Developer Tests Decision AI Jev Solving Rubik's Cube Like a Human
On September 22, blogger gdechichi posted a series documenting his experiment having @typesafeai's decision-making AI "Jev" solve a Rubik's Cube on its own: not only did Jev succeed, it solved the cube layer by layer like a human, rather than using the graph-search approach typical of computers. The whole solution runs directly in the engine, with a WASM payload of only about 200KB. gdechichi, who was a deep-diving cuber in his childhood, was motivated precisely by the difference between how humans and computers solve the cube.
Confirmed
- Computers have long been able to solve Rubik's Cubes, but via graph search, entirely unlike humans, who solve layer by layer in roughly 4 to 20 moves per step.
- The author encoded the human layer-by-layer solution steps as instructions and tested whether the decision-making AI Jev could make the right decisions accordingly—Jev solved it.
- Early attempts fed the cube state directly to the model and had it output probabilities for every possible turn, failing badly; the key turned out to be data representation—describing the cube state clearly in natural language was what enabled the model to decide correctly.
- The entire cube solution is integrated into the engine, with a WASM payload of about 200KB.
Why it matters
This experiment demonstrates the potential of decision-making AI to mimic human strategies on structured reasoning tasks: the issue is not model capability but how state is represented. For developers, the hard-won lesson—"translate the state into natural language"—is broadly applicable.
2026-09-22 ~ 2026-09-22 · 5 related posts
Primary sources
- Decision-Making AI Jev Solves Rubik's Cube Layer-by-Layer Like a Human — gdechichi · 2026-09-22
- [source] AI agent Jev solves a Rubik's cube layer-by-layer like a human — gdechichi · 2026-09-22
- [source] Making an AI Solve a Rubik's Cube Came Down to Describing State in Natural Language — gdechichi · 2026-09-22
- The Rubik's-Solving Setup Ships Inside the Engine as a ~200KB WASM Payload — gdechichi · 2026-09-22
- [source] Encoding human layer-by-layer cube strategy for a decision-making AI — gdechichi · 2026-09-22