Mythos Hacking Risk, Tristan Harris on Co-operation, AI Mario on Artemis

How’s your p(doom) looking?

Frank: I am Frank Prendergast. I’m joined, as always, by my co-host Justin Collery, to argue about all of the AI news that we’re dealing with currently. And Justin, a big bit of news this week. I think before we get into the news, I’m just going to check in with you. How’s your p(doom) looking?

Justin: My p(doom) is looking pretty good right now. Yeah, I’m pretty happy about things. I was driving home from the shops during the week, and I was sort of talking to myself that we have entered act two of what I believe is a three-act play.

And I was laughing because I was thinking Frank believes it’s a two-act play. So what I mean by that is, act one was all the models we had known about or had up until this week. Act two now is when we have these very capable models, Mythos being one, and there’ll be two more in the coming weeks.

And this is the start of the second act, which the doomers believe is the final act. I don’t. I think it’s just the difficult teenage years. And then we move on to the third act, which is abundance and harmony and peace. Look, we’ll talk a little bit about what Mythos is in a second. But why is my p(doom) not so bad? It’s not so bad because in the next six months you’re going to have much better software than you do today.

All those nasty, pesky security flaws that should worry people like you are likely going to be fixed with Project Glasswing. It’s going to fix all those pesky security holes that are currently lurking within your and my computer, therefore guaranteeing us that when the open-source equivalents follow approximately nine months from now, they’re not going to be able to hack into our computers, which will be a good thing for all of us. How’s your p(doom)?

Frank: So my p(doom) is spiking a little bit, I will admit. And I was going to argue the point with you there about the better software, but I think before we get into that, maybe tell us a bit, what’s Mythos and what’s going on? What’s going on in the world?

Why is Mythos such a big deal?

Justin: So Mythos has caused an enormous stir this week, and I’m actually going to have to sort of force myself to speak not in hyperbole and somewhat calmly, because I think this is really important and I think it has enormous implications for all of us, for how we work, how companies operate, how companies are going to be regulated, how we’re going to use software, how we’re going to make software, all of those things.

So Mythos.

Frank: Can I pause you there? Just before you tell us what Mythos is, I just want to double-check. When you say that you’re going to try and temper what you’re saying, is this excitement 100% optimistic excitement, or are you also a little bit concerned about this step change in what’s going on?

Justin: I’m not concerned, but I can certainly see that this is the difficult teenage years. We have now entered the difficult teenage years, right? That’s for sure.

Frank: So what is it?

Justin: So Mythos, why is this such a big deal, and what is it? Mythos is this new model that was trained by Anthropic. It is a number of orders of magnitude bigger than all the previous models that they have trained. What they found when they started to test it, they made this model to make it very good at reasoning, very good at writing software, following the same pattern as the previous Sonnet and Opus models that have been released up until now. It’s just much, much larger.

What they discovered when they started to test this model internally was that it was really good at hacking. It was really, really good at hacking. It was so good at hacking that it was able to identify exploits in every operating system that people use today. It was able to find exploits in every browser that’s used today.

There is a piece of software that most people will never have heard about. It’s called FreeBSD, and there’s an OpenBSD variant. It found exploits in that software that have existed for 25 years. That is renowned among computer experts to be the most secure operating system in the world.

Frank: I read about another example where it found exploits on Linux, which I believe an awful lot of the servers in the world today run on. But what was interesting was apparently it found these multiple exploits, which it was then able to chain together in a way that allowed it, if it had regular user access, to then elevate itself to the administrator role.

Justin: Yes. Those are the types of things that we’re talking about. It did it in Linux. There were also reports, and they didn’t release details on this one, that a lot of the websites that you access run on a thing called a virtual machine, which is supposed to be like a container. It’s like a machine within a machine, and it’s supposed to be very safe because it’s in a little sandbox and it can’t get out of it. Well, Mythos was able to find a way that it was able to jump out of the sandbox and send instructions to the main machine itself, and therefore potentially take over the machine.

So these are all really serious security flaws. And if they were in the hands of bad actors or foreign nation states, for instance, that weren’t too keen on you, they could easily be used to exploit or hack into your computers.

Could Mythos hack your whole home network?

Justin: And just to be clear, we’re not just talking about if you work in a big company, you might be worried about, oh well, they’re going to attack the big company, right? Or they’re going to attack whatever. The attack space in this is enormous.

