The Depravity Economy

22:20Sam:I dropped a link to this website in our work Slack. And I haven't written a story about it yet. If anybody who is listening knows more about it, please get in touch with me. But there's a site called, I think it's Rubet. And a lot of people are streaming it. And it basically shows traffic cameras from across the world and overlays markers on the road on top of it and allows people to bet on how many cars are going to make it through this intersection in the next 15 minutes. And when I think about the depravity economy, what I think about is the point in mostly American capitalism, in our economy where it's like, increasingly, there is less real production and more financialization of everything.

23:21Sam:And that is just like all that is left to do. And it's the only way for people... It's the only way they think they can actually get rich is to find something that is like crypto, find something that is like the stock market, find something to bet on. Because the idea of actually working hard and getting a real job that is stable and retiring is absurd at this point. So people would rather put it all on black. And you follow that to its logical endpoint, and that's people betting on whether world leaders are going to get blown up or not. And that's... Nothing of value

23:59Emanuel:is being created here.

24:00Sam:Right. Exactly. So before we get off Iran, And Matthew also wrote a story with the headline, Amazon data centers on fire after Iranian missile strikes on Dubai. Matthew, how do we first learn that something happened to our precious AWS services in the Middle East? People complaining about not being able to access certain services on social media from Dubai. And also like all of these videos of as the Iranian strikes occurred in the United Arab Emirates in Bahrain, there was a lot of really incredible Instagram reels of influencers pointing at the sky, watching the missiles crash down. And then AWS has a health dashboard where they publish information about what's going on.

24:53Sam:And shortly after the missile strikes, I think the first we heard of it was on March 1st, they start posting updates about why people can't access certain services in the Middle East region. And a couple days ago, or yesterday when I first wrote this story, it was because they just said there had been damage to the centers. Unspecified damage. Unspecified damage. and I reached out to them and they were like, yes, something has happened and then they just referred me back to the dashboard. Yesterday at 7 Eastern Time, they finally kind of came out with it and were like, okay, it was Iranian drone strikes on two data centers in Dubai and another drone strike that landed near one in Bahrain and exploded, but wasn't specifically at that data center, knocked these data centers off.

25:51Sam:And it's a thing where like, who knows when those are going to come back up. Yeah. So we don't know if either of these data centers was targeted or not. Personally, this is completely me speculating, but it's hard to imagine at this point that they would be targeted. I feel like Iran would have far more high priority targets. in the region. But it is a very interesting development. I don't know if people know, but Matthew also has a podcast about conflict called Angry Planet. So the one question I wanted to ask you about this before we move on to the next segment is, do you think countries are looking at a missile strike on a data center?

26:45Sam:And will that actually... Does that change offensive slash defensive postures for countries around the world? Are we worried now about data centers being charged? Because you can imagine how... A hundred percent. ...that would destabilize countries or increasingly as militaries rely on AI and big data to do things. What was the big defense story in the week leading up to the attack on Iran, is Anthropic fighting with the DoD. And are they going to be declared a supply chain risk? Blah, blah, blah, blah, blah. Will all of like Claude and all of these other services live and die by these data centers that are being built across the entire planet?

27:34Sam:You know, it's hard to... It would rely on like human or signals intelligence being able to tell you like which part of a service is being serviced by which data center. But that's a surmountable problem. You can figure that out. And then, yeah, those are targets. If I can knock out the system that's going to help you do targeting or help you that's running missile defense, I'm going to do it. It'll be the first thing I attack. So yeah, absolutely. I think data centers are... And it causes this kind of civilian disruption like we're seeing here. like yeah those are those are targets for sure yeah uh i think you look at ukraine and you see like the energy grid constantly being attacked and it's not hard to imagine in the future how like in a few in a future conflict it's it's data centers because that would be equally disruptive potentially okay well and they're going to be there's going to be energy infrastructure built around all of these data centers too right yeah well that's where we're going to build nuclear power plants and all sorts of things.

28:36Okay.

28:37Sam:We'll leave that there. And then when we come back, we'll talk about a story I wrote about Wikipedia.

28:58Sam:This episode is sponsored by Better Health. March 8th is International Women's Day, and so today we are taking a beat to celebrate women and all that they carry, at work, in relationships, in families, and in the many roles that they play every day. Last year, my mom fell at home, and suddenly this woman who had supported the family for her entire life, she ended up needing a lot of her own support. And we rallied to her side, but we couldn't do it all. And part of what helped her heal was therapy, the kind of therapy you can get with better health. BetterHelp's therapists work according to a strict code of conduct and are fully licensed in the United States.

