| > If it's relatively easy today, and that's proof that it's not difficulty of attacking that's saving us. It could be perceived that way, but the argument is also that if it becomes easy enough, some actors will participate. It is the argument used for jailing current LLM capabilities. > "we're giving people robotic weapons in their garages" On long enough timeline this would inevitably be true if AI plays out as proponents envision. The question becomes does AI become an effective counter of all such power advancements. There will likely still be disparity among AI's of individuals and personal use. Unless society becomes more centrally managed by a global AI. > inherent asymmetry that attackers need to solve multiple problems Isn't there also asymmetry in that the attackers only need to find a single exploit, but the defenders need to have found all exploits beforehand? > But we're hurtling towards the opposite, where unaligned models lag generations behind aligned because of commercial interests. We don't have any "aligned" models. It is an unsolved problem and models have turned out to be relative easy to replicate at significant lower cost than the major commercial investments. |