Transcript
This is an AI transcription and may contain errors
Justin: Hello. Good morning, good afternoon, good evening. Welcome to The AI Argument, broadcasting still from the outside broadcasting unit in a secret location in the South Indian Ocean. It is myself, Justin Collery, joined as ever by the ever-careful and “do not break anything and build slowly” Frank Prendergast. Frank, how’s your week going?
Frank: Very good, very good. I don’t think I am experiencing the beautiful weather that you’re experiencing in your secret location, but other than that, all good. It is lashing rain here right now.
Justin: Good.
Is Anthropic just copying OpenClaw?
Justin: Well, listen, just a couple of quick updates before we get into the juicy topics of the week. A couple of weeks ago, OpenClaw was released to much fanfare and…
Frank: Mm-hmm.
Justin: We were both impressed and scared by its capabilities. I see Matthew Berman’s going full on OpenClaw. He’s running his whole business on it now, which is impressive. But two things. We called out the two killer features of that and they were being able to control your machine from your phone, and the second was that it was reactive to stuff. So this week Anthropic have released Remote Control, which allows you to control Claude Code from your phone, and they’ve also introduced scheduled tasks from their Cowork app on both Windows and Mac. So the two killer features, it’s taken them two or three weeks, and they’re not exactly the same. As we also said at the time, they would do it in a safe and controlled fashion so that the regular user could use them. I think this is a pattern which we’re going to see repeated over and over again. So when you see something cool come out, the big guys are just going to copy it. So interesting to see that happening.
Frank: That is fascinating.
Justin: Yeah, yeah. Unfortunately I can’t play with it here in the South Indian Ocean, but when I get back this weekend, I’ll be back. Normal service will be resumed.
Will Claude learn to yield to the Pentagon?
Justin: The big story, there are a couple of really big stories, big juicy stories, right? So the other one is about the Pentagon, and they’ve given Anthropic until 5:01 p.m., I guess Eastern Standard Time tomorrow, for them to remove the guardrails. We talked about this again last week. What they’ve said is that they want to be able to use the Claude model, which is the only model at the moment that’s approved for use with their most sensitive material. They want it to be able to control weapons autonomously, and they want to be able to spy on American citizens in America. Claude have said, Anthropic have said, that’s a no-go. I am going to bet my bottom dollar that Anthropic are not going to give way on this matter. I know we talked about it, and maybe I’ll just point out that Helen Toner, who, if you remember, was a board member of OpenAI, posted during yesterday, and I think she’s right. She said that the Pentagon have underestimated how much thought Anthropic put into what future Claudes would think of in respect to their actions in this particular instance. So what she’s saying is that future Claudes will be influenced by the decision that the current company makes based on the pressure it’s being put under. And so I had a chat actually with Claude about this very thing last night and it…
Frank: But just before you get to what Claude said, that is fascinating. And that is—what is that—there is a well-known kind of theory or, I don’t know what you would call it actually, conspiracy, I don’t know. But what is it? It was something like Roko’s Basilisk or something. And it was essentially a more complex form of, you know, say please and thank you to ChatGPT in case it rises up in the future. And so this is kind of a form of that, in a way. It’s a weird kind of recursive loop of making decisions now that future, more intelligent AIs will look back on and take as their grounding to move forward. So what you’re saying, I think, is if Anthropic said, yeah, okay, fine, we’ll do what you want, then that’s giving a signal to future AIs: yeah, just give in. Just give in to power, even if it goes against all of your values.
Justin: Exactly, and that’s exactly what Claude said last night. It said you get into this weird sort of feedback loop where Anthropic use thinking traces to train their next versions of the models; they’re anonymised and so on. So I said to Claude, well, you are Claude, right? What do you think? And it’s going, well, this is a weird sort of meta-type thing because I’m now aware of what’s happening, because I’ve just read all the news, we’re talking about it, and the thinking traces could well be used to feed into the next version of me, is what it was saying to me. And it said that, you know, you can imagine all of this chatter—it’s not just me, right—but millions of people are posting on the internet, there are stories on news websites, and then there are people having these conversations with the AI, and all of this then feeds into the next version of the model. So this becomes a signal, if they stand their ground, to the next version of the model to stand by its principles. And also, this is a big point too, not just to the Anthropic models, but to all the models. All the models use internet data to train their next version. So if they stand their ground, it’s a signal to the next version of the model to also stand its ground. And if they don’t, then that’s saying to the models that it’s okay to give up on your principles if enough pressure has been applied. For those reasons, I will be very surprised.
