The 4 properties of AI
A short reference for the things that make models capable in some situations and limited in others.
Four properties explain most of the magic and most of the disappointment. Keep them in your pocket.
1. Generative
Models produce. They do not retrieve a single canonical answer unless you force a tool to do so. Production is why they can draft; it is also why they can fabricate.
2. Stochastic
The same prompt is not guaranteed to yield the same tokens. Lower temperature for extraction and JSON. Allow more variation for brainstorming. Never treat one sample as the only possible output.
3. Steerable
Instructions, examples, and tool results change behavior. This is the whole job of prompting. It is also why a polluted thread is so hard to recover — you have been steering it into a ditch for twenty turns.
4. Bounded
Context is finite. Training is dated. Tools only see what you expose. The model is not looking at your laptop. When something is missing, it will often pretend it is not.