Kimi K3.1 and a Rubik’s Cube Solver Demonstration

要約

30 秒でわかる

Matthew Berman demonstrates a virtual Rubik’s Cube solver associated with Kimi K3.1. He says it uses a solving algorithm rather than reversing the scramble, and judges the displayed result a pass; the short clip does not inspect the implementation or show repeated trials.

1A test still worth tryingBerman says he expected models to pass easily, but considers open-weight models to have struggled.
2Scramble and solveThe virtual cube is scrambled and Berman says a solver uses an algorithm rather than reversing the moves.
3An animated resultThe interface plays the calculated turns and displays a solved cube; Berman calls it a strong pass.

ポイント

Berman says some open-weight models still struggle with his Rubik’s Cube test.
The displayed virtual cube starts scrambled and plays a sequence of turns to a solved state.
Berman says the program calculates a solution using an algorithm rather than simply reversing the scramble.
A single demonstration does not establish the solver’s reliability across all possible cube states.

トピック評価（Evals）コーディングエージェント Open-weight models

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

The Rubik’s Cube test

Berman revisits a test he expected models to pass easily. He says open-weight models have still had difficulty.

I thought the Rubik’s Cube test had become too easy for AI models, but in my experience open-weight models still do not do well on it. This short demonstration is associated with Kimi K3.1; the clip does not provide a systematic comparison across models or trials.

Solving the scrambled cube

He starts the solver and describes its approach. The virtual cube is shown turning through a solution.

スライド 1 / 2The scrambled virtual cubeThe demonstration begins with a scrambled three-dimensional cube.

The virtual cube is fully scrambled. I say the solver uses an algorithm to find a solution rather than simply reversing the scrambling moves. When I start it, the interface plays the turns one by one in three dimensions; I describe the sequence as calculated in advance. The video does not inspect the underlying code to verify the method.

Result and assessment

The displayed cube reaches a solved state and a celebration appears. Berman considers it one of the best passes.

スライド 2 / 2The displayed solved stateThe interface reports a solved cube at the end of the demonstration.

The displayed cube reaches a solved state and a small confetti animation appears. I call it an excellent pass, one of the best I have seen. This demonstrates the result for the shown scramble, without establishing performance on every possible cube state.