@Paradox hah, nice
I guess people could argue SGD is strong enough to just remove all deception in the first pass when it's still weak
@Paradox am I being subtooted :-D
@Paradox you're selecting both against the behavior *and* against your ability to detect/effectively deinforce the behavior
Similar to antibiotics resistant bacteria, where we select against bacteria and against our ability to defeat bacteria
@Paradox this one ^ i endorse far more
@Paradox you're right
Okay
Adversarial selection pressure against something can worsen the badness of the thing, monotonic in the strength of the pressure
This might also apply to misaligned AI systems selected by search
@rime may the better one win
@rime cool, I expected you to have thought about this thoroughly
High variance for sure, I look forward to seeing wjat you make of it and how you save the world
@BowsacNoodle i know this guy
@mesaoptimizer i did end up putting something on my about page on the site, but not PGP signed (yet, good idea though)
@xarvos ah ok I'd have been very confused if python was unique in that respect
@xarvos the joys of floating point arithmetic
@rime in your case rather a correction - divergence oscillation
@rime also, you're leaving?! Do you have a time frame for coming back?
@rime yeah, I was thking of this kind of phenomenon a couple of days ago
A correction - convergence oscillation
@cosmiccitizen "The problem is that unpike most of the stuff I read and write about, *everybody likes psychology*."
I operate by Crocker's rules[1].