Frank: I think the midterms are good news for AI regulation in the US

Justin: this idea that they're gonna slow down or pause AI, that's just, it's, it's not gonna happen. Hello. Good morning, good afternoon, good evening, and welcome to the AI Argument. I am your resident techno over-optimist, Justin Collery, joined by the ever careful, ever monitoring, ever watchful, and I would say pretty happy this week, Frank Prendergast. It's been a pretty good week for you. Let's get straight into it.

Will the midterms slow down AI?

Justin: So, uh, look, as the markets have gone south, things are in- my poor pension fund is sweating at the moment, and this has all been brought about by the midterms in America, where a whole load of people have been voted in that think that AI should be regulated as AI. In fact, there's a, there's a lot of calls for, in fact, a pause, a pause in AI development, and they'd like to see this whole thing slowed down.

Um, this is not a political show. We don't actually care about, uh, who got elected and who didn't. All we care about is, am I gonna keep getting my toys every week like I am at the moment? Um, what's your view on the, the, the impact of the midterms on AI?

Frank: Yeah. So look, normally I know very little about American politics apart from, you know, the occasional, um, presidential election. So it was fun to try and go deep into the midterms and figure out what this was all about.

What did pro-AI super PACs buy?

Frank: Um, I thought TechCrunch had a really interesting story about just how much money was spent trying to influence AI policy in the midterms.

Um, a lot of it-- Well, uh, sorry, the, the story-- Well, the story focused on a crowd called Leading the Future, and they're backed by OpenAI President Greg Brockman,

Justin: Yes

Frank: also the VC firm Andreessen Horowitz. So I always find Greg Brockman a fascinating character because, you know, OpenAI say, "Oh yeah, we want regulation.

Every-- Please regulate us. Please regulate us before we kill everybody." And then meanwhile, Greg Brockman is putting millions into these super PACs that are anti-AI regulation. Um, what's interesting as well about Leading the Future is like they back both AI-friendly Republicans and Democrats, so they actually have two different arms, uh, to do so.

And between the two, they spent about 34.6 million, and that was just one of those super PACs.

Justin: Ok

Frank: One of the things that I found really interesting about this, right, was as I was reading all these articles about these super PACs and the way that they support these AI-friendly politicians, they don't talk about AI.

What they do is they run ads that are, um, basically talking about, you know, whatever the popular issues for each candidate might be. They do not talk about AI. Whereas on the other hand, you also have lobbying groups for AI safety, and they do frequently call out the dangers specifically of AI and data centres.

And so that, uh, that imbalance I found very interesting because the pro-AI lobbyists don't try and sell this dream of the abundant AI utopia because basically, I think nobody trusts or believes them

Justin: Well, it's great. Yeah, it's interesting. So the-- I do think that the concerns on the anti-- Let's call them anti-AI. Is it okay to call them anti-AI? Let's call them anti-AI.

Frank: For simplicity

Why are voters wary of data centres?

Justin: Yeah, the concerns on the anti-AI side are unreasonable, actually, some of them. So, and what-- The, the ads that I'd seen was they were campaigning on the grounds of, you know, no more data centres unless, you know, the gu- the water within the area will not be impacted.

Now, I don't think that's a big deal, right? I think that's a, I think that's urban myth bunk, right? I don't think modern data centres actually use as much water as people think. So yeah, whatever, fine. We'll, we'll guarantee you that one. Then the other one they campaign on is electricity. No increase in electricity prices, uh, for regular users as a result of AI.

I mean, I don't think that's unreasonable, to be honest with you. I mean, that's, that's okay. Yeah, that just points to the need to build lots of nuclear power stations. We should be doing the same thing and then sticking huge data centres beside them, but that's an argument for another day. Um, so, uh, so but the, the concerning-- Well, the-- Okay, and I'll give you the good and bad from what I saw, right?

