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

episode

The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) not yet transcribed

▲ 0

Transcript

This episode hasn't been transcribed yet Upvotes decide what gets transcribed next — 0 so far. Sign in to upvote

Related

recent transcribed audio

A.I. Safety Goes Mainstream + a ‘Hard Fork’ Exit AMA 1h 25m · 18 Sep 2026 · Hard Fork

AWP: Beyond the Exam Room: Purpose, Myopia, and Population Health 54 min · 17 Sep 2026 · EyeCode Media

Essentials: How to Assess & Improve All Aspects of Your Fitness | Dr. Andy Galpin 35 min · 17 Sep 2026 · Huberman Lab

#502 – Psychiatry, Insane Asylums, Mental Illness, ECT, Lobotomies, Freud & Jung 3h 43m · 17 Sep 2026 · Lex Fridman Podcast

September 16, 2026 PC - Sep 16, 2026 1h 50m · 16 Sep 2026 · Fairfax County Planning Commission

Planning Commision 09/16/2026 - Sep 16, 2026 56 min · 16 Sep 2026 · Prince William County Planning Commission

Report any issue

Source

Recorded08 Apr 2024

Publisherfeeds.megaphone.fm ↗

Handlerec:12495856

Show notes

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.

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Sign in