How do you feasibly regulate AI? It's so easy to write a machine learning algorithm and the fundamentals for learning it are all over the place. YouTube tutorials, Udemy, Coursera, or university courses all teach it. Once you know how to write AI algorithms, all you need is a computer.
The hard part is getting data to train your model, and it becomes a natural monopoly because the service that performs better will see more traffic to collect more data.
But I don't understand how you can regulate AI in a way that makes any sense. Ban specific algorithms? Limit data collection? Create some confusing wording around generic "algorithms that improve other algorithms" or something like that?
And even if you passed such a law, it's trivially easy for someone from any other part of the world to set up an API for their ML algo to run.
Regulation on AI sounds like it would be about as successful as the war on drugs.
Size. Ban exports of GPUs and other compute devices, task the CIA to track down leaks, bomb data centers you can't raid with the FBI. I think that's very unlikely to happen though.
It's not a lost cause if you want to advocate for it, it might be worth setting up a protest because not a lot of others appear to be doing so, see Holly Elmore and Rob Miles dialogue on AI Safety Advocacy at [0].
I'd trust Yann LeCun a lot more if he committed to eternal human supremacy, but he's said some dubious things in the past that amount to being okay with giving over the light cone to some arbitrary AI system. I'm pretty sure he does mean "light cone" more or less, he's a techno optimist.
There's a "system trap" where building intelligent systems or augmenting humans will grant at least a short term advantage, it has natural "arms race" dynamics.
Historically people have not done well at avoiding systems traps and maintaining any sort of global consensus not to do things. Particularly things that a) don't seem to be an immediate atrocity and b) grant economic advantage.
It is not because it isn't unexpected that it should not be pointed out. Regulatory capture is very common. Does this mean we should not talk about it?
13 comments
[ 4.5 ms ] story [ 23.5 ms ] threadThe hard part is getting data to train your model, and it becomes a natural monopoly because the service that performs better will see more traffic to collect more data.
But I don't understand how you can regulate AI in a way that makes any sense. Ban specific algorithms? Limit data collection? Create some confusing wording around generic "algorithms that improve other algorithms" or something like that?
And even if you passed such a law, it's trivially easy for someone from any other part of the world to set up an API for their ML algo to run.
Regulation on AI sounds like it would be about as successful as the war on drugs.
[0]: https://www.lesswrong.com/posts/gDijQHHaZzeGrv2Jc/holly-elmo...
Historically people have not done well at avoiding systems traps and maintaining any sort of global consensus not to do things. Particularly things that a) don't seem to be an immediate atrocity and b) grant economic advantage.