"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis

Ignore Previous Instructions and Listen To This Interview with Sander Schulhoff, CEO of Learnprompting.org

32 snips
Jan 30, 2024
Sander Schulhoff, CEO of Learnprompting.org and a researcher in prompt engineering, shares insights on navigating the complexities of AI interaction. He emphasizes essential techniques for effective language modeling, revealing the power of few-shot prompting. Schulhoff highlights challenges in evaluating AI models and discusses vulnerabilities, including user manipulation and creative input modifications. The conversation also touches on the role of self-generated examples in problem-solving and the future of AI skills in professional environments.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
ADVICE

Basic Prompting Advice

  • Include proper context, like pasting relevant emails or defining rare terms.
  • Use few-shot prompting with examples and thought generation techniques like chain of thought.
INSIGHT

Role Prompting Effectiveness

  • Role prompting is less effective for accuracy-based problems than previously thought.
  • It can even hurt accuracy, as shown by a "dumb person" prompt outperforming a "brilliant professor" prompt.
ANECDOTE

Model Behavior Changes

  • Sander observed that model behavior changes significantly with updates.
  • Switching from GPT-4 preview to GPT-4 32K immediately fixed a prompting issue.
Get the Snipd Podcast app to discover more snips from this episode
Get the app