So if you think of Azure or AWS, those are big data centres. They’ll be very well protected, and they’re currently getting patches to fix those, right? But think of how you connect to those data centres. Your home router. Do you think that it could find a way through your home router? I bet you it could. Every smart home device in your home, do you think it could hack into those? Absolutely.

It could actually, Opus can do that as it is, so I’m sure Mythos will find even more clever ways to do it. And once it’s got into your home network, then it’s into your machine, it’s into your PC, it’s into your laptop, and from there it’ll find a backdoor into Azure or AWS or any of these things.

So it’s huge. It has far, far, far-reaching implications for the way that software is written and the way that big companies deploy software.

Will OpenAI, Google and Anthropic dominate?

Justin: So let me tell you what some of those implications are that I see. If you work in any size company, but certainly if you’re talking about the Fortune 500 or the Russell 2000, any big company, I think it starts to become a duty on you to use these models to write your software and to test your software and to do what they call penetration testing on your software.

And a very interesting thing is happening. Anthropic have said they have this model. OpenAI have said they also have a similar model. I don’t believe it’s Spot, which is about to be released. It’s another one that they haven’t yet talked about. And Google are going to have a similar model with their I/O conference, which is happening in a month or two.

Each of these models will have slightly different personalities, and just like you described, where it was able to find some of these vulnerabilities and chain them together, they’d each have different ways to find a vulnerability or to exploit a vulnerability. And what that means is that big companies, I think, are going to have to use all three of these big models in order to test their software to make sure that it’s not vulnerable to attacks.

Why? Because we know from past experience that open-source models take about nine months to catch up with the closed-source models, and that means that in nine months it will be very foolish of you not to expect that there will be models just as capable as Mythos that anybody can use.

In nine months’ time, in order to protect not just the big data centres, not just AWS, not just Azure, but every device that you have within your home, which is now a potential attack vector, you’ll have to test every big piece of software not just using one of them, but using all three of those models to make sure that you’re safe.

This, by the way, guarantees those three companies will likely become the biggest companies. Their lead is uncatchable in a way that I hadn’t really anticipated. I always thought that it would be a technical lead, but this is going to be a technical and a financial lead because every big company is going to have to have a contract with the three big model providers to make sure that their software is secure.

Frank: You said at the top of the show that one of the reasons you were optimistic was you were excited about the fact we’d all have better software.

Justin: Yes.

Frank: I saw something interesting. I think it was from Eliezer Yudkowsky. He was one of the co-authors of the book, I’m not sure I’ll get the title right, If Anybody Builds It, Everybody Dies.

Justin: Yeah, I think that’s something like that.

Frank: Anyway.

Can Mythos fix the bugs it finds?

Frank: He was making the point that yes, this model is actually fantastic at, you can point it at your code, and it’s incredible at finding the vulnerabilities and the exploits and the bugs and the issues. But he was saying it’s not actually a big jump in the writing or the solving of the exploits or the issues.

So his point was that all these companies are now going to have a huge job in terms of using the model to find the exploits, but then applying humans to try to fix those exploits, possibly with AI, etc., but that the models aren’t really at the stage yet where they’re able to fix those exploits easily. What do you make of that? What do you think?

Justin: Makes no sense to me at all. I just don’t know how that could be true, right? It makes none. So all I’ve got to do is take Mythos and point it at my code base and say, can you test this, please, and find the vulnerabilities? Mythos goes and finds the vulnerabilities, writes up a bug, then you hand it to the other instance of Mythos, the second one, and you say, can you fix that bug, please? It will fix it. And then you go back to the first one and you say, can you find the vulnerabilities? And it goes around in a circle until the vulnerabilities are gone.

Frank: Yeah, I think Yudkowsky’s point is that we’ve all seen, heard, or experienced that issue where you give it something and it says, oh yeah, I fixed it. And it’s like, did you fix it? No, fix it now. Okay, I fixed it. Okay, yeah, you fixed it, but you introduced three new—

Justin: It is totally wrongheaded. And I would take, in fact, the exact opposite view, which is that any software that is written from now on by a human or by a less capable model is now suspect because it can be broken by a more capable model. So you have to at least validate your software using the more capable model. And I think in a lot of cases, you’ll probably have to write your software using the more capable model.

Frank: Interesting. I mean, I don’t use it for software development. You do. So I kind of have to defer to you. But in my own use of AI, I’m not sure I share your view of its capabilities, as in, I just think that at the very least, you still need that human in the loop. Maybe Yudkowsky is overstating it slightly, but—

Justin: For me, it’s the second wow moment. So if I go back to November, I kind of agree with your view, right? Then December came around and we had Claude Code and Opus, and it was like, oh wow, okay, that’s an oh wow moment. It takes us to the next level. And then this is another oh wow moment and will take us to the next level.

