Neural Net Go Problems

Neural Net Go Problems is a site with Go problems maintained by lightvector, who also maintains KataGo, the AI engine for playing and reviewing.

The website's UI makes use of WGo.js.

The about pages says

How these Go problems were made:

This site arose out of experiments with neural nets (...) to see how well they could model what players of different ranks understand. The problems on this site were created using neural nets to automatically extract positions for each rank from high-level games where the neural net thought the next move would be instinctive for a pro but might be educational or non-obvious for players of that rank (...).

The problems have NOT been curated by any human and probably have more inaccuracies on average than published Go books or problem collections. The neural nets were also trained primarily on historical human pro games rather than trained through a self-play process. So they will not reflect the latest joseki innovations and may not always be as sharp or accurate as the neural nets from the most recent wave of strong Go programs.

However, these problem sets also contain far more open-space tactics than most other problem sets, a highly under-covered topic in the Go literature. On the whole, they should cover a wide sampling of the standard good shapes that form the building-blocks of strong play. (...)

Review

Dieter: as the creator says, the problem sets indeed cover a gap in traditional problem sets: open-space tactics, often middle game exercises. A sample of 1d exercises shows that I can solve about 80% and the other 20% are instructive (so is the reinforcement of the 80%). This site is an absolute gem!