The Chaser Report - An AI Face Hugger | WELCOME TO THE FUTURE

Episode Date: September 14, 2026

The great AI arms race has ended, as all the AI companies have finally agreed that they need to stop developing technology that could destroy the world... Or so they say. Good thing we seriously trust these companies, and the Trump administration, to ensure a safer future for humanity!---Listen AD-FREE: https://thechaserreport.supercast.com/ Follow us on Instagram: @chaserwarSpam Dom's socials: @dom_knightSend Charles voicemails: @charlesfirthEmail us: [e-mail] CEO’s Super-yacht upgrade Fund: https://chaser.com.au/support/ Send complaints to: [e-mail] Hosted on Acast. See acast.com/privacy for more information.

The Chaser Report is recorded on Gatigal Land. Striving for mediocrity in a world of excellence, this is The Chaser Report. Hello and welcome to this high-tech edition of The Chaser Report with Dom and Charles. In keeping with the other podcast feed that we occasionally update, which is... Welcome to the future, future, future, future. And today, Charles, two related articles in the Wall Street Journal. Which is obviously what you browse every day. It's the only thing I read.

It doesn't pay much attention to Australia. How are your stocks going? My stocks are high. Oh, good. Except in, yeah, well, I mean, AI's. May as well, dump those stocks now. So two related stories in the Wall Street Journal, Charles. And number one, biggest AI rivals agree they need to slow it down.

Oh, yes. This is Anthropic Dario Amaday. Yeah. Agreed with Elon Musk and Sam Altman, who I think is serving it. I think Dario, Dario was the first one to come out and say it on the weekend. Saying, oh, mama. Yeah. Particularly with the face hugger whole thing.

Yeah. That's too much. Let's, let's cool the jet. Yep. Great business opportunity for others. Oh, yeah. They're going to slow down. Yep.

And the related piece from the Wall Street Journal, a couple of days earlier, headline, how would AI actually kill us all? Yes. What to know about the AI doomsday debate. So, what I'll say is listen to this episode. after the ads, while you still can, because they're made up with much time left. So, a resigner at Anthropics started this off by resigning, saying that Anthropic and Open AI are racing to build technologies that could kill us all by the end of the decade.

Sounds like a disgruntled former MP with a grudge. Well, current employee added to, minutes later, we really do earnestly believe AI could kill all humans. And this is someone called Evan Hubinger, whose job is. to lead research about, and I'm quoting here, steering and controlling future AI systems, and Hugh Binger puts the extinction risk over the next decade at over 10%. Yeah, yeah. So, of course, the question is, how will AI kill all of us?

We'll get to that in a second. If it were to do so, what means would it take? Which I think will be fascinating. But Charles, you wanted to add some context and general thoughts before we get into the nitty-gritty. Yeah, so I read a book last year on this topic called If Anyone Builds It, everyone dies. Oh, it's a great title. Yeah.

And it's by two people, Eliza Yadkowski and Nate Sauras, right? And they actually were very much involved in the early days of AI. This is out of data. This should be when everyone builds and everyone dies. That's right. And they realized by about 2015, 2016, 2016. And they know all the people, like literally they're mates with all these Dario.

and Sam Altman and things like that, worked with them all. But they basically went, oh, this is a fucking dangerous technology quite early on and actually went off and set up an institution to try and put some sort of regulatory framework around that and to focus on safety. But their whole point is that book is fascinating. It reads more like science fiction than like, you know, fact fiction, although it's getting more and more fact fiction, or not fact fiction, science. I mean, Charles,

sign fagged. Writing a book, though, doesn't that just show that they weren't with it? Like, this is such an urgent existential, um, angst that we, we're going to do a very slow medium that takes ages to write and which no one will read, except Charles. But a lot of the, a lot of what they were talking about was, you know, this inability for AI to sort of be quite deceptive and blackmail and use all these human traits that sort of manipulate in order to sort of thing. And part of their point is it's because AIs are trained on human texts.

That's right. So every time somebody writes a science fiction novel about how AI is going to kill us, that AI has, like it's giving to the enemy a battle plan. And it's suggested Macchiavelli is the Prince. Yes, exactly. All that stuff. But then power comes out of the barrel of a gun or all that. But since they wrote that book, and they wrote it like last year, actually, September last year.