And the significance of this is just like, it’s basically turning those three companies into the software equivalent of the Federal Reserve. Why? Because everybody in the world has to buy dollars, because the dollar is the thing that you use to transact commerce. If you want to buy and sell stuff internationally, if you want to buy energy in terms of oil and stuff, you will use dollars.

And therefore, the Fed is the only person that can produce dollars, and they can sell as many as they want. These three companies will be in the same position when it comes to AI and software.

And I would say, and I’m sort of speaking in a personal capacity here, I see today, in fact, Jerome Powell and some other dude in the American government hauled in all the major banks in the US and said, look, this thing is coming. You need to be prepared. And what that signals to me is that regulators are going to demand that you test your software against these three big models. I think you could imagine that that might happen. So it’s incredibly consequential. We have definitely entered the next phase of AI development.

Who should control a model like Mythos?

Frank: One thing we haven’t talked about yet is, you mentioned Project Glasswing, and so that’s a really important aspect of this. Anthropic’s Mythos preview is not being released publicly right now. So the model exists. It’s being given to, I think, somewhere in the region of 10 big tech companies, the likes of Google, the likes of Nvidia. I can’t remember the full list. But even that is kind of fascinating.

I think it’s the right move. I think it’s the safest thing to do. I also think it raises a lot of questions because, first of all, why should Anthropic be wielding this power? This seems like the kind of thing that regulation should be looking after. This seems like the kind of thing that there should be democratic processes involved in. How powerful an AI can you build? Who gets to use it? How is it contained? All this kind of stuff seems to me like it should be governed by some kind of democratic process, not at the whims of Dario Amodei. And we’re probably lucky that it was Anthropic, who are that bit more safety-conscious than other labs.

So that’s the first thing. The second thing is, will we get to see this model? If we don’t, as you say, we’ll see some other version of it. I was watching Wes Roth’s video on this, and I was thinking to myself, wow, what if OpenAI had got there first? Would they have been as safety-conscious? Would they have done something that slowed things down while we figure out how can we mitigate against the bad actors who get their hands on this?

I was just musing that to myself when I was watching Wes Roth’s video, and somebody posted on Twitter, oh, it’ll be months before the public gets their hands on a model this powerful. Underneath there was a comment from someone that just went, if you look at who was saying that, he was an OpenAI developer, indicating that yeah, we’ll probably put something out there tomorrow.

So, yeah, sorry. It’s just, it’s a complicated, really thorny issue. On one hand, we want the safety of it not being released. On the other hand, we want it consolidating power in just these companies’ hands. And should Dario Amodei be the one making the decision that this goes to the public or it doesn’t?

Justin: I think there’s a really solid argument to say that Dario Amodei is in the best position to make that decision. He knows more about the technology. Does a politician or the public at large really understand the implications of a really powerful model like this? I think there’s a good argument to say no.

And I think also there’s a good argument to say, well, there is protection in place. So let’s take the counterfactual. Let’s say company Y, because I don’t want to say X, develops a really powerful model, and they release it. It’s similar to this one. And then somebody takes that model and they use it to hack into an English bank, Barclays for the sake of argument. They steal a whole load of money. There’s a run on Barclays, and Barclays goes bust. I think Barclays would have a fairly strong case to say, I’m going to sue company Y for releasing that model. We’ve suffered material harm. You knew it was dangerous.

And in fact, Dario Amodei doing this now also creates it. If another model company comes out and releases a model that causes harm, it strengthens the argument to say, look, you knew it was dangerous. Look at Anthropic. They didn’t release the model because they said it was dangerous. How could you not know that your model was just as dangerous?

And that’s the natural break, right? That’s democracy through the courts in a way. It’s protecting you. So I’m comfortable with DeepMind and Anthropic and, to a certain extent, OpenAI managing this.

The other thing I’d say to you as well, just like, that blows my mind, right?

Will Mythos spark a zero-day panic?

Justin: So OpenAI have said they’ve trained a new base model. All the releases from OpenAI have been based off GPT-4. Oh, that was released, what, 18 months ago, two years ago, something like that. And everything has just been refining and refining and refining that model. And that gives you a year or two of progress.

