Inside the Hunt for the Discord Leaker + Twitter Chaos Updates

30:23Okay, what happened? So what happened was, as I said, these posts to your circles are supposed to be limited to a group of people that you've put on an allow list, okay? But then some poor souls browsing the Twitter feed started to see posts on the For You page. The For You page is the algorithmically ranked feed of posts that Twitter might think you would enjoy for any number of reasons. Right? It's ranked, but it could potentially be seen by anyone. And so essentially, you had private posts that were being not just made public, but recommended to the public. So the best analog I can offer here is imagine that you posted something really heartfelt or embarrassing or vulnerable on your close friend's Instagram story, and then a stranger opened up the explore page on Instagram and saw it promoted at the top of their feed.

31:14That's what just happened on Twitter circles. And in a way, it's like funny to me that this hasn't gotten more discussion because this is a truly catastrophic privacy breach. Right. I mean, it's sort of like if your DMs were just showing up on people's Twitter timelines. It's like that level of a privacy scandal. Yeah, absolutely. You know, Elon did an interview this week where he boasted that a lot of folks said Twitter would be dead by now. But it's not dead because like his standard for his Twitter alive, I think, is like, can you still log into the website? But this is a great example of how when you get rid of a huge number of people that worked on your technical infrastructure, weird stuff is just going to start breaking in ways that can have quite a terrible effect on your user base.

31:57Yeah, I mean, I've been wondering whether there will be some kind of mass DM leak that stems from all these engineers getting laid off. That hasn't happened, but this Twitter circles thing does seem to be pretty close. Like, these are supposed to be private, but all of a sudden started showing up on people's for you feeds. Oh, yeah. And when the layoff started, Twitter engineers started messaging me saying, like, don't use DMs anymore. Like, they're just not a safe product. So I don't use Twitter DMs anymore for anything remotely sensitive. But, yeah, watch out, man. Wow. So do you have any idea of, like, what went on behind the scenes to make these Twitter circles posts start showing up for people they weren't supposed to be seen by?

32:35To my knowledge, they have said nothing about it, which does not exactly inspire confidence. So put your spicy tweets somewhere else because they are not safe in Twitter circles. OnlyFans is now accepting spicy private subscription only posts. When are you starting your OnlyFans? Well, we'll see how subset goes. Okay, that brings me to my third kind of catch-all category for what I am calling clown shit. This would include things that Elon Musk has done in the past couple of weeks. that just seemed like he's just kind of playing for laughs or not behaving in a serious way. He briefly changed the Twitter logo to the Doge image, the Shiba Inu dog that is also the face of Dogecoin.

33:21That lasted for a few days. It was not on April Fool's Day. No, I believe it launched on April 3rd, and everyone was joking like Twitter tried to do an April Fool's joke, but it took like an extra two days to launch. Which seems plausible to me. Yeah, totally. He also painted the sign on the Twitter building in downtown San Francisco, or he had it painted, so that it just now reads Titter. I truly don't believe that even an eighth grader would find this stuff funny. I can't imagine anyone is observing any of this and thinking this is so funny and cool. Right. This is the man that I want to entrust my company's advertising budget to.

34:02He also changed his display name on Twitter for a day or two recently to Harry Balls. Harry spelled H-A-R-R-Y and Balls spelled B-O-L-Z. This man is not well. Can we just say, the man is not well. Yeah, what is going on? Look, I think that driving a social network can actually drive you crazy. It's very difficult. You're running a platform that is hosting this huge portion of the world's speech. You're getting criticism from absolutely every corner. And that's hard enough, even if you are a very good CEO with kind of a good plan for where you're going. I think if you come into that with no background in the subject matter, and you've committed $44 billion, and you are losing money every month, and none of your ideas are working.

34:56Oh, and also you're working on a handful of other giant companies that have their own problems. That is just not a recipe for success, right? And so many people when Elon took over said, well, look, this guy's going to be different, okay? He's special. He builds rockets. A social network? Yeah, I think he can handle that. And then you fast forward to April 2023. Guess what, honey? It didn't work. He's like, give me the rockets back. That's right. Right. This isn't rocket science. It's actually harder. It is harder. But is there one person who wouldn't be relieved if Elon Musk says, I am stepping away from Twitter and I'm just going to work on rockets full time?

