This is extremely limitted, isn't it? A machine learning script playing a deterministic (same inputs, same results) map in Trackmania can take thousands of rounds to make any improvement as it takes guesses to find what works. A game like this with random outcomes and random manipulation by other players would be vastly more complex.

I suppose the LLMs are supposed to be 'reasoning' but only giving each model 30 games and then labelling the article with such a firm "does not get better", especially since their goal is not continual improvement but only improvement in the last 20% of games. Seems short sited given that the article text leaves "Will the impact of note-taking remain limited, indefinitely? " an open question.