So I th- I think it was Bernie Sanders was like, you know, "We need to stop this madness now. We need an AI pause. We need to ban super intelligence," all of this sort of stuff. And then a whole load of new incoming, uh, senators were saying, "Absolutely, yes, totally agree," all that sort of stuff. They're not going to ban super intelligence.

It's not gonna happen, right? So that's a pipe dream. And, uh, and the other thing that was interesting, right, which is where it's all very well saying something, but actually doing it is different. So Texas, uh, Texas went blue from being red, whatever, and, um, uh, you know, there would've been-- Elon Musk, of course, is building a whole load of data centres, um, in Texas, and these people were campaigning, um, on stopping the data centres.

But after they got elected, they clarified that they're not gonna stop any data centres that have already been approved. They're not gonna stop any data centres that are currently being built. It's just that any future data centres are going to go through a more rigorous, um, application process in order to get the data centre approved, and in that, they're gonna make sure that it doesn't impact water supplies, electricity prices, all that sort of good stuff.

Which is, I think it's fine, right? I, I don't really have a-- I, I'm fine with people having a say in what gets built in their local area, so that's not a concern. The-- But just the, this i- this idea that they're gonna slow down or pause AI, that's just, it's, it's not gonna happen.

Frank: Well, we'll see.

How did data centres fare at the polls?

Frank: I, I think there was some interesting stuff along the lines of what you were talking about there. So this crowd Leading the Future, they, you know, they back people who are pro-AI on both sides of the fence, and it seems from what I was reading, right, they mostly backed like safe candidates who probably would've won anyway.

But in the places where they were, um, backing people going for like con- you know, hotly contested seats, they, uh, they did back some losing, mostly Republican candidates. 'Cause it, it, it's not fair to say that this is like a Republican-Democrat, um, argument. I think there are people pro and anti on both sides actually, but a lot of more anti-AI, uh, or more safety, more AI safety Democrats ended up getting seats.

So, um, Leading the Future put more than 2 million into a South Carolina candidate called Darlene Graham, who's apparently known as Data Centre Darlene, and

Justin: it.

Frank: she lost to someone called Annie Andrews, who was proposing a two-year moratorium on data centre construction, like, just like you were talking about.

Then we had Republican Laurie Bookout, she lost to a Democrat who also had anti-data centre policies. Aaron Flint lost to Sam Forstag, who was a Democrat who advocated for strict federal oversight of AI. And overall, so The Washington Post basically characterized it as voters rejecting data centre construction in their backyards and electing Democrats who are outspoken about catastrophic risk from AI.

Apparently, there was 41 data centre-related issues in total on the ballot across the United States, and 40 of the results were basically against building further data centres or expanding them

Justin: I think, right, uh, so what's your view? Because I think that in the public zeitgeist, the, uh, the issue isn't the catastrophic risk of AI really. I think people are probably more c- exercised about a great big data centre being built right next door to them, and the noise that that generates and the electricity it's gonna use.

I'm not even sure they care about much else. Like, it's, it's, it's, it's a form of NIMBYism, I would've said

Frank: So fif-- the, NBC did exit polls and there was 56% said AI and data centres played a major role in their vote. That could still be mostly the data centres, um, as kind of as you're saying. But 57% believed AI would do more harm than good to society. So I think, I do think there's more here than just the data centres, and I think that, um, just ju- judging from some of the kind of individual comments that I read that people were saying as they exited the polls, I do think people believe that there's a lot of corruption going on and a lot of cozying up of the billionaire tech oligarchs with the current administration, and they're concerned about that as well

Is Anthropic speeding up its own AI research?

Justin: All right, let me explain why it's not gonna happen. Um, so Anthropic, uh, who are supposed to be going public, but I guess with all the mar-market turmoil that, that might be postponed for a certain amount of time. Um, Anthropic let this company called Meter in to do an evaluation of how, how much the use of AI has sped up the development of AI.

So M-E-T-R, Meter, they're the company, and they release loads of really good reports about time horizons and how good AI is at doing long-run tasks and so on. They were given access to everything within AI p- uh, within Anthropic, right? They were given access to models which were not released. They were given access to people.

