unit-of-trust.
An interactive essay about reviewing AI decisions, drawing on my recommender, agent-log viewer and small language model. Still a draft.
Heads up: this is still a draft. I'm actively iterating on the writing and the scroll interactions, so the live version will keep changing.
What it is
An interactive essay at unit-of-trust.vercel.app about what I need to see before trusting an AI system's decision. It brings together three projects where I built ways to inspect how a result was reached.
The three projects it pulls from:
- recommend-arena — a recommender where I wanted to understand why an item ranked where it did.
- observe-ui — groups an agent's event log into decisions I can inspect.
- wine-gpt — a small GPT-2 I built from scratch. It led me to experiment with keeping intermediate steps visible and delaying the final answer.
Why I made it
I wrote about these projects separately before noticing how often I was building tools to explain a system's behavior. This essay puts those ideas together. The scroll interactions let me show the examples as I discuss them.
Some of the sections still feel unfinished to me. I'd rather put it up and keep editing than sit on it.