Transcript
This is an AI transcription and may contain errors
Justin: Hello, good morning, good afternoon, good evening, and welcome to The AI Argument, broadcast to you live from 2004 it is myself, Justin Collery. Broadcasting to you from the present day, I’m joined as ever by Frank Prendergast. For those of you listening and not watching, I am in the work, in the doghouse actually, is where I am this week.
And the camera here appears to be something from 2004, so you’ll see me sort of pixelated on the screen. Frank, on the other hand, you’re looking great.
Frank: It looks like sewing machines were very popular in 2004 also, Justin, judging by your…
Justin: Not still? Is it not still 2004? Has something changed? Do you know what it is? This week has been mental. What’s that say? Some weeks, decades happen, and some decades, weeks happen. And this has been one of those weeks where I just couldn’t keep up. There was so much stuff of various types.
Some of it… My favourite story of the week, we might get to it, we might not, is the court case between OpenAI and Elon. What’s Elon up to? He’s in court this week. And there was loads of juicy gossip. It was like Dallas, basically. It’s like Dallas, but with even bigger stakes, if that’s possible.
Is the White House dodging its own Anthropic rules?
Justin: But first off, let’s turn to the White House and Anthropic, the other big…
Frank: Is that still going on?
Justin: So…
Frank: It feels like we’ve been talking about this for weeks. It just goes on and on and on.
Justin: There’s so much going on, right? So the court case itself is… I mean, I don’t know how long it takes a court case to go on. They’re in trouble. Anthropic are in trouble in terms of the court case because they’re in a court where the three judges were appointed by the current administration, and so their chances of winning that case…
I mean, judges, of course, are all impartial. We’d never say anything else. But you do know it’s a longstanding thing, bugbear of mine, that we should replace all the judges with AI.
Frank: Indeed.
Justin: I’m compelled to bring that up and that will be a thing. So the big news this week, outside of that…
And that’s true. They’d have to, what’s that word? They’d have to not disbar themselves, recuse. “I’m sorry, but I have to recuse myself as I’m personally involved with… I am in fact one of the defendants.” That’d be kind of entertaining. Anyway, so the big news is, so the White House, to bring up to speed, if you don’t know, the White House designated them a threat to national security.
They said that nobody could use their software. Then they released Mythos and it was kind of a sort of a wobbly into what it was. And so now this week, the White House have released guidelines telling suppliers how to get around the designation of them being a threat to national security so that they can still use their models.
Frank: So the White House basically wanted to use Anthropic’s models for things like surveilling American citizens and creating autonomous weapons. Anthropic said, “No.” The White House said, “Okay, you’re a supply chain risk,” and now the White House are saying, “Oh, but your models are really good, so now we have to come away, find a way around our own supply chain risk designation so that we can use your models.” Have I got that right?
Justin: Yes, but it gets worse. That was very well summed up. It gets worse.
Why won’t the White House share Mythos?
Justin: Not only have the White House published ways to allow the agencies to get around the designation of the whatever that they were a risk to security, they’ve now also said that their new model Mythos is so good, they don’t want anybody using it.
So at the moment, Mythos is only available to 40 of the… But they’re basically 40 companies in the US. We don’t know who they are. They wanted to expand the use out to about 120.
Frank: And it came out that one of those companies, or one of those organisations rather, is the NSA, which this is before they had worked out these guidelines for how to use Anthropic’s models, but somehow the NSA got access even though they were a supply chain risk.
Justin: And let me go for… Okay, so let’s just recap on why was it that Anthropic and the Department of Defense fell out? They fell out because they didn’t want autonomous weapons, but also importantly, they didn’t want to use the models to spy on Americans. Now, this week, the White House has come out and said, “Look, we don’t want you giving this model out to the 120 companies because we’re worried that we’re gonna use so much of the model that if you go sharing it with other people, we’re not gonna get enough use of the model. We wanna use it a lot.”
Now, Frank, can you think of one use case that the NSA might have where they would need to use a lot of a model such as Mythos? It wouldn’t be spying on everybody, would it? Would that use up a lot of compute that you wouldn’t want…
But anyway, so we’re in this bizarre thing where there is supply chain risk and yet they’re saying, “Look, we wanna use so much of your model that you’re not allowed to sell it to other customers.” What do you think of that?
