Games
Games are worlds with state, actions and consequences. Most games encode those consequences by hand. A world model can learn how the world moves, then judge a move by what it expects to happen next.
It can ask:
- What changed because of the player's last move?
- What has become possible or impossible now?
- Which strategy fits the state of this particular world?
- What consequences should unfold next?
- Is the player actually changing the world, or only accumulating points?
Current work: Mandate of Heaven
Mandate of Heaven puts you in charge of a new AI lab in January 2020. You begin with a $12 million seed round and six years of real AI history about to unfold around you. Each quarter you choose a move. A trained world model reads what happened and judges what your lab becomes.
The model learnt from 10,066 events across the AI industry. It tracks capability, adoption, compute, talent and credibility while the authored game supplies the rules around money, costs and odds. That split matters: the game defines what you can do; the model decides how your strategy changes the world.
Where this goes
A world-model game does not need every consequence written in advance. The designer creates the world and its boundaries; the model learns its dynamics. That opens up strategy games whose systems keep surprising the player, historical games that respond to alternate choices, and simulations that are useful even when nobody calls them games.