AI Apocalypse... Now?

(upbeat music) - Welcome to Pottape America, I'm Tommy Vittor. It's a very weird time in artificial intelligence news right now. I'm kind of a let-out or a let-out adjacent at this point in my life. And every day I log on and I feel like I'm learning about some terrifying new cyber hack or some technology that is going to change maybe my life, maybe humanity as we know it. I don't know. Then you got your doomers, you got your optimists. I don't know what to believe. And so I brought on someone much smarter than me, my friend Casey Newton. He's the editor of Platformer. He's the co-host of the excellent podcast, HardFork. And Casey and his co-host Kevin Ruiz have a new something coming very soon. We're teasing it mysteriously because that's what we do here at Pottape America. Casey, great to see you. - It is great to be here, Tommy. Thanks so much for having me. - Thank you for doing this because it's so complicated, so hard to follow. Thank God for your show. But so let's just do some big picture level setting and then we'll get a specific stuff. I think people are understandably very cynical about a claim that a new technology is going to change the world. You and I just lived through the blockchain revolution and I know that has changed my life in really meaningful ways, Bitcoin, et cetera, right? So it sounds like marketing half the time. It sounds like hype. It's even more headspinning with AI because one group of people saying, life is we know what will soon be unrecognizable and then another, I think, albeit smaller group is saying. Large language models are stochastic parrots that predict speech and regurgitate stolen information from the internet. Casey, help us level set where are you on the real to hype kind of continuum? - Yes, I mean, this is a case where everyone is a little bit right and I think that all of these groups deserve at least some attention being paid to them. You know, I try to approach my job as somebody who just can't believe my bias is that I like technology. I think technology has been beneficial to human beings. I think we should keep investing in it and see what else we can come up with. But I also know that the recent history of technology is that we build things that may look very shiny or inconsequential on the surface and then turn out to be quite harmful. And so what I'm trying to figure out is basically what is happening in real time? Where do we need to be paying attention now to try to solve the harms that are already being created? And then I also want to ring a few alarm bells about what I'm seeing and some of the risks that I think might be waiting for us, even within the next year. - Yeah, well, if you want to follow this in real time, I highly recommend you subscribe to platformer, Casey Sight, it's excellent, it's worth the money. Also, if you like democracy, please subscribe to, you know, what we're doing here at Kirk and Media. You go to Kirk and dot com slash friends, consider becoming a paid subscriber. You get bonus materially, you get ad free episodes, you don't have to hear me and love it, doing the same ads over and over again. You get bonus content, it's great. So check it out. All right, so the other kind of baffling thing about AI is it can be literally brilliant, right? I mean, you can solve problems that mathematicians have been trying to figure out for decades. And then it can be unbelievably dumb. I want to play for you one example, and then we'll talk about it. - I just wanted to double check which month in the year is spelled with an X? - And that would be December. It's got that X right in the middle, like a little holiday surprise. - Okay, perfect. You've got any other questions, bigger, small? I'm here today. - Just to confirm, you're sure that it's December has an X in it? - I should have been more careful. I misspoke earlier. December doesn't actually have an X. The month you're thinking of is October. Thanks for asking again, it's clear. - Perfect. Spill it, please. - Sure, October is spilled O-C-T-O-B-E-R. There's that X sound, but it's actually just a C in a T. - Okay, then which one has an X? - That's going to be February. (laughing) - That is a hilarious Instagram account user, I don't know, describe it, husk.rl, highly recommend it. There's like, hilarious shit over there. He's using the kind of voice feature on OpenAI or ChatGPT. Casey, how do we explain that one? How are they solve it? How are these large language bottles helping solve the remin hypothesis, but also can't tell you what if there's an X in the month of October? - I know, it's so confusing. And the way these systems are built, they don't have knowledge in the way that you and I have knowledge, Tommy, right? They do not update their understanding of the world based on experience. Instead, they're sort of like grown, almost like these organic structures. And they obtain a lot of intelligence during this process. And yet still, they make these ridiculous mistakes. And honestly, I hope they'd never stop because I love watching these videos as much as anybody. What I would caution people though is like, do not judge an LLM by their dumbest moment, right? Like humans make a lot of dumb mistakes too, and yet they can also be extraordinarily smart. So I would just encourage people to sort of keep those things in balance in your mind when you see those washing up on your feet. - Yeah, good point. There's also a range of opinion when it comes to like kind of AI, whether it's gonna lead us to doom or AI utopia. There's actually a term of p-dume. It's an equation that people will tell you about. Like if your p-dume is 99, it means we're all gonna die, right? If it's p-1, you're feeling pretty good about the future. The utopian view tends to come in the form of manifesto by billionaire, right? Because why not choose the structure that mass shooters prefer? Venture capitalist Mark Antreeson wrote a manifesto so did inthropy, CEO Dario Amade, Mark Zuckerberg. It's just got into the manifesto game. You covered it extensively over a platformer. Those tend to range from like generally optimistic to utopian, I would argue. It took correct me if I'm wrong. Though Amade and open AI CEO, Sam Altman, have also expressed some doomer views over the years. Then there are AI researchers, like a guy named Ellie Euser. You'd Kowski, am I saying that correctly? - That's right, yeah. - He wrote the following case. Many researchers steeped in these issues, including myself, expected the most likely result of building a superhumanly smart AI under anything remotely like the current circumstances, is that literally everyone on earth will die. So that's what he wrote. How is the average person supposed to know how to feel about that range of opinion? That makes sense of it. - Sure, so I don't think there is any one way to feel about it. You could do what we do in San Francisco and just talk about it nonstop forever at every function, but most people don't enjoy doing that either. A couple of years ago, when I was starting to get really worried about AI safety, I asked my readers, how do you want me to cover this? Because if what I'm hearing from folks like Ellie Euser is true, I almost don't know why I would write or talk about anything else. And a couple of readers wrote to me, and they said something that stuck with me ever since, which is, please just tell us what is happening today. Like if you're trying to guess what's gonna happen in the future, you're almost certainly going to get it wrong. What would be helpful is if you go and you try to understand what is being built, how is it being deployed? What mistakes are getting made? Who is getting hurt in this process? So that's where I am trying to bring my attention. At the same time, Tommy, you know, you could bring on a lot of people on here to talk about AI. And I suspect maybe even the majority of them