The way I see "safety" isn't really about what AI "can do", but about how we allow it to be used. E.g. an AI that's used to assess an insurance claim should be

The way I see "safety" isn't really about what AI "can do", but about how we allow it to be used. E.g. an AI that's used to assess an insurance claim should be fully explainable so we know it isn't using racist biases to deny claims based on skin colour. If the AI can't give that guarantee, it isn't fit for purpose and its use shouldn't be allowed. Same with killer robots (or whatever it is people are afraid of when they talk about "AI safety"). As long as we can control who they kill, when, and why, there's no real difference with any other weapon system. If that control is even slightly in doubt: it's not fit for service. Does this mean that bullshit generating LLMs aren't fit for service in many areas: it probably does. But maybe there steps can be taken to mitigate risks. I'm sure this will involve some bureaucratic overhead. But it seems worth the hassle to me. Being against AI Safety is a stupid hill to die on. Being against some concrete declaration or a part thereof, sure, that might make sense. But this smells a lot like the tabacco industry being against warnings/filters/low-tar, or the car industry being anti-seatbelt.