They were given access to processes, the whole shooting match, the whole gambit. The report which they released was heavily redacted, but here's the most important part of that report. What they evaluated was the speed of AI development using the latest tools, so the latest models and the latest harness within Anthropic.

And what they discovered was that the speed of AI de- uh, development in the last six months had doubled as a result of using those models and harnesses. So if you can imagine a graph, right? 'Cause they have these graphs, they measure with a, it's, it's, what is it? AIEC, I think it is. It's the, it's a measure of intelligence of the models.

And if you plot that over time, it's like, it's a beautiful straight line, just like the scientists had predicted. But then over the last six months, they have these new, they call it Model two, Model three, and Model four, which are models that we haven't seen yet in the public, and the slope of the graph changes, and it's gone up, and it's a two X increase.

And, uh, you can-- Right? So that's happened in the last six months. So you can imagine the next six months, that two X will become four X because they'll use those ones to increase it again. And six months after that, that four X will become eight X because they'll use those same model. This is recursive self-improvement.

And so the thing that we had thought about many years ago, right, we had talked about this many years ago, and everybody thought it was gonna be OpenAI, was once one company gets out into the front, because they're able to use their models to improve their models, they would then be uncatchable. And this is now happening.

It's taken two or three years, which is longer than I think any of us expected. So this is now happening, right? And so if you have a pause for, even for a year now and say, "No, no more development," you're handing that lead. It's a, it's, it's, you know, everybody knows that can be done now. So all the other research, not in Europe, obviously, 'cause we don't do AI in Europe, but all the other, uh, AI companies throughout the world will be sort of aiming for the same thing, knowing that it can be done, which is always much easier if you know it can be done.

And if they get there while you're in your pause, God, I was gonna say something rude there, to that during your pause, right? You will never catch up again. You have ceded the lead. You've done an OpenAI to the rest of the world. You had the, a lead and, and, and, and you've handed it over to your competitors

Frank: You know what I'm gonna say though?

What can Congress do about AI safety?

Frank: I, I'm going to say that, um, I think that the midterms are good news, and we'll look at exactly why in a minute, but I, I think the midterms are good news for AI regulation in the US, which, uh, as you know, I think has been sorely missing and lacking. Um, but I think beyond the US, hopefully it also means that there will be more efforts for international cooperation, because as you say, that's what's gonna be vital.

And I'm not saying it's gonna be easy. It's not. Um, it's gonna be incredibly difficult, and I don't think that, you know, international pinky promises are gonna do it either. I think we need to look at just as much as we've been accelerating the AI along the terms that you've been talking about, we need to accelerate the research into how to go beyond pinky promises and look at what are the ways that we can detect when somebody is breaking the international agreements on building a superintelligence.

Um-

Justin: Good luck with that.

Frank: Thank you. We will need it. I admit we will need it.

Justin: I just, I mean, even, so even if you had an international agreement, right, that doesn't stop people fr- you, you still end up in the situation where if you turn ar- and I just, it's just, I just cannot see this happening, right? I, I don't see in any world that this happens. So let's say through some miracle that they say, "Okay, well, in, in the US we're gonna pause, um, development of AI for the next 12 months to get a better handle on safety."

That's not gonna happen, but let's say it did happen. The other g- the other people in the world, like the Chinese aren't gonna stop development for six to 12 months. They're gonna go, "No, we're behind. We're gonna try and catch up, and if we

Frank: But that's why we--

Justin: how bad?"

Frank: that's why we need the international agreements and the technical agreements to, you know, have things in chips or have detections in data centres, et cetera. But look, we started talking about the midterms, so, uh, let's just come back to the specific, like what does it, what does it actually mean now that we have some people in the US government who are actually interested in AI safety?