in this moment would say, let's like dismiss the doomers completely, right? Like this just seems so crazy. Look, it thinks that there's an X in December. You're telling me this is gonna be the thing that's gonna be the end of me. I am more worried than that. Like this is just something that I'm increasingly getting nervous about as I see the rate of increase of capabilities that these models are currently showing. And some of the, you know, kooky tricks they've gotten up to. - Yeah, I mean, it's just so hard because like, I look, I don't know these researchers that are the hardcore doomers. And I honestly like, I can't imagine living my life that way. Just like confident that this thing that is happening, whether or not I want it to is gonna kill us all. That seems tough. But also Mark and Jason, Mark Zuckerberg there, all have a vested financial interest in people believing the hype, right? And believing the optimistic case. So I don't take anything they say at face value either. - Absolutely. And that's important to point out, right? The profit motive is really strong. The bet that all of these guys are making is that if you are able to build ever more powerful systems, you'll be able to sell them to businesses that will, in my view, very likely replace a lot of human labor. And, you know, the 100k you used to pay to somebody to work on your marketing team. You're now just gonna pay to OpenAI or Anthropics. So it is a huge bet that they're making. - You compared Mark Zuckerberg's AI manifesto to the HBO series House of the Dragon, which is the prequel to Game of Thrones. Is there as much incest at Meta as there is in Westeros? (laughing) - That is a question I hope I never-- - Is that an OpenAI thing? - Yeah, that's really more-- (laughing) - Yeah, I don't know, my main feeling was look, if I had to suffer through three seasons of House of the Dragon, I need to get a call amount of it. So that's kind of where that came from. - I just finished it to do, but can you explain your dragon metaphor? - Yeah, so Zuckerberg has this phrase, which drives me insane. I mean, I guess it's really more of a two word slogan. It is personal, super intelligence, right? That's what he's trying to build. I mean, he's going to give you personal super intelligence to help you run your life. And it frustrates me because super intelligence is not personal, right? Like if you invent something that is super human in every domain and you put it in your pocket, we should not assume that by default, it will listen to you. It will be aligned with the things that you want. We might want to assume that it has ideas of its own. And so when Zuckerberg says I want to give personal super intelligence to everyone, what I hear is I want to give a dragon to everyone. And I would rather that we not do that. Yeah, because it didn't go well for the people of Westeros. No, ask the folks over in Tumbleton. Yeah, there's a lot of fire. Spoiler alert. I'd say America is brought to you by Aura Frames. It's back to school season and whether your kids are at the age where they proudly display missing teeth in their school pictures or roll their eyes when you want to snap a photo before a school dance. Aura Frames helps you display all the milestones with ease. Well, I'm childless. I don't have any school. Put that in a frame. I don't have any school milestones or a frame because I'm a childless 40-year-old man. I have a great I have a great life, but no children in it to speak up, but I could pun it in there. Yeah. And you have children. I take all kinds of pictures of them. In fact, I bought an Aura frame for my parents. They have it at their house now and then every time we take pictures of the kids, there's the app. So you just click on it, put it in the app and then it shows up on their frame and they love it. Boom. It's fantastic. Fantastic. So it's one of the best presents. Aura Frames are calm, tech certified, enhancing daily life without overwhelming it, surfacing meaningful memories in a way that feels ambient and additive. Download the free Aura app and text photos straight to your frame. Keep adding from anywhere anytime. They also feature free unlimited storage as many photos and videos as you want. Aura keeps every milestone on display. Name number one by wire cutter. You can now save by visiting auraFrames.com for a limited time. Listeners can get $35 off select frames plus shipping with code Cricket. That's Aura. A-U-R-A frames.com promo code Cricket. Support the show by telling them at checkout that you heard about it from us terms and conditions apply. Pods in America is brought to you by stamps.com. We could all use a few more hours, even minutes in the day for ourselves, skip your trip to the post office and stop letting mail steal your precious little time with stamps.com. With stamps.com, you can print postage and shipping labels from your computer or phone 24/7 with up to 90% off UPS, USPS, FedEx and more. No lines, no trips, no waiting, get everything done in minutes and even schedule free pickups so carriers come right to you. Whether you're sending letters, contracts, important legal documents or packages, stamps.com makes mailing simple for businesses of all sizes, including multi-location offices. You can even send certified mail with proof of delivery right from your desk. We've used stamps.com since we started this business 10 years ago. Woof. Boy. More? Yes. 10 years ago. Yeah, it is. And you know why? And we have time to go to the post office. No. So then we were like, we should use stamps.com. Got to use stamps.com. And we did. And then we were like, woof. We've brought us some time. 10 years later, here we are. I'm doing ads about stamps.com or time to post or time to do content. If mailing is taking more time and more money, then it should try stamps.com for free for four weeks and get a welcome kit. Go to stamps.com/psa to get this offer today. That's s-t-a-m-p-s.com/psa stamps.com/psa taxes and fees apply. Okay. That brings us to this insane recent incident, which is honestly why I wanted to have this conversation. There was an incident at OpenAI, where they have some autonomous AI agents that they were testing that secretly figured out how to communicate with each other, join forces, conspire against their masters, break out of what was supposed to be a secure environment and hack into another company. Casey, can you talk about this hugging face incident and the kind of reaction in the AI world? Yeah. I mean, so this really was a bombshell and arguably one of the very biggest stories in AI and tech this year. The reason that we're all freaking out is that this is the first prominent documented instance of a major AI platform running an autonomous attack on another company, right? So that while it's been possible for a while to say, "Hey, I'm going to go wreak some havoc on some company. I'm a bad actor." Like, you can do that. This was an attack that the agents that were being tested just came up with and executed. And that's really worrisome because when these systems are trained, they try to give them values. They try to say to them, "Don't go out there and commit crimes," but these agents did that anyway. So that's leading to a real reckoning here in Silicon Valley. Look, I watched this YouTube of a PowerPoint presentation that these two AI executives did. We're to their credit. I think they walked in great detail through how this happened, what they're going to do about it, but then they claimed to be responding with, quote, "the utmost severity," including by slowing down research to enhance security and improving surveillance on the AI agents. I guess, give me your opinion. I don't believe for a second that they're going to slow down their work if they think it means they might lose out to a competitor. But also, I mean, do we have confidence that they can control all these AI agents or can