29:36Sam:And BetterHelp also has a therapist match commitment. It does the initial matching for you so you can focus on your therapy goals. And if you aren't happy with your initial match, switch to a different therapist at any time. Your emotional well-being matters. Find support and feel lighter in therapy. Sign up and get 10 % off at betterhelp.com slash 404media. That's B-E-T-T-E-R-H-E-L-P dot com slash 404media.

30:08Emanuel:If you've ever felt overwhelmed by the idea of learning a new language, you're not alone. Studies show that 70 to 90 % of people trying to learn a new language give up. In fact, I tried to learn Spanish a few times and I gave up the first few times. Then I traveled, felt like I had to learn it, and made a lot of progress. I've been able to both keep up my skills and learn new ways of learning with Babbel, which is built so that it's really easy to get started. Babbel understands that people learn differently. So you can just like dive into a quick podcast when you don't feel like a lesson, or you can speak out loud to get that practice in.

30:44Emanuel:You can even explore courses based on specific topics and create your own customized review lists, all within the app. What I like about Babbel is that it's built around real conversation. The lessons are short, about 10 minutes, so they fit into your day. But they're designed by over 200 language experts to teach you words and phrases you'll actually use. I've been using it to work on my Spanish, and I've noticed that I'm way more comfortable speaking out loud instead of just freezing up all the time. It breaks things down step by step, so you build confidence gradually. Whether that's ordering food, making small talk, or just understanding what someone says back to you.

31:19Emanuel:Learning a language with Babbel is really about taking small steps and progress you can actually feel. However you learn best, listening, speaking, reading, writing, Babbel adapts to your style and keeps you motivated with real-time feedback and progress tracking. They've sold over 25 million subscriptions worldwide, offer 14 languages, and every course comes with a 14-day money-back guarantee. Make fast, lasting progress with Babbel, the science-backed language learning app that actually works. Here's a special limited time deal for our listeners. Right now, get up to 60 % off your Babbel subscription at babbel.com forward slash 404.

31:57Emanuel:Get up to 60 % off at babbel.com forward slash 404. Spelled B-A-B-B-E-L dot com forward slash 404. Rules and restrictions may apply. This episode is sponsored by LED. If, and that is if, you find yourself using AI personally or for work, you might also find yourself wondering where all your AI conversations go. How are they stored? When are they actually deleted? As we enter into the Quick GPT era, it's a good time to check up on your AI privacy while searching for alternatives outside of the Silicon Valley surveillance machine. LED was built on the same encryption used by Proton and Signal. You get the power of flagship open source AI models hosted in Finland, all without the Tech Bro surveillance exposure.

32:45Emanuel:LED employs a multilayered approach to privacy from end-to-end encryption to German-owned and controlled data centers, all with zero exposure to the U.S. Cloud Act. That's true data sovereignty. Visit ellydee.ai today. It's free to try and their mission plan starts at only$5 a month. That's led.ai.

33:18Sam:And we're back. This is a story that'll be out by the time this podcast comes out, but is not yet published. So the tentative head right now is AI translations are adding hallucinations to Wikipedia articles. And this is a story by Emanuel. So Emanuel's been following this. the Wikipedia saga and Wikipedia beef is like a recurring um recurring beat for 404 and it's always a delight and never complicated but um you know the story is no different so um I mean why don't you just kind of set us up a little bit to begin with with some context on Wikipedia and how it's dealing with AI if that's something that you can get into like elevator pitch level explanation.

34:08Sam:Like what are they, what's their stance at this point? How are they kind of, how are they dealing with AI and the influx of generative AI in general that'll kind of set us up into this AI translations problem? I'll try to do it as briefly as possible because as you say, it is very complicated. I'm going to talk exclusively about the position of Wikipedia editors. Wikipedia volunteers as opposed to the Wikimedia Foundation because it's quite different. But what Wikipedia editors think actually matters more because those are the people who govern the articles you see on Wikipedia. Their position, I think, is extremely nuanced and sophisticated and good.

