Look, there is no AI safety advancement without AI capability advancement. I think we van learn fuck all about AI safety if we don't try to actually build those AIs, carefully, and play around with them. AI safety is not an actual field of study when you don't have AIs of corresponsing level to study - otherwise there are zero useful results.
Sure, but you snuck an assumption in there. Just because AI is possible, or someone else will do it, doesn’t obligate us to build it. If we can’t make AI without risk of significant or existential harm, then we shouldn’t do it at all.