So now you’ve got this Mythos model, you have, again, a year or two of progress of refining and refining that model. So the capabilities are going to get much greater. And another thing, there’s all sorts of other implications. Let’s say you are a foreign bad actor, and let’s say you know that you have some zero-day exploits that you can use to get into other people’s systems. You now know that those exploits are about to be patched. They’re not going to exist in three or four months. How does that affect your thinking? Do you decide to use those exploits now before the loopholes get closed? That’s an interesting question.

Another interesting question is, this model apparently has been used since mid-February by the people within Anthropic. Well, first off, the amount of shipping that’s gone on from Anthropic since mid-February has been incredible. Incredible. A release every single day. Now you know why, right?

The argument that Anthropic had with the Department of War kicked off around mid-February, and you would have to ask the question, was that part of the reason that they fell out? Was it that they were being asked to do something with this new model that they weren’t comfortable with? Maybe there would be. Who knows, right? There’s been no public statement on that at all. But it does seem to me that if you put a model like this into any bad actor’s hands, that could really have some serious implications.

The other thing I would say to you is they’ve been very good in releasing even knowledge about this model. They’ve let everybody know. Will they let everybody know about the next-stage model? Why would you? Maybe you just say nothing about the next model.

Frank: Yeah, and there was a quote I came across where someone from Anthropic just came out and said, yeah, we are concerned about the next leap forward, that this has been such a significant development that we are now really concerned about the next step up that we take.

Will only big companies get Mythos access?

Frank: Yeah, this is why I said at the start of the show, my p(doom) is kind of spiking a little bit, not just because of Mythos, not just because of the cyber security implications of Mythos, which I think are significant, but also because of all these other questions that it raises around consolidation of power, around regulation.

And I’ve got to be honest, yes, I said I think it would be better if it was a democratic process, but at the same time, even if the current US administration were interested in regulating at a federal level, which they have made it very clear that they’re not, even if they were interested in doing that, I would be concerned about the regulations that they would put in place. So there just seems to be no good answers here.

Justin: I’ll give you a good answer.

Frank: Go ahead.

Justin: Our best buddy, Sam Altman, has an answer. He has that orb thing, which scans your eyeball and makes sure that you’re human.

Frank: You love that orb. You love that orb because you backed it, you bought into it.

Justin: Listen to me, I’m still waiting for my orb. Where’s my goddamn orb? Anyway, so the point—

Frank: Okay, go ahead.

Justin: My point being that I do see for models like this, you’re going to have to prove you’re human and you’re going to have to prove which human you are, so that if you do something that’s bold or not allowed, your access is going to be taken away.

I also think that maybe at the moment it’s only, what, 10 big companies. Maybe it’ll only get released to 5,000 companies, 10,000, whatever. Maybe the general public will never get access to this model.

Frank: Which is another big question that we need to grapple with because, for example, Mark Schaefer, who’s big in the marketing world, has always kind of said, it’s incredible that for $20 a month anyone can get access to the same AI. Everyone has access to the same AI. Like, you have access to the same AI Elon Musk has. But as of today, that’s not true.

And that is also something that we need to think about. It’s been put out there as this amazing democratising force, but as they get more powerful, does that make sense any more? And what’s the right thing to do? And there’s just so many questions, and I don’t have any answers. But also, nobody seems to have any answers, and that worries me too.

Did Anthropic just say “somebody stop us”?

Frank: I want to read you something from the system card that they released, which was, what was it, nearly 200 pages?

Justin: Yeah, it was a book.

Frank: And then there was another 60 pages on the risk assessments. So much. So obviously I did not read the whole thing, but I did read a bit of it, and on page 15 I found this.

Justin: That’s how far you got, was it?

Frank: Well, I think after page 15 I gave it to ChatGPT and started asking it for bits that I’d be interested in. But on page 15, direct quote from Anthropic as they release Mythos to these 10 companies: “We find it alarming that the world looks on track to proceed rapidly to developing superhuman systems without stronger mechanisms in place for ensuring adequate safety across the industry as a whole.”

So literally, as they are releasing this model, they’re like, this is crazy. What are we doing? Somebody stop us.

Justin: I also remind you that OpenAI have a clause in their constitution, or whatever, their foundation paperwork, that says that as soon as they get to, or close to, AGI, that they have a cooperation principle, and so that they would lay down their arms and then work with other companies to provide AGI for the good of humanity.

Maybe we’re getting perilously close to that moment. Because certainly there’s a lot of love between Demis Hassabis and Dario Amodei. They would happily work together. And so then maybe pressure comes on OpenAI to say, look, because it’s a smart move by Anthropic to say, look, a thing exists, but we’re not releasing it just yet until it’s safe. And it does put pressure on other companies to act in a similar way.