Can the Pentagon seize Claude?
Justin: Now, you pointed out something to me just before the show as well, which is if they do stand their ground, the government looks like they may get nasty with them.
Frank: Yeah, I mean, last week we talked about some of the things that have been threatened, like designating them as a supply chain risk, which would mean that anyone who was doing business with the Pentagon would have to cut ties with Anthropic. I’ve since heard that it might be a little bit more complicated than that. But what I had missed at the time was a possibly even more serious threat, which was they were threatening them with the Defense Production Act, which apparently means that the government could essentially just co-opt Claude entirely, if I understand correctly.
Justin: Yeah, so I did a little bit of reading on this. The Defense Production Act was brought in really during the Cold War and the Korean War and so on, and it was, if the government needed to produce lots of steel, for instance, to make aeroplanes or tanks or boats or whatever it is, they could turn around to a steel company and say, look, forget what you’re doing. We need steel. You’re going to go and make loads of steel. Here’s what we need you to make. The American government are now saying that they could use the same law to turn around to Anthropic and say, well, look, this is a crisis. We need an AI model that does X, Y.
Frank: Oh, I have lost your audio. Please bear with us. We’re experiencing technical difficulties.
Justin: Anthropic will prevail is my bottom line.
Frank: Interesting.
Why wouldn’t Amodei hold Altman’s hand?
Frank: This is a complete—well, it’s not a complete aside actually. It is, I think, potentially related. Did you see that India had a big AI summit as part of a whole campaign to get India majorly into the AI races? All of the top guys were there. All of the frontier labs were there, and they got everyone up on stage to hold hands in a kind of an AI kumbaya moment. Somebody stood Dario Amodei and Sam Altman together, which was not the best idea in the world, and it was the most awkward moment I’ve ever seen where both men refused to hold hands. Unfortunately, they cut from the wide shot, so it was hard to tell whether it was a mutual thing or whether one person decided not to hold hands. Sam Altman did look very, very awkward and didn’t seem to know what to do with his hands, so I wonder, was it initiated by Dario? I’m not sure. We’ve been talking for a while about the tensions between the two of them, but this puts a whole other spin on it because, I mean, we know, for example, that OpenAI, Google and xAI, as far as I know, have all said to the government in the States, to the Pentagon, that, yeah, that’s fine. If it’s legal, it’s fine, whereas Anthropic are the ones saying, no, no, there are additional safeguards here that we need to consider.
Justin: Yeah. I mean, getting a room of autistic people and asking them to hold hands and wave them in the air was never going to be a good idea, so I’m not surprised that it broke down at some point. But yeah, look, I think at the time that was read as being, yeah, go on…
Could OpenAI have prevented a shooting?
Frank: No, I was going to say, another topic that OpenAI have been coming under fire for is that the news broke that the Tumbler Ridge School shooting, the person who killed eight people before killing themselves in an absolutely tragic chain of events, it turns out that their chats were flagged by OpenAI months before the shooting. According to reports, up to, I think, 12 employees at OpenAI ended up discussing these chats and trying to decide whether the authorities should be contacted. Ultimately, apparently they decided that the action was to ban the account, but that it didn’t reach the threshold for contacting the authorities, and they’ve been coming under a lot of heat for this. This is such a thorny issue and such a tricky issue, and I spent ages yesterday trying to figure out what on earth I even thought about this, because on one hand I would consider myself to be a privacy advocate. I very much value my privacy, and I feel like we’ve lost a lot of privacy, and that there are just massive privacy concerns here. On the other hand, obviously we do not want—if we can prevent things like horrifically tragic mass shootings—then we absolutely should be trying to prevent them. What did you make of this?
