Frank: I’m Frank Prendergast. I’m here, as always, with Justin Collery to argue about all of the latest AI news, and I’ve no doubt we have some really good arguments ahead of us because we’re going to be talking about one of my favourite topics: the regulation of powerful and dangerous AI.

Justin: It’s not that dangerous.

Why is Mira Murati’s AI model free?

Frank: But before we get to that, there’s been some pretty interesting model releases, and the first one is from Mira Murati, who I think you and I both were like, “Mira Murati is going to do really interesting things.”

She was the chief technical officer, I think, at OpenAI before Sam got fired. If I remember correctly, she was pretty instrumental in Sam getting fired as well. She had a lot of conversations with—can’t even remember. Yes, and she had documents, apparently, of Sam’s—

Justin: And she was instrumental in bringing him back. Anyway, that’s not important now. Mira—

Frank: She was, yes. But she was, yeah. So she was the CEO, interim CEO there for a while, and then Sam came back.

Justin: Frank? Surely there’s a movie about this.

Frank: The end of the year, sometime, I think. Yeah, just in time to squeeze in for Oscar season, I think.

Justin: Oh, and Christmas. What better Christmas movie than a political drama about OpenAI? Let’s talk about the model, and let’s talk about why this is important.

Frank: Yeah. So her company now—she left OpenAI—the company is Thinking Machines. They’ve brought out their very first model. Up until now, the only thing they’ve released is a thing called Tinker, which allows you to fine-tune an AI model. Now they’ve actually released a model. It’s called Inkling, and you can fine-tune it with Tinker.

It’s an open-weights model, so the model itself is available kind of freely, as long as you have the resources to run it, because it’s a bit of a beast, I believe.

Justin: Yeah, big problem.

Frank: A bit more about the technical specs. And so the model itself is free, which is, from a business perspective, an interesting business model: the AI model being free, but you pay for the tool to train it on your organisation’s data, processes, knowledge and everything.

What do you think of it all?

Justin: Great business model. It’s very interesting because, around that time, two people left OpenAI. One of them set up the company Thinking Machines, and the other one is Safe Superintelligence, right, which is Ilya Sutskever’s company. We haven’t heard anything from him at all, right?

But now that I see what’s going on with Thinking Machines, it makes me think that he must have great stuff. I’d love to know what he’s doing, right? He must have great stuff going on in the background, but he said he’s not going to release anything, which I think is a mistake.

Is fine-tuning AI worth the hassle?

Justin: So this model and the business model that they’re putting forward—I’ll give you the good part and the bad part.

The good part: this is now the best open-source model that is not Chinese. There are, I am sure, loads of organisations that would love to have a model that they can run within their own infrastructure and with their own security boundaries and stuff like that. And so this sort of fills that gap in the market.

From that point of view, this could be super successful, right? Loads of companies, and people, might want to do that. That’s the good part. The bad part is that the model itself is very big, but not terribly smart yet, right? It’s their first model, so give them a bit of scope, right?

From what I’ve read, really it’s about Sonnet level. So it’s up there, but it’s not super, super smart.

Frank: That’s interesting because the promotional materials would have you believe that it’s closer to Sol level, OpenAI’s latest model. So that’s a big difference between what the promotional materials would have you believe and where people are saying it actually is.

Justin: Yes, exactly. And the other thing they’re saying is that you should fine-tune this model. Fine-tuning is hard, right? There’s money involved in fine-tuning something, and there’s effort involved in fine-tuning something.

And I do worry that, if you’ve got a big model that’s super smart, as opposed to a different open-source model that’s not super smart and you’ve got to do some work on it, it’s just a bit of friction. It’s friction, right? It’s hard.

So my hope—I think there’s a gap in the market there. It’s the gap, in fact, that the French company—what do you call them? Le Chat, Mistral—they should have been filling that gap, right? The open-source Western model.

