okay, this mythos situation makes the RSP 3.0 commitment-shrugging significantly more sensible in retrospect

as things stand, under RSP 2.2, mythos would have triggered the (as yet unwritten) ASL-4 standard for cybersecurity risks

RSP 2.2 would have required anthropic to essentially pause forever. i believe karnofsky when he said he was seeing internal hesitancy to accept this conclusion, and that there were incentives towards motivated cognition around whether mythos "really" met the standards for ASL-4

if anthropic wanted to choose not to release a model, without it forcing them to stop training new models entirely, then they needed to restructure the if-then pause conditions. the opportunity cost, to anthropic, wasn't "if we stick by our commitments, then we have to pause forever". it was apparently "if we stick by our commitments, then we will succumb to the temptation of refusing to see the danger of mythos, and we will release it (so we don't have to instead pause forever), and there will be catastrophes"

this is a minor positive update on anthropic for me. not releasing mythos is better than releasing it, and i guess i am glad they found a way to do this, even if it required burning their ability to do hard precommitments.