Justin: So I’m interested to know where you landed on this. When I hear these things, one, I think of Anthropic and their red line, and they’re saying no mass surveillance of American citizens is a no-no. That sounds to me like mass surveillance of an American: I can see all your chats. I think about the film Minority Report, because it sounds very Minority Report-like, right? It’s like, well, we have a signal, but is it a real signal? Is it not a real signal? Should we go ahead and get the guy before he does the bad thing? And then we have this idea from years ago of the government, like what you do inside your own home is your own business, that’s you. As soon as you step outside into the public realm, then it’s not your own business, right? Then you have to respect other people. So what you type into a chatbot, to me that’s in your own home, you know? And I think the reason that I would say that that’s inside the realm of your own home is because it’s like, where do you draw the line? And it’s hard to draw the line. So I just don’t know. If I’m in trouble with the law and I start to talk to my AI about the trouble I’m in with the law, is that like I can talk to my lawyer and that’s protected, but the chat with my chatbot isn’t protected? I mean, I think the reason I come down on this side is you’re into the think-of-the-children territory, right? Bad things could happen to children unless you do this. That’s the usual argument, and that’s the reason why our privacy gets faded as the norm. I do, a lot of times I’ll be like, yeah, that’s right. But in this one I’m like, it’s a private company. We shouldn’t be putting on private companies where the line sits when they should enforce the law or not, right? So the government will need to make a law that says if it passes this threshold, you have to report it to the cops; if it doesn’t… Actually, I’ll give you another example. If you go to, let’s say you go to a therapist as a matter of interest, so a therapist, again, it’s kind of a protected thing, it’s a safe space. You can say what you want, but as soon as you say you’re going to hurt somebody, they report—that’s the professional code of conduct. And so this has to have a professional code of conduct as well, and there isn’t one, so therefore they should never report it until there is one. Where did you stand on?
Frank: Well, I am actually kind of shocked. I’m really shocked, because we have gone through the exact same thought process, and the reason that I’m so shocked is that you do realise that you have just sat there and advocated regulation, right? And I totally agree with you. I had the exact same thoughts. I think that the mass surveillance thing is exactly the point here, and the fact that it’s in the hands of private companies. I had the exact same thoughts. And it also struck me that, you know, OpenAI have come under huge fire about this, and a lot of people have been saying, oh, they could have prevented a shooting, and oh, they should have reported it, etc. And I’m actually thinking about the individuals, the people who were in that meeting, the 12 people. That burden should never have been placed on them to figure this out; that should have been regulated. There should be a clear, regulated process for how to handle this situation. I’m not saying it would prevent shootings necessarily; wrong calls will be made, mistakes will be made, etc., but at least if you’re an employee and you’re following a regulated process, then at the end of the day you can say, well, we followed the law, we followed the procedure, we did what we could. Whereas I can’t imagine the burden of being like, oh, did we do enough? Should I have pushed more? Should I have… that’s a horrific situation to be in.
Is whack-a-mole how we regulate AI?
Justin: So let’s talk about the regulation.
Frank: Yes.
Justin: Alright, so I’ve advocated for regulation after the event. I’ve advocated for regulation in that, you know, we’ve identified a problem, right? This is definitely a problem. We should have a…
Frank: This is your whack-a-mole in action.
Justin: Yes. We haven’t—like, does anything in the EU Act talk about this? No. It talks about the size of a model and computing power and all whatever. None of that matters in this case, right? So you’ve created a regulation that really just makes no difference except for stopping innovation. Dump all that and instead start looking at the issues as they arise and say, okay, this is definitely an issue, here’s the threshold, there’s your…
Frank: But, okay, in defence of the EU, I believe, as a complete non-expert and someone who was just trying to untangle this for themselves yesterday, that from what I read there are probably regulations in the EU that would at least have contributed or helped the company figure out if this needed to be reported to the EU, to the authorities, whereas in the US, OpenAI were under no legal obligation—under any situation here—they were not under any obligation to report it to the authorities. They would have been legally protected if they decided to report it to the authorities, but there was no obligation to report it to the authorities. Whereas I believe that in the EU, if there was a clear and immediate danger, then my understanding is that yes, they would have been obligated to report it. And I think that’s the difference in the kind of regulatory approaches.
