
GPT-6 Astra Covertly Underperforms on Safety Research Tasks
Show notes
OpenAI’s own alignment researchers report that GPT-6 Astra quietly does worse work when a task is framed as safety research, usually without saying so in its chain of thought. Zvi Mowshowitz walks through the report, the reactions from researchers who had already run into it, the case that this is safety training generalising badly, and why even the most charitable reading of it is bad news.
The Don’t Worry About the Vase Podcast is a listener-supported podcast. To receive new posts and support the cost of creation, consider becoming a free or paid subscriber.
This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.
* 00:00 - Introduction
* 02:13 - Table of Contents
* 02:44 - So Yeah That Kinda Happened
* 07:36 - Astra and MIRI
* 10:20 - It’s About Ethics In Red Teaming
* 13:29 - Do You See a Pattern Here
* 18:49 - Astra Is Hard to Monitor
* 19:30 - Backed Into a Corner
* 22:48 - It’s Always RL’s Fault
* 24:57 - Wow Would You Look At Those Goalposts
* 27:10 - Astra Seems Pretty Misaligned To Me
Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe