Grokking (en.wikipedia.org)

Feels like a rebranding of "eventually the inductive bias wins": train long enough and you can jump from memorization to a simpler rule that generalizes. Interesting, but I’m not sure how often this shows up outside toy algorithmic tasks.