35:27Like, everyone would be overjoyed. I think even his fans would be overjoyed. Completely. Because, you know, he's been getting blowback from some people who previously have supported him. People like Paul Graham, who is a big Silicon Valley investor. He's a co-founder of Y Combinator. And he's been criticizing Elon recently for basically just taking his eye off the ball. You know, like, wasting his time running Twitter instead of working on SpaceX or Tesla. So even the people who are sort of predisposed to support him are not fans of what he's doing at Twitter. I want to ask you about one more thing that I would put in the clown shit category, which is what is going on between Twitter and Substack?

36:05Well, so, you know, ethics disclosure, I publish my newsletter platformer on Substack, and they've helped me in various ways over the years. And I have a note on my website about that. But Substack decided in the midst of all this chaos to— And Substack, for the people who may not know what it is, it is a newsletter platform. You can subscribe to newsletters, either free or paid. Substack is just kind of the infrastructure that powers that and allows people to pay you to read what you have to write. Yeah, that's right. And they saw everything going on at Twitter, and they said, well, gee, it seems like this thing is having some trouble.

36:41Maybe we could build something similar of our own. And so they built this product called Notes. It launched this week. It lets you post text and links and pictures, and you can reshare those. They call that restacking. And when Elon got word of this, he really overreacted. And for a time, he blocked users on Twitter from engaging with tweets that had links to sub stacks in them. If you click the link, it would say this website is spammy or dangerous. Which is true in the instance of platformer. Platformer is a spammy and dangerous newsletter. But for other Substack newsletters. It's better. Unfair.

37:25Yeah, it's unfair. So they marked Substack links as, quote, potentially spammy or unsafe. They also blocked searches for Substack. So if you search the word Substack, it would just show you results for the word newsletter. So it's like, oh, you're looking for a Substack? Well, we're not going to show you that, but we will show you newsletters. So when he got asked about this, Musk said that Substack was, quote, trying to download a massive portion of the Twitter database to bootstrap their Twitter clone. And I went over to Hacker News to see what they were saying about this comment, which I found completely inscrutable.

37:59And the Hacker News people, this is a community of engineers and some of the most delightful nerds you'll ever meet. And the Hacker News people were just laughing their heads off because they were like, what does this even mean? It's like, you can't download it. It just didn't make any sense whatsoever. And so, you know, as usual with him, there's not a relationship with the truth there. It's just kind of like, you know, spewing words, and it was all nonsense. I mean, what I found most interesting about this beef between Elon Musk and Substack is that one of the things that animated him when he took over Twitter was that he thought that the previous management of Twitter was sort of throttling or shadow banning sources of news that they didn't like, right?

38:37That's right. He was very upset about what happened with Hunter Biden and that story during the 2020 election where Twitter, you know, blocked people from sharing that New York Post story about Hunter Biden. Right. And then he Hunter Bidened Substack. He made it so that you could not share Substack links or comment on them or like them or search for them. He disappeared Substack from the platform. And Substack is where a lot of the people that he is ideologically allied with, these sort of Twitter files, journalists and other opponents and skeptics of the mainstream media, are going. He was basically cutting off his friends and allies from one of their main sources of distribution.

39:17It truly is the funniest possible outcome to this story. And I just think the more serious point is this is not a principled man. This is someone who talked about nothing except his speech principles and his desire for bipartisan dialogue and let's move to the center. And he spoke about it in the most grandiose terms. And the minute he felt the slightest bit of discomfort, all those principles went out the window and he said, I am going to punish my enemies. And look, some people are very upset about this. I can't get upset about it because every day people learn what I truly have been saying all along, which is like the emperor has no clothes here.

40:01And you can just go on Twitter every day now and see it for yourself. But one of the other things that you have said is that this sort of chaos at Twitter is going to result in people moving. Yeah. So far, that hasn't really happened. There has been no mass exodus of Twitter users to another platform. You know, you're very active on Mastodon now. I think it's safe to say, like, that has not really threatened Twitter in a massive way. Now people on Twitter are talking about Blue Sky, which is this sort of decentralized alternative to Twitter that was started by Jack Dorsey. But none of these alternatives to Twitter have really gained momentum.