Um, so I had a look into what people were saying, and the things that we can expect are, I think probably the most immediate thing is we're going to see, uh, hearings and we're gonna see the, the, you know, AI executives and CEOs test- uh, testifying on safety incidents and model testing and things like that.

And they will happen because they don't require the president to agree to them. They, they can just-- they can, they can go ahead. Um, we then, in terms of safety legislation, that's a bigger ask because it'll definitely be pushed for more. However, I think, um, I think Donald Trump has already been posting images of the word veto on his, um, Truth Social.

So he's, he's gonna veto anything that, that they try to push through. Um, but

Justin: Can I just say, sorry, can I ju-just say on the, on the-- So yeah, so as I said, right, I don't think anything will change 'cause yeah, you've got a big fat veto going on there. But the previous point, right, about hearings and about, you know, if there's an incident, right, they can bring pe- I think that's totally cool.

I think that's a good idea, right? If we have security incidences and, you know, the mo- And I think the companies themselves, by and large, are getting much better at talking about, you know, the stuff that happened in Hugging Face. They released a report, they got an outside body in, right? So they, I think they get it, right?

I think the, I think the, uh, the model companies understand the risk. I think where the blind spot in all of this is, it's not the big model companies that you have to fear. They've got vested interests in making sure that these things are secure. It's the smaller companies, it's the bad actors, it's the people who download open weight models and then sort of do things to them to make them, you know, more dangerous than-- They're the people that are gonna bite you in the ass in the next six to 12 and 18 months, and those models don't typically come from the US.

So it doesn't matter who you haul in front of the, the Senate or whoever it is that they haul them in. It's not the people that are doing the damage. But having said that, I still think it's good that in, if, you know, one of the big US companies have an incident, that we all get to find out about it

Frank: Yeah, and even in terms of public awareness and understanding of what's going on, I think it, it'll be huge. Um, and I think that one of the last important points, 'cause I mean, we already know obviously data centres and AI safety, they're generally gonna become more politically relevant. Um, but the other interesting thing I thought was that they might now be able to resist-- Remember, remember the, the current administration was basically saying, "You cannot regulate AI at a, at a state level.

It has to be federal." And so some of the analysis I was reading was saying, well, that will now be resisted and it's quite possible that individual states will be, will retain their ability to regulate AI individually.

Could Musk put data centres in space?

Justin: I did note as well, actually, there was a tweet from Elon Musk, uh, after, 'cause again, he's building out his data centres in Texas, and his comment was that data centres in space just got a whole lot more interesting, which I I guess is true

Frank: which Google, Google, I think we were talking about on the show there recently, Google have already started exploring, well, what would it take? And their initial estimate is it'll be like 100 times more expensive than an earthbound data centre. And I think their, their comment was, "Yeah, let's explore every opportunity on, on Earth before we go to space."

Justin: A little bit of nerd out actually, just a little bit of nerdy detail, right? So the big challenge on data centres in space is that space is an insulator. So if I'm hot, right, and I want to get cold, the, the heat goes into the air. Like I, I've-- whatever that is, convection, conviction, I can't remember, right?

All those things from junior cert science. But anyway, you sort of, you-- in space, it's very hard to get rid of heat, right? Because space is a vacuum and there's nothing to transmit the heat, so you have to emit it as radiation. And people are like going, "Yeah, it's a stupid idea." You know, these GPUs generate loads of heat.

But actually, Elon's point is that you can radiate it as heat, but the surface area to get the energy in the first place is four times as big as the surface area to, uh, get rid of the heat. So if you create a huge solar array, right? And then on the other side of it, you've got the thing which gets rid of the heat, the solar array is always gonna be bigger than the thing that gets rid of the heat, and therefore it does make sense to put the, the GPUs and the data centres in space.

That's his argument. I'm not a rocket scientist

Frank: Yeah. Okay. I'll, I'll just-- I, I'll believe him this one time.

Justin: Okay, cool.

Can we keep up with OpenAI’s Linux bug hunt?

Justin: So updates on secur- yes, security. Let's talk about security, right?