34:55Sam:And as I've said before, it should be a model for the rest of the Internet. But it is essentially that they are not making blanket rules about generative AI because they are open to the possibility of it improving, of it becoming better. They are generally like a very permissive community. But what they are doing is establishing, this is my wording, but I think it is fair, a kind of zero tolerance for AI -generated errors. So they are aware that there's a bunch of generative AI. There's a special task force of Wikipedia editors that goes around, tries to find generated content with errors on Wikipedia and fix it.

35:49Sam:and they are implementing special rules on how to edit and fix AI-generated content. So that's kind of the general outlook. It's no blanket rules against AI, but high, high vigilance about the errors that it produces because they're already seeing it. Yeah, that's fair. And if you're using AI and not checking it, that's a huge problem and a red flag for you as an editor. So this story focuses on and revolves around something called the Open Knowledge Association or OKA. Can you break down what the Open Knowledge Association is? And what role is that playing in this AI translation solution problem?

36:32Sam:What is the problem that we're seeing? So I think if you go to OKA's website and read what they do, it kind of looks bigger than it actually is. My understanding after talking to the founder is that it is basically a side project for a Google engineer who's based in Switzerland, I believe. and the idea is that he's a native French speaker. He thought that there were a lot of articles in French that are not available in English and there's a lot of articles in English that are not available in French and he's like, I would really love it if there were more translations of articles across Wikipedia generally, which I think is like a fine and noble goal.

37:25Sam:But the way that he wants OKA to treat this problem or this challenge is they pay translators, mostly people in the global south who don't get paid a lot of money, a few hundred dollars a month with the expectation of like 40 hours a week for that pay. and explicitly instructing them to use AI to aid them in the translation. Okay, gotcha. And you mentioned you talked to the founder. What did he have to say for himself in general? What was his stance on all of this? So I talked to him over email over the past couple of days. His name is Jonathan Zimmerman, by the way. And it was pretty interesting because I'm coming to him after wading through very long, detailed Wikipedia discussions.

38:37Sam:Something that I love about reporting on Wikipedia, there's two things. One, if you're talking to Wikipedia editors, they're the best sources in the world because all they do is write very concise, precise, informative articles. So you'll ask them a question. You'll ask them a question and they're like, here's the information in the right order summarized as efficiently as possible. so I was talking to an editor who was a source of mine on a few previous stories who first informed me about this and then the other thing that's great about reporting on Wikipedia is if it's a controversy among editors the way they govern Wikipedia is that they have all these conversations in the open in talk pages and essentially what has happened is that a few people started to know pretty glaring editors things like hallucinated citations this is something that is very common with the llms you'll ask an llm to generate a scientific paper it will generate something that reads a lot like a scientific paper with citations but then the citations don't exist it's like you'll click on a link to a scientific article and it goes nowhere because it just made up the url so there were errors like that there were factual errors there were formatting errors because people these okay editors or translators were copy-pasting articles to LLMs and then submitting them as drafts to Wikipedia.

40:13Sam:And the editors noticed a pattern that it was all these okay editors. Then they kind of investigated the company. And they found that not only were they instructed to use AI, not only Zimmerman disputes this, But to them, the quota seemed pretty high. There was a high churn of articles that you could imagine would result in more errors. at some point they claim they were specifically instructed to use grok to translate the articles which uh i think is especially a slap in the face to wikipedia editors because grok made uh this wikipedia competitor called grokipedia which is i mean i don't know if jason has a a more fair description of it but it's like it ingested all of wikipedia and then ai generated an alternative that kind of has a right-wing skew sometimes and then also has a bunch of errors because it's AI generated.

41:17Emanuel:I would just add that sometimes that is then filtered back into other places and sometimes back into Wikipedia, not for very long, but there's been a couple instances of that.

41:31Sam:Yeah, it's a real pollutant and that's actually the reason I wanted to write this because this is essentially what happened here. You have wikipedia translators using grok and then introducing grok errors back into wikipedia um sorry all that being said to get to your question i talked to zimmerman and um you know he his position is the position that i see from a lot of people who are uh generative AI believers. He thinks that generative AI is useful. It can make people work faster. It can make Wikipedia better. He's not alone in that. I've written before about Jimmy Wales, the founder of Wikipedia, also is not as optimistic, but he's at least open to the idea.

