alt Hacker NewsWe have invented a new tool that can cause great harm. Do you see any value whatsoever in promulgating safety guidelines for humans to use the tool without hurting themselves or others? Do you not own any power tools?
Replies
I think in order for "AI safety" to be achievable and effective, we need to have a shared agreement on what "safety" means. Recently, the word has been overloaded to mean all sorts of things and used to justify run-of-the-mill censorship (nothing to do with safety).
Safety should go back to being narrowly defined in terms of reducing / preventing physical injury. Safety is not "don't use swear words." Safety is not "don't violate patents." Safety is not "don't talk about suicide." Safety is not "don't mention politics I don't like." As long as we keep broadly defining it, we're never going to agree on it, and it won't be implementable.
Of course there is value in promulgating safety *guidelines*.
But we cannot guarantee those guidelines to always be followed.
> Do you see any value whatsoever in promulgating safety guidelines for humans to use the tool without hurting themselves or others?
Are all the tool users required to train your safety guidelines and use it in a context that reminds them they are responsible for following them?
Because if no, then no the guidelines are useless and are just an excuse to push blame from the toolmakers to the users.
I see value in promulgating safety guidelines for power tools, sure.
There's another comment comparing LLMs to shovels, and I think both that and the power tool comparison miss the mark quite a bit. LLMs are a social technology, and the social equivalent of getting your hand cut off doesn't hurt immediately in the way that cutting your actual hand off would. It's more like social media, or cigarettes, or gambling. You can be warned about the dangers, you can see the shells of wrecked human beings who regret using these technologies, but it doesn't work on our stupid monkey brains. Because the pain of the mistake is too loosely connected to the moment of error. We are bad at learning in situations where rewards are immediate and consequences are delayed, and warnings don't do much.
I guess what I'm really saying is that these safety guidelines are not nearly enough to keep us safe from the dangers of AI that they're meant to prevent.