What is the A.I.’s existential threat to humanity?

1 points by boraoztunc ↗ HN
Today's newspaper had a section about how AI should be considered a societal risk, with statements from creators of these large language models.

> Executives from top artificial intelligence companies have warned that the technology they are building should be considered a societal risk on a par with “pandemics and nuclear wars,” according to a 22-word statement from the Center for AI Safety that was signed by more than 350 executives, researchers and engineers.

> Eventually, some believe, A.I. could become powerful enough that it could create societal-scale disruptions within a few years if nothing is done to slow it down. Those fears are shared by numerous industry leaders, putting them in the unusual position of arguing that a technology they are building poses grave risks and should be regulated more tightly.

I don't understand how these exes are leading these conversations, as they are responsible creating these products with potential harms in the first place.

13 comments

[ 3.2 ms ] story [ 24.5 ms ] thread
Concerns: Some argue that A.I. is improving so rapidly that it has already surpassed human performance in some areas, and that it will soon surpass it in others.
Sure. I can buy a car that can go faster than I can sprint, too. That's not an existential risk; that's just a better tool.

How do people realistically get from "it can outperform us" to "therefore we all die"?

Threat to upend our society? Sure. So did the industrial revolution, and the agriculture revolution. And most of us don't want to go back.

>How do people realistically get from "it can outperform us" to "therefore we all die"?

It's the year 2023. The "it can kill us all" narrative is 2030+ range as far as I know.

>Sure. I can buy a car that can go faster than I can sprint, too. That's not an existential risk; that's just a better tool.

Not exactly the same. The nuclear bomb can do a lot more damage than a human can physically. Nuclear bombs are an existential risk. Some people equate the development of AI as the same as nuclear bombs.

>I don't understand how these exes are leading these conversations, as they are responsible creating these products with potential harms in the first place.

They're basically trying to form a consortium, hoping to create an environment for AI labs and regulators to mitigate the dangers of AI. I suggest you go to their website. Their mission and reasons are stated there.

For example, one of the core ideas is that AI labs are competing in a high stakes race, and many of them will do dangerous things with AI in order to try to come out on top. This consortium is trying to prevent this sort of thing.

> "I don't understand how these exes are leading these conversations, as they are responsible creating these products with potential harms in the first place."

Imagine a caveman figured out how to make fire. Their tribe sees how they can use it against their enemy tribes or to clear brush or whatever. They decide to hold a meeting to discuss the fire and how to not let it burn down their own village or their whole forest. Probably the guy who figured out how to make it would be involved in that meeting.

That was a nice way to explain, thank you! And also was a reminder for me that living daily with all these technological developments, one loose the perception of that we are the current cavemen of future societies. After 100,000s of years, all these will the first paintings on the walls.
The danger is induced by the possibility that AI will surpass the humans in a few years leading to disruptions in working and personal life. Then, the whole economy needs to be overthrown and alternatives for taxes and such needs to be found. This endangers the social communities imho. One solution could be to tax upon the work, not the salaries of workers.

Also, if one takes a closer look who is saying that.. of course, a lot of people, but the Execs say it (or sign letters to halt the development) just because of they securing their stakes.

Big companies will get rid of new actors entering the market by laying strict rules upon new actors, that will slow down development and hold new ones viable, but for themselves, they need just to change their technologies.

The ones who sign letters to stop development for a year are keen to develop their own and are in need to slow down the big competitors. Elon Musk, f.e.

So, it's a danger, but Internet was for some actors in the beginning, too, a big danger;)

"the possibility that AI will surpass the humans in a few years"

Geoff Hinton it has been saying for a couple of months now, "we are already there". The "digital intelligences" have already surpased human level intelligence.

Many complex explanations follows usually, some related to energy efficiency / consumpsion vs. processing capabilities (spoiler alert: LLMs are better than us);

other stuff is related to another "not AGI yet" arguments,

i.e. the hallucinating thing in LLMs (if you train a LLM -or you fine tune it - with only data with precise truths, not the whole Reddit posts, you could create a precise not hallucinating ever LLM),

i.e. the "human neurons still rule". Hinton says, you cannot instantly copy your whole memory to 10.000 other humans - just born humans - and these things can, heck, the whole gang, 10.001 LLMs can just instantly share between them whatever they learn, instantly. Hence, the human neurons just got deprecated, but few have realized yet.

Ilya Sutseker has said some stuff, not as much alarming as Hinton, but explaining how LLMs work, and saying, in the most polite way possible: these things are otherworldly for us humans.

I think both guys didn't say the current LLMs (Palm, GPT-4 upgraded to 2023), are AGI already, just because there is no way to prove they are general intelligences, just as there are no way to prove otherwise.

I'm with you here. Hallucinations are the first sign for creativity. Creativity is seen as one part of intelligence, like in animals who use tools and also combine them. We're glad, that no one has invented a 'trouble maker' now, an technique to ask and keep other models busy. That's what's in our head, it never rests. Fast times we live in :) I never thought to experience a pandemic, a war, rise of the AI, and other things to come, in my life
Q: "What is the A.I.’s existential threat to humanity?"

A: a scam. The existential threat is us, see https://mfioretti.substack.com/p/you-are-all-getting-this-st...

Probably one of the worst Substack posts I've ever read. Waste of time.

The whole article can be summarized as "AI can't kill us because it can't access the physical world without human help. You are stupid if you think AI can kill us."

As to why A.I. is being held up as a threat, well tracking and loss of privacy should be top of the list due to the threat of over analysing the population and acting on the data. But IMO at the moment there is a push on behalf of whomever to try and put A.I. back in the bottle ... for a time at least ... but anyone half smart knows most of the companies already involved, will be working overtime to make it such that when it was allowed back out, it would be a controlled product, purposely hard for the average user to get information and help easy, without paying much money.

For a real end of a person's existence ... if one was to give A.I. the ability to push the button or control traffic or any scenario it was in control where if something small could go wrong resulting in death ... there's always the possibility for it to be abused.

Presently I think a case could easily be made that the present piss poor web search results available to the general public, are more of a threat to humanity.

Poorer education levels are more likely the bigger threat to humanity though.

> Poorer education levels are more likely the bigger threat to humanity though.

Already doing a great damage, especially in third world countries, where manipulation from authorities are superior due to lack of education.