Beat Saber automapper fine-tune appears to reuse language-model patterns

A user says they used a logit lens to examine a friend's Beat Saber automapper fine-tune, which had a new token head. In the original vocabulary, arcs came out as (, while walls appeared to reuse “people as subjects” machinery. The post quoted another user's claim that there are many obvious ways to get more training tokens, challenging the idea that the data wall is a problem.