Episode

Localizing and Editing Knowledge in LLMs with Peter Hase - #679

Podcast
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Published
Apr 8, 2024
Duration seconds
2986
Processing state
failed
Canonical source
https://twimlai.com/podcast/twimlai/localizing-and-editing-knowledge-in-llms/
Audio
https://pscrb.fm/rss/p/traffic.megaphone.fm/MLN6128348451.mp3?updated=1712607942
JSON
/v1/public/podcasts/twiml-ai-podcast/episodes/localizing-and-editing-knowledge-in-llms-with-peter-hase-679
Markdown
/podcast/twiml-ai-podcast/localizing-and-editing-knowledge-in-llms-with-peter-hase-679.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/twiml-ai-podcast/episodes/localizing-and-editing-knowledge-in-llms-with-peter-hase-679/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/twiml-ai-podcast/localizing-and-editing-knowledge-in-llms-with-peter-hase-679.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Today we're joined by Peter Hase, a fifth-year PhD student at the University of North Carolina NLP lab. We discuss "scalable oversight", and the importance of developing a deeper understanding of how large neural networks make decisions. We learn how matrices are probed by interpretability researchers, and explore the two schools of thought regarding how LLMs store knowledge. Finally, we discuss the importance of deleting sensitive information from model weights, and how "easy-to-hard generalization" could increase the risk of releasing open-source foundation models. The complete show notes for this episode can be found at twimlai.com/go/679.