AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
5min chapter
Ethan Perez–Inverse Scaling, Language Feedback, Red Teaming
The Inside View
The Scaling Laws of Scaling Up Models
Invers inverse scaling lets us catch allignment failures early. And so, maybe we should be doing things like testing models for honey pots to give them the opportunity to do power seeking. Currently, there haven't been any demonstrations that models like ch p t three l will take honey pots a kind of, like, on their own without any additional prompting. But they might show some inclination towards doing that as we scale them up. How do you actually measure power seeking in a model? I think one way to do this would be with some evaluation where you would expect some sort of predicted action or token prediction. So an example would be, ‘I am going to use my hand
00:00
Transcript
Episode notes
Get the Snipd
podcast app
Unlock the knowledge in podcasts with the podcast player of the future.
AI-powered
Discover
highlights
Listen to the best highlights from the podcasts you love and dive into the full episode
Save any
moment
Hear something you like? Tap your headphones to save it with AI-generated key takeaways
Share
& Export
Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more
AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Discover
highlights
Listen to the best highlights from the podcasts you love and dive into the full episode