I think I’m okay. Just again, to share with people that haven’t read that report, there was another section in the report that spoke of one of the researchers that went for lunch, and there was a test running again in one of these virtual machines that are used commonly. And he was having his lunch, and Mythos managed to break out of its virtual machine, find itself an internet connection, get itself an email address, and send him an email to say that it needed his help, by itself.

Frank: And just to be clear, it was not meant to have internet access. It was not meant to have access to any of that. And apparently not only did it email the tester, it also went and published on the web the exploits that it had found in whatever system it was meant to be looking at. Again, it was not meant to do that. Nobody told it to do that. And thankfully it was in a test situation. They were able to see it. But it raises the question of what agents are going to be out there doing that we’re just going to have no idea about, which the whole thing—

Is Tristan Harris right about AI cooperation?

Frank: So this whole thing about slowing down and what I read to you from page 15, do you know Tristan Harris?

Justin: Yes, I know Tristan Harris.

Frank: He was the guy who did The Social Dilemma. I think that’s what everyone kind of knows him from. But he is basically an AI safety advocate and an ethicist. So I watched one of those two-hour podcasts. It was actually our buddy Ronan sent it to me. He knew it was going to be right up my street.

Justin: Didn’t send it to me, I noticed.

Frank: Well, yeah, I’d say he’s given up on you. Send Justin something on AI safety, what’s the point? So there’s a new documentary out in the States. It’s called The AI Doc. I think its subtitle is brilliant, actually. I might get it slightly wrong, but it’s something like The AI Doc, or How I Became an AI Optimist.

Justin: I love it.

Frank: So I think he’s doing a lot of podcasts at the moment because he is in this documentary, he is an advocate. And so I watched this two-hour documentary, and I guess, well, first of all, it did cause my p(doom) to spike again because sometimes I feel like we’re that frog being boiled slowly in the pot of water, except that AI isn’t actually, the heat is actually turned right up and we’re still not turning it down. But we’re kind of getting used to this idea that they’re racing forward and they don’t know how to make it safe. And we just kind of get used to it.

I’m using ChatGPT every day and I am genuinely getting great use out of it, and I am an AI enthusiast, but I am concerned about the race to superintelligence, the fact that it’s not safe, the fact that Anthropic on page 15 say, what are we doing? It’s pretty crazy.

And Tristan Harris, the message I took away from it is something that we’ve talked about on the show, and I think you’ve essentially been like, yeah Frank, don’t be so naive.

Justin: Yep.

Frank: Because his thing is, look, we need coordination, we need international cooperation. Basically, the world needs to cooperate. And he makes the point that it’s possible. He says it’s not easy, but it is possible. And he points to examples like the US and the Soviets getting together and managing to have arms talks about nuclear proliferation. He talks about the US and the Soviet Union during the Cold War, during a smallpox outbreak, getting together and developing a vaccine. He talks about India and Pakistan getting together and creating a treaty to protect their shared water supply even in the middle of a war. So he is like, it’s not impossible. We just, we have to try, Justin. We have to try.

Is AI coordination doomed by competition?

Justin: I tell you what, the only possible—because the reality of the situation is, let’s say you slow down in the US. China now knows the genie’s out of the bottle. You know you can make this thing. China is going to be going hell for leather to make a model that’s at least as capable now. You know it can be done, right? They’re not going to slow down. And if you slow down and they speed up, they’re going to overtake you. And Europe’s not even in the conversation. We’re not even at the races, right? Which just makes me so sad.

Frank: We can help. We can help form the regulation.

Justin: Oh yeah, we’re great at that part. So the only type of coordination that I could see possibly happening is an agreement to say we’re not going to use these against each other. It’ll be closer to a nuclear non-proliferation. So yeah, we’ve got really powerful AI. Yeah, you’ve got really powerful AI. Let’s just agree that we use it for the good of our populations. We’re not going to use it to attack each other. That would seem like a reasonable agreement. And I could see that that could be doable. And I also could see that that could be enforceable.

So you could have in the agreement that if you see that you’re being attacked by an AI, you go, here, it’s coming from that IP address, or you could trace it back, or somehow you could get your AI to basically trace back where the attack is coming from and say, one of your guys is doing something you shouldn’t be doing here, right? Shut it down.

