Yep. I don’t like your proposed test (what’s going to define “progress”?), but yes.
My main purpose for this post wasn’t to make amazing AI safety researchers though. It was to offer people who want out of the inner doomsday trap a way of exiting. That part is a little more tricky to test. But if someone wants to test it and wants to put in the effort of designing such a test, I think it’s probably doable.
Yeah, the test has to be set up with all the normal caveats in advance (including being specific enough to measure, but broad enough to avoid people having good excuses to ignore whatever its conclusions turn out to be).
Yep. I don’t like your proposed test (what’s going to define “progress”?), but yes.
My main purpose for this post wasn’t to make amazing AI safety researchers though. It was to offer people who want out of the inner doomsday trap a way of exiting. That part is a little more tricky to test. But if someone wants to test it and wants to put in the effort of designing such a test, I think it’s probably doable.
Yeah, the test has to be set up with all the normal caveats in advance (including being specific enough to measure, but broad enough to avoid people having good excuses to ignore whatever its conclusions turn out to be).