Sixteen months ago my CNN beat HastyBot 51.7% of the time. The new transformer value net, macondo-tf, wins 57.65% over 100,000 game pairs, level with a 2-ply Monte Carlo simulation while thinking more than a thousand times faster. Here’s everything we tried to get there, including the long list of things that didn’t work.

I’ve trained a convolutional neural network that beats our long-standing equity engine (HastyBot’s static evaluator) in direct play - 51.7 % win rate over 250K self-play games - making it, as far as we know, the strongest pure static evaluator in Scrabble today. In this post I walk through the feature set (board image, rack, bag, tempo, power-tile counts, more), the training pipeline, results, and why a value-only CNN is already changing how we build BestBot’s Monte Carlo search. Code, weights, and data scripts are all open-sourced for anyone who wants to tinker.