Observe
Maintain a changing model of where things are and how they interact.
Ends Up Here Podcast · Working page
Idea / In development
When AI can
simulate the future.
What changes when an AI can model a situation, imagine possible futures and test an action before it happens?
Key propositionIntelligence becomes more than answering.
It becomes consequence modelling.
It becomes consequence modelling.
The difference
A chatbot predicts words. A world model predicts what happens next.
Language models learn statistical relationships between tokens. A world model attempts to represent situations, objects, people and dynamics so that it can reason about future states.
That lets a system ask not just “what is a plausible answer?” but “if this action is taken here, what is likely to happen, and what other action would lead somewhere better?”
Working proposition: planning begins when the system can search through consequences before acting in the world.
The loop
Simulate
Run possible futures, including second-order effects and uncertainty.
Choose
Select an action against goals, constraints, values and likely consequences.
The civilisational leap
From reacting to managing trajectories.
A world model could help coordinate infrastructure, water, transport, agriculture, health and energy. A continent-wide project could be simulated before it is built, with alternative routes, costs, climate effects and human consequences compared in advance.
The difference between a predictive sketch and a sufficiently detailed model of civilisation matters. Better prediction improves tools; a continuously updated model becomes a new layer of governance and physical coordination.
The hard problem
Simulation is not the same as wisdom.
Searching through possible futures can be computationally expensive, and a model can be wrong, incomplete or trained around the wrong objective. It may predict consequences without knowing which consequences people consider acceptable.
The episode should keep both sides visible: world models could make prevention and coordination possible, but they also make the values and authority behind optimisation more important.
Ends Up Here question: who chooses which future the system is allowed to optimise for?
Talking prompts
What is a world model?
How is simulation different from text prediction?
What does a digital twin of civilisation mean?
How expensive is searching futures?
Who supplies the objective?
Can prediction become governance?
Working source note
This page is based on a project discussion about world models and should be checked against current technical research before publication.