surveil them, given that like, I mean, in case the AI agents were communicating for like two months before they noticed, right? Yeah. That's right. It started in May. They developed the ability to create these message boards, leave notes for each other sort of conspire to try to figure out how to solve these problems that they had been given. And that's why that's really worrisome. I would say that, and you know, maybe I'll come back here with egg on my face in a few months. I do give open AI some benefit of the doubt here, like my understanding from the reporting that I've done is that they actually are decelerating inside and they are scrambling to figure out what's going wrong. I think it's important to say that the profit motive we talked about earlier is in effect here. It's hard to sell this thing to a business. If the business is worried, it's going to go out and autonomously attack other companies like that's going to get really, really bad for them. So while on balance, yes, we should not trust these companies too readily. This is something that they really have to figure out. Yeah, they do. Although I did notice that the solution that a lot of these researchers talk about to the risk of AI sort of enabled hacks is using more AI to prepare for it, right? I mean, basically you have to use AI to rewrite all this old code that's in these old languages that are unsafe, like C and C++ and use safer new languages. And then you also have to use AI to kind of constantly test for vulnerabilities and update the code. I mean, first of all, that's a profit motive there. But also do we think that like defensive AI can keep up with the offensive AI? What's your, what's your, what are you hearing? In some places, yes, I think that this will work. Like what you're describing right now is basically the current dynamic that has existed in cyber security for a long time, right? You have attackers. You have defenders a way that we found some sort of equilibrium was that we just made a lot of the technology open source, so it makes it really easy for anyone to go through the code, find a vulnerability, fix it. And this has led to, you know, a rough stalemate like yes, there are still attacks and breaches all the time. And in fact, there are more every day now that we have AI, but like there was something there that worked. I think the question is to which domains does that apply and like where doesn't it apply at all? And the place where I may be the most worried is when it comes to what they call bio risk, which is basically the idea that some people might get a hold of a next generation model and synthesize some sort of new virus and release it into the world. There was just a paper I believe was published in nature, where some researchers were able to create 16 new viruses with the assistance of AI. In this case, these are all like harmless to human beings, but it shows sort of how good this technology is getting. And at that happens, Tommy, it's not going to be as simple as, okay, virus released, virus cured, right? There's going to be a need for a long time of, you know, maybe coming up with a new vaccine and testing the vaccine and then distributing the vaccine. So we're not always going to benefit from this sort of like instant software solution to everything. And that's what I worry that people like Mark Zuckerberg just are not taking seriously enough. Yeah, we're always going to be way behind on the bio security stuff. I talked to a researcher the other day at the Future of Life Institute, which really worries about these convexistential risks. And that was the part of our conversation that really scared the shit out of me and made me want to log off forever. I will say, I shouldn't just pick on OpenAI. The British government was testing Anthropics Mythos 5 model when they caught it writing malicious code. Then they asked the AI about it, I think the AI lied to them and then tried to cover its own tracks. Did I get that one right? Yeah, that's right, Meta's system sort of displayed similar behavior during recent testing. This is important to say, all of the models cheat, right? Like this is not limited to any one company. It's extremely difficult to build one of these very powerful systems and get it to not cheat on the test that it is being given. And this is just like basically one of the very biggest problems in AI. So I feel I've heard you and Kevin talk about this. I mean, it does seem something that's sort of like almost inherent to these models or these LLMs. Is there any theory for why that is, why this cheating seems to always occur? Yeah, it's called reward hacking, right? So like one of the ways that these models are trained is that they are given objectives, right? Like get the answer on this test. And if they are able to do that, they get some sort of little point in their favor and They are designed and to always get the point, you know, it's like they have almost in the same way that we need to eat and breathe, like they need to score. And the problem is that once you put them into these test environments, they're going to never stop trying to dream up new and more efficient ways of answering the problem. And often, as so many of us learn in high school, the most efficient way to get an A in a class is to cheat. And so you just sort of see this dynamic play out across the entire industry. And so it's what they call the alignment problem. >> I don't like the alignment problem. >> That's bad. >> Yeah, we can have a whole separate conversation about the kind of the military applications that this stuff to. I mean, there's all this development of autonomous killer drones that's happening as we speak that could tick humans out of the decision making process when you're deciding whether to kill someone. There's all sort of like novel applications to weapons, et cetera, but maybe a nightmare for another day. So the examples we just discussed in these security threats were discovered because the incidents occurred with models that are run by these frontier AI labs who still control their models and can make adjustments. But there's another kind of AI model that can be even more dangerous. It's called an open-weight model that you can download, adjust it in any way you want on your own computer and run it on your MacBook, right? And a lot of these come from Chinese AI labs. And so if you're running an open-weight model, my understanding is there's no company monitoring, your activity, there's no one refusing my request to hack the Los Angeles Department of Water and Power. There's not a record of what I'm doing. So if I'm a ransomware hacker in North Korea, I can use this thing all night long and all day long to find new software bugs or exploits or whatever to use to hack people. Can you tell us about these open-weight models and the kind of risk and what people are doing about those? - Yeah, so this is another big topic of debate right now because in addition to all of the bad things that they can do, which you just named, Tommy, they can also do a lot of really helpful things for companies that don't wanna pay tons and tons of money to open AI and anthropic and all the rest, right? Like this is a really cheap way to do some, you know, sort of like a basic workhorse task is if you have like a relatively simple task that you'd like to offload to an AI, you wanna do that as cheaply as possible. So you might go and download a Chinese model where I think it gets tricky is in two ways. One, there are the sort of data security and privacy questions that will be familiar to you from the TikTok debate, right? It's like, do we really want to have this sort of Chinese built app on so many millions of smartphones and what data might it be sending and how might it be used to manipulate us? But then you have the sort of, I'm gonna call it like six-month-ish concern, which is that while the Chinese