@teknium corrected this: "This just trains the model on the specific samples that most closely relate to the problem right before doing inference though – it is not generalizable at all, and does test-time specialization of the model"

→ View original post on X — @kimmonismus