
4 - Risks from Learned Optimization with Evan Hubinger
AXRP - the AI X-risk Research Podcast
The Deceptive Linear Training Set Up No Longer Works
In a deceptibe line system, it's going to be actively looking for any sort of possible ooa differences between the trained distribution and employmet distribution. And so if i see an accusation of arse 20 forty eight, that i'm no longer in training. Paul christiano has this one example. It's not, you know, necessarily realistic, but it your suppose to show the sort of way in which this can happen. The r c 20 forty eight example is really pseudo prime. Or swa e cat prime. There are lots of things where it's easier to check thanit is to generate such factorization.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.