Frank: yeah, so last-- I heard, I heard last month referred to as Hacktober

Justin: And this is going to be, oh my God, November. Um, yeah, so OpenAI this week have done everybody in the community, um, a big favour. So they turned their latest and greatest security model at, uh, Linux, which is Linux is the operating system which runs pretty much the entire website or the entire World Wide Web.

So when you connect to a web server, it's probably running Linux. So they turned their latest security model at Linux and said, "Go find all the bugs, all the security holes, you know, and, and fix them." Normally, you might find 10, you might find 20. This particular model found just under 1,000, 948 security vulnerabilities on the software that runs the entire internet, and they fixed them.

So that's cool. Now, but obviously that raises the question, like, how are we going to patch all this software? Like, what's going to happen? Like, there's all

Frank: was going to say, so

Justin: out there

Frank: so they've, they've fixed them in terms of they've come up with the solutions, but now they have to make sure that those solutions get to everybody

Justin: Yes. Now let me just explain to you how vital these fixes are. So these aren't like bugs where, you know, you get your app and your app crashes or whatever. It's not those types of bugs. These are bugs where I can send messages to your server and suddenly I've got what's called root access to your server, and then I, I can control your server.

I can access it. I can do anything I want to it. Let me give you an example of one of the security flaws which was found. It was beside this one, right? But it's similar. Same idea, right? So this is in your home router, the one that I always worry about. Your home router runs like a little Linux server within it.

That's how it works. They found... This was separate, right? But a bug was found where if I wanted to access your home router, the username is called root, by the way. That's the user that can do everything on your home router. So if I tried to log in as root and I gave it a random string, so a random password, and so long as it was more than 256 characters, I get in. It's the simplest hack that nobody-- And it's been in... Yes, and that's been in the software for 15 years, that particular bug. Nobody found it yet. Now, it's not on every Linux server, right? It's not on your web server, but it's on a sp- specific flavour of Linux. These are the types of... And that was not a, that wasn't written by AI.

It's not AI stop. That was written by a human

Frank: But I'm sorry, this information is now out there, but the patch hasn't necessarily been rolled out everywhere?

Justin: Uh, that's part of the problem, right? So Ubuntu is a very popular Linux operating system. There is a known, uh, problem with Multipath TCP, which I had a company to do with many years ago, and so I'm very familiar with it, and all of the Linux, uh, distributions have patched Multipath TCP except Ubuntu.

Everybody's known about this bug for about a month now, and it allows we-- Multipath TCP, by the way, is the thing that makes, um, FaceTime work. It makes-- It uses your cellular network and your Wi-Fi at the same time to make your internet faster and better and all that sort of good stuff. So, so the, so the issue-- And this comes-- Okay, so here's my big thing, right?

Is cybercrime turning us all into preppers?

Justin: Here's my social commentary on all of this, right? So I was a bit freaked out this week. So we've been getting these messages. So everybody in Ireland got this leaflet about a month ago, which was, you know, "Be disaster ready" or, "Be prepared." It was like, "Be a good Boy Scout," right? And, you know, have three days of food and three days of water and all that sort of good stuff.

And I go, "Yeah, whatever. Grand. I don't..." Some, some civil servant has decided this is a good idea. Whatever." I didn't really pay much attention to it. And then I saw the same or a similar leaflet went out in France, and then another one in Germany, and then another one in Italy. And I'm going, "Hmm, that sounds a bit more coordinated."

And then I'm scrolling through Twitter this week, and the European Union has said I better have a go bag ready, and in my go bag, I need to have three days of water and three days of food and whatever it is. I go, "This is serious." Like, so what-- And the scary thing, we actually-- It came, it's in the zeitgeist because my kids were talking to me about it, and it's kind of what is, what is everybody-- The, the thing that's most worried about this is they don't tell you what the thing to worry about is, right?

