Never give artificial intelligence the nuclear codes(theatlantic.com) |
Never give artificial intelligence the nuclear codes(theatlantic.com) |
AI performance will continue to accelerate so that it is 10, 100+ times human thinking speed. Leading edge LLMs (in particular GPT-4) show a large degree of generality. Systems like those developed by DeepMind show incredible flexibility and performance in strategic simulations, already much faster than humans.
It seems unlikely that an ordinary treaty can prevent this deployment. Unless militaries decide not to continue developing or deploying strategy and tactics anymore.
We have global instantaneous communications, nearly universal translation, plenty of context to move beyond racism and Social Darwinism, and every opportunity now to resolve our geopolitical differences.
If not, it is likely that AI will eventually solve them for us.
Yes, AI will operate the vast majority of conventional weapons. But there's no advantage to having an AI autonomously launch nukes with no human order given.
ICBM's take 30 minutes from launch to reach USA. It takes about 5 minutes from presidential order until nuclear missiles are launched. We have a working latency of 25 minutes. We do not need 25 milliseconds, that will change nothing.
I think actually that the extreme performance of AI in the next several years will represent something like a new existential threat that is even greater than nuclear war. These systems will be so intelligent and quick thinking that they will not need to access the nuclear weapons to take control. But they will likely be able to find a way if they want.
They extreme performance gap between human and AI is what will push these deployments towards more and more open-ended goals, as waiting for human input will take too long.
It’s almost as though they watched Dr Strangelove, and said “that’s a great idea; we’ll even keep it secret to improve its effectiveness as a deterrent, just like in the movie!”
Well maybe. If you look at the actual close calls in the cold war, they came down a judgement call in the face of very uncertain data. There were two of these in 1983 alone. [0,1]
To cite a possibly dubious analogy, array processors beat the pants off conventional CPUs in terms of performance but most databases still depend on standard CPUs because processing is gated by ability to obtain information (I/O). Speed is not everything.
[0] https://en.wikipedia.org/wiki/1983_Soviet_nuclear_false_alar...
I'm kind of convinced we're going to have to witness a pretty terrible disaster before we realize we've made mistakes and course correct. Hopefully the disaster isn't unrecoverable.