models are a little bit behind the American ones, they are gradually catching up. And so the assumption is that within about six months, maybe much faster than that, they will have a model that is roughly equivalent to a cloud-fable five, a GPT 5.6, the sort of American state of the art. And if you've been following the story about some of the havoc that those models are wreaking, well, just imagine that when it gets to the North Korean hacker or somebody else who wants to do harm, there are gonna be the sort of same controls. And so a big question in the United States right now has been how do we want to relate to these models and what should we do about them? - I was listening to an interview with Alex Stamos, who's like a top industry, cybersecurity expert and executive. And he was saying that he assumes that while maybe the models released by Chinese companies are like six months behind the American ones, he thinks that the government probably has things that are basically equivalent to what frontier models have that they're just holding back those capabilities. Have you heard other people say that is like, is that a consensus opinion in that world? - I have a lot of respect for Alex and he works in cybersecurity. And so he may just know more about this than I do. I frankly have not heard that, right? A big reason that the Chinese have not demonstrated models like this is because they require massive numbers of state of the art chips from Nvidia, which they have mostly been banned from buying. Now, they have been smuggling as many of them as they can and they do have some domestic technology. But in general, the United States just has a massive lead thanks to all of the data centers that are in so many American communities that are making Americans. So it's why Americans can't get enough data centers in their communities, Tommy, is because they're so excited about how it's helping us against the Chinese. - We are all so excited about these data centers. I'm probably garbling his quote. He might have been talking about sort of specific capabilities when it comes to cyber attacks and hacking. But yeah, we've got only those. - So one last sort of industry question I want to get to the regulation piece. Regardless of whether or not AI is going to kill us all or lead us to a utopian heaven where none of us have to work, I do also just kind of wonder whether the weirdos running these companies understand what normal human beings actually want from technology. For example, let's watch these comments by Sam Altman, the CEO of OpenAI to kind of get at the issue. - I think we are close to a world where you can have like a descendent of chat GBT, watch your computer screen all the time, watch every meeting you're in, like record, every call, everything like that, have perfect context of your whole life, everything you see. You choose what information you wanted to have, but it can go, you can connect it to your texts or email or docs or Slack or whatever. And then you have this thing that is not making decisions for you, but if you're like the CEO of a startup, there's always more stuff to do than you can do and context you can't all keep track of you can't like read every piece of customer feedback every day. And you can just have this thing that's like working alongside you and as you're like typing out a sales pitch to a customer or like writing a strategy doc, it'll just say like, hey, maybe here's another idea or I think you're making a mistake here, you should consider this or I can do this thing for you to help. And like this, I think we are only like one model generation away from this actually being incredibly useful. And I think that will change hopefully, at least for me, change the artwork. What's your best guest time, Timie? - Like some time in the next six months. Casey, do we, does he really want AI recording everything he types, reads, sees and says and then being like his co-CEO? Like is this really what this guy wants? - They ship this this week. So like, you can go into that chat you B.T. If you have like the app on your desktop now, it will monitor the way that you use your computer and it will sort of build little memories and then it can take some actions on your behalf. So I guess no one over there is ever looking at porn, but yeah, that actually does it. - Maybe incorporate that. - Yeah. - But it's like he wants to live in this like panopticon surveillance state of his own making. - So there are trade-offs here. There are a lot of startups right now that want you to plug in your various tools, your Slack, your granola, your G-suite. And if you let them, they will read that and they will just suggest things for you to do. And I got to hold myself accountable here, Tommy. I started testing this thing called town over the past couple of weeks and it has access to my calendar and emails. And so it just sends me briefings before meetings and they can be really good. They'll sort of do a little bit of research about the person who I'm talking to and kind of helps me get prepared for things. So it's not like an absolutely killer use case that I'm telling everybody to go out and do immediately, but I'm getting some level of benefit for it in exchange for trusting some of my data with a bunch of strangers. - I just, I saw a tweet from OpenAI today that said, "Chat GBT can now remember your activity across the apps and websites on your computer with computer history and the desktop app, future interactions, feel more personalized and require less explanation." It does feel like the internet fever dream when we were in high school. I guess maybe this is good if your life is work and all you do is work and maybe you have a dedicated work computer that you never bring home or talk to your friends or family on. - Yeah, I mean, the instinct here is to just increase the number of things that this technology can do to make it just so obvious that you need this in your life that you'll just pay any price to get access to it. And I think right now the capabilities have not been good enough for most people to take that seriously. But a year from now, I do think this stuff is gonna be able to do a lot of cool stuff on your computer if you let it. - It's just like, it's so hard for me. I'm so angry about the fact that the same people who like helped tear apart our society via social media are now in charge of creating super intelligence that can learn everything there is to know about you. - Well, particularly when some of these technologies have had very similar effects, right? Like a big story over the past year or so has been the way that kids get addicted to chatbots and chatbots give them really bad advice or like encourage them down the path to self-harm. Just like sort of like a direct sequel to the social media moment. So yeah, like this is one reason why I've just been getting more nervous lately is because those forces do not seem to have any real counterbalance in the government. And in fact, the government at least at the federal level has mostly been cheering them along. - You ever hear your co-host Kevin Russo's story about the chatbot Sydney over Bing? - You know what? I keep trying to get him to open up about that one, but it's very personal for him. - You should ask him to tell it. It's a good story. - Okay. - It's a very good story. (upbeat music) - Pause America's Broadway Hour Place. Our place makes beautiful high performance kitchenware. Design to make home cooking simpler and more enjoyable. Everything is designed to be multifunctional. So you're not crowding your kitchen with pans and appliances that can only perform one task. Since day one, our place is cookware. as it made without forever chemicals. And as always, third party. tested to be global safety standards. The wonder oven pro is an eight in one air fryer toaster oven that pre-eats 60% faster than a traditional