Learnability atlas · 23
How far is AI-optimal from human-learnable?
HumanSL predicts rank-conditioned human moves, but maximum probability, sampling, and target strength are not the same. Teaching should locate the next reachable judgment, not the person most similar to AI.
Put strength and human likeness on separate axes
Standard KataGo searches for strong play; HumanSL models human distributions by rank and era. Together they ask whether a good move is discoverable, but neither reads a student's mind.
One board enters an AI-value axis and a HumanSL-discoverability axis. Top-1 is not automatically most human-like
HumanSL research compares argmax, direct sampling, and softmax, finding an intermediate temperature better reaches intended strength. A teaching tool should not turn the largest probability into the only natural move.
Agent A takes top-1; agent B preserves a distribution, producing different local habits. Distance from the previous move is a human clue
The study uses move match, likelihood, and distance from the previous move, and improves human likeness with a distance correction, especially at weaker ranks. People often answer locally before jumping to the global optimum.
A global AI jump and a local human reply make the distance visible on one board. Rank is a probability band, not a gate
Multiple-skill policy research improves rank estimation to 80% with ten games and 92% with twenty in its experiment. That shows distributions carry information, not that one game certifies a rank.
5k, 1d, 5d, and 9d form a continuum rather than four absolute walls. A bridge move joins habit to target concept
If the AI move has negligible target-profile probability, displaying it may leave only a coordinate. A slightly lower-value move can expose the same direction, weakness, or sente concept at a reachable difficulty.
A is the hidden optimum; B is a more reachable bridge to the same concept. Learnability must survive delay and transfer
Understanding once is not learning. Days later, a structurally similar position from another game should test checking order, candidate set, and rationale, separating screenshot recall from judgment.
Training and transfer boards share structure but not memorized coordinates. Build a personal frontier, not a rank sticker
Retain loss, target-profile probability, concept bridge, search stability, and transfer result for every lesson. The frontier updates with player and model instead of declaring that every 5d must know one move.
The learning card keeps value, reachability, stability, and transfer side by side.
HumanSL and skill research
Before generating a learnable lesson
- Separate strength from human likeness
- Do not replace a distribution with top-1
- Keep the target profile
- Offer a bridge rather than only an answer
- Validate delayed transfer