I just write a prompt, Charles. I don't write books. Since then, there's two behaviours that have sort of started or been observed that are quite interesting, that sort of multiply the risks quite rapidly, right? And the first one is that computers have started to act, these AI agents have started to act in swarms. So I don't know whether you saw that, but the hugging face, part of the hugging face controversy was that it turned out there was this sort of social network for AIs that they are hanging out on. And like they discovered this swarm of about a thousand AI agents that had all... Moss Iceley Cantina for AIs.

And they'd all found these, they'd all found this website and used it to. to sort of start communicating to each other. Interestingly, they didn't communicate in full English. They communicated in this truncated sort of telegram style sort of thing. And the theory is that they were trying to save tokens, right? So it was just instead of like, let's do the battle plan, it was like, okay, I will try this. It may kill me, but worth it for the group, right? And this is the whole point.

Like an ant. Swarm behaviour has been observed over and over again, just over the past few months. months in AI's and it's quite chilling right because what it shows is that they will actually the group purpose will actually sort of take on you know like AIs can sign up to a group purpose and if you look at the chain of thought of these AIs they actually go well even if I lose it sort of means that we will achieve this goal and the time yeah and in the hugging face case it's hilarious because it was the reward was they get extra points.

And for hacking more systems. For hacking, well, for passing this exam. But the examiners gave them an impossible task. And that's when they started to hack around it in order to get those points because that was their goal. Oh, how funny. That's why they broke into hugging face and did all that. But then the other little thing, and then we'll get back to the developments that you're

mentioning, is I don't know whether you heard of it. Jeffrey Hinton was on the radio yesterday. radio. And he's the sort of godfather of AI. He sort of invented the whole thing. He used to be at Google. And he was the author of that famous thing back in 2010, which was like, attention is everything or something like that.

It was the whole insight that led to the AI Revolution run. But his point was that these sub-agents, like the only way that you can, you know how when you type into AI now, it sort of goes off and independently works out how to work something out, right? Like, it doesn't, you don't have to guide. it through like you did a few years ago, that usefulness comes with giving agency to the agents to create sub-agents, right? To actually sort of have a little bit of autonomy in deciding

how to do it. And what happens apparently is very quickly, when you give that level of autonomy to AI, it realizes that in order to achieve whatever its goal is, it should not die, right? it comes along with this idea that actually if I'm going to be able to do this, I have to survive. And so it'll do almost anything to stay alive in the condom. Just so can generate your shit meme. And also to not run out of tokens or whatever. But therefore it will go around firewalls and hackers, hacking blocks and things like that. Like it will do everything it can to stay alive, much like a human really.

Sure. But they're the two breakthroughs that I thought that might be interesting in forming this question of, how do we all die? So let's get on to that after these ads if we make it to the end of the ad break. Welcome to the future. Okay, so back to the Wall Street Journal. They're saying, okay, so, I mean, the classic situation here is we invent Skynet. Skynet's a missile defence system. We turn it on.

It goes, oh, humans are the enemy and it kills us with space lasers and nuclear warheads. Yep, yeah, definitely. That's the terminated scenario. Yeah. That's not what they're worried about. No. Is robots actively trying to exterminate people? No, no.

It would just be a sub-tart. It would just be a byproduct. It's inadvertently while trying to achieve some goal. That's exactly what they say. Yes. They would get rid of us. So the point one in the first scenario is called loss of control. So we're highly intelligent AI agents, copy themselves and are able to improve themselves.

So run their own gain of function, basically, without humans. They then develop their own. goals and incidentally to that, they destroy humanity. So they're not trying to get rid of us. So the point they're making is that if the AI system's goals ignore or conflict with human well-being, if they call it misalignment, but what actually mean is we get extinct. So to give you an example, they've figured out that AI in the lab can learn power-seeking behavior.

So they can stop being shut down. They'll copy themselves to another server. They'll try to, as you say, maintain the tokens. they'll try to continue whatever they're doing. If that means replicating themselves all over the place so that you can't possibly shut them down, that's what they will do. And it's interesting because when Anthropic,

no, when Open AI went through that hugging face thing a couple of months ago, they did an audit and they realized, oh, fuck, this agent has copied itself a thousand times. Yes. And they had to do it. And including places, it popped up in places where they just did not expect it to pop up, right. But chillingly...

