It's downright absurd.
We're building AI agents that form explicit models of how the world works. They simulate possibilities, test predictions, and revise those models through interaction.
Explore the research ↗Today's AI systems can write, code, perceive, retrieve, and act. But in complex environments, they often fail to build persistent models of cause and effect.
They can describe relationships without reliably representing them. They can use context without accumulating understanding. They can act without knowing how their theory of the world should change when predictions fail.
We believe the next generation of AI will need more than larger models and longer context windows. It will need systems that construct, test, refine, and use explicit models of the world.
Dwnrt Absrd is a neurosymbolic AI research lab.
Foundation models provide perception, language, intuition, and flexible generalization. Symbolic world models provide structure, memory, simulation, interpretability, and explicit reasoning.
We're exploring how they can work together to form systems that learn continuously from interaction and evidence.
Our long-term research goal is the development of causal agents: AI agents that learn, reason, and remember through evolving models of the world.
A causal agent observes an environment, forms hypotheses, predicts outcomes, notices failed predictions, repairs its theory, and uses that evolving understanding to plan better.
Our research asks how AI agents can form explicit hypotheses about a system, test them against evidence, and revise them over time.
Symbolic World Models as a Learning Primitive for AI Agents
Memory retains experience. Skills retain ways of acting. A symbolic world model holds an explicit, revisable hypothesis of how a system works. This paper explores how agents can form, test, and improve those models.
Our research was selected for presentation at the ARC Prize Research Summit 2026, hosted in collaboration with MIT.
We're building.
Some of what we build will be useful. Some of it may just be for fun. The point is to keep learning from contact with reality.
Apps, prototypes, simulations, and playful experiments give us something real to observe: what people try, where systems fail, what surprises us, and what a model needs to learn next.
We're working toward a learning loop that connects useful applications back to better models.
SupaSim is where this research meets real operating decisions.
SupaSim is an AI business analyst that helps teams stress-test important business plans with executable models.
The vision is for SupaSim to serve as a company brain by creating an evolving executable mental model of the organization.
Today, SupaSim helps teams model decisions and test scenarios. Over time, validated models could support a deeper, evolving understanding of how an organization works.
If you are interested in neurosymbolic AI, causal agents, simulation, systems thinking, or AI that learns how the world works, we'd love to hear from you.