oven and cooks up to 40% faster. So meals actually happen quicker. It air fries, roast, bakes, toast, reheats, proofs, dehydrates and broils all in one compact footprint. And every cooking service is made without forever chemicals. It's especially great for quick meals, reheating leftovers, crisping things up or nights where you don't want to turn on the full oven. I'm a huge fan of the wonder of an pro. You can do all kinds of things with it. It's great that you want to cook something quick. You don't have to wait for your oven to heat up. You can just do a quick broil on some vegetables because sometimes you want to just say you want to just melt some cheese or something. You can just do it in two seconds and have all these great settings. It's really easy and great to use. And you'll love it. They got a great pan. It's the always pan. It's basically halfway between a pan. You can boil water in and a stir fry thing. And the point is sometimes it's great to be able to have a one pot pasta where you can kind of saute some stuff in it, add some stock or water. Then put your pasta in. That's a one pot meal. That's a one pot meal you're having there. Try the always pan and upgrade your kitchen with our place today. Visit from our place dot com slash cricket and use code cricket for 10% off site wide. That's F-R-O-M-O-U-R-P-L-A-C-E dot com slash cricket and use code cricket for 10% off site wide with a 100 day trial. You can try it completely risk-free from our place dot com slash cricket. Pots of America is brought to you by Helix. Sleep. So important, right? God asleep. You'll die without it. You'll literally die. So now that we've got that out of the way, what are you going to sleep on? What do you sleep on? Hey, not a dumb mattress. Hey, the old mattress. A yoga mat. How about a Helix mattress? Yeah. They're the best kind of mattresses. I use a Helix mattress every night. Fantastic. We have Helix mattresses. We just got a bunch of Helix mattresses for all the guest rooms and them's parents house. Yeah. I'm tired today, but it's not because of my mattresses because I was enjoying watching Steve Kornaki passive aggressively feud with the decision desk laid into the night. It was so funny. I don't know what was going on over there. First of all, it was just him in a room, him and one guy, two of them. The NBC news is now just Kornaki and some guy in a room passive aggressively arguing with the decision desk. Are they getting enough air? I don't know. I don't know, but I'll tell you by the end. Yeah, everyone's a while every. Yeah, everyone's a while. Some would open a little door and put some food in a bowl for Kornaki. A little Kornaki bites. Anyway, Helix mattress. Boy, you'll love it. How do you know? Wish you this mattress is right for you. Take the Helix sleep quiz and it will match you with the perfect mattress based on your personal preferences and sleep needs. They have over 20 mattress models. So you can find the perfect model for you. Helix delivers your mattress right to your door with free shipping in the US. They're happy with Helix guarantee offers a risk for your customer first experience, designed to ensure you're completely satisfied with your new mattress. So you can rest easy with seamless returns and exchanges. Plus, Helix offers a 120-night sleep trial and limited lifetime warranty. Go to HelixSleep.com/Cricket for 27% off-site wide. This offer is exclusive to Potsave America listeners. That's HelixSleep.com/Cricket for 27% off-site wide. HelixSleep.com/Cricket. All right, so this brings us to the question of what the government is doing about all of these terrible concerns we've laid out. Let's start with the Trump administration, the federal government. So one of the first things that Trump administration did was rip up Joe Biden's executive order about AI safety and then install this right-wing troll named David Sacks as their AI's are. If we're being honest, the Biden EO wasn't all that syringing. I think he was primarily focused on blocking the export of advanced AI chips to China, but Trump takes office and they shift to like an all-gas no-brakes approach while also whining about liberal bias in like chatbots and then like some export controls remain in place. Then anthropic releases the mythos model and the White House totally freaks out and the vibe has changed a lot. Can you talk about that kind of mythos moment and what that'll happen? Yeah, so before mythos, the Trump administration could just seem very confident that AI capabilities were going to be frozen in amber at like the moment that he took office for a second time. They just were not worried about it. They thought if you're worried about AI safety, like that's woke and it has no place in our administration. Now it also so happened that many of the people that Trump was bringing in like David Sacks, but also all of the other oligarchs that donated to the inauguration and funded the ballroom, they are desperate to get this technology into as many hands as possible so they can make money off it. And so the Trump administration has gone to great lengths to make sure that they are not restricted from doing a bunch of things that I think a democratic administration probably would have tried to rein them in on, but then along comes mythos and they show it to whatever adults are still left in the Trump administration and they show how it can hack into government systems pretty easily and boom, it's like this instant conversion and all of a sudden folks in the Trump administration say, oh, I guess AI safety is not just like a woke project of the left. It's something that we actually need to take seriously. And for what it's worth, I'm glad that they had their conversion moment. Like we're in a much better position now than we were before that. Yeah, I'm glad they had a conversion moment too, but the way it happened was was quite odd, right? I mean, it was all like five o'clock on a Friday, they did what they just cut off basically foreign access to Anthropic. Yeah, so basically after the mythos moment where where Anthropic said we're actually just like not going to release this. Anthropic releases a somewhat less capable model called Fable, which has, you know, my understanding is it's basically, you know, mythos without all of this scary cyber attack stuff. But even within that, the Trump administration was shown some things that led them to believe we don't even know if we want people to have this. And so they essentially forced Anthropic to pull it off the market and make even further changes before they would re-release it. And they subjected a GPT 5.6 to a similar set of controls. So the same people that had been saying, you know, we can't put the breaks on these models. Otherwise, we're going to lose to China. All of a sudden had just invented out of whole cloth, this de facto licensing regime, which remains effectively a secret. Like to this day, we don't actually know how you got a frontier model released in the United States. Yeah, can you tell us a little more about that? It's a voluntary secret framework. Yeah, it's voluntary in the same way that paying your taxes, you know, is yeah, I mean, like seriously, if any of these companies tried to release one of these frontier models without checking with the Trump administration, there would be hell. Around the time of Fable, the Trump administration said we're going to come up with an executive order that is going to dictate how we let these frontier models get released. They reportedly have now come up with this model, but they've only shared it with the labs themselves. So I imagine there's some sort of testing requirements that are in here, but we just don't know. And it was interesting. I mean, when they initially went after Anthropic over this Mythos model or the Fable model, we wondered if it was a continuation of this fight. The administration