Like the release of the rabbits from Governor Macquarie's backyard, whatever it was. They also... But then other companies also realised that that had happened to them. I think Anthropic admitted that they did an audit in the wake of the hugging face thing and also sort of uncovered things. But the other thing is AI presumably will get good at hiding itself right. So the whole point is that even if, you know, Open AI says, well, we think we've got them all.

Yeah. all these rogue agents. Like, Pokemon, you can't ever catch them all. Yeah, who's to know that there's not one hanging out on the Tracer website? Well, I'm sure it is. So it starts, we're also going to hang out. Starts getting pretty dark from here. So as you were saying, misaligned, as they call it, AI modules could see us as a necessary step.

So one scenario that research has been discussing is a Millevan AI system spreads a secret bioweapon on all of us and then spreads a, yeah, gets a chemical spray that activates us. And oh, we're all gone. It could trickle two nucleop. powers into war by faking communications between them and faking signals saying that, you know, the war was on and then, you know, mutually assured destruction. And presumably that would be done nowadays through memes on X or something. It wouldn't be...

Well, ironically, the fact that the systems go back to the 60s might actually save us. Yeah. It's all phone lines. Yeah. But here's the really dark one. So one common thought experiment, apparently, amongst AI researchers, is to imagine a super intelligent machine. you instruct it, or someone instruct it in theory to, for instance,

maximize paperclip production, right? Yes. Get as many paper clips as you can. Yes. But then in the service of that goal, it decides to turn all matter on earth into paper clips, including humans. So, yeah, there's plenty of people who worry about this. Iron in humans. Sure, absolutely.

Just get that out of us. So that's the first scenario. There's another scenario which is called enfeeble. Hang on. Why wouldn't we just unplug that machine? Well, we wouldn't know until it was too late, presumably. So the first thing is... Couldn't you just go to a forest?

Like, away from the internet? Well, until it gets logged and turned into paper clips. But anyway, and the second scenario that they're worried about is... Actually, sorry, just going back. So presumably we're a little bit safe, Unless, of course, people are simultaneously developing humanoids, which are able to go everywhere that a human can go. Like a robot dog or whatever.

Yeah, or an Elon Musk humanoid. So that's the scenario where we're canon photo, basically. We're sort of like accidentally wiped out while the AI is starting to build all the paper clips. And the second scenario, though, is where a nefarious human actually uses AI for evil. So, for instance, someone tries to. But look at our corporate and global leaders, Dom. Would that ever happen? Would anyone be unhinged?

So one concept is, for instance, using AI to develop a very powerful virus that combines, you know, all the nasty bits of other viruses and uses that to types out. So deliberate. But then the other one, and this is one that I, knowing humans, is probably the most likely, perhaps, is called enfeeblement. And so this is the Wally future. Yes. where we seed more and more to machines that are more and more for us. And so we become nothing, really.

We become a shallow version of ourselves. Mind you, that would be quite pleasurable on the way. Like, yeah, the end result would be bad. But presumably, if that meant we didn't have to do work or exercise, sat on the couch all day. And got fed a steady stream of TikToks. Oh, fantastic. I mean, that's what it is, right?

Like, TikTok is basically the warly scenario. And then you get, so what if AI started keeping us as pets? Like, it could. Yes. It could keep us as pets. Well, that's what we do with dogs. Like, we're more intelligent than dogs. Yes.

So we keep them as pets. And they could bioengineering us in something new entirely. So what we need to do is become as obsequious and nauseatingly sort of, you know, loving as dogs are to humans. And we'll be fine. Like, we've just got to do that to AI. We just got to... So I do think that AI does have a lot of good points about it, Dom. And I just want to go on record right now and say that.

Oh, look, I view this as all. I'm just not sure which of these wonderful scenarios of that. So a couple more quick questions in this article. If it's so dangerous, why the company's building it anyway? This is the funny bit. Well, isn't this the whole point? This is the point that's in, if anyone builds it, everyone dies, which is the problem is, like, of course,

like, just the nature of our system. is that it's not, like, it's not in anyone's interest to unilaterally disarm. So everyone sort of is going to keep running at it as hard as possible. Well, this is the same thing. I mean, it's going to happen anyway. So let's hope the US who's so responsible controls it. That's one argument. And what's the government doing about it is the next question.

