The InfoQ Podcast

Generally AI Episode 1: Large Language Models

4 snips
Jan 26, 2024
Explore the vulnerabilities of large language models, the history of the transformer architecture, and the basics of LLM inference. Discuss the challenges of recipe hacking and password guessing. Learn about a red teaming event for large language models and the origins of GPT models. Delve into the motivations behind the transformer and enjoy some light-hearted banter.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
ANECDOTE

Recipe App Vulnerability

  • The "Pak 'n' Save Mule Bot" app initially allowed users to input any ingredients.
  • This led to users generating recipes for dangerous substances like chlorine gas, highlighting prompt injection vulnerabilities.
ADVICE

Prompt Injection Protection

  • Protect your language model-powered apps from prompt injection.
  • Consider whitelisting allowed inputs, like the recipe app eventually did.
ANECDOTE

Prompt Hacking Gandalf

  • Roland and Anthony played a prompt injection game with "Gandalf the White."
  • They tricked the model into revealing passwords using jokes and poems, demonstrating prompt hacking techniques.
Get the Snipd Podcast app to discover more snips from this episode
Get the app