How does it work

Being able to reliably predict future states of complex systems is a bold claim. How do we actually do it?

Executable world models

At the core of Prorok lies the idea of defining laws of the universe as code.

This is a different approach to traditional AI world models, where future states are approximated by manipulating continuous mathematical vectors in a hidden latent space (e.g. Yann LeCun's JEPA).

We believe our approach works better when trying to look deeper into the future, as we don't suffer as much from exponentially accumulating random variance, effectively providing a less "blurry" view into future states.

Our world models generally allow for much more flexibility (e.g. self-modification at runtime). We're also able to construct them with much less data than it would take to train a traditional world model.

Neural frontend

Prorok world model creation can be aided by specialized AI agents. This is a crucial innovation over the computational models and expert systems of the past.

Using a combination of pre-made and ad-hoc models, we can augment AI agents' reasoning in real-time by letting them formulate and test hypotheses about complex environments they find themselves in. Based on the results, agents are able to understand their place in the world and make plans before taking action.

Simulation-based inference

Making a decision, diagnosing an anomaly, planning towards a complex goal

• all of these require a good understanding of the domain in question.

This is difficult since we're usually starting out with a non-perfect observation of the world. For example we might start off with a vague description of the problem domain and a few data documents.

We start our forward-search loop with an observation of the real world. This step is generally

Formulating the experiment hypothesis

Large-scale, dynamic simulation

Similar to how it turned out to be with LLMs, we believe the key to success is sufficient scale.

Historically, agent-based simulations were always incredibly narrowly defined systems, often custom-built for speed if targetting larger scale.

We're taking a stab at modernizing things, introducing multiple cutting-edge features simultanously, such as: distributed computation, dynamic rearranging of objects across cluster based on collected metrics, polyglot logic definition, nested simulation, runtime model mutation.