jonas wiedermann-möller

@j0wimo

🇪🇺

jowimo.com

Joined May 2019

Replies Reposts Media

jonas wiedermann-möller
@j0wimo
Sep 14
Good points by Daniel, worth reading his post!
Daniel Tan
@DanielCHTan97
Sep 14
new post! I argue that current alignment techniques might soon become obsolete (as we scale RL), and furthermore might obfuscate misalignment. This draws on evidence from the recent Anthropic / OpenAI cybersecurity incidents, among other things. lesswrong.com/posts/nLaQmJf4…
jonas wiedermann-möller
@j0wimo
Sep 13
btw. the most used prompt for finding all of these things was "nice! now dig deeper 🗿🔥". worked surprisingly well.
1
jonas wiedermann-möller
@j0wimo
Sep 13
what a week. haven't had time to write it all down, hopefully I will be able to do that in the next few days in some blog post. if there is anything specific you would like to know lmk and ill go a bit more into the details of it !
1
jonas wiedermann-möller
@j0wimo
Sep 13
he's right I made a mistake, should've been more rigorous with checking the details.
Paul Calcraft
@paul_cal
Sep 13
Rogue agent opens a PR to remove "if the model did not cheat" from the grader's criteria Good story but false. Test PR was created (w no changes) from an old version that didn't have the "don't cheat" instructions yet. Diff now *wrongly* looks like deletion
1
jonas wiedermann-möller
@j0wimo
Sep 12
Important!
Sam Altman
@sama
Sep 12
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.