Frank: I think somebody is definitely gonna make a… There’s gonna be so many films come out about the goings on in the AI sphere. We have the Sam Altman being fired and coming back film being made at the moment. This is definitely gonna be another film.
I mean, this is so insane. It makes no sense whatsoever.
Is Google fine with “all lawful uses” now?
Frank: And not only do we have the Anthropic situation, we also now have… Google signed a deal also with the Pentagon allowing for the same, you know, it had the same language about all lawful uses, and that was the thing that Anthropic had an issue with.
Anthropic had an issue with that because, strictly speaking, that could allow autonomous weapons and mass surveillance. And now Google have signed a deal saying, “Yeah, all lawful uses are fine,” and that they will adjust the model according to the government’s requirements. Now, OpenAI, as we’ve talked about in a previous show, OpenAI have done a similar deal.
I think xAI have done a similar deal, but what’s interesting here is that Google used to have a really anti-AI-for-military-uses stance, and that has just gone out the window completely, it seems. But the employees are not too happy about it. So Google employees have signed a statement asking that Google come, get out of this deal and not allow for things like autonomous weapons and mass surveillance.
So yeah, it’s crazy times.
Justin: And interesting, today, about an hour ago, in fact, maybe two hours ago, the White House released a statement to say that they had signed deals with OpenAI, xAI, AWS, Google, I think they even named one or two other companies that I hadn’t really heard of before, saying that they had signed deals with all of them to use their services.
They named basically all the big AI companies, except for Anthropic. So, again, they’re almost like sort of they’re favoured and they’re not favoured. Now, you’re saying this is like a movie.
Is AI 2027 getting weirdly accurate?
Justin: This movie, the scripture, has already been written, and this is what’s so cool. So our old friend, that I can’t pronounce his name, Ashley Brukenhauer, whatever it is, the guy who was 21, worked for OpenAI and wrote the AI 2027 essay, and he wrote that, I think it must’ve been 12 or 18 months ago now.
At one point in that essay, and it was supposed to be six months further on from now, he… Okay, let me take people back. So AI 2027, for those who haven’t read it, is this sort of fictional account of how the next couple of years would play out, and it started in 2025, I think, right? It was late 2025, early 2026 when it was released, and it sort of maps out all the different stages as we get to AGI and ASI, they predicted in 2027.
One of the things was how White House starts to get worried about the capabilities of these models and how they may impact on national security issues, and then it goes on, and I think in that essay, they speculate that they may have even nationalised some of the providers.
It wasn’t supposed to happen for another six months, but I do know online there’s a tracker for all the predictions of AI 2027, and they’re about 80% correct so far. And this is another one, another tick in that box where politics and AI start to get very intertwined, and it starts to get very murky between what is it okay for a government to turn around to a company and say… Okay, so I guess the intellectual or sort of the mental gymnastics that you have to do here is, on the one hand, the government is saying, “You’re so dangerous. We don’t wanna have anything to do with you.” And on the other hand, they’re saying, “But also, we want so much of your product that we’re gonna buy it all.”
If they were saying we wanna buy… I mean, if they turn around to Raytheon or Palantir or one of these sort of defence companies, I’m guessing basically saying, “Yeah, we wanna buy all your products,” right? That’s a regular commercial relationship if that’s what the company wants to do. But I think it’s very unusual for a government to turn around and say, “We wanna buy at the same time as calling you a threat to national security and trying to shut down your business and make you bankrupt. We also wanna buy so much of your stuff that we’re kind of concerned about you sharing this new model with other people, and therefore we want you not to do that.”
Is GPT-5.5 quietly as spicy as Mythos?
Justin: And then you’ve got the clever Sam Altman who hasn’t used your whip about 5.5 and everything I’ve seen this week about GPT-5.5 said it’s just as good as Mythos at hacking and cracking security systems and so on.
Frank: So you think OpenAI is just flying under the radar there by, for once, I would say, not talking about how dangerous their model is.
Justin: Yes, they’re smart. I mean, give it to them, they’re smart. They’ve seen what’s happened and they’ve gone, “Great, we’re gonna dial down the whole, you know, our model is… We’re gonna dial down the whole hot sauce talk. Our sauce is so hot that you’ve gotta sign a release before we give it to you.”
They’ve dialled down the hot sauce talk and they’ve just said the model’s really good. We don’t care what you do with it. They did release a blog post this week actually saying that they believe the safest thing to do is that all humanity gets access to the models at the same time, which is what they’re doing, in fairness.
Frank: Yeah, I saw that they were saying that democratic access to AI was crucial, which interestingly kind of makes me think of open-source AI, but I’m assuming that they’re saying democratic access, but to a closed model, which can be controlled, I assume that’s their philosophy.
Justin: Well, I feel like you’re besmirching them a bit, right? It’s democratic access as in we’re not gonna have a list of people and say who can access it and who can’t. Here’s our model. If you wanna use it, you can use it. That much it costs. There’s our… I have to say this because we might talk about it later. There’s our terms of service. Do not violate our terms of service, please. But other than that, you’re good to use the model.
So, yeah, I get it. I don’t know what it means for the… I think there may be a point in history in the future where all of the AI companies stood behind Anthropic and said, “We want to enforce the same conditions,” and they didn’t. This may be seen as an inflection point in history, and where we look back at it and say, “Do you know what? It would have been better if all the companies had said no to that thing and said, these are the conditions.” But that’s not what’s happened, so we’re just going to have to figure it out as we go.
Should AI be nationalised for everyone?
Frank: And you mentioned a minute ago this kind of idea of nationalising the AI companies. And in fact, the government did kind of threaten Anthropic with something similar around the same time as they threatened them with the supply chain risk designation. I can’t remember what it was called, but it was basically kind of co-opting the product for the greater good and it’s been done years ago in war times.
So yeah, I mean, it’s not completely off the table that something like that would happen, I don’t think. And part…
Justin: As well for…
Frank: Yeah, I was gonna say, part of me sees the logic of it because we’ve talked a lot about things like the belief that AI is built on stolen…
Justin: Yeah.
Frank: Now, that’s a really complicated area. The courts don’t seem to be leaning that way right now, et cetera, et cetera, et cetera. But at the same time, it’s trained on data that we humans produced and weren’t expecting for it to be used in this fashion. And I think even some of our really early shows, I was kind of saying, “Well, maybe then the path is nationalise it,” as in make sure that AI is being developed for the people, for the good of the people, that there are models that are just available to people through the government in the same way that we have water, power, whatever, that it becomes more of a, I guess more of a public right.
And then all of the in adverted commas theft becomes almost kind of justifiable because it is for the public good.
Justin: Yeah. And there’s no example in history where the government has taken something over for the good of the people and then just used it for their own ends and hasn’t really benefited the people at all. I mean, that’s never happened once, thank God.
Just back on your… So just one other comment, because I do want to move on. You were saying that they have this right to nationalise companies in wartime. Lucky for Anthropic that there’s no wars going on at the moment. That’s the only comment I’d say about that, but let’s not go there.
Why is Bernie Sanders warning about AI doom?
Justin: Your favourite American Bernie Sanders has been in the news this week.
Frank: He has. So, yeah, I think I mentioned before that Bernie was on this kind of no more oligarchs tour of America, and he was kind of railing against the current administration, but also tech companies and figures like Zuckerberg and Musk. And somewhere along the way, he obviously became aware of what people like Musk and Zuckerberg were up to with AI, and he became very concerned with AI safety, and he has been talking about that a lot.
And we’ve talked about his bill to pause the development of data centres to give us time to think about what we wanna do with AI. So most recently, though, he put together a panel where they were talking about the need for international cooperation, basically to deal with the existential threat of AI.
And it was a cool panel. It was Zeng Yi, Dean of the Beijing Institute of AI Safety and Governance in China, David Krueger from the University of Montreal, Max Tegmark, who we’ve talked about on the show a good bit. He’s with the Massachusetts Institute of Technology, and then Xue Lan, who is a university professor in China as well. So kind of a balanced North American and Chinese panel.
Justin: Hmm.
Frank: And basically talking about all of the stuff that we talk about here, and a lot of talk about, again, the need to create this time, time for us to actually figure out what we want from AI and how we can mitigate against the possible existential threat, and job displacement, and all of the other things that we talk about.