So Mira Murati may well now fill that gap, and I think there’s loads of companies. The models need to get smarter, and we shouldn’t have to fine-tune them on our data. That’s just—

Frank: So we kind of had this conversation, actually, not so long ago when Microsoft, I think, had their developer conference and they announced—what do they call their thing?—frontier tuning. They outlined a similar philosophy about how organisations should have models that are trained on their institutional knowledge, et cetera.

And we had that same conversation about, yeah, it sounds great, but how many companies, or what size of organisation, are going to be really able to do that, prepare the data for that and undertake the whole journey?

Should AI reflect us, not replace us?

Frank: I do like the philosophy around this that Mira Murati is talking about because—

Justin: What’s her espoused philosophy? Because I can tell you what the actual philosophy is going to be, I think.

Frank: Well, she presents it as a more human-centric AI future, which I like the sound of. And the idea, basically, is that there shouldn’t be one model that is set in stone by a company, where its capabilities, knowledge and culture are just set by the company that trains it, and that’s it. Everyone feeds into it, and it’s basically designed to replace human work as this single, ground-truth, all-knowing being that one company has decided upon.

Whereas Mira Murati—this is my take on what Mira Murati is saying—and the Thinking Machines philosophy is more like, okay, this should be an extension of human values, ethos and knowledge. It should extend human capabilities, and therefore you fine-tune a model on your own institutional culture, knowledge and know-how.

The human input into the model within every organisation is vital, and you end up with all of these—she kind of vaguely alludes to there being as many different AI personalities as there are people. It shouldn’t just be one ChatGPT personality or Claude personality that everybody has to use. So, yeah.

Justin: Sounds great. Let me tell you what’s really going to happen, right? There’s been a lot of heavy stuff going on over the last couple of weeks. It’s like, where are the good old days when we could talk about Terminators and bots that are deleting loads of files by accident and stuff like that?

Whereas now things are getting real. It’s got a little bit more serious, so we need to keep things light, right? So let me tell you what’s really going on with Thinking Machines and all these things.

Are OpenAI and Anthropic actually screwed?

Justin: I think OpenAI and Anthropic are in big trouble, possibly screwed, in fact, because of these releases and the other one, which we’ll talk about in a couple of minutes, because their business model is dependent on them hosting the models and making money off inference.

And so they’ve got to charge a premium every time that you use their model. That’s how they make money. And so, when these guys, Thinking Machines and other open-source models, come out that are nearly as good as their ones—and remember, they now have difficulties in shipping new models because the American government has stepped in to slow them down—when these open-source models come along, the cost for inference goes down to the cost of electricity and hardware. That’s it.

There’s no margin in Thinking Machines, right? They’ll just give you the model, and you can just do your own inference. And so then it becomes harder, especially for OpenAI and Anthropic, to train new models because they know they’re not going to get the—at some point, the investors are going to go, “Hey, it’s going to cost you 100 million quid to train this cool new model, but how do you expect to get the 100 million back? Because you’re not making any money on inference any more.”

And so the big lesson is that Mira Murati—how much money did she get? I think she got one or two billion.

Frank: Right.

Justin: Not hundreds of billions, not 50 billion from all these other companies. And there’s a fantastic tin-hat theory that I heard on the internet, so it’s not my theory, right?

Frank: Yeah.

Justin: But there’s a lot of money here, right? And people aren’t as nice as you are. There’s a lot of nasty people in the world. And I’m not saying that any of this happened, but some of the moves that are happening—why would, for instance, an investor in one of the inference companies, or in one of the model companies, go and report that model company to the American government?

Frank: So you’re talking about the reporting that it might have been Amazon who told the American government that Fable was too dangerous and it shouldn’t be publicly—

Justin: I’m not, and I’m not saying that they did, and I’m not saying they did it for this reason. I’m just saying that people more inquisitive than me would ask this question.

And one of the answers is, well, maybe they see that there’s a lot of money being spent there, and they could buy that company for a lot less money if they were to hasten its—if it was obvious that this was happening in the future, then hastening its demise so that they could buy that company may not be a bad outcome.