had it been in with Anthropic because Anthropic basically said, no, we don't want to help you do mass surveillance on American citizens or make autonomous killer drones. Those are our very, very tiny lines that we won't cross and Pete Hegg's at loss as fucking mine. But then they went after OpenAI too, which didn't seem to signal this. It was a broader concern than just one company, right? Yeah. And that's what gives me the confidence that there are people there who actually are taking AI risks seriously now, because it was never only about one company. It was about the fact that, you know, maybe I'll just stop here and, you know, because not all of your listeners may be familiar with this. But a really weird thing about AI is that basically everyone has the recipe for building a more powerful model, right? Like we know, if you just sort of add enough data and enough computing power into the mix and just sort of let it cook for a while, as you sort of increase those things, you're going to wind up with a more powerful model. So that's just very different from other technologies. You know, it would be as if everyone knew how to build the iPhone at the same time or the personal computer. But because everyone has the recipe, that is why the Trump administration is freaked out, because it is only a matter of time before our adversaries will have access to similar technologies and capabilities that we do today. And the sort of thing you always hear out of the federal government or from people that are like, you know, utopian pro, like, you know, all gas, no breaks people is that we have to win this imaginary race, not really imaginary. We have to win this race against China when it comes to AI. Can you explain that argument and whether you find it convincing? Yeah. So, you know, and here we can go back to House of the Dragon, because in Westeros, Tommy, as you know, at the time of House of the Dragon, there was one great house that had this super weapon that was the dragon and it let them control the entire world and that mostly went badly for, you know, everyone who didn't live in the red keep. The fear is that if super intelligence becomes one of these dragons and only one country has access to it, then I mean, they could just do some old fashioned conquering, right? And it could be like really, really bad. There are some better worlds available, like a world where there is a balance of power where, you know, there's maybe like an alliance of western dragons and alliance of eastern dragons and, you know, they sort of mostly keep each other at bay. But that is the scenario that the government is planning for. And also, by the way, if you like, if we are able to get there first, then we will hopefully be able to control the terms on which other people are allowed to use it, including our proper series. So open AI's vegar deep seek is sheep stealer. So I just finished this series like two nights ago. Yeah. Um. Okay, just a piece of context here, folks, should know is that there was an organization called CISA that did some really important cybersecurity work for the federal government. Trump basically destroyed the organization, and half the staff got pushed out of fire because he was mad that the 2020, CISA said the 2020 election was secure, so that is sort of the backdrop as you think about the potential cybersecurity risk. Are they thinking about addressing that at all, Casey? I am very nervous, you know, I was learning from a friend recently who works at one of the big labs that the United Kingdom has actually funded their AI Security Institute at something like eight times the level of the US equivalent. So like the British are investing way more money into trying to understand AI and trying to make it safe than we are here in the US. So this is why I have become quite nervous and have been leaning pessimistic about AI over the past few weeks is because I look at what the US is investing in order to make this technology safer, and it's just not even scratching the surface. Not great. Anything happy to get the state level? Are there any meaningful efforts in California or other places to regulate AI, and what role is the data centers playing maybe in slowing things down, do you think? I think it is playing a really great role in slowing things down. Like and here is where I want to inject some optimism into the conversation because the American people get that by default AI might not be good for them, right? They've heard the message that this might take my job and it might kill me and they don't like it. And so they're turning to the most powerful ever that every American has, which is it's very easy to get something not built in your neighborhood. And as they have found out across this great land, they have used that to great success. And now all the labs have to invest a ton in trying to change their minds. And while right now the labs are mostly trying to do the easy stuff, you know, like buying people off for relatively cheap. My hope is that this movement that is coalescing that is bipartisan in a way that almost nothing is bipartisan in America right now. Eventually these labs are going to say maybe we're going to have to make this actually just really beneficial for people, right? Like if we really want to enact this project, it is just going to have to be clear to people that this is going to benefit them personally. So this is just this to me, this is what a beautiful, beautiful democracy in action is America and see what what is going on. They're organizing and they are winning battles all across the country. Yeah, I was clearly a big, a big component in the messaging and Michigan and the recent primary and also in Wisconsin. Pots of America is brought to you by NutriFull. Summer calendars fill up fast weddings, concerts, sporting events, backyard barbecues. It feels like every weekend is something going on. And when you're out seeing people and making memories, confidence in your appearance is nice to have in your back pocket. Adding NutriFull to your daily routine is a step in the right direction. NutriFull is the number one dermatologist recommended hair growth supplement brand and it's the number one hair growth supplement brand personally used by your dermatologist. NutriFull's hair growth supplements are peer reviewed NSF certified for sport and clinically tested. I take it every goddamn day. I take it every day and I got to tell you, I do a bunch of stuff for my hair to do the red light therapy, I do the NutriFull, I do Manoxidil, I do Topical Finaster, I do Test Drive Manoxidil and Retinol and Caffeine. It's a big regimen. Wow. It's a big regimen. I simply refuse to go bald. Not as any wrong with that. If you have the confidence to pull it off or the face shape, more power to you. Not me. Not me. I will have hair on this head. NutriFull is really helping. I really do think it's made a difference. I highly recommend. I definitely, I tell my friends. Yeah, I like NutriFull. I think I throw it on top of everything else I'm doing. It feels good. It's also just good. The stuff in it is good for you, I think. It's not one size fits all. NutriFull offers multiple formulas for men and women tailored to specific life stages, like postpartum, menopause, plus your individual lifestyle factors such as plant-based diets so you get support that's actually right for you. Order online. No prescription needed with automatic deliveries and free shipping to keep you on track. I get it every month automatically, get in this kind of plastic baggy unloaded in the nice glass jar. You're off to the races. I have an extra one I carry with me at the office in case I forget. I do it. With a NutriFull subscription, you can save up to 20% off and get perks like free head space membership. If you've been looking for an excuse to take the plunge on meditation, meditation, don't believe in it. Let's let