40:35So what is it going to take for something to pop up and replace Twitter? Or is it just kind of like the entire concept of Twitter in our media ecosystem sort of dies off and is replaced by, I don't know, group chats and Instagram accounts and whatever else? So there's two answers to that. There's the product answer and there's the NPR answer. I want to talk about the product answer first, right? So you just named some products that compete with Twitter in various ways. And you're right. None of them has supplanted Twitter in the popular imagination. But again, let's remember, Elon has only owned this website since October.

41:10We are six months out from when he took over. We first started talking about this on Hard Fork. The reason that these products are now being built is because there is an appetite for them. They are growing. And frankly, there are just now so many of them that even if one of them hasn't, like, entered that period of hockey stick growth quite yet, I do think it is probably going to happen for someone. So you're going to have to give it some time. Someone is probably going to have to invent some fun new little mechanic. There's going to have to be some weird, silly new toy that just kind of captures our imagination.

41:46But I do think it's going to happen. So first answer is these products are still getting built, but they are getting built. And that was not true before Elon Musk took over, right? Nobody smelled blood in the water until the past few months. At the same time, Kevin, some important people actually are starting to leave. And I want to deliver one of our signature hard fork tributes this week to a little media organization called National Public Radio. You know these guys? Yeah, NPR. NPR, they were just minding their own business. They were just tweeting links to their stories. And then one day they wake up and they've been slapped with this label, state-affiliated media.

42:23And state-affiliated media was a label that the old Twitter created to identify essentially propaganda. Right. It was like a little label that went under the profile of essentially foreign propaganda outlets just to let you know that the tweets that you were seeing from this account have state affiliation. Yeah, it's just trying to give you a little bit more information so you know where the account is coming from. They allowed them to post, but they wanted you to know that. Well, all of a sudden NPR, which does receive some minor amount of government funding, but is completely editorially independent of the United States government, they now have this label as well.

43:02So it effectively said, this is propaganda. That was the message that I believe the people who added this label to NPR's account wanted you to receive and messages are exchanged. And eventually the state affiliated media label gets downgraded to the government-funded media label, even though it gets less than 1 % of its funding from the corporation for public broadcasting. So the entire point of this was not to serve users of Twitter in any way. It was just to slag NPR and to try to make it look less credible. Right, it was a dunk. It was a dunk. And on Wednesday, Steve Inskeep tweeted a picture of John Lansing, who is the CEO of NPR at a morning edition meeting, morning edition, one of the flagship shows of NPR, and said that NPR will de-emphasize Twitter.

43:56He said that aside from the misleading label, Twitter isn't used by most Americans, drives little traffic to NPR, and no longer has the public service relevance that it once had. And to that, I just want to say, slay. Well said. This is what I have been saying, and pretty soon a lot more people are going to get it too. You think media organizations are going to leave Twitter? This is what I have been waiting for, is for somebody to wake up and say, I do not want to come to a platform every single day and get punched in the face. You know, we know what I've been thinking about lately, Kevin. Do you remember during the Trump campaigns when there would be these rallies and in the center of the rally, there would be a pit for the media and a signature moment of every rally would be Trump pointing to the people in the media so that everyone at the rally could say, boo, we don't like the press.

44:44that is what Twitter has become. It is the press pit where a bunch of people are standing around you in a circle, jeering. Adding this state-funded media badge was one of those steps. But I'm barely joking when I say that I think eventually every reporter who is still on the service will have a clown badge next to their username. And you just have to decide if you still want to be there when that happens. So you are still on Twitter. Well, in the way that I started with. So in December, I was like, I got to change the way that I'm using this website. And so I basically started using it as an RSS feed.

45:16Whenever I have a sort of Twitter-like thought, I put it on Mastodon. And when I have some new piece of journalism to share, a newsletter, a podcast, I put it on Mastodon, and then I repost it to Twitter. I'm not in there having fun. I'm really not replying to people. I'm reading it because it feels important to know what is being said on Twitter, and it's important to know how Twitter is falling apart. Social networks are my beat. I have to pay close attention to it. But I am not really doing free labor for Elon Musk anymore. Hmm. Fascinating. But you still are. Let's talk about that. Well, I want the clown badge.

