An article was posted a few months back about a competition for the best rock-paper-scissor program, where submitted programs played each other. Point was to build the program that would be able to predict best the opponent program's moves (which itself is trying to predict you program's moves).

It was a very interesting read, and it explained a really "simple" strategy which I found very memorable: build 3 strings representing "all previous moves + one of the 3 possible next moves", run a text compression on them, pick the move that yielded the smallest compressed result. That's the one where the compression algorithm identified the strongest pattern.

I found it really smart, instead of modelling the game it went one step higher in abstraction and works directly at the level of pure pattern recognition.

That keeps coming back to mind when thinking about things like randomness, the "bitter lesson" and LLMs.