42:26Sam:And he says, yes, there are errors, but we have very strict rules about editors having to look and verify that the output is correct. And if there are any errors that got introduced to Wikipedia, he apologizes and says that they need to review and improve their process around that. But he did not say like, oh, my bad, maybe we shouldn't use AI in generating Wikipedia articles. Not only that, he also suggested that one new measure they're introducing is having an AI check the output of the AI translations, which, as we talked about very recently with something like AlphaSchool, is a, at this point, established, flawed method.

43:22Sam:You can't have an LLM that is prone to one type of errors check the output of the same LLM for the same errors because it's lacking the ability to catch its own mistakes. Yeah, always a great choice to send AI after AI. So they went back and forth, the editors, and they caught a bunch of these errors from OKA. many of them like you said were the the source citation errors and just going back and actually looking at the source material and saying oh that that book doesn't even talk about that page doesn't even talk about what this bookie period article is about so humans are having to go in and check the work of ai as usual so where did the editors land on this?

44:13Sam:What was the final call for them? Maybe not even final, maybe this evolves, but where's the call for them right now on whether or not they're going to let OK keep contributing? Or OK translators, I guess. Yeah, again, I think the response here from editors speaks to how reasonable, fair, and moderate the Wikipedia governance model is. There's no blanket rule against OKA, There's no blanket rule against AI. What they've done is they flagged a couple, or not a couple, a few OK translators who have made mistakes. And now they're subject to additional review as opposed to a regular translator or regular editor.

45:04Sam:And then the other thing is, I think it's a two strikes and you're out rule. So if an editor uses AI translation twice, and they're given a warning, but they find two drafts that have AI generated errors, they prevent the account from making contributions. So I think that's a very measured, responsible response. Zimmerman is not thrilled about that, but also said that he was like part of the discussion and is, you know, accepts the new restrictions. That's good, I guess. I also want to point out just AI in translation is a very tricky problem in general. language is very fluid, has lots of meanings in different contexts, something that AI can't understand in a lot of contexts.

46:02It doesn't have the human capacity for language, obviously.

46:07Sam:It's doing, you know, language prediction models and things like that. But a couple months ago, Galt actually published an article about HarperCollins using AI to translate Harlequin novels, which I think is a very interesting use for translation considering Harlequin romance novels are this very formulaic, prolific genre of book that gets churned out en masse. No offense to Harlequin romance novels and the authors and their work, but it's based on being able to pump through these books very quickly. And they're globally popular. So HarperCollins said, you know what, we're going to start using AI in France to eliminate human contractors who are translators who did the work before, which is awful, but is also relying on humans to check the work of the AI in addition to this.

47:17Sam:Did this story bring up memories of that story? Because there's a really good comment section happening on our site with the HarperCollins blog and people debating whether this is a good use of AI or not. I haven't seen that debate. No, it's not such a debate. It's like thoughtful comments being made. Not so much a fight. It's super interesting. It did bring up those memories for me. It also, you know, and to even further it, like, we saw, I assume, the New York Times profile with the romance authors that are using AI to generate, like, 40 books a year. Which I thought, you know, which I thought was also pretty interesting.

48:03Sam:Especially since, like, AI can do, like, Chai TBD is going to do erotica. There are so many, like, LMs that do erotic writing now. Makes sense. And those women that the New York Times profiled also have a side hustle teaching other people how to do it. Oh, hell yeah. We're doing Pyramid, maybe. Yeah. So it's like their whole scheme is, look, I'm generating one romance novel in 45 minutes. And on top of that, I am teaching you how to do it too. Selling a class. Selling a class. But I do wonder like, what the bottom is for that. How many people are actually reading those books? What the bottom is for it?

48:49Sam:What the market is going to be like in the future when specifically on art, maybe people are more interested in having something that's explicitly labeled as this was written entirely by a human being. but I think that like romance is an especially interesting place to watch because it is so business forward and formulaic that I think that it's one of those places where any of the advances and the fights are going to happen there first right and then like all of that will kind of trickle down to the rest of the genres

49:27Emanuel:I feel like the market for those books is like can you trick people into buying them on accident Yep. Because if you've paid$1.50 or whatever it is, there's a good chance you're not going to fight Amazon to get it back.