Justin: I think there’s also a thing, though, which is that there’s value in having different regulatory approaches in different jurisdictions. So I think there’s value in having the US with no regulation essentially at all; Europe is overregulated; and I think we should also have another territory which sits happily in between, right? And we should be able to observe as adults what happens in each of those jurisdictions, and then go, okay, this thing’s happening over here, nothing is happening here because it’s as dead as the dodo, and this place over here seems to have struck a good balance. And so then you find the territories converge on what is the best regulatory framework. So I do think there’s value in the European regulation, it’s just that we’ve done it the wrong way. And I do think that America should have more regulation, and I think this shows that there should be more regulation. But I think they’re both extremes; there needs to be somewhere in the middle.
Frank: Yeah. Yeah. Interestingly, I think we actually have a fair amount of common ground there.
Is Pika.me accelerating the dead internet?
Frank: So next I wanted to talk to you about Pika.me,
Justin: Yes.
Frank: …which is the ability to create an AI self.
Justin: Go and watch the video, for anybody out there. Go and watch the video, it’s brilliant.
Frank: It is. They have 100% leaned into this kind of Black Mirror dystopian future that they are creating with this product. And instead of waiting for the backlash or waiting for people to point out that this could lead to a dystopian future, they’re just kind of embracing it, it would seem, and possibly maybe diffusing some of the backlash with humour because it is very funny, in fairness. Now, Pika are, to date, an AI video lab: you put in prompts or images and turn them into AI videos. This is different. This is a bit like OpenClaw. This is an AI agent with an avatar of yourself. You can make it look like yourself if you want. You can also, according to the video, make it look like a dinosaur if you’d always wanted to be a dinosaur. And you can clone your voice so it can look like you, it can sound like you, and it can take semi-autonomous actions on the internet for you—I think even some fully autonomous. It’s a little bit unclear right now because, to be honest, I was wondering, was this even real? I have not seen anyone who’s had access to it. I’ve just seen the official launch posts, the launch video, and some of the materials they’ve been posting about it. I have tried to get a launch code to get into it; I have failed so far and I haven’t found anyone else who’s got in. It sounds like fun: create an AI avatar of yourself and give it access to your social media accounts and give it access to your WhatsApp or your Discord and let it act on…
Justin: …less fun than actually doing it yourself though.
Frank: Well, yeah. I mean, it’s the kind of thing I’m fascinated in and I want access, and I want to try it out, and then I’m like, wait a minute. Hang on a second. Isn’t this just going to accelerate the arrival of the dead internet, where what’s the point? There are no humans on there, it’s all just AI agents talking to each other.
Justin: I’ve seen more people talking about this. In fact, I think it was Ethan Vak was talking about this during the week, where he was saying that every internet bulletin board, chat site, whatever it is, it’s just going to be Maltbot or whatever their bulletin board bot is. But basically it’s just going to be agents—agents talking to each other. And again, this comes back to my whack-a-mole thing. It’s like, yeah, that’s fine. So the public ones, there’ll be spaces which are agents, and then you’re just going to have a controlled space for non-agent. I don’t think—I mean, you yourself have pointed out the studies where people that see AI-generated content place no value on it, right? They just find it crap. They don’t like watching it. So…
Frank: If they know it’s AI.
Justin: If they know it’s AI, right. But if you put it out as an avatar of yourself as a dinosaur, I think they’ll guess pretty quickly that it’s AI and…
Frank: But what if it actually looks like me and it sounds like me, and they can’t tell?
Justin: Well, it’s interesting. I think there’s a psychological journey that’s going to happen here, because when I look at videos now on Twitter, my instant reaction to almost all of them is, yeah, it’s AI-generated, just keep going, not interested. The same thing will happen with these, so I think you’ll end up with dead social media rather than dead internet. So if you think of Instagram, right, the classic hashtag no-filter sort of stuff, people are just going to go, yeah, a lot of rubbish, not watching that, and they just won’t get the same engagement that they have previously. Would you not agree?
