gwern comments on Memorizing weak examples can elicit strong behavior out of password-locked models