The problem space
The layers of an AI system against the failures they meet, one line per pair. A pair drawn in ultraviolet has a page; the rest are coming.
Model
the weights you run
Data
the examples it learns from
Harness
the software around the model
Context
what it reads before it answers
Devices
the machines it runs on or moves
Agents
models that act in steps
Evals
the tests that score it
Monitors
the checks that watch it work
Memory
what it keeps between sessions
Tools
the apps and add-ons it can use
Prompts
the instructions it reads before your message
Incentives
what the people around it are rewarded for
Objectives
what its training rewards
Internals
the activity inside the model
Updates
every change after launch