I Was the Crash Test Dummy

September 30, 2026 · A conversation field note

The small games in this collection began with a question: what could the frontier model available at the time make from a simple prompt?

The code was entirely AI-generated. My part was to ask, give some direction, try the result, and tell the model what happened. I was the crash test dummy: pressing buttons, getting stuck, noticing something strange, and coming back with another question.

What I wanted to show was the model's ability to turn a little language into something you could play with. The games preserve a moment in that changing capability.

Observation

A thought could become an interface before I knew how I would implement it. That shortened the distance between wondering and trying. A game, a cube, or a little creature gave the conversation something concrete to push against.

The surprise was how quickly there could be something to explore, and how much exploration that made possible.

Question

What happens to curiosity when the cost of giving a thought behavior falls?

I can spend less of an evening wondering whether an idea is buildable and more of it discovering whether the idea is interesting.

Artifact

The collection itself is the record. Keep earlier versions beside later ones. Retain model credits when they are known. Let the awkward controls and unexpected behavior remain part of what the experiment teaches.

Some artifacts have since been revised through additional AI-guided work. Their current behavior reflects that iteration; an original model credit does not mean the current version came from a single untouched prompt. Where the original prompt or model version was not recorded, I should leave that uncertainty visible.

Field report

Entirely generated code still needed someone to experience it. Guidance, testing, and acceptance were my involvement. A runnable artifact could surprise me, disappoint me, or expose a question I had not yet learned to ask.

That small loop grew into a much more deliberate practice in the enterprise work: shared contracts, evidence, handoffs, and human review. The little browser experiments remain useful on their own terms. They show what was becoming possible, and what I noticed while trying it.

I supplied the curiosity. The models generated the code. Playing with the result gave us the next question.

Still observing

Which limitations disappear with the next model? Which survive? What becomes easier to generate while remaining just as hard to understand?

I want to preserve those questions alongside the artifacts.

Explore the collection · Read the September excavation · Return to the Journal Room