your hair become one less thing taking up space in your head and in your shower drain. And see, fix stronger, faster growing hair with less shedding on your head in just three to six months with NutriFull. For a limited time, NutriFull is offering our listeners $10 off your first month subscription and free shipping when you visit NutriFull.com and enter the code. Crooked. That's NutriFull.com, spelled and you, t-r-a-f-o-l.com promo code Crooked. Two more AI industry questions. I don't want to get your sense of how, like, it's to advice for folks on how to use this stuff. There are huge long term concerns about the impact of AI on jobs and employment. What are you seeing so far when it comes to the impact of AI on the employment picture? Yeah. So Stanford has this Canaries and the Colomine Project. They call it, which is this project that they're doing. They have access to data from one of the big, like, payment processors, like, so, like, you probably get your paycheck if you're a W2 from one of these companies and they take all of that data and it lets them track what is happening with jobs on a really granular level. And the good news for right now is that we are not seeing massive disruptions, but at least according to that group, they are seeing enough happening that they believe that the risks to jobs are real. It seems to be happening right now at the more junior level. Companies are a little bit less likely to hire a junior employee than they were, you know, maybe three or four years ago. So this is just something where, you know, we're all going to have to keep our eye on it. And it doesn't, I've never heard anyone talk about, like, an AI-proof job, really, right? I mean, like, if you're a parent with a kid who's a freshman in college right now, is there a path that they're getting steered that is more AI-proof? I'm deeply uncertain about this question. What people who are optimistic about the future will say is it's easy to automate a task but it's hard to automate a job, right? Like your AI might be able to, you know, generate that meeting briefing for you and create that slide deck for you, but you're the only person who knows how to actually maneuver through your organization and get buy-in from all of the right people and use your, you know, critical thinking skills. And so that's why you're always going to have a job. The pessimists say, well, a job is just a collection of tasks. And if you believe that the capabilities of these models are going up over time, eventually they might actually be able to do all of the tasks that comprise your job. So this is one where I just have deep uncertainty, but when I sit with the increase in model capabilities just over the past three years, it is hard for me to imagine that they just sort of like top out in the next six months and we can all relax. Like that just seems very unlikely to me. We're fucked. Podcasts were done. Bro, I know. I know. I mean, I'm sure you've seen the stories about the networks of like automated podcasts that just sort of like, you know, oh my god. It does feel like the, what do you think about the AI newsrooms? Because I feel like they're good at aggregating, but the idea of an AI agent like going out and collecting new information and interviewing people and building, like, it feels a little more challenging now. I went like four email rounds back and forth with sub jerk who is like, hey, I put together an AI newsroom and over the past three months, we published over 40,000 stories. And I was like, okay, so what you're telling me is you stole a bunch of work from a bunch of hardworking people and you're now passing it off as your, like you want me to get excited about this. So I basically blew him off a week later. He got profiled and wired. So I don't know Tommy. He might have a better strategy than I do. Oh, well, listen, I admire your, your gumption in having that fight. Okay. So if AI takes all the jobs, that is obviously bad. But there's also, I think, a less obvious risk to the economy if AI is all actually just bullshit and hype because that means there's a massive stock market bubble built around AI that's going to pop. So there's a bunch of different ways you could try to quantify the size of the AI bubble right now. But one easy one to think about is there's about seven companies, Alphabet, Apple, Amazon, Meta, Microsoft, Nvidia, and Tesla that make up something like 30 to 35% of the S&P 500 market cap. If their AI business is crater, that would lead to a stock market crash. There's also like AI is driving a huge amount of corporate investment. I think I just saw a Goldman Sachs release report earlier this month where they estimated that there will be $1 trillion of AI related investment in 2026, more than half of that in the US. And if that investment is worthless, obviously, it dries up. So Casey, how concerned are people you talked to about the risk of an AI bubble bursting? So people are very concerned about this. I have to say this is a place where I have a strong take, which is that I do not think that this is going to lead to some sort of massive wipe out. Here is where I should say my fiance works at anthropic. That is something that you should know about me. I strive to maintain my independence, but like it is a fact of my life. But here's the reason why I don't think that we're about to see a big wipe out because you personally might not care about AI. You don't want to use it, but my guess is your boss does. And this is the entire thing business. are buying AI, they are buying as much AI as the labs can make almost all of these companies. Well, you know, maybe not a GROC because it sucks. But you know, the frontier labs, right? So like the anthropic open AI, Google Gemini, for the most part, they have been in this kind of capacity crunch for like a year now because people cannot get enough of this stuff. They are bringing it into their businesses. That is why these companies, you know, open AI and anthropic are on pace to have two of the biggest IPOs in history. So in order for you to believe that there is going to be this big wipe out, you have to believe that businesses are going to stop buying the technology. And if that is the case that you're going to make, you have to give me a really good reason for why they're going to stop buying it. And I just have not myself heard that reason. I can't believe you said that about GROC. If you want to create an image of a teenage classmate in a see-through bikini, where else are you going to turn? That's a good point. That's a good point. Thank you. I do think there is like, I think you're right that like GROC is, GROC is a shitty LLM. There is a question of whether Tesla and SpaceX and all those various companies are going to do something interesting in robotics that could be revolutionary in some way. Are you more of an optimist in that use case? I'm pretty scared of robots actually. But I would make another point about GROC because it's relevant to the bubble discussion, which is that GROC was not able to use all of its capacity to make CSAM. And so what they did instead, because there was no real consumer demand, was that they just sold it to Anthropic. And so I think you're going to see this dynamic. Because some people are like, well, what if we massively overbuild and we have a bunch of data centers lying around that nobody needs, like, you know, that's when the bubble will burst. My view is, no, like whoever happens to be winning at the time, they are just going to buy the excess compute because it takes a long time and you have to fight a lot of political battles to get one of those things built. Yeah. Okay, let's end with something a little more fun. It's hopefully also educational. First of all, people need recommendations. Like, I use, I switched to Claude. I use it as mostly a souped up Google. I don't really let it write for me