45:49That would be fun for me. Okay, so Twitter is in clown town right now. But next, I want to talk about a different town. One that is populated entirely by AI. Uh-oh. We'll be right back.

46:15Thank you.

47:00Casey, did you watch Westworld, the HBO show? Yeah, I watched the first few seasons, but I gave up as it became increasingly incomprehensible. Okay, so I also watched like one season and I think it fell off because I just didn't get it anymore. which I attributed to me being dumb but I think maybe the show just also got worse well no here's how it became one of those shows where it was like at first it was very exciting to learn that that person you thought was a human was actually a robot but then that was just the entire show it was like well guess what other human is a robot every member of the cast right so for people who haven't watched Westworld the premise is there's this like theme park where there are humanoid robots that you can go and interact with and they'll talk to you perfectly lifelike Perfectly lifelike.

47:41But when the guests leave, the robots' memories all reset. And so the next time you go, you can just kind of have like brand new interactions with them. Except what Westworld asks is, what if they didn't actually forget? Exactly. So I want to talk about this next story because I think it is actually moving us in quite an interesting direction that is a lot like Westworld. So a paper came out last week by researchers from Stanford and Google, and the title of the paper is called Generative Agents, Interactive Simulacra of Human Behavior. Another one of these catchy clickbait titles from the academics.

48:19And I want to talk about this paper because it's really fascinating, and I think it actually points to the next phase of what's going to happen in AI. You know, we've been talking a lot on the show about these large language models, these chatbots, things like ChatGPT, and how you can actually, like, use them as a kind of persona. Like, you can say, you are my tutor, or, you know, you are my speechwriter, or you are my editor, or you are my friend. Like, talk to me as if you are that thing. And what these researchers wondered is, could you create a city, like a virtual simulated environment, and fill it with AI-generated characters?

48:58A simulated city or SimCity, if you will? Right. Basically, could you make The Sims, but The Sims aren't coming up with things on the fly? So The Sims is this, you know, very popular computer game. But in The Sims, like all of the interactions between characters are scripted and programmed. So what these researchers were wondering is, could you do The Sims with no scripting? If you just took a set of AI-generated personas, which they're calling generative agents, and placed them into a virtual environment, like a city, and just pressed play, what would they do? I have to say, very interesting question.

49:40So the way this works, they created 25 characters. they put them in a virtual environment that they called Smallville, which is kind of like a kind of Animal Crossing style video game. And they gave all 25 of these characters short personality descriptions. So one example was a character named John Lynn. And the thing that the researchers told the AI, and in this case, they did this using the ChatGPT API. So all this is happening with the help of ChatGPT. So the thing that they put in as the sort of description of this character was, John Lynn is a pharmacy shopkeeper at the Willow Market and Pharmacy who loves to help people.

50:21He is always looking for ways to make the process of getting medication easier for his customers. John Lynn is living with his wife, May Lynn, who is a college professor, and son, Eddie Lynn, who is a student studying music theory. John Lynn loves his family very much. Sounds very sweet. So they build these sort of like pre-programmed personalities for these characters. And then they set them loose. They just say like, you know, we're going to observe you for two days. And we're just going to see what these characters will do. And within the context of this game, like are there objectives? Are these characters moving throughout the world?

50:59Are they just sort of bumping into each other on the street and striking up conversations? Or what is happening? Yeah, so the way it works, the researchers who are controlling this experiment can sort of prompt the characters. They can say, you are going to the store to get some food. Then the language model sort of fills in what happens next. Or you're a character, you're bumping into another character in the park, and you're going to have a conversation. What does that conversation say? And then the language model sort of fills in the rest of the conversation and any subsequent actions. Got it.

51:31So I can actually show you what this looked like. Let's see this. I'll share my screen here. All right. So this is a sort of replay of the Smallville. So you can see it's sort of like a blocky, like pixelated map. Yep. With some various rooms. And then you can scroll down here, and here is a list of characters, 25 characters. I mean, they're these sort of like very cute little sprites. Right. And so you can click on any of the characters and see a log of all of their actions. So here's a character named Ryan Park, and Ryan Park's current action is sleeping, and he is in his apartment in the bedroom.

