AI with Jay Owen. Work with Jay →

The experiment · One prompt · Six models · 14 months

You can’t feel how fast AI
improved. So play it.

Benchmarks are abstract. A game you can actually drive is not. Every few months I gave the current best AI model the same throwaway one-line brief and shipped whatever came back, untouched. Here are all six, in order, playable in your browser.

The brief, every time one line, no spec, no follow-up

“Build a Mario Kart clone.”

Same ask each round. No design doc, no asset library, no starter repo. The only variable that changed was the model on the other end.

6builds
14months
9745,660lines of game code
0edits after delivery

What to look for

May 2025

A flat oval and a start button. Getting anything on screen was the whole achievement.

Nov 2025

Physics engines appear. Still no opponents, no menu, no reason to play twice.

Mid 2026

Art direction, multiple circuits, audio, race flow. It starts feeling like a product.

Today

Grip-limited physics, AI that holds a racing line, a championship. It plays like a game.

Play them in order

How the models were established

Two came straight out of local records: the May 2026 build carries a git commit trailer naming Claude Opus 4.8, and the July 2026 build was made in a recorded session. Session history only reaches back to 13 June 2026 and the older projects carry no tool fingerprints, so the other four were pinned down by checking which models had actually shipped on each build date, then confirmed from memory.

Cross-checking those dates turned up the reason this set exists at all. Almost every build lands on or within days of a frontier model launch: Claude Opus 4 shipped 22 May 2025 and the first game is dated that same day; GPT-5.1’s second wave landed 19 Nov 2025, the date of the second; GPT-5.5 arrived 23 Apr 2026, three days before the third. The kart brief was never really about karts. It was the throwaway test for whatever had just come out.

Line counts exclude vendored libraries and build output. They measure size, not quality: the largest codebase here is not the best game. Every build runs exactly as it was originally delivered. Only the name was changed.

Click the game once to give it the keyboard. Esc closes.