It's just like, "Be prepared." For what? "We can't tell you." It's like, well, that makes it even worse. And so there's loads of theories going around. My working hypothesis, and I could be totally wrong, but my working hypothesis is that, you know, governments and people who are in charge, uh, people who look after us basically, um, know that OpenAI have discovered 948 vulnerabilities.

They know, and they see all the other things that are happening at the moment, and it's like the, you know, software is in a dangerous place for the next six or 12 months. There could be a hack on supermarkets like we had this week in SuperValu. There could be a hack on a power station. There could be a hack on your bank.

You know, keep some cash around because maybe the ATMs will stop working, and it could happen for a couple of days. We can't tell you what the danger is because we don't know where that hack would come from. Like, it's, it's, the vulnerabilities are everywhere, and they're very diffuse, and until everything gets patched, just be prepared.

But they aren't saying that, so I could be wrong. But you should be worried, Frank.

Frank: are saying, they are actually saying that though as well. I mean, they, you know, they weren't super specific, but Ir- so it turns out Ireland's leaflet was one of the least specific leaflets that went out. And this is, you're right, it is coordinated 'cause it is part of a European preparedness operation, whatever the word is, programme.

Um, and most countries, uh, or sorry, not most countries, but, um, a lot of countries were actually given this, like, huge long list. I came across one, I think it was Finland. Let's see if I can find it. Finland had, uh, 19 different possible types of emergencies, and then their advice on their website, you could actually filter the advice by the type of emergency you were most worried about.

Justin: No way, really?

Frank: Yeah.

Justin: ridiculous. Like that, that's, that-- Okay, that's worse because that's like you don't know what's gonna happen, and so you're just, "Ugh, I don't

Frank: The Ire- the Irish government did some, um, uh, they put together their first leaflet and then they got, um, you know, a, a, a focus group to tell them what they thought of it, and the focus group were like, "It's too scary." So, so the I- so the Irish government toned it all down and we got this, like, ridiculously vague thing. like, you're right, like cybersecurity is 100% one of the laundry list of things that Europe is concerned about right now

Would criminals pay to train hacking AI?

Justin: So, but just to, to the point, 'cause we were talking earlier about, you know, safety concerns, the American midterms, all this sort of stuff, right? So this week, Young Sung, who's a researcher from South Korea, and she does research on AI and all that sort of stuff, doesn't work for one of the big labs, but easily could, right, is a brilliant researcher, reported that she had been offered obscene amounts of money on the d- from people on the dark web to train models to offensively attack systems, right?

So these would be, you can imagine, they would download a model, um, from the internet, one of the big open source models, and then they would do things to that model to make it better at cyberattacks, and then launch those cyberattacks to do things like the, the Binance hack that happened recently, and also this week there was a French bank, what was it called?

Credit Mutuel, uh, got hacked this week as well. And so the prob- the people that you need to be afraid of are not the big US labs. They're not about to do you any harm. The people you need to be afraid of are the people in the shadows, the people who are taking the technology and using it offline to develop better models to go and attack things like power stations and banks and whatever, supermarkets, as it happens, um, this week.

Um, so yeah, lookit, we are in the midst of the difficult teenage years. We are at the foothills of the singularity. This is going to happen. It will all be fine. Interestingly, right, I'll say one more nerdy thing, and then I'll let you come back. Uh, interestingly, I've been reading some researchers-- So I said Linux is the thing that runs the entire internet, and, uh, people, this seems very true to me.

This rings true. They were saying that, you know, 12 months from now, we might have a totally new operating system that runs the internet, because Linux came from a system called Unix, and Unix was built for universities in the 1960s. So the whole security architecture and how the sec- the thought around security went into that was for a totally different environment than that which exists today.

And actually, the task of patching every single hole might just be too big, and it might be that, you know, we just basically start from the ground up again and say, "Here's a new operating system, looks a lot like Linux, but isn't, and is secure, and this is what you should use instead." That could happen.