52:12And that's goals. Cheers to you, Ryan Park. And here is another character named Sam Moore. Sam is taking a walk around Johnson Park. He's stopping to take pictures. Oh, and he's having a conversation with another character named Adam Smith. And all of this is being generated by the language model. And along with these characters, the researchers actually developed something really interesting that is sort of like a memory, like a kind of virtual memory for these characters. Because one of the problems with language models, one of the limitations, is that they don't remember things. So you can have a conversation with ChatGPT.

52:51You can tell it some things. And then if you come back a different day and start a new thread, like, it's not going to remember what you talked about. You have to kind of re-instruct it to be a therapist or a teacher or a friend or a negotiating counterparty or whatever. So they sort of get around this with something that they call the memory stream, which is basically a natural language list, like a text list of what a given character has been doing and experiencing. Any conversation they've had gets sort of stored in this memory file. And then periodically, these memories get kind of synthesized into inferences and reflections, sort of summarized.

53:31So you basically have a system where each of these 25 characters can refer back to a file that contains sort of a synthesis and a summary of everything that has happened to them in this simulation. And they can refer to that in different contexts. So, you know, if one character is walking through the park and sees two other characters talking, they can, like, infer that those people must be friends and must know and like each other, and then they can use that information later on in the simulation. So it's sort of a neat technical trick that simulates a kind of memory. Okay, so these researchers, they let this experiment run for two days, and they just observe what these characters do prompted by this large language model.

54:13And what they found was really interesting. So they did a sort of control group experiment where after this two-day period was up, the researchers asked the generative agents a series of questions to kind of assess their self-knowledge, their ability to make plans and remember things. They also gave that test to a group of humans, 25 humans, who had access to all of the same information about these characters. And they basically gave them a replay of what the characters had been doing for these two days and all of their memories. And so the humans answered those questions, and then the researchers had these other humans, these evaluators, go in and rate how believable they thought those answers were.

54:53And the generative agents were evaluated as being more believable than the humans. So that's bad news for humans, so go on. So that was one finding. The other very strange finding that they had was that there were these emergent behaviors that they did not expect and did not intend to happen between these generative agents. Okay, if you tell me that one of these things started asking if it was part of a simulation, I'm going to need a drink. No, they started leaking military documents. So one of the emergent behaviors was that these generative agents gossip. What? But they shared information with each other without prompting.

55:33You're telling me that these Sims are spilling the tea on each other? These Sims are spilling the tea. So in one interaction, the researchers told one character, this character named Sam Moore, that he was going to run for mayor of Smallville. Sam Moore then shares this information with the rest of the town and starts a mayoral campaign. And the other agents start talking about this without, like, interacting with Sam Moore. They're just like, they have learned this information. I mean, this is admittedly big news in a town where, for the most part, people are just sleeping and talking on a park bench.

56:09Like, I could see why this would sort of be the buzz in Smallville. So, but not only are they talking about the fact that Sam Moore is running for mayor, but they're sharing opinions of him. So one agent at one point says, I'm still weighing my options, but I've been discussing the election with Sam Moore. What are your thoughts on him? And another agent says, to be honest, I don't like Sam Moore. I think he's out of touch with the community and doesn't have our best interests at heart. Wow, go off, King. And not only are they sort of gossiping about each other, but they're also coordinating and planning together.

56:41So in one case, the researchers instructed an agent named Isabella to throw a Valentine's Day party. That was the instruction that they gave this character. And this character, Isabella, starts spreading the word. And by the end of the simulation, 12 characters know about this party. And how many of you do you think showed up? I mean, in the real world, a lot of people would flake. But I'm hoping that these AI agents are more respectful. So I want to hope that all 12 showed up. Well, you are wrong because these are believable entities. And much like humans, many of them flaked. Seven characters flaked on the party.

