AI Risk, as Seen on Snapchat
The news tiles on Snapchat are utter drivel. Looking at my phone, today’s headlines are: “Zendaya Pregnancy Rumors!”, “Why Hollywood Won’t Cast Lucas Cruikshank”, and “The Most Dangerous Hood in Puerto Rico”. Essentially, Snapchat news is the Gen-Z equivalent of the tabloid section at a Walmart checkout aisle.
Which is why I was so surprised to hear it tell me the arguments for AI risk.
The story isn’t exactly epistemically rigorous. However, it reiterates points I’ve heard on this site, including recursive self-improvement, FOOMing, and the hidden complexity of wishes. Here’s an excerpt of the story, transcribed by me:
Given enough time and the ability to self-generate improved versions of itself, it wouldn’t take long for a fully autonomous general AI to achieve superintelligence, a level of cognitive processing power so many times stronger than our own that its abilities would appear godlike to us, and the pace at which it could improve would be lightning fast.
Superintelligent AI could think through problems in a few hours equivalent to what the world’s smartest people could do in a thousand years. How could we possibly control or out-think such a powerful intellect? [...]
Being the first superintelligence, the AI could see any competition as a threat, and could choose to remove [all humans]. [...]
When in the possession of such superintelligence, scientists may pose it questions, like how to solve previously-impossible mathematical equations. To solve the problem, the AI may choose to forcibly convert all matter on Earth into a supercomputer to handle the processing of the equation. It could solve the problem while killing the species that asked the question.
We could also make the mistake of giving the AI too vague a problem to solve, such as “end human suffering”. It could then decide the best way to do this is to eliminate all humans, and therefore end their suffering.
Why does this matter? The channel, Future Now, has 110,000 subscribers(!), and the story was likely seen by many others. This is the first time I’ve ever seen anyone seriously attempt to explain AI risk to the masses.
The fact that Future Now ran the story implies that the masses are, in fact, capable of understanding the arguments for AI risk, so long as those arguments come from someone who sounds vaguely like an authority figure.
I don’t think people here realize that this is an incredible opportunity, since the arguments are now going to be field tested with tons of people. It was written pretty efficiently here, too.
But mainly, it’s horrifically bad news, because AI safety is now in the hands of extremely nefarious shadow people using it at large scale for god-knows-what. Left-Right Polarization is the main bad outcome, but it is by no means the only one. There’s Russian bots out there, and they’re as bad as any nihilistic “for the lulz” hacker from that one trope.
Hopefully it peters out without doing too much damage, e.g. setting attitudes in stone in ways that will last decades.
The public will end up with a somewhat garbled and politicized picture. Which is still more accurate than their current picture based on some hollywood scifi.
Miri will carry on being Miri. The russian botnet creators aren’t smart enough to make an actual cutting edge AI, and aren’t as good at creating compelling nonsense as hollywood. Most people have a kind of ok perception of climate change despite the botnets and oil companies and green nuts.
The masses are capable of understanding AI risk even without that. It’s really not hard to understand; the basic premise is the subject of dozens of movies and books made by people way dumber than Eliezer Yudkowsky. If you went to any twitter thread after the LaMDA sentience story broke, you could see half of the people in the comments half-joking about how this is just like The Terminator and they need to shut it off right now.
Maybe they’re using the “wrong” arguments, and they certainly don’t have a model of the problem detailed enough to really deal with it, but a sizable amount of people have at least some kind of model (maybe a xenophobia heuristic?) that lets them come to the right conclusion anyways. They just never really think to do anything about it because 99% of the public believes artificial intelligence is 80 years away.
I’ve seen a lot of countersentiment to the idea of AI Safety, though:
(I have a collection of ~20 of these, which I’ll probably make a top-level post.)
Totally agree; but the very basics (if we make something smarter than us and don’t give it the right objective it’ll kill us), are parse-able to what seems like a large fraction of the general population. I’m not saying that they wouldn’t change their minds if their favorite politician gave a “debunking video”, however the seeds are at least there.
I was surprised by this tweet and so I looked it up. I read a bit further and ran into this; I guess I’m kind of surprised to see a concern as fundamental as alignment, whether or not you agree it is an major issue, be so… is polarizing the right word? Is this an issue we can expect to see grow as AI safety (hopefully) becomes more mainstream? “LW extended cinematic universe” culture getting an increasingly bad reputation seems like it would be extremely devastating for alignment goals in general.
Reputation is a vector not a scaler. A certain subsection of the internet produces snarky drivel. This includes creationists creating starky drivel against evolution, and probably some evolutionists creating snarky drivel against creationists.
Why are they producing snarky drivel about AI now? Because the ideas have finally trickled down to them.
Meanwhile, the more rational people ignore the snarky drivel.
Cool, but the less rational people’s opinions are influential, so it’s important to mitigate their effect.