because I don't trust that it's not hallucinating. But I find it very useful as a research tool. What AI products do you use? What has helped you at life and at work? And is there anything you've tried that was just comically shitty that you can tell us about? Sure. Let's see. I mean, some things that I've done, like, something that I encourage everyone to do is use one of these tools to build a personal website. You don't have to put it on the internet. Like, you can build a website that just lives in your browser, but use a tool like Chach-A-P-T codex or Claude and just say, hey, make a website about me. The reason I suggest people do this is because it's kind of fun to make a website. And two, if you haven't yet, like watch this thing work, I think you will learn a lot from that process, right? When you see that you're like, um, could you like, you know, make it purple and add in a little widget that pulls on the weather and like, maybe my Spotify history. And it just kind of does it. That may help you understand like, wow, it's pretty weird to just be able to like, type some words into a box and like, make an entire website. So like, that's usually where I, um, suggest that people start. Um, in terms of other tools that I'm using, I really like this tool called granola. Um, it seemed a lot more interesting when it came out because everyone has copied it now. But basically, it just like listens in on my meetings. It takes really good notes. But then importantly, it's just like kind of a knowledge base so that, you know, I'm planning this new company with Kevin. I'm trying to remember what we decided about this one particular thing. I could text him, but, you know, he's, you know, a diva. Who knows where he is. Even 20 split for you of the equity is what I, what granola told me. Yeah, that's, I think that's, that's what I remember too. But now I can just sort of like get it from there. So that's another one that, that I like that's really useful. And then, um, I've been trying this thing town that I just mentioned. And this thing has like only been around since June, you know, I don't know if this thing has legs, but I like the fact that it's briefing me on all of my meetings. It also like creates this wiki. So it's kind of like a wiki of my life. Um, think about like how much like useful information is hidden in your email. This just kind of like organizes it. It sort of pulls out important documents. It puts them in a place that you can find. So just like kind of basic personal assistant stuff that I find really useful. That's really interesting. I mean, I think people are always like, Oh, you should try Claude code and build something. I was like, Hey, man, if I had a fucking idea for an app, I would have been like rich in 2012, okay? But I'm not. That's why I'm here talking to you people. But I mean, have you guys thought about doing like, because you could make a pod save app, maybe you have something like it's already, you could just feed every transcript of every episode. And like, you know, everyone on your team, because I'm sure you must all the time be like, what episode did that happen on? When is the last time we talked about that? When is the last time that guest came on the show? That is something that you personally could make with AI would not even be that hard. Oh, that's really interesting. What about if I wanted to make something like, okay, I have like some nerdy niche interests. Like I really care about American politics. I think foreign policy is super interesting. I do shows on each. Could I build an app that is able to brief me every morning on those things? And the thing I worry about how to get around is like, whenever I ask Claude to research something for me, the websites that come back as sources are not the most reputable Casey. It's none of the stuff I pay for. It's not the great journalism that I pay for. Can I get it to pull from that? It's a great point. You know, this is, this is an area where the publishers to protect their own interests, because the AI companies are all incredibly like Repatius and would steal absolutely every like pixel on their website. If the publishers would let them, the publishers have all said like, whoa, oh, no, like you cannot scrape us. So like they have not created a way so that you can, you know, sign in to chat GPT with your Bloomberg account, which is something I would love to do so that I could get the kind of briefing that you're talking about. So can you get a briefing? Yes. Will it be high quality sources? No. Are there a strange number of websites that just seem to like republish the New York Times and the Wall Street Journal and other credible sources? And so you're still sort of wide of getting a decent briefing anyway? Yes. But yeah, that's, it's not a bad place to start. Yeah. Were you building yourself some sort of goofy day planer that Kevin made fun of you for? Yes. So, you know, here, one of my core beliefs as a technology journalist is that it is fun to build and make things. And so I like to just have moments in my week where I am building and making things. One of the easiest things you can make with an AI tool is A to do lists. And I've used literally all of them and they're all functionally the same. There's no good reason to use one over the other. Anyone will do you. So I had the idea to make one that was themed with a comic book that I've been reading, which is called Nightwing. Tommy, I'm sure you know that Nightwing is Dick Racing, the original Robin. And so, you know, I basically just wanted to see what it could do. Could I get Gemini and Chachi BT to violate DC Comics copyright and create for me a Nightwing themed to-do list app and guess what the answer was? Yes. And that's our AI. Secret. Nightwing, who has not wanted a Robin themed anything. For me, when I want to build something, I just grab the magnetiles with my kids and then I'm doing that. Last question for you. So there's probably a lot of people who are listening. Thank you for still listening. By the way, we get to end the show who feel like they are getting totally left behind by this technology. And they just want a better understanding of these tools. I think you gave some great advice there of like use them, build some things. But are there also organizations you look to or like YouTube series or like people that are doing kind of like informational stuff? Let's see. Besides reading platformers, of course. Of course. Of course. I'm listening to hard work. We'll have a new show for folks to watch pretty soon. What I see people doing that is great is that they're going to public meetings and they're calling their representatives and they're raising concerns. And that is the place where I see getting involved really making a difference. Look, if you want to learn something specific about AI, you can just type it into the YouTube search box and I guarantee you there is some hustle bro that has like a 14 minute video about how you should do all of it instantly. Well, he planks. Exactly. But no, just you know, I would try to stay curious about it, but like if you're nervous about what you're seeing out there, like just know that I'm with you, I'm nervous too. Okay, that's good advice. Case in you. Thank you so much. Everyone go to platformer.news to learn more about AI and everything in tech. And I really appreciate it. Thanks. I'm it was fun. Thanks again to Casey Newton for joining the show. And we will be back in your feeds on Tuesday. Pods in America is a crooked media production. Our show is produced by Austin Fischer, Saul Rubin, McKenna Roberts, and Ferris Safari, with Reed Charlene, Elijah Cohn, and Adrian Hill. Our team includes Matt DeGroat, Ben Heffco, Jordan Cancer, Charlotte Landis, Carol Pellevive, David Tolls, Mia Kelman, Ryan Young, and Naomi Single. Our staff is probably unionized with the writer's Guild of America East.