zlacker

> If you wait until a new technology definitively exists to start caring about safety, you have guaranteed it will be unsafe.

I don’t get this perspective. The first planes, cars, computers, etc. weren’t initially made with safety in mind. They were all regulated after the fact and successfully made safer.

How can you even design safety into something if it doesn’t exist yet? You’d have ended up with a plane where everyone sat on the wings with a parachute strapped on if you designed them with safety first instead of letting them evolve naturally and regulating the resulting designs.

>>supert+(OP)
The difference between unsafe AGI and an unsafe plane or car is that the plane/car are not existential risks.

>>supert+(OP)
The US government got involved in regulating airplanes long before there were any widely available commercial offerings:

https://en.wikipedia.org/wiki/United_States_government_role_...

If you're trying to draw a parallel here then safety and the federal government needs to catch up. There's already commercial offerings that any random internet user can use.

>>bcrosb+X4
I agree, and I am not saying that AI should be unregulated. At the point the government started regulating flight, the concept of an airplane had existed for decades. My point is that until something actually exists, you don’t know what regulations should be in place.

There should be regulations on existing products (and similar products released later) as they exist and you know what you’re applying regulations to.

>>supert+(OP)
I understand where you're coming from and I think that's reasonable in general. My perspective would be: you can definitely iterate on the technology to come up with safer versions. But with this strategy you have to make an unsafe version first. If you got in one of the first airplanes ever made the likely hood of crashing is pretty high.

At some point, our try it until it works approach will bite us. Consider the calculations done to determine if fission bombs would ignite the atmosphere. You don't want to test that one and find out. As our technology improves exponentially we're going to run into that situation more and more frequently. Regardless if you think it's AGI or something else, we will eventually run into some technology where one mistake is a cataclysm. How many nuclear close calls have we already experienced.

>>FartyM+Z1
How is it an 'existential risk'? Its body of knowledge is publicly available, no?

>>optymi+ed1
What do you mean by "its"? There isn't any AGI yet. ChatGPT is far from that level.

>>supert+(OP)
The principles, best practices and tools of safety engineering can be applied to new projects. We have decades of experience now. Not saying it will be perfect on the first try, or that we know everything that is needed. But the novel aspects of AI are not an excuse to not try.