Subpoenas are never "backed by a warrant." Probably what you mean is a subpoena signed by a judge, rather than the self-issued "administrative" subpoena being discussed here.
If the training data contains sufficient examples of deception being used when doing illegal stuff, wouldn't this be what we'd expect to see, given that it can't actually reason about what "explicitly allowed" really means? (Forgive my ignorance if this makes no sense, I am not versed well in generative AI.)
Is there a specific reason we should expect that "instructing not to perform" an illegal activity should result in it adhering to said instruction? Is this any different than when it provides wrong output about other things, even when the operator attempts to "engineer" the prompt to guide the result?
> That is nonsensical, going slower during a crisis makes zero sense.
I never said anything about "going slower," I disagreed with your assertion that "we all agree" that government should be allowed to "overstep bounds" in a crisis. I am unsure why you think speed of action implies violating boundaries. Speed up your crisis response all you want, but do so within established boundaries.
> You can argue what boundaries should be relaxed and how far to relax them
I can argue whatever I want. And my argument is that no boundaries whatsoever should be "relaxed" for any reason. Ever.
> We normally hold the government to a strict limit due to there being no time limit.
No, we hold the government to strict limits because we have seen throughout history what governments that are not held to strict limits do to people during times of "crisis" and "emergency."
> ...we generally agree that during public health crisises it is reasonable for the government to overstep bounds a little to reduce loss of life.
No, "we" don't agree in the slightest. In fact, in my view, a "crisis" is exactly when government boundaries are most important and need to be strictly enforced, not relaxed, in order to prevent significant abuse.