57:24Three of them said they had other plans and the other four just didn't show. It's like setting up a grinder day. This is horrible. But five characters actually did show up at this party. And this was not something that the researchers had predicted. So why are we talking about this? You may be asking. Well, I assume that each week you show up and try to terrify me with some new glimpse at our coming dystopia. Is there another reason? So I do think this has dystopian implications, but I think the most obvious application for this is something like gaming. Absolutely, yes. So in games, a lot of them, especially role-playing games or open-world games, there are these things called NPCs or non-playable characters.

58:05And they're the kind of characters that you come across and they tell you some line or give you some hint. How would you describe an NPC? Yeah, an NPC is essentially an object in the game that exists to give you an objective or to advance you along a quest, but they are scripted to make you feel like you are interacting with something sentient. And of course, you know it's not. But the question that this poses is, well, what if word got around about what you were doing in this virtual world? What if that character could ask you about that or maybe change their behavior? This sort of thing does already happen, but all of those interactions are scripted, right?

58:48What you're sort of suggesting here is that we might not have to script these things in the future. They can just happen. Right. They could just act as the language model tells them to act. So one example that was sort of a joke, but I think someone said in response to this paper that you could be playing a game now and you could kill an NPC in the game and you could return back to that same spot later and all the other NPCs could be holding a funeral. And like that's not something that the game designers would have to necessarily script. I did talk to the lead author of this paper, who's a Stanford researcher named Joonsung Park.

59:22And I sort of said, like, well, beyond games, like, how could this be useful? And he said that one of the ways that this could be used was to simulate behaviors in online groups. So, for example, if you were going to start a new subreddit devoted to, like, celebrity gossip, And you wanted to see, like, okay, how would a set of rules given to users of this subreddit, you know, encourage them to react? Would they become rebellious? Would the thing sort of devolve into a toxic cesspool? Like, how could you basically simulate that without actually having to start the subreddit? Yeah. And this experiment suggests that you could actually just start a simulated subreddit and make some simulated users and see what happens.

1:00:08Totally. You know, another thought that comes to mind is I think we both believe that AI companions are going to be a really big deal. There's going to be some kind of virtual agent that you're talking to, whether it's on your phone or your desktop. And in order for that to work, they need to have a memory. They need to remember what kind of day you were having yesterday. They need to know the names of your friends and family and other people who are important to you. What I think is interesting about this is it suggests what if you actually have multiple different companions for different purposes, right?

1:00:36You're going to have your personal trainer and your therapist. And of course, at first, none of these are going to be as good as humans, but they probably all get better over time. And what if they talk to each other and share information? What if you have a virtual doctor who remembers your entire health history and doesn't make you write down your birthday every time you go to the office, right? So I think there is a huge amount of potential here, and that is a very fun glimpse into somebody already running the experiment. Well, so yeah, so this is the experiment that got a lot of attention this week.

1:01:05But there's also this sort of related thing that is happening in the AI world right now, which is that a lot of people are trying to make these large language models into agents. So turning them from chatbots into things that can act autonomously without human intervention. We talked a fair bit about, like, the need for AI vocabulary. And I'm telling you, agent is just one of those words that you are going to start hearing absolutely all the time. Yes. So in the last couple of weeks, there have been a number of projects that have tried to basically cut out the human. So right now, if you go to use ChatGPT, like you prompt it, you type something in, it types something back, you type something in, it types something back.

1:01:43It's like a kind of ping pong. And what these experiments with AI agents are asking is like, what if the AI model could essentially talk to itself and come up with plans and execute those plans and then generate AI agents to carry out different parts of this task? Yeah, we are really just determined to wipe out humanity, are we? Well, some people literally are. So right now there's a lot of research and energy in the AI world around turning these large language models into what are called agents. One of them is called AutoGPT, and AutoGPT is sort of a tool that someone hacked together that allows you to, instead of just like typing to a chatbot and it types back, you actually give it a task.

1:02:25and the AI can then sort of talk to itself and kind of recursively execute that task or try to execute that task. So you could tell, theoretically, this auto GPT, like start a drop shipping business. And you could then sort of talk to itself and say like, okay, for that, I'm going to need someone who's an expert in shipping. I'm going to need someone who can talk to the big e-commerce companies. I'm going to need someone who can set up, you know, maybe a Stripe account. I'm going to need someone to code up a website. And also I'm going to need someone who can go on Twitter and advertise this business.

