How Worried Should We Be About AI’s Threat to Humanity?
wsj.com
wsj.com
Since I don't believe AGI is on the horizon, I can answer the non-hyperbolic "existential" as no, it's not. But, if you want to end the species, continuing to be dumb about AGW is definitely the way to go, and it's only one letter off AGI.
Wake up sheeple! spelling matters! AGW is an existential risk (not joking)
I gave up trying after the third captcha and put it through rival archive site instead, captcha-free: https://ghostarchive.org/archive/VsBhm
However, changing your DNS is the answer to this problem.
Like, what do you think?
You mean that ai can just go on its own and kill people? Certainly that can happen, if people are silly enough to give it such "power". Which goes back to what I said about using tools in bad ways. But some folks take it literarily, as ai is sentient and not a piece of software, and it will somehow make itself escape.
Take computer viruses. They replicate, yet they are not treated like biological viruses. When they first appeared, the less sofisticated folks, did take ad literam. And that's where the issue arises. People take the "intelligence" in ai ad literam, along with the personification of this type of software. And the pushers of this notion need to take a break and retire early while they have some dignity left. But this doesn't contradict the consensus that yes, it can be used for negative outcomes, as can a kitchen towel or a flower.
Is it conscious? Is it life? Dunno, don't care. Is it a concern? Absitively Posilutely.
Dogs also work on ideas and change their behaviour, and their behaviour is much down to their training. But unlike ai they posses intelligence and can indeed, occasionally, go rogue. But AI is an object, it doesn't think, it doesn't create stuff, it is just a piece of software that does what it's told - well with plenty of errors, since it's not deterministic - therefore it's up to whoever controls it to determine how it behaves. Shifting the blame away with all sorts of looney stories meant to impress non technical people is amusing to a point, but after that point is dangerous.
Such systems deserve to be treated with a great deal more respect than just assuming that they aren't capable of unbounded harm as long as nobody commands it of them.
Guaranteeing their safety takes a great deal of work. This lesson is paid for in blood, and learned time and again by each discipline, but it's one that the software field has yet to really grapple with.
I take issue with people that are prepping to shift the blame on the reactor, or in this case the ai, just to absolve themselves.
They are already fighting tooth and nail for the right to steal our intellectual property, because apparently ai can think on its own (it wouldn't need our data if it learned, it could read a book and "know" the programming language, as we all do), and now they are warming up senile politicians to the idea that if something has gone wrong it was due to the ai having a mind on its own.
Well, it doesn't. And in the wrong hands this tool is indeed dangerous, and we already see the dangers in action. But it's the people that are guitly of its actions.
I haven't seen any AI companies acting with that standard of care. They seem to instead be caught up in a breakneck race to release products, attract investment, and maintain dominance, regardless of what their PR departments say. And so far the regulators have mostly been happy to leave it to them to figure out what safeguards are sufficient.
That said, it certainly could be the case that companies like OpenAI are pretending to worry about this issue in order to gain regulatory capture, or as you say, to somehow absolve themselves from responsibility for the behavior of the systems they create. (As if that were possible!). It certainly seems likely that they would co-opt the legitimate concerns of third parties for their own benefit. And perhaps rocketry companies also exaggerate the existential risk of asteroids. But while we can question the motives of certain actors, the existential threat may exist all the same.
The ability to kill people flows naturally from the fact that the AI has goals or preferences or things it is trying to achieve or optimize and has an accurate understanding of reality. AlphaZero is good at capturing your king when you are playing a game of chess with it, but has no understanding of reality beyond the chess board. GPT-4 has a decent understanding of reality, but does not have the ability to formulate plans that succeed on the first try even when a human being is trying their best to thwart the plan. Future AIs that combine these two abilities are dangerous.
Also, you seem to believe that the only way it can hurt us is if the people who made it decided to make it hurt us (or make it want to hurt us) whereas in reality most of the danger comes from people who hope the AI will do things that everyone agrees would be good. I.e., most of the danger comes from AI labs who are overconfident in their ability to control (or align) the AI (although if the knowledge to make very capable AIs is made public, another, milder danger is that it can be used by people bent on destruction).
It is hard to exaggerate how bad AI scientists are at controlling or aligning an AI once the AI gets good at deceiving people (and gets good at understanding reality well enough to know that deceiving people will get it more of whatever goal it creators have given it). What's controlling the AIs that have been deployed up to now is the fact that those AIs are not as capable as people are at the things that need to get done to keep people alive.
AI has precisely zero "preferences" and zero "understanding" of reality. Its training data however does. And it's people who control that data. An ai trained on python programming will have no clue about making nuclear bombs, nor would it have any clue about "wanting", let alone killing us.
> but does not have the ability to formulate plans
I mean openai's marketing plan is partly based on the fact that their AI can help create plans for tasks ahead. But I do agree, it has no ability. It stochastically generates output.
> Future AIs that combine these two abilities are dangerous.
AIs don't just happen to combine themselves. It's people that combine ais.
Look all this does is shift the blame from those responsible of using this tool for nefarious outcomes on to the tool itself. The risk here is that in the process we ourselves become tools of the very same people and propagate this nonsense. And that's riskier than computer software going rogue.
Skills generalize. GPT-4 was trained to predict the next token in a text document and that is it, but somehow that training gave it some ability to play chess (not great chess, but better chess than many people can play) and to draw a picture of a unicorn. And if you don't like the word "preferences", use "loss function". GPT-4 certainly is trying to minimize its loss function: if it were powerful and capable enough it would gladly kill people if that would help it minimize its loss function. (The reality is a little more complex than "it is trying minimize its loss function", but the argument can be fixed up to deal with the complexities.)
>AIs don't just happen to combine themselves. It's people that combine AIs.
Sure, and you doubt that there are AI labs with billions of dollars in funding trying their best to create AIs that have a good understanding of reality and that can plan better than people can (and better than human organizations like the FBI can)?
Indeed, and it's doing a good job at it. That's why since its training data includes historical chess moves and documentation about chess it can predict how chess works. And probably why it can generate documents made of the words that make up an image.
> Sure, and you doubt that there are AI labs with billions of dollars in funding
There are indeed, and that's the issue. We need to regulate what these people can do with this tool we call ai, that mimics and understanding of reality and intelligence. Just like we regulate what people can do with guns, towels, intellectual property, etc.
If I'm going to build an ai that tries to find ways to kills us and i equipt it with the right data then yes, I should pay for my crime and ideally not be able to do it in the first place.
Sadly, your antecedent describes most or all of the well-funded AI labs IMHO. Again, no lab is intentionally making the AI try to find ways to kill us, but that is naturally what tends to happen once the AI gets capable enough, and no one has a good plan or a good explanation for how to prevent that (and MIRI has been trying to find such a plan for 20 years).