---
title: "The gap is trust"
date: "2024-12-19"
summary: "Two years in, the models are brilliant and the agents are real — what's still missing isn't intelligence, it's trust: grounding, provenance, memory"
tags: ["ai","trust","building","agents"]
---

Two years. That's how long it's been since the first chat window, and tonight I'm sitting with what's left unsolved.

The intelligence question, for me, is closed. This year I watched models stop blurting and start planning — slowing down, holding a long context, reasoning toward an answer instead of reaching for the nearest one. Whatever I doubted in 2022, I don't doubt now.

So here's what nags at me as the year ends.

The model in my editor is genuinely brilliant and completely indifferent to me. It doesn't know my studio's conventions, the decision we settled in a Slack thread last March, or which of my own libraries it should reach for instead of inventing a worse one on the spot. Every session, it meets me as a stranger. Every session, I re-explain.

And when it does hand me something, I can't trace it. I don't know what it drew on, whether the answer stands on my actual documents or was improvised with confidence. The output is fluent. Its provenance is fog.

So the missing piece isn't more IQ. The labs keep delivering on that, faster than I can absorb. The missing piece is trust. Grounding, so the thing stands on real ground. Provenance, so I can see where an answer came from. Memory, so it carries me forward instead of forgetting me between Tuesdays.

I keep sketching these on the same napkin. They're not three problems. They're one shape, and I don't have a name for it yet.

Twenty-five years of building taught me one thing: the interesting work is rarely the flash. It's the part that holds weight while everyone photographs the view.

I can feel the thing I want to build. Next year I'd like to know what to call it.