Copying could underpin both evolution and intelligence, one user suggests

One post describes an induction head as an AI attention mechanism that encounters a token, finds an earlier occurrence and copies the token that followed it. The author says these heads can emerge in transformers with two or more layers, but not in single-layer models, and credits Anthropic’s interpretability team with linking their emergence to a bump in Kaplan et al.’s scaling-law plot. In a reply, the same author speculates that copying, stable states such as memory, and randomness might be the basic ingredients needed to kick-start both evolution and intelligence—while acknowledging they haven’t fully thought the idea through.