactive Simulacra of Human Behavior”. This approach creates a collection of individual AI agents with their own personalities, goals, and beliefs, allows them to interact, and then observes the emergent behavior.

Interactive Simulacra of Human Behavior

This paper introduced “generative agents”: AI-powered software agents that simulate believable human behavior by combining large language models with memory, reflection, and planning systems that allow agents to persistently act, form relationships, and make decisions over time. The team at Stanford tested these agents inside a sandbox world inspired by The Sims (a video game world where people simulate real lives in virtual environments), where 25 autonomous agents organically developed complex social behaviors.

The key finding was that if you use a new architecture focused on memory, reflection, and planning, complex social behavior emerges without being explicitly scripted. For example, when one agent was given the goal of hosting a Valentine’s Day party, the agents organically spread invitations, formed new relationships, coordinated plans, and arrived at the party together at the correct time.

The team took this technology and founded Simile.ai, which just raised a $100m Series A round led by Index Ventures. Today, they are focused on “How and why customers, employees, or populations respond to change”.

From People To Entities

The Stanford paper proved that AI personas can help forecast human behavior via persona simulation, but in doing so, it also suggests that this could help predict behavior for more complex systems: entities.

If a simulation can be used to create emergent behavior for individually simulated personas, why could it not also be used for individually simulated entities? Instead of 25 individuals with specific goals and incentives who emergently created a Valentine's Day party, why could it not be 25 local coffee shops or competitive parts manufacturers, simulating competitive positioning and strategy?

To understand if this type of entity-level simulation is possible, you have to look at the architecture that was used in the original paper. Emergent behavior was possible due to three main functions:

1. Memory: The agent’s ability to know what has happened in the past
2. Reflection: The ability to synthesize memories into higher-level reflections
3. Planning: The ability to set goals for the future

When translating this to an entity, memory would need to reflect all previous business decisions and the inputs to those decisions over time, both data-driven and otherwise. The reflection process would need to synthesize memory in a way that preserves key decision-making data. And the entity agent's planning would have to reflect each entity's real-world goals and objectives in the simulation and align with management's perspective over time.

This is a daunting task. The amount of data required to inform the memory stage could be enormous and is likely not all accessible in a clean format. Decisions that live in the minds of management teams would need to be captured and codified. Deep domain knowledge would need to be written down in markdown files that capture a range of edge cases. On top of that, it's important to factor in that organizations are collections of people that do not always act logically or in the best interest of the organization as a whole.

And even if this was all possible, in order to put any weight on the decision, robust backtesting would need to be in place to validate the outputs.

Lessons From The Evolution Of Digital Twins

While simulating the actions of an entire competitive landscape may still be years away, the technology is likely to emerge in a similar way to digital twins.

1. Early Stages: The concept of the digital twin can be traced back to NASA in the 1960s where they attempted to simulate spacecraft and troubleshoot issues in real time. This idea continued to grow and was most prevalent in aerospace and manufacturing through most of the early 2000’s.
2. Tech Breakthroughs: With the proliferation of the internet of things (IoT), cloud computing, and 5G, access to real-time data became more accessible and allowed for broader adoption.
3. Platform: NVIDIA Omniverse, one of the most high-profile digital twin platforms, was released in 2021. Per Jensen, Omniverse was, “Building on NVIDIA’s entire body of work, Omniverse lets us create and simulate shared virtual 3D worlds that obey the laws of physics. The immediate applications of Omniverse are incredible, from connecting design teams for remote collaboration to simulating digital twins of factories and robots.”
4. System Integrators (SIs): Platforms like Omniverse have simplified the process of creating digital twins and enabled system integrators such as Accenture and Ansys to provide custom white-glove services to all types of companies.
5. In-House: Today, depending on the level of complexity of the models, digital twins can now be brought in-house.

So the earliest stages of digital twins were not scalable, Omniverse (the platform created to make it scalable) was built by a large incumbent, and the teams implementing the platform at large corporations were from large IT Services firms. This process seemingly does not leave much room for venture-backed startups, but there are a few companies that have achieved venture-scale returns, and each can teach us a lesson about what can make a strong stand-alone company.

• Incumbent Formats / Data: Another problem facing startups is building on top of existing platforms. For example, Cesium integrates with other CAD platforms, whereas companies like Exodigo or Twin Health (Full-body digital twins) provide a suite of tools that capture first-party data. Twin Health just raised a $53m series E at a $950m valuation and Exodigo just raised $96M at a $700M valuation.
• Scalability: Companies like Mujin and Passive Logic avoided being acquired by system integrators by actively resisting the complex, services-oriented work that system integrators gravitate towards. By creating broadly accessible tools that do not require a team of PhD’s, they ensured that their business was both venture scalable and incompatible with the processes of system integrators. Passive Logic just raised a $74m Series C at an undisclosed valuation.

Neither of these takeaways can be viewed in isolation and each is a double-edged sword. Despite running the risk of a smaller outcome, utilizing incumbent formats reduces friction, and non-scalable work is how many startups are forced to begin.

Takeaway: Simulation is a broad category but it is developing rapidly, with everything from world models to agent-based simulations. We are already seeing proof points that agent-based simulations can predict individual behaviors, and there may be an opportunity to take this a step further by predicting the behaviors of companies, regulators, and competitors. While this is still a ways away due to data and backtesting requirements, adoption is likely to take a similar format to how digital twins emerged over time. Simulation companies that own their own data and can move beyond providing white-glove services are likely to be the next unicorns in the simulation space.