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.

How we implemented mathematically rigorous Bayesian inference in Macondo to let BestBot reason about its opponent’s rack, and why it only barely edges out going first.

A tournament report for the most prestigious NWL tournament of the year - the SPC 2025. Lessons learned and so forth, as well as a shameful (or shameless?) request for help.

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.

I just released a new web app for learning words called WordVault. Read more about it here.

Scrabble endgames are computationally difficult to solve exhaustively. I used some chess programming techniques to greatly speed them up.

I have found that in the literature about games and AI (artificial intelligence) that Scrabble is invariably referred to as something of a “solved” game, in the same way that Chess, Go, etc are “solved” by AI —in these latter games, there exist AIs that have superhuman performance and can basically beat any person.

I’m going to log more of my progress on the apps that I wrote about in an earlier post, in an attempt to:

1. make myself more likely to work on these apps
2. write a log for me and others and drum up some excitement! ;)

Monte Carlo simulation is basically working on Macondo. I expect that since I just got it working, that I’ll discover some bugs and special cases, and there’s so much more I want to do with it, but for now I’m excited that I got it working. As a nod to Scrabble expert Ron Tiekert, who might have been the first person to invent Scrabble simulations back in the 1980s, I plugged in the rack that he used to explain this concept, AAADERW.

The article here sums up pretty well the debacle that North American tournament Scrabble players have been faced with since the introduction of the new Tournament Word List in April of 2015:

Major Scrabble Brouhaha: Can You Copyright a List of Words? _In the 1980s, when Brian Sheppard created a computer program that played Scrabble, he typed in a lot of words-more than…_www.slate.com

Basically, Hasbro or Merriam-Webster or both decided to take the volunteer work various Scrabble players did in putting together a new word list, slapped a copyright on it, and made it impossible for players to obtain a digital version of the word list. This rendered various study tools such as Quackle, the best Scrabble AI, close to useless. I’ve heard a few players argue: who cares if Quackle doesn’t find every last word? It will still give you pretty good plays. I’m sorry, but playing Scrabble without all the acceptable words is like playing chess without knowing how the rook moves. Not quite, but it almost feels like that. Other players argued that we kids had it easy and that back in the day people studied by poring through the physical copy of the book. This is a ridiculous argument; computer study methods such as Leitner spaced repetition or even my own site Aerolith make it much easier to search for and remember words than looking through a giant book.