Frank: But in the meantime, everyone should put their go bag together because we all have to be preppers now

Justin: Uh, actually, it's, the weird thing is, this is like I am the most optimistic techno futurist whatever amount there, and I'm not going, "Yeah, we probably should get a couple of bottles of water and some tins of fish and stuff like that." It's weird. Like, what a weird time to be alive, but it'll be fine

Frank: It'll be fine.

What other stories deserve a mention this week?

Frank: So Justin, we have a new segment of the show. We're gonna go into fast mode, fast mode where we turn off, we turn off our reasoning models and, uh, and we just give the bare facts about a couple of stories that we thought were big enough to include, but not big enough to get into a big debate about.

Justin: Go for it. What do you got?

Frank: The first one that caught my eye was this thing where GPT-6 Astra, so that's OpenAI's biggest model, apparently, um, these researchers were able to use it to diagnose and repair a problem inside another model. Uh, it was a Qwen, a Chinese model, a Qwen model, uh, by directly modifying some of its weights.

So essentially, Qwen, when it was in reasoning mode, would get caught in a reasoning loop and just go round and round in circles, and Astra was able to figure that out, go in, change the weights, so no t- no further training needed. Um, just go in there, change the weights, fix the problem. So right now, this is pretty preliminary, but what people are saying is, well, imagine a world now where you have a model and instead of having to like train a big new model, this like could be an indication that yes, we could get to recursive self-improvement, where a model is actually changing its weights on the fly to address problems, um, or, or one model changing the weights of another model to address problems and improve the model without needing to train a new model.

Justin: Super exciting. US company BioComputing Company has released a new image generation called Mischief. Now, this one is cool because what they do is they get a chip, they put it into a computer, but in that chip is real live human brain cells. And then what they do is they feed images through this sort of bio chip, and they watch how all of the neurons fire.

They study that, and then they take that logic of how the brain cell neurons fire, and they implement that in code and silicon, and that creates an image generation model. The model they created is 80% cheaper and five times faster than any other image generating model on the market. Go for the win

Frank: Wow. Well, I love AI image generators, that's good news, but what a weird world we live in. Uh, copyright in the EU, so this is-- I, I find this fascinating. A Hungarian website operator was taking, uh, or is taking Google to court about copyright infringements in both the outputs of their AI and in the training of the AI.

So they're basically saying, "You're infringing on our copyright when you talk about our materials, and you're infringing on our copyright when you put our materials into your training data." So, uh, a, an advisor to the EU's highest court said that when, in terms of the output, they might be infringing on copyright.

It's okay apparently for the AI output to include facts, but it said that, in inverted commas, "the form of expression is protected." So essentially, if it goes beyond facts, and if the person would never have to visit the original article, then it's probably infringing on copyright and it's not okay. But the advisor said that in terms of the training data, and this is where it gets really interesting, he basically said, "No, the EU's text and data mining exceptions mean that the training on copyright materials like that is permitted."

Justin: Cool. This one, you're getting to get two for the price of one, and this one is super cool. So citizen researchers on the internet this week, right? When we're talking here, there are planes flying around us all the time. And so what they did was they took audio from people on the internet, and they also then used this, uh, website called Flight Radar, which tracks all of the planes.

And what they did was they listened for the very faint signal or the very faint sounds of the planes in the background, and using where the planes were, they were able to identify where on the globe you were just by the background noise of the planes. Isn't that cool?

Frank: wow

Justin: an even cleverer researcher then took that research and said, "Well, that's pretty cool, and there's a lot of planes in the sky, so generally there's a plane around you somewhere," right?

If you listen very carefully, you will probably hear a plane. And so the absence of a plane probably indicates that the video was AI-generated. And so they searched for the presence of planes in the audio of AI-generated videos and used that to determine if the video was real or AI-generated, both of which are cool projects

Frank: That is very clever. And of course, everyone now is going to put a plane track into the background of the audio of their

Justin: The AI

Frank: AI.

