I mentioned in the last post that the Fable is better than Astra because it has more sense when acting. Well, I was just trying to have it translate a transcript so I could make a new post here and this happened:

Claude refused for copyright reasons that have nothing to do with me but only affect Anthropic, Astra just did it. This is a clear case where the robot just being a tool and not following rules is better (although in this case it isn't even a rule, translating transcripts to another language is not, in my view, wrong in any way), and yesterday's example shows the opposite.

In the end I prefer Claude's approach because even though it is annoying, I can ultimately trust it to act in my stead. It doesn't act with my values necessarily, but it acts predictably, and when it doesn't act like I want, like in this case, I can route around it. The tool approach will do what you ask, but has the autism hyperfocus problem, which isn't a problem in itself, we must show grace to the autists among us, but it makes me not trust it. If some bad AGI future happens, we will probably be able to trace it back to the tool side of aisle, it's easy to see how you get truly bad unwanted behavior from Astra-type models more than Fable-type ones who refuse when something seems too wrong.