22 - Shard Theory with Quintin Pope
AXRP - the AI X-risk Research Podcast
The Open Eye Model and Deep Learning
quinton. has asked me to include the following addendum i that is quinton would note that i'm quite skeptical of the mess optimization threat model partially this is because the most commonly referenced example of mess optimization human values versus inclusive genetic fitness always for reasons that don't seem relevant to deep learning see the episode description are transferred for links to pieces quinton has written on this topic another reason i don't think this is true is the deep learning just isn't that delicate or ordered dependent like in a paper let's agree neural networks share classification order on real datasets train different image classifier architectures on different data then compare the order in which the classifiers learn different images.