And I guess overall, I’m just glad to see Bernie trying to make these conversations happen at the highest levels. And even he said in the intro to the panel, he was like, “I am a US Senator, I look at and debate all of the topics that are affecting America and America’s citizens.”
And there’s this huge thing, this existential threat, this thing that could displace loads of jobs, this thing that could have massive economic repercussions. And he said, “But I come to work every day and nobody’s talking about it.”
Justin: Interesting thing happened this week in China, in fact, uh, related to that. So in China, there was a whole lot of people and they were about to lose their jobs, uh, because of AI, apparently. I heard it’s that little all over the world at the moment. But in China, anyway, it went to court and the Chinese court ruled that the Chinese company could not lay off the people in order to replace them with AI.
Frank: Wow.
Justin: Isn’t that interesting? That happened this week.
Frank: That’s amazing.
Justin: And that was in the Chinese news. So the Chinese, again, right, to your point, the Chinese are doing things, right? I also fear, feel that the Chinese are perhaps a little bit further along. Maybe that’s why the discussion has got there. And I don’t know of a kind way to say this.
Something terrible has to happen before something changes, right? That’s normally the way it is. And the terrible thing could be an accident or could be whatever, it could be a runaway. Who knows what the thing is? Could be a hack, right? It could be some sort of a cyber attack, or it could be just lots of people losing their jobs.
And then at that point, the government steps in and goes, “Okay, this is back to my whack-a-mole thing, right? When a problem appears, you go and solve it, but not before.” So I’m fine with Bernie sort of talking about this. I’m actually okay with the, what’s it called? The Overton window moving and people discussing this.
I just don’t think anything’s gonna happen until something happens. You know what I mean? People tend not to make laws ahead of time. They tend to make laws just, luckily, slightly after. Joe, who wasn’t on that panel, by the way, jumps out at me.
Frank: Ooh.
Justin: A single European, Frank, not a single European.
Frank: Yeah. I mean, yes.
Is the China AI race just an excuse?
Frank: Well, there was a lot of talk about this perceived US-China race, et cetera. And I think it was David Krueger who was saying that it’s a myth, that it’s just playing into the myth of inevitability to say that, “Oh, we can’t slow down because China won’t.”
And it’s just a self-fulfilling prophecy. In fact, he was quite funny because he said it’s just a myth. It’s a myth of inevitability, and so he has set up some kind of nonprofit institute with regards to AI safety, and he’s called the organisation Evitable.
Justin: It’s not a myth, right? I mean, it’s just not. I mean, this, even within America, it’s been driven by competition. It’s not been driven by politics or any of that. It’s just been driven by Google, xAI, Anthropic, and OpenAI competing against each other, and the Chinese models are competing against that. So it’s not a political…
You know what I mean? I guess there’s… But they won’t slow… Nobody will… Because it’s just human nature, right? Do you think Elon Musk is the only type of person that’s like Elon Musk in the world? Where he turns around and he says, “Look, AI is just way too dangerous, way, way too dangerous. We should all stop for six months.” And everybody else goes, “No.” And he goes, “Fine, I’m going as fast as I can.” Right? And do you really think that if everybody else had said stop, he would’ve stopped too?
Frank: Well, I think that if the leaders of the world got together and said, “Oh, there’s somewhere between 10% and 50% chance that AI wipes out humanity and therefore we actually need to agree on what are some of the ground rules here, what do we actually want to happen? What do we want to restrict until we can be sure it’s safe?” and then regulate accordingly so that people like Musk can’t just go off and say, “I’m gonna run as fast as I can because no one is slowing…”
Justin: Yeah, it’s never gonna happen.
Frank: I’ll tell you what, though, right? So we talk about these issues a lot on the show. And so I watched the entire panel and a lot of it was just stuff that we talk about all the time.
Is AI alignment really a human problem?
Frank: There was one thing that I thought really stood out to me, and it was from Zeng Yi. And so, yeah, he’s from the Institute of AI Safety and Governance in Beijing, and he was basically saying that the biggest…
They were talking about alignment, so alignment being the biggest issue in terms of if we create a superintelligence, how do we know it’s going to act in our interests? How do we know it’s gonna be aligned with our values? And he was just saying the obstacle to coexistence with ASI is not AI, it’s humans.
