Testing a GPT-5-Generated Rubik’s Cube Simulator

要約

30 秒でわかる

Matthew Berman shows a short test of a Rubik’s Cube simulator generated in response to a GPT-5 coding prompt. He demonstrates scrambling and solving on 3×3, 5×5, and 20×20 cubes, noting slower rendering on the largest. The visible results appear solved in the presented examples; the clip does not establish how the solver works or whether it handles arbitrary cube states.

1The coding promptBerman asks GPT-5 for a Three.js simulator supporting interactive cubes up to 20×20×20.
2Scramble and solveThe 3×3 simulation scrambles, then its visible faces appear solved after he activates the solve control.
3Larger cubesHe tries 5×5 and 20×20 versions; the larger simulation runs at a lower frame rate.

ポイント

The prompt requests an interactive HTML and JavaScript cube simulator using Three.js, with sizes up to 20×20×20.
The displayed 3×3 and 5×5 examples appear to return to a solved state after Berman uses the solve control.
The 20×20 demonstration shows a drop in frame rate and ends with the visible faces in a uniform state.
The brief demonstration does not independently verify every cube face, all possible inputs, or the solving algorithm.

トピックSoftware engineering 評価（Evals）Generative media OpenAI

全文

チャプターごとに整えた全文を読む

元の文字起こしと対照

文字起こしのみ

Prompt and 3×3 test

Berman asks GPT-5 for an interactive Three.js cube simulator. He scrambles the default cube and observes the solve animation.

スライド 1 / 4Three.js simulator promptThe prompt requests a dynamically sized interactive cube up to 20×20×20.

I have been trying GPT-5 and want to see what it does with a Rubik’s Cube challenge. I ask it to write a complete HTML and JavaScript program using Three.js for an interactive cube simulator, with sizes up to 20×20×20. The demonstration tests the generated program rather than a physical cube.

スライド 2 / 4The 3×3 scramble animationThe generated simulation shows cube layers in motion during a scramble.

The initial 3×3 cube looks good, though I have seen other models render a cube that falls apart during rotation. Here the scramble animation moves the layers, and when I press Solve, the colors begin to group together. The visible faces return to uniform colors in the shown example; the short clip does not inspect the algorithm or every face.

Testing larger cubes

He tries the same controls on a 5×5 cube and then a 20×20 cube. The largest cube slows down, but its visible faces return to uniform colors.

スライド 3 / 4The 5×5 solve animationThe larger cube is shown with layers rotating as the solve action runs.

I switch to a 5×5 cube and run the solve action. Its layers move and the colors come together, and I describe the result as solved. This shows the behavior of the presented example, but it does not establish that every possible cube state can be solved.

スライド 4 / 4The 20×20 cube during solvingThe visible colors become grouped while the larger simulation runs more slowly.

Finally, I build and scramble a 20×20 cube. The frame rate visibly drops, but the solve animation continues and the visible faces end up in uniform colors. It is an impressive result to watch, though the video does not show how the solver derives its moves or independently verify all sides.