1:03:01Now, just to be clear, there's no evidence yet that AutoGPT can actually do this stuff well. It's still not very good. And there are still some pretty big technical hurdles standing in the way of these agents being able to actually do things autonomously. But that's the direction that some people are trying to push these AI systems. Yeah, it's like when you're fighting a boss in a video game and they're like, and you think you're finally whittling them down and then they just summon a bunch of other demons to eat your face. That's what we're building. Right. And the examples of this auto-GPT stuff that have been floating around are mostly just kind of people trying to do these business tasks.

1:03:36But there is one project called Chaos GPT. Oh no. Where someone basically used auto-GPT and gave it the explicit instruction to try to destroy humanity. Well, the world is still intact as we record. So what happened? So yeah, this has not worked, as you can see by the fact that we still exist. I mean, my world was destroyed when Taylor Swift and Joe Alwyn broke up. But other than that. But there is a mode on AutoGPT called continuous mode, which basically allows it to, in theory, run forever, just doing these kind of like recursive task loops. And when you turn on continuous mode, it does actually pop up a little warning saying like, you might not want to do this because it is, quote, potentially dangerous and may cause your AI to run forever or carry out actions you would not usually authorize.

1:04:25But the mode exists. People are using it. It's out there already. And I've talked to some AI safety researchers in the last few days who are very concerned about this shift from thinking of these AI language models as tools to using them to create autonomous agents. And what they'll say is like, look, these experiments, these Smallville, these generative agents that are just going to parks and talking to each other and planning Valentine's Day parties. Like, these are not dangerous because, you know, they're contained in this virtual environment and they're, you know, pretty safe in the things that they are talking about.

1:05:00But you actually don't want AI agents operating without human oversight. You don't want them carrying out their own goals. You don't want them, you know, doing long-term planning and coordinating with each other. That's just like not a capability that you want to build, right? Yes, and while I'm glad that some of this research is being done in the academic realm, the fact that these kind of experiments are already being run does make me hope that our government is paying attention and is pretty soon going to start to have some ideas about whether and how all this ought to be regulated. Yeah, I mean, I will say I'm pretty skeptical of regulation in this space, at least the proposals that I've seen so far.

1:05:36But when I saw Chaos GPT, which stated that its explicit goal was to destroy the world, I was like, yeah, that should probably be illegal. Yeah, I mean, there's, yeah. Like, and this is the thing, like all the scenarios of AI doom out there, like imply that the robots are gonna have to like become power seeking and, you know, greedy for some resource that it will kill everyone to get. But like, actually it may be that like some troll on the internet is just gonna like give it an instruction to go destroy humanity. And as soon as that's done, you know what the first thing they'll do is? They'll go on Discord and say, hey, look at this, guess I won.

1:06:11Right. So the last concern that I heard from AI safety researchers about these AI generative agents is that these things that start off as being very benign and not all that capable tend to get more powerful over time. Like the Hartford podcast. Exactly. And so you can imagine hooking some of these generative agents up to physical robots. Like that is a thing that can be done. Like Megan? Yes, you could take one of these generative AI personas and tell it to control a hardware robot. And like, it could do that. Oh, you know what? My dream of a sassy Roomba has never been closer to reality. So I really just think this is an area to watch and that I hope we'll come back to on this show.

1:06:59Because this use of AI to create these autonomous agents, I think, poses some really interesting opportunities. I do think there are applications for this beyond gaming. And I think it's an area where there is the potential for a lot of harm and mischief. So stay tuned. We will be giving you more updates on the AI agents. In the meantime, I think I want to go check out Smallville, actually. I want to see what the buzz is. Are they talking about us? Are they listening to the show? I don't think so. But, you know, I'm going to check out another character. Maybe we'll find something here. Latoya Williams is currently taking a yoga class, doing standing poses.

1:07:37Everyone in this town is having a more relaxing day than I have, and I'm jealous. Let's go to Smallville. All right, let's do it. All right.

1:08:03Thank you.

1:08:42Thank you.

1:09:01Jeffrey Miranda. You can email us, as always, at hardfork at nytimes.com. But please, no top secret documents. Save those for your Discord server.

1:09:36Thank you.