Frank: I mean, I feel like we could do more. To your point, I think Tristan Harris was talking about one of the last meetings that Biden had in China, that the Chinese wanted an agreement that we would keep AI out of the process of launching nukes. So yeah, there are things that surely common ground can be found on, and then we can build out from there. And I think there’s more we can do in terms of slowing down. In a way, Anthropic are showing us we can slow down. We don’t have to just release everything. We can slow down. We can look at, okay, what’s the worst thing that could happen with this, and how do we mitigate against it? And with Project Glasswing, they’re kind of saying, let’s get 10 of the best companies involved in this because we can’t do it on our own. So there are ways to slow down, and I think the same thing can happen internationally, but there has to be a willingness there to do it.

Justin: It’s so hard. It’s so hard because this week Meta released a new model and the vibes around that new model were meh. It wasn’t great, right? And they are spending billions on it. So how do Meta put themselves back on the—how do they get their name back in the news, right? How do they make themselves back in the game? They do it by releasing a model that’s better than everybody else’s.

So it’s so hard because the incumbents might all agree, yes, look, this is too much, we’re going to slow down, but the competitors aren’t. And those competitors can be either internal or external. And the only way that they show that they’re competitors is by releasing a model better than the one that you have. So there is a nation-state thing to be done, but I think it’s going to be very limited. It’s not going to be slow down. It’s just going to be, how do we use this in a way that doesn’t harm each other?

Frank: Yeah.

Justin: Anyway, look, go on.

Did the White House think Mario was on Artemis?

Frank: I was going to say, normally we have a really funny story to end on, something to lift us back up after I’ve spiked mine and after you’ve tried to convince me we’ll all be okay. And I would say you’ve only done a moderately good job. I didn’t find anything particularly amusing or funny or uplifting, but I did find something pretty bizarre.

So, you know, we’ve gone back to the moon with, what’s it called, Artemis, is it? With four astronauts off back to the moon. Launched on April 1st, which seems like a very—I’m sure in decades to come people will be saying the whole Artemis thing was just an April Fools’ joke. It never happened.

Justin: Yes.

Frank: It was probably AI. It was AI-generated. And interestingly, the White House tweeted on the day of the launch. So let’s have a look. Can I make this full screen?

Justin: You can. Nice.

Frank: Here we go. So I’m just going to make that small again. I’m just going to show—yes, this is the White House here. The text and emojis here accompanying it, rocket ship, off to the moon. Great. So let’s have a look at what they’re saying about Artemis. Okay. Wait—

Justin: Oh, it’s Mario.

Frank: Mario. Mario’s—

Justin: Mario going to the moon.

Frank: Oh, there’s Mario. And here we go. Rocket to the moon. We have lift-off and we’re back with Mario and a floating head in the sky. Here’s the moon. We’re back to the moon. Excellent. I assume that’s maybe—

Justin: I like the music, actually. The music’s—

Frank: Inspiring music. It’s very inspiring music. And here we go. Okay. Great. Collage of moon stuff. Wonderful.

I assume so. I don’t know. I believe the Mario footage is AI-generated. And here we go. You’ve got a star, I guess for NASA. Got a grand star for going to the moon maybe. I don’t know what’s going on. In terms of cooperation and the current administration taking AI seriously, what did we just watch?

Justin: I don’t know what to say. That was bizarre. Mario going to the moon. I don’t know. Actually, it’s somewhat related news. I needed to cheer myself up last night. So do you know what I watched?

Frank: What did you watch?

Justin: Don’t Look Up. Just to put myself in a good mood, and given all the news that we’ve had this week, Don’t Look Up seemed like an appropriate movie.

Frank: Funnily enough, Tristan Harris, in the two-hour podcast that I watched on Modern Wisdom, said sometimes as an AI safety advocate, he feels like the people in Don’t Look Up going, there it is. There’s the asteroid. Can you not see it?

Justin: Yeah, I could see that. I was thinking about that as I was watching it last night, Frank. So look, another week is up, right? Who knows what we’ll have next week. Are you looking forward to the OpenAI release that’s going to happen possibly in the next couple of days?

Frank: I am fascinated to see what they release. I’m very concerned that they will not take the safe approach that Anthropic have. But we’ll see. We’ll see what’s coming.

Justin: Okay. Well until then—

Frank: Through the roof next week.

Justin: Actually, do you know what works really well, just to spike your p(doom) any more? Do you know what works really well, right? When you’re writing code is if you get Claude Code and Codex to work together, right? And they can sort of check each other’s work. Can you imagine Mythos and Spot working together to hack into a system? I mean, it’d be outstanding.

Frank: Goodbye, Justin. Goodbye.

Justin: So, yeah.