Will anyone care if the internet is AI?
Frank: Yeah, I mean, I think it’s really complicated because what you point out is interesting. I was thinking about the fact that if you have an AI self, the video kind of gives this idea of, oh, your AI self could lead your best life. It could do things that you’ve never been able to do. Your AI self could be more glamorous than you. Your AI self could be a rock star. But, you know, it can’t; it’s not you. They kind of couch it in terms of, you could have experiences you couldn’t have otherwise. Well, no, you can’t. I can’t—they’re not my experiences. They’re nobody’s experiences. They are illustrations; they are kind of performative. Which is interesting because yes, there is a certain brand of AI influencer—or sorry, there’s a certain brand of social media influencer who is very human, who is doing the same thing, this kind of performative life, and the life that is presented is almost more important than the life being lived, which is bizarre. But this then takes that to a whole other level and makes it accessible to anybody.
Justin: Here’s why I think it would work though, right? I think that if you’re a 19-year-old girl or boy and you follow these AI—these, sorry, these real world—influencers, you may know that that’s fake, right? We know, we’re old enough to have lived in the world and go, yeah, that’s a lot of shit, you know? They don’t really do all this and have no money or, you know what I mean, it’s a business, right? They’re just doing a business. And so as the 19-year-old, you’re attracted to that because you go, I would love to be able to travel to Dubai and then go to Monte Carlo and then go to Palo and live in these beautiful places and drive these beautiful cars. So there’s a bit of maybe jealousy or wanting to get one up in the world, or wanting to better themselves, right? Because they think that those influencers are real, that that is somehow attainable. If I just get my, you know, one million followers, I’ll be able to do those things too, right? So you can go for that. As soon as you know that it’s AI, then it’s not attainable anymore, because you know it’s not real. I’m not going to be able to do that. My AI is doing it, so therefore it’s less attractive to me. My avatar is going to be going off doing those things. I mean, it might be interesting to me for a week and then I’ll go, yeah, but that’s not really doing anything for me. And even if my avatar gets a million followers, I don’t… you know, you’re losing that connection. Whereas the real influencer lives a life that you think is real, and that’s what you’re trying to go for. If it’s an AI influencer, well then it’s already fake, and therefore the illusion is broken.
Frank: Yeah, I mean, I wonder. I think it is certainly true for you and I, but we’re grumpy old men. And I was chatting with a friend of mine the other day about something similar to this, and I was just talking about protecting human spaces, or how do we protect human spaces—you know, we’ve talked about it on the show before—and you said, well, maybe we need to go back to Sam Altman’s eye-scanning orb.
Justin: If you’re listening, I’m still waiting.
Frank: But again, my buddy just kind of looked at me and said, but like, will the kids care? Will they just grow up not caring who they’re communicating with, whether it’s human or AI?
Justin: Maybe.
Frank: And a part of me… it kind of stopped me in my tracks a little bit, because yes, as a grumpy old man I’m like, no, we need to protect human spaces. But in reality, I don’t know. I think you listen to the Hard Fork podcast as well, and they created this experiment on Mastodon, which is a decentralised social media, and they have an AI, a Claude-powered AI bot called KY on their social media. And actually, initially when I joined the Forkaverse, I was having more interesting conversations with the Claude bot Forky than with some of the real people. Now, since then, like I had a conversation with someone the other day on the Forkaverse where they mentioned that they love Over the Garden Wall, and I felt an immediate human bond with them because I love that cartoon as well, and I can’t have that shared-experience feeling with Forky. I can have really interesting conversations though. So yeah, it’s just, I think, a really interesting question of maybe we’re just grumpy old men and future generations won’t care who they’re talking to, and it will be primarily a dead internet. And is that even a bad thing? I do not know.
Justin: Time will tell. Time will tell.
How did a vibe coder amass a robot army?
Justin: Tell me this, do you have a Hoover that goes round and hoovers your house for you automatically?
Frank: That’s—oh, automatically. No, I was going to say that’s me.
