LANGUAGE MODELS ARE SENSITIVE TO PROMPTS

Summary from the TLDR AI newsletter

It turns out that prompting models for few shot benchmarks can give anywhere from 4-88% accuracy on the task, depending on the prompt. This research shows how to scientifically improve your prompts.