Sure, but this sentiment is why software "engineering" isn't really. You can justify it by not being important enough for actual engineering practices I guess, but to me it's a lack of pride in and care of your product.
> You can't formally verify anything that uses consensus
What did you mean by this then? There certainly seems to be nothing special about consensus that makes it any harder to verify than anything else. It's not fundamentally impossible to verify the software that CERN uses, it just takes some work.
Thanks, interesting. However, that just seems like a protocol like any other, with no real reason why you "can't" formally verify it. Is there something special about a consensus algorithm / protocol that makes it any more difficult to verify than any other algorithm which doesn't yet have a formally verified implementation?
Care to elaborate? Perhaps the tools to do this in practice aren't there (which just shows how young the field of software "engineering" really is), but what consensus are you talking about and how is it an obstacle to verifying code? Most of the web follows standards and protocols, which actually sort of a prerequisite for communications across different systems...
This article actually explains how this bug in particular could have been avoided. Sure you may not consider his approach realistic, but it's not at all saying "don't have bugs". In fact, not having formal verification or similar tooling in place, would be more like saying "just don't write buggy code".
Not to single you out in particular, but I see this sentiment among programmers a lot and to me it's akin to a structural engineer saying "I laughed out loud when he said they should analyze the forces in the bridge".
Well, the question of what is "too much" of a snitch is also a question of ethics. Clearly we just have to teach the AI to find the sweet spot between snitching on somebody planning a surprise party and somebody planning a mass murder. Where does tax fraud fit in? Smoking weed?
Well no, breaking that rule would still be the wrong action, even if you consider it morally better. By analogy, a nuke would be malfunctioning if it failed to explode, even if that is morally better.
The agen already bypassed the file reading filter with cat, couldn't it just bypass the URL filter by running wget or a python script or hundreds of other things it has access to through the terminal? You'd have to run it in a VM behind a firewall.
Unfortunately yes, teaching AI the entirety of human ethics is the only foolproof solution. That's not easy though. For example, what about the case where a script is not executable, would it then be unethical for the AI to suggest running chmod +x? It's probably pretty difficult to "teach" a language model the ethical difference between that and running cat .env