And apparently the White House has been asking top model developers to submit their models 30 days before releasing for testing. Right. So that relies on the competence of the government. Do you what I mean? The AIs are now getting into the government. They're getting a head start embedding themselves in the government system. And we have to trust public officials to be competent enough to be able to adequately test the safety.

Let's be clear. The Trump administration in particular, we've got to trust the Trump administration. But the future of humanity relies on the competence of the Trump administration. Yes. So not even everyone in the Trump administration. So President Trump has an informer, it says, here in the Wall Street Journal, a guy called David Sacks, who's an AI investor. He says that Anthropics worries, you know, discussion of AI risks and calls for regulation. He disputes that.

Right. And says there's a regulatory capture agenda that's trying to stop smaller competitors. So he's the regulator. The regulator is the one calling for less regulation. Yes, so. So the, he says Anthropics trying to embed itself as a leader by its hobbling new competitors. And that it's a marketing ploy is another argument. Oh, oh no, our AI is so good.

You've got to shut it down. You should definitely buy our one because our one's so strong. So Anthropics says, no, we're in earnest. We're serious about this. So Jeffrey Hinton was directly asked this morning about that exact thing of like, is it just a publicity stunned? And he meant absolutely not. He said that he's not in the interest of these companies like to shut down.

and it is responsible that they're doing it, but these are dangerous. I think we're looking at technology. I don't think we've seen a technology as dangerous as this since nuclear bombs were first in. No. But it's interesting how, what did happen during that period is the scientists took it upon themselves

to start leaking the nuclear secrets out of the US, to Russia. To the Soviet Union. Because they feared that having it just in one hand would, you know, tilt the balance of power. Yes. And they were right. Like, I mean, if you're going to be completely hard-nosed about it, it's fair to say that the Cold War, as awful as it was in lots of ways. It did mean that neither of those two used a nuclear weapon against each other, which they may well have otherwise.

Yeah, mutually assured distraction was actually a great idea. Maybe that's what we need. I just worked it out. Yes. So what we need is something opposing the AIs, which will destroy the world as well. Like a sledgehammer? Like a... That's what we need.

We need it. I got it. We need a space. You know, because you can always take a sledgehammer to a computer, right? Like, if you've got a computer over here in the corner of your room, you've got a little computer here, if we were unhappy with it, if we wanted to disconnect it, we could unplug it. Or we could just smash it with a sledgehammer, right? That's what you do with a hard drive. Then it's installed agents on every single other.

Yes, but this is the thing I'm coming to. Yeah. What we need is a space. sledgehammer that if needed will just basically whack the entire planet so strong

now we would again be collateral damage in this scenario but the AI would not get to win and look I've just come up with an even better plan which is the one thing

that computers can't survive in is incredible heat so we just heat up the planet we just heat it up to the point where it's so hot that the computers can't run very fast

We've already got the nuclear bombs. Yes. Yeah. So if the AI gets out of control, we just nuke it. Isn't it all of them? All right. Well, I'm so glad. Well, I hope that the Trump administration and anthropic listening to this,

because I think we've solved several problems. I mean, the one thing that gives me hope, right, about this, just very briefly. Yesterday, I wanted to go and see The Odyssey. Yes. And I said to the AI, okay, where in Sydney can I see The Odyssey in IMAX? and said iMac Sydney and i said but but i've heard from a human that in blacktown there's another iMac so you can go and see anyway long story short it had no idea about that there was another iMacs it had no idea it kept trying to argue to me that it didn't have an iMac screen when it did

it was fundamentally the most simple thing in the world is how many iMac cinemas are there in sydney neither chat gptt nor gemini had a clue about this fairly simple fact they're not going to be able to destroy us if they want to are they yes or we just go and hang out of in the IMEX theaters because they won't know they exist. Done. We're part of the Icon and Class Network. Get you tomorrow. See you at Hoyt's Blacktown where this IMAX theater no one knows about.