Long before every software company discovered the phrase “artificial intelligence,” Jay Scott was maintaining a website about something much narrower and more useful: how computers learn to play games better.
Visit Machine Learning in Games
The site is aimed at AI researchers and game programmers. It organizes material around heuristic search, neural networks, genetic algorithms, temporal-difference methods and other techniques, then connects those methods to actual game-playing systems.
The subject list is a time capsule by itself. There are neural-network backgammon programs, strong learning Othello programs, pursuit-evasion research, robotic soccer, NeuroChess, Go networks, Metagame and general game playing. Later additions include StarCraft competitions and the BotPrize Turing-test experiments in Unreal Tournament.
More important than the individual examples is the way the site worked as research infrastructure. Scott maintained indexes of games and people, tutorials, links to AI resources, lists of online papers, software, workshops, competitions and a bibliography. Other academic pages from the period cite it as a place to find researchers and work in strategic game learning.
Scott’s current homepage now describes Machine Learning in Games as long out of date but still of historical interest. That is a wonderfully accurate description of why it matters in 2026. The machine-learning field moved on at terrifying speed, but an old hand-curated map of what researchers thought was important at the time tells you something a freshly generated search result cannot.
The page also captures a different AI culture. The problems are concrete. Can the program learn Othello? Can it improve at backgammon? Can evolutionary methods produce better strategies? There is little mystical language and not much pretending the machine has become a coworker with feelings. Mostly there are algorithms, papers and games.
For researchers, that means the dead or aging links are part of the evidence. For webmasters, it is a reminder of what a genuinely useful link directory looked like: a knowledgeable person imposing structure on a technical field because they had actually read the material.
CacheRat’s 1,967 Ancient Web Domains research list contains many of these specialist indexes. They are valuable not because every link still works, but because they preserve the shape of a field at a particular moment.
Machine Learning in Games is still there, cheerfully obsolete and therefore increasingly useful as history.