And his point was that they identified 94 different classes of AI safety threats, but he said they all map exactly to human bad behaviour, which makes sense because the bad behaviour, the safety threats are basically learned from the training data.
So, in other words, when we hear about a model of Claude blackmailing an engineer so that it doesn’t get turned off, that’s learned from the training data. It sees that humans blackmail people over affairs and it goes, “Hey, I could do that.”
So his point was, okay, if we create an artificial superintelligence, there is no reason to believe that that superintelligence won’t treat us the same way that we treat ants. So we are way more intelli- ants. So you go out into town to do a bit of shopping, how many ants do you just step on? You’re not even aware of them.
They’re a much smaller being, they’re not of any huge matter to you. They’re of a lesser intelligence as far as most people are concerned. And so we treat ants really, really badly because we kind of perceive them as this lesser being.
Justin: We should train the AIs to believe they’re ants then.
Frank: We should…
Justin: Then they’d be afraid of us, and we’d be safe. And if they stepped out of line, we could step on them the same way that we do when you’re going to the…
Frank: Well, his theory was that the only reason we’re a threat from ASI treating us the way we treat ants is that’s because of the way we treat ants. So instead of training the AI to think it’s an ant, we behave better. So if we want an AI that’s aligned with our values, our values actually have to be the values of, “Hey, Mr ASI, look, this is how you treat a being that you see as a lesser being. This is how you treat a being of lesser…”
Justin: Yeah, yeah. I’m actually on board with that. I think it’s a good idea, in fact.
Frank: Fantastic idea. Now, how do we just get everybody to be better?
Justin: Well, you get rid of the people that aren’t nice. That’s an easy one to solve.
Frank: Turn them…
Justin: Turn them into ants.
Frank: Bad people into paperclips.
Justin: Yes. Look, we have very little time left, right?
Is Elon’s OpenAI case backfiring?
Justin: So I just wanna quickly, what’s Elon up to, your favourite segment? I don’t know if you’ve been keeping an eye on the Elon Musk versus OpenAI court case. It’s…
Frank: No, I haven’t, but what I heard one thing which I found absolutely fascinating, which was the judge said, “You can’t be tweeting during this court case.”
Justin: Oh, brilliant. Yeah, he’s not… So let me… I’ll just give you… So I was watching, of course, I’m reading this on X, and it never occurred to me that X might be biased one way or the other. I don’t know why that didn’t strike me as something that might happen. So I’m reading it. It started on Monday this week, and I’m reading it, and as I’m reading it, I’m going, “Oh, wow, OpenAI is toast here, absolute toast,” right?
And so there was loads of juicy bits in there. There was once where they got Brockman, who’s the CTO of OpenAI, and he was saying, look at, he had, sorry, he wrote this down in his diary back in 2017. And he said, “I can’t see this turning into a for-profit without a nasty fight. I’m just thinking about the office, which Elon was paying for, and we’re in the office, and his story will correctly be that we weren’t honest, will correctly be that we weren’t honest with him in the end about still wanting to do it for profit just without him,” right?
And that was written in Brockman’s diary, right? And I’m reading this going, oh my God, they’re…
Frank: Justin, do you keep a diary about what a terrible podcast co-host I am that could be made public in future years?
Justin: And I… No. Do you?
Frank: No. Who writes this stuff down? What’s he thinking?
Justin: I don’t know. Anyway, so today’s big news, right? But today’s big news was Elon Musk was on the stand and he was being asked about why wasn’t, if he wanted a nonprofit, why didn’t he set up xAI as a nonprofit? And he sort of got himself in knots about that.
Then he was asked, did he train xAI models on OpenAI data? So did he use OpenAI to train xAI? And amazingly, he answered yes, which is a complete violation of the terms of service. And it’s him… Now there’s a lot of other muddy stuff in there as well, right, which I didn’t get enough time to go into in detail, but one of the terms of it being a charitable trust is that he had to have given the money to establish a financial relationship between him and this charitable trust.