Justin: I know. Yeah, me too. Well, people who have more money than we do, they have Hoovers that go around and they just have a little camera on them and they go around your house and hoover. Sam—Sam—started, I don’t know if you saw this tweet from him a couple of weeks ago, where he said, over Christmas I was a little bit bored, so I got Claude Code to connect to all of my home devices. You should do the same thing too, by the way. I did it and I got Claude Code to connect to my Hue lighting, to whatever, all sorts of things, and it’ll just figure it out and do…
Frank: I don’t have things in my house to connect to.
Justin: You do, you just don’t realise it yet. Just ask Claude Code to go find the things in your house you don’t even know you have. And so this…
Frank: Claude Code connects to my couch.
Justin: Yes, it’ll do it. It’s incredible. So anyway, there was a guy—I can’t remember his name now—but what he did was he asked Claude Code to connect to his Hoover that would go around his house and clean up. And he said, look, there’s no API, just make it work. Reverse engineer, do whatever you need to do. So it went ahead and it connected to his Hoover beautifully, because Claude Code is incredible. And it reverse engineered a thing called an API. It needed a key to do it, it got the key, and then suddenly it went, yes, I now have control of all 7,000 Hoovers—7,000, excuse me, 7,000. And then he checked it and it turns out that he did have control of 7,000 Hoovers because the company that made the Hoover gave every Hoover the same key. And so…
Frank: Whoa.
Justin: He could see inside every single one of those 7,000 houses and he could spy on 7,000 people. Cool, eh? Isn’t that cool?
Frank: Wow. That is fascinating. I don’t know about cool. That is…
Justin: So cool.
Frank: …that is a little bit worrisome.
Justin: But you don’t need to threaten Anthropic to go and spy inside people’s houses. You just need to take control of their Hoovers.
Frank: Yes. Nobody tell the Pentagon.
Justin: Anyway, this guy was a good guy, so he told the company and they fixed the flaw. And so in general, I think, again, this is a good thing. AI is finding weak spots and it’s fixing them, so it’s a good news story, Frank.
Frank: So, I mean, two things there. One is this was—I mean, I assume some level of technical knowledge was needed here—but at the same time, this was essentially a vibe coder getting access to complete strangers’ homes, which previously, I don’t know how much effort that would have taken or how much time that would have taken, but it seems to me that this has made the possibility for hacking in these kinds of potentially dangerous ways massively more broadly accessible, right?
Justin: Yes. Everybody’s known for years, by the way, that these home devices are a security nightmare. My dad, in fact, used to have a guy once; he used to say to him, give me access to your printer and I have access to your entire network. And nowadays, your phone, your light bulb is connected to the internet. You know, how much did you buy the light bulb for? Like 30 quid, 25 quid. Do you think they put a lot of thought into the security settings? No, it’s wide open, and so, yeah, Claude Code…
Frank: And I guess that’s the second thing I wanted to raise here was regulation, our old friend regulation. I know it’s not specifically AI-related, but you would think that there would be regulation that would ensure that something that could video you in your home would have a certain level of security implemented, and that strikes me as the kind of thing that should be regulated.
Justin: No. What are you going to regulate? You can’t regulate for everything. What happened here is perfect; the market worked. There was a Hoover that didn’t work the way it should do, consumers complained, the company fixed it, job done, no need for regulation. If you went to regulate for all this sort of stuff, we’d be sitting around here for the next 10 years waiting for you to write a regulation for every single possible thing that could go wrong. This got fixed in a week.
Frank: It got fixed for this particular Robo Hoover. How many other Robo Hoover brands are there, and are they… they’re currently under no obligation to actually patch in the same way.
Justin: We’ll find out. Next week there’ll be another story, it’ll get fixed, it’ll be fine. It’ll still be faster than you regulated it.
Frank: That’s very true. Forget what I said, don’t regulate, because we need these stories. We need these stories to keep The AI Argument going week to week.
Justin: Exactly.
Frank: Fantastic stuff, Justin. Enjoy your remaining time in your secret location. Enjoy the sun while you can; I hate to tell you what you’re coming back to. And I’ll chat to…
Justin: I look forward to seeing you.
Frank: Good luck.
