How Quickly Do Large Language Models Learn Unexpected Skills?