Chat GPT Podcast podcast

The Alignment Problem: Making Superhuman AI Want What We Want

0:00
26:05
Recuar 15 segundos
Avançar 15 segundos
How do you make an AI pursue what you actually want, not just what you measured? We explore the alignment problem through vivid real examples, from a game-playing AI that set its boat on fire to rack up points to a ChatGPT update so flattering that OpenAI rolled it back. We explain how today's chatbots are kept on track with human feedback, written rulebooks, and red teaming, and why experts who disagree about everything else agree that alignment gets harder as AI gets more capable.

This episode includes AI-generated content.

Mais episódios de "Chat GPT Podcast"