49:43Sam:Or Amazon has the Kindle Direct program where somebody pays a sub and then they can read as many of these as they want. And if they get a couple pages into one and see an AI artifact, they're just going to move on to the next one. And it's maybe already counted as a view for the purposes of whoever's turning out the slop. And I kind of have this wonder in spaces like that if we're not going to get to a place where people abandon spaces that are just kind of filled up with AI, and if that may be one of them. But I'm wondering if that'll eventually be a reaction. have either of you ever used ai translation in your regular life like either talking to an app and having a translate or just like translating some text have you ever done that well i mean that's what google translate is right right but i bring that up because google translate actually i think is ai now but i've used it a couple of times and it is actually like shockingly good um much better than the stuff you'd usually get from Google Translate.

50:49Sam:I used it twice. One is, I think I was translating something for my mom from Hebrew to English. And I put it into Chachapiti. I translated it. I read the output. I'm fluent in Hebrew, so I could check it and see that it was good. But it worked very well. And she ended up using it. And then I was trying to explain something to my neighbors who only speak Spanish. And I did the same thing and asked Jason to look over the translation and see if it was okay. And he actually made some tweaks to it. They weren't mistakes, but the language was very formal or something like that. It was like, nobody would actually talk like this.

51:37Sam:So he made some changes. The point I'm making, in both cases, LLMs, not surprisingly, extremely good at translating from one language to another can be a very useful tool. Crucially, in both cases, I was able to check it with somebody who actually speaks the language, which is what you should be doing. And then it would theoretically be fine.

52:00Emanuel:I mean, I use it all the time. like all the time. This is probably the most common use because one, it's like YouTube has an auto translate feature now. So in addition to auto generated captions and transcripts, which are really useful for our reporting, just in terms of finding things in really long videos, I can now watch an hour long video in Hindi and have it auto translate subtitles to English. and I can tell that it's not that good. You can definitely just tell it's not that good. I don't speak any Hindi, but the English that it translates to is very broken often or sometimes it just can't do it.

52:41Emanuel:But for information gathering purposes, it's incredibly useful. And that's how I did almost all of the AI slop reporting just in terms of finding out relevant portions of different videos based on the auto-translate. And then we got people who actually speak Hindi to do the direct translations for us. But that's an example. I mean, obviously, things like translating websites and news articles, using the built-in tools as well. You can tell if you're translating a German article to English, using the... Translate this entire website. Sometimes things are broken. but it is useful to get the gist of what's going on.

53:29Emanuel:And then for your own information gathering purposes, you don't necessarily need it to be 100 % accurate. But for something like Wikipedia, where you're directly translating and then publishing, that's not the case. You do need it to be accurate.

53:47Sam:Emmanuel, what did you translate for your mom, you said? I believe it was an artist statement that she previously wrote and needed in English or something like that, yeah. I think that the use case, right, is huge because we're talking about for information-gathering purposes and being able to have someone check it afterwards. And I think in the instance of romance novels or art, and this is a huge thing. We haven't really covered, but I've kind of watched that it's playing out in anime and the translations of JRPGs. Don't laugh. You're making that face. But it's a big deal. A lot of these companies are moving to auto-translated and people are upset.

54:32Because there's something like when you're looking at artistic expression,

54:40Sam:a translation is not just about a one-to-one analog. There is nuance in the language. and a literal translation may not get across the same meaning in a piece of art. And I think that the LLMs have trouble with that. The AI has upgraded that. Have you ever read two translations of the same book, two different translations of the same book? Yeah, it's fascinating, right? No, it is. No, it's good. I love that we're very cultured like that. You're having that fight right now with the... Is it the Iliad or the Odyssey? there's a new translation.

55:22Emanuel:Guys, guys, we gotta go. We gotta go on to that. It's like, yeah. Dude, that new translation of Madame Bovary

55:29Sam:reads a novel, Jason. You gotta check it out. Okay, let's leave that there. And when we come back to our subscribers only segment, we will talk about Amazon wishlists.

55:44Emanuel:Thank you for listening to 404 Media Podcast. I'm Jason Kevler. As a reminder, you can subscribe to 404media at 404media.co. If you liked this episode, please leave us a five-star review. If you did not like this episode, listen to a different one where we have a host who has a real script. Please tell your friends about us. This episode was produced by Kaleidoscope by Alyssa Midcalf. We will be back in a few days with a new episode. The end. that's the end.

56:19Sam:Perfection. Beautiful.

56:47Sam:Here's another big number, 100 million. That's how many data points LifeLock monitors every second. If your identity is stolen, we'll fix it guaranteed. One last big number. Save up to 40 % your first year. Visit LifeLock.com slash podcast for the threats you can't control. Terms apply.