Research series

Organizing AI Capability

A current essay series on how long-lived AI systems represent, organize, validate, and revise effective capability across models, tools, skills, workflows, and changing environments.

Part 1 · Published

What Can This AI Actually Do Right Now?

Why long-lived AI systems may benefit from a persistent, derived representation of their current effective capability rather than reconstructing capability organization from scratch on every task.

Part 2 · Published

Possessing a Capability Is Not the Same as Being Able to Use It

A capable model can fail in deployment because someone—or something—must still recognize, invoke, and integrate its ability. The difference becomes visible when we examine what an evaluator has supplied and what the AI system has actually done.

Part 3 · Published

Memory Is Not State

An AI may retrieve a superseded rule and dozens of accurate past successes even while the current rule is in context. The unresolved question is what information still applies, and what may legitimately determine the next action.