Justin: yeah. There's a name for that. It's a law which is that which you measure ceases to be a useful measure. I can't remember what that is.

Frank: yeah, yeah.

Can an AI teacher spell seventeen?

Frank: Uh, so our last-- for our last fun story of the day, have you come across this guy, uh, what's his name? Husk? Is it Husk? Yes. Uh, he has a br- he has a brilliant social media where he just constantly engages with AI, usually in voice mode, sometimes in videos, and, uh, he, he gets AI to act in insane ways.

And he had one recently where he got onto an AI teaching avatar, uh, that was meant to help you with your spelling

Justin: Very good. Could do that myself

Frank: uh, he asked it about spelling, uh, the word seventeen, and and at first it absolutely insisted that seventeen was spelled S-E-V-E-E-N, and therefore had three E's. Let's have a look at a little clip because I think just think that, uh, this particular AI avatar is like something out of a horror movie.

Um, so this is, this is just after he has correctly spelled seventeen to the, um, to the AI avatar

Justin: Allora, adesso look

AI (screenshare): No, no. Oh. No. Oh,

Husk (screenshare): okay. Sorry.

AI (screenshare): Don't apologise. Oh. That's how we learn.

Husk (screenshare): So but seventeen is two words?

AI (screenshare): No, no, no.

Husk (screenshare): Okay.

AI (screenshare): That is not-

Husk (screenshare): Okay, let's

AI (screenshare): No.

Husk (screenshare): Yeah. calling back.

AI (screenshare): The number seventeen is spelled S-E-V-E-N-T-E-E-N.

Husk (screenshare): Okay.

AI (screenshare): So if you count the... Well, I was going to say three, but you cut me off.

Husk (screenshare): I didn't even cut you off.

I just nodded my head.

AI (screenshare): Oh, I see. I thought you were interrupting me. Well, the answer is three.

Husk (screenshare): Would you be mad if I did?

AI (screenshare): Oh, I wouldn't be mad at all. I actually think it's quite char-

I am not going to shut up. Okay. You are my s-

Husk (screenshare): I was just testing if you'd get mad if I interrupted.

AI (screenshare): No, I won't get mad.

Husk (screenshare): Shut up.

AI (screenshare): I am not going to shut up. Y-

Justin: She is scary

Frank: she is very scary. So like for anyone listening and not watching, her ex- her facial expressions when she's correcting him and when she's saying, "No, no, no," she's getting quite intense. And then I find that, that what's really disturbing is the switches between that intensity of like, "No," to suddenly this weird smile and the widened eyes. It's like her

Justin: The eyes in particular. Well, for me, I was looking at the eyes like,

Frank: Well, the s- the smile reminds me of that horror movie. Was it, I think it was actually just called "Smile," was it? Yeah.

Justin: it was good.

Frank: Um, I- what I thought was interesting though was the company behind the avatar were call- are called Anim.a- AI or yeah, I think that's it, Anim.ai. I was wondering had they some Irish connection with that name, but, uh, 'cause Anim is the Irish for name.

Uh, but they're a London-based company apparently, and, uh, they leaned into this. So they leaned into this because obviously people were having similar reactions to me and saying, "Whoa, her expressions are terrifying." And they were like, "Well, she spelled 17 wrong." She did spell it right eventually, but she did spell it wrong.

Uh, but she was super expressive when she said it, and that's their whole business is expressive AI avatars. So they now have a page on their website dedicated to the fact that they got 17 wrong, but she's very expressive. And

Justin: very expressive and very, and very good at convincing people but don't get the facts right, surely that makes them perfect for senior management

Frank: She would definitely, I mean, she would definitely convince me that I'm now-- I'm just gonna spell 17 as S-E-V-E-E-N now if, as long as she's d- as long as she doesn't get mad

Justin: Yeah. It's this AI aligning the humans. Ooh.

Justin: Frank, pleasure as always. Look forward to speaking to you next week

Frank: Brilliant stuff. Chat to you then, Justin