Every system has a set of states, and limits on how it moves between them. Those limits are rules, and the rules together form the transition function of the system's state machine. The transition function is, in effect, the system's decision-making.
Fixed and evolving
A transition function can itself be represented by a model, bounded by the states the system can reach. That gives a useful split.
- When the model is fixed, the system is a world. It behaves according to rules that do not change.
- When the model evolves, the system is a mind. Its rules change with what it learns.
And a mind, given enough data, can learn to generate the transition function of a world. That is a fair description of what we ask of intelligent software: watch a domain long enough to write its rules.
The model of understanding
An intelligent agent needs a decision-making system. It can be centralised or spread out, but every ability of the agent, sensing and acting alike, connects to it. That is where its intelligence lives, because that is where its knowledge is used.
For knowledge to be used, concepts have to be organised so they can be retrieved, applied and updated. That organisation is the agent's model of understanding.
Picture one article and two agents, each with its own model of understanding. Both read it. Each turns it into different information, and each updates its own model with what it took away. The information each one forms is the article's meaning relative to that agent.
An abstract model
Now the interesting question: is there an abstract model of understanding that both agents' versions are instances of? If there is, it opens a door. It means there could be a language for creating abstract concepts, the way alloys, binary or DNA encode structure. A language like that would make possible a framework, a domain-specific language, for creating agents. From a genetics point of view, it would be something like a genetic protocol for minds.
What it turned into
These notes were written long before there was code for them. They describe, in plain terms, the parts that now matter most when building agents: a world with declared rules, an agent whose model of that world changes with experience, memory that keeps what it learns, and a language in which both the world and the agent are written down.
Drafted in 2022
Updated for site in 2026