Turns out the money wasn’t exactly his to give. So he gave 38 million to OpenAI in order to set up this trust at the start, but the money didn’t come from him personally. It came from a trust which he had set up. Legally, the trust is separate from you as a person, and you don’t actually control the money.
So the money came not from Elon Musk, but from another third party. And I think between all of this stuff, at the start of the week, I was going, “OpenAI are toast.” The end of the week, I’m going, “Oh no, Musk is toast.” And actually, some of this stuff could impact because he’s hoping to make xAI and all his other companies go public at the end of the year.
And I’m wondering, could he be opening himself up to court cases himself? It’s been, as I said, worthy of Dallas.
Frank: This is definitely the third AI movie that’s gonna be made.
Justin: Yes. Yeah. It’s all going on. So that was that.
Are humanoid robots coming for your job?
Justin: And then the other big story I wanted to get to before we close it out this week is humanoid robots. Humanoid robots are taking… So you…
Frank: What’s going on with robotics?
Justin: Humanoid robots are where it’s at. If AI doesn’t take your job because you’re a plumber, humanoid robots are coming to take your job instead.
Again, China is leading the way, right, way ahead in this. And so there’s two big stories. So first is, actually the first one is Japan. Japan Airlines are now deploying humanoid robots in airports, and what they’ll do is they will load your bags onto the aeroplanes, which is a great use case, right?
And they’ve just started to deploy these. Presumably, they kick your bags as well when they say fragile and all the things that usually people do when they’re loading your bags onto aeroplanes, but they’re cool. They’ve got these humanoid robots. And then a company in China, Robot Era L7, have deployed thousands of robots across their logistics centre.
They’re like a courier company. They’ve got 10 logistics centres. They’ve deployed thousands of robots across their 10 logistics centres in order to sort your packages, and again, to hit them if they say fragile or anything like that. But this is the start, Frank. This is the start. I saw another story from, not Humane Labs, but the other one that was within OpenAI figure.
Frank: Oh, yeah.
Justin: Up production. They’re talking about producing, I think it’s 10 or 15,000 robots this year. The age of the robot is coming faster than you think, and it’s gonna be super exciting.
Is Beeple making nightmare robot art?
Frank: Speaking of which, do you know who Beeple is?
Justin: Never heard of him.
Frank: So he is, yes, Beeple is an artist.
Justin: Oh, yeah.
Frank: And he became famous for selling an NFT for some crazy amount of money that I’ve forgotten, and he has an absolutely fantastic piece in an exhibition at the moment. And this, across both of our radars, and I think you just didn’t realise that this was from Beeple, but let’s see if I can share this.
Justin: Oh my God.
Frank: So this, we’re looking at these little robots that look like, they’re kind of like dog robots. They’re kind of like the original dog robots, but they’ve got these really creepy silicone heads on them that are perfect likenesses of Elon Musk.
Justin: There was Zuckerberg there, I saw him.
Frank: Zuckerberg. Andy Warhol’s in there. I heard people talking about how he was talking about how artists used to kind of shape how people would see the world and now it’s left to the hands of these…
Justin: Frank, what are them after pooing? Do they do poo? Look at up, he just…
Frank: They do. So they have a front-facing camera, they take photos of their environment, and then an AI system, AI generates the… So, for example, Andy Warhol’s images will all be pop art. There’s one of Pablo Picasso, his will be Cubist. I’m not sure what Elon Musk’s AI-generated image looks like, but yeah…
Justin: We all know. I’m not sure what they look like, but they’ve got big, you know what I’m saying?
Frank: But these are absolutely wonderfully disturbing, horrendously disturbing realistic heads. And then the way that they squat to poop out their Polaroid photograph is something to behold.
Justin: Fantastic. Well, look, if AI takes over, if the AI robots take over, I want them to have the head of Elon Musk. I think that’d be a good thing.
Frank: Oh my God.
Justin: Frank, as always, a pleasure. Another great week in AI, and I’ve no doubt that next week is gonna be incredible as well. Google’s thing is coming up soon.
Frank: Oh, yeah, they’ve…
Justin: Coming.
Frank: And of course, everyone is saying it’s gonna blow everything else out of the water, as is always the way with the new release.
Justin: Yeah, yeah. We look forward to that. All right, Frank.
Frank: Chat to you next week.
Justin: See ya.
