As a teenager, I had this BASIC Computer Games book with a game called HEXAPAWN. Lines 900-970 were just cryptic numbers that made no sense to me. Finally figured it out decades later.
Turns out it's machine learning from the 1970s! The AI learns by literally deleting bad moves from an array. After losing ~10 games, it becomes unbeatable. Just 19 board states, no neural networks, no fancy algorithms.
Martin Gardner (who wrote about it) also mentioned MENACE - a tic-tac-toe learning machine made with matchboxes and beads. Same principle, physical implementation.
Made a JavaScript version if anyone wants to try. The AI really does get better.