LessWrong (Curated & Popular) cover image

“Narrow Misalignment is Hard, Emergent Misalignment is Easy” by Edward Turner, Anna Soligo, Senthooran Rajamanoharan, Neel Nanda

LessWrong (Curated & Popular)

00:00

Intro

This chapter explores the intricacies of model alignment, highlighting the differences between narrow and emergent misalignment. It examines how fine-tuning on specific harmful examples can unintentionally affect a model's broader behavior, while discussing training techniques and the stability of solutions in machine learning.

Transcript
Play full episode

The AI-powered Podcast Player

Save insights by tapping your headphones, chat with episodes, discover the best highlights - and more!
App store bannerPlay store banner
Get the app