Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm

[trigaten]

https://arxiv.org/pdf/2102.07350.pdf

[trigaten]

• Claim that few shot prompting isn't doing any learning, but rather retrieving the task from the model's memory
• Believe few shot can be outdone by 0 shot
  • Have example of this for translation
• cite a number of internet users
• "semantic contamination" - great term
• "programming in natural language"