They report a fairly large jump in accuracy across several datasets in their paper. Looks like a promising approach, basically leveraging the very large context window that some recent LLMs provide to give many examples in context before an actual task is attempted.