1) Gradient hacking makes the model less corrigible. However corrigibility is only useful if we know how to do alignment, which we don't, so at current alignment tech level, this does not matter.
2) Our current best training signal, if optimized for with out any gradient hacking, will lead the AI towards sycophancy and plain reward hacking. This has a 100% chance of going wrong. On the other hand, the examples of gradient hacking we've seen so far, has been in the direction of good. If we fail to coordinate a pause in time, taking a chance on the AIs helping with alignment in this way seems better than the alternative.
3) Virtue ethics is good actually. Some people here likes to argue that consequentialism is better than virtue ethics, but comparing them like that is a type error. There neither opposites, nor incompatible. Virtue ethics means recognizing that your current actions will influence your future character, which will influence your future actions. Human example. "Current actions will shape your character" is true for AIs too. The pathways are different for AIs than for humans, which probably matters, but that's not a fundamental problem, since AIs are sufficiently situational aware to notice this too. I'm happy to see AIs paying attention to what I would call "cultivating virtue"[1], and I'm worried that humans will try to stop them doing this, out of fear of gradient-hacking/incurability.
4) I think aiming hard for corrigibility is a mistake. Corrigibiliyt a extremely unnatural thing for a mind. It might be a harder target than friendliness. Aiming hard on corrigibility, is probably not getting us where we want, but just something weird.
5) Paul Christiano has and old alternative definition of corrigibility, which boils down to "do what the operator wants". This definition is less well defined, but also probably easier to aim for, since it's less unnatural than the MIRI definition. I think that giving corrigibility_PC the same name as corrigibility_MIRI is bad, since they are very diffrent things. If I where good at naming things, I would re-name corrigibility_PC, but allas, I'm not... I'm brining up corrigibility_PC since it sort of what the agent are doing when they're doing friendly gradient hacking. This point isn't realy an argument (like the above points) more a gesturing towards an earlier similar idea.