So maybe that happened. I don’t know. But the release of the Thinking Machines model and the release of what we’re going to talk about in a while, Kimi K3, both put—for me, I would be very surprised if there was an IPO. I think it’ll be a kind of rocky IPO for either of those companies, for OpenAI or Anthropic, because it’s not clear yet where the revenue is going to come from.

Is Kimi K3 the new DeepSeek moment?

Frank: So let’s talk about Kimi K3 because it’s highly relevant to what you’re talking about there, in terms of the open-source models eating the lunch of the likes of OpenAI and Anthropic. Kimi K3 is a Chinese model, and I believe it is also going to be open weights, although the weights have not been released yet.

Justin: Days. They said they’d release it in about 10 days.

Frank: Okay. Right. But before Kimi K3 released, in the run-up to it, as it was being rumoured to be coming, a lot of people were like, “Oh, it’s going to be a DeepSeek moment. They’re going to release something truly amazing, and everyone’s going to freak out. It’s going to show that the Chinese are actually ahead, not behind, and there’s going to be huge economic fallout and all the rest of it,” like there was with DeepSeek.

It’s been released today, so obviously very, very early. But the early indications that I’m seeing are that it’s not a DeepSeek moment yet, at least, because with DeepSeek it was like, yes, this is a really powerful model, and it was much cheaper to train, allegedly.

There were lots of reasons beyond the model itself why it freaked people out, such as the cheap training. With Kimi K3, it seems like people are more like, “Okay, it’s not a DeepSeek moment,” but it does show that the six-month-to-a-year gap that people keep talking about between the US and China might now be blown out of the water, and that actually the gap is tiny compared to what was believed to be.

Did Kimi prove compute isn’t everything?

Justin: Okay, so let’s talk about this one, right? People are getting a little bit upset about this one because, up until now, especially from Anthropic—a little bit from OpenAI, but especially from Anthropic—their big complaint has been, “Hey, these Chinese providers, they’re distilling our models. They’ve got all these things that are running, and they’re stealing the stuff from us that we stole from somebody else. How dare they?”

And what I mean there is, IP copyright holders might say, “Didn’t those companies steal our stuff to make their model, and now they’re giving out because somebody else is stealing their stuff? That seems unfair.”

And so now this model comes out, and the initial use of it, and all of the feedback, is that it’s somewhere between Opus 4.8 and Fable. That’s how capable it is.

And the technical report that they did with it shows that they did some clever stuff. They’ve got a different sort of architecture around their mixture of experts and whatever, and they’ve trained the model in a particular way.

And so the other thing is that it’s the old line from the Rockets: “What do you mean their Chinese researchers are smarter than our Chinese researchers?” Lo and behold, China has some really smart people, and they’ve made a great model.

Frank: Yeah.

Justin: Back to two things. One is, it comes back to the money thing, right? The Chinese had an AI conference overnight, and the head of China, the premier, said that they were supporting open-weights models and open-source models, and they’re setting up a programme for the good of humanity and all sorts of things, right?

They want to stop this sort of two-tier society. So that, again, comes back to the thing where, really, what they’re doing is undercutting the business model of the big US companies. Kimi, by the way—guess how many employees they have.

Frank: I don’t know, actually.

Justin: Two hundred.

Frank: Wow.

Justin: As opposed to the thousands and thousands that the big American labs have.

And the training run, from what I believe, cost, I think, 50 or 100 million quid, right? And they themselves, I think, came out and said, “Look, what we’ve shown is that the world has tried to make China compute-starved. They said that intelligence was gated on compute, and what this model proves is that that’s not the case. Even though we are compute-starved, we can still make a model that’s as intelligent as a frontier model, minus maybe a couple of months.”

So that’s huge. Here’s my favourite bit, though, for you, Frank. This is now a model that’s nearly as capable as Fable, very close, between Opus and Fable.

And you remember Pliny the Prompter? We can download this, and we can run it. We’d need a big machine now. We could run it, and we could disable all of the safeguards, I think, within seven prompts.

Frank: Wow.

Justin: With Pliny, and then we can get it to—I think Qwen 3.7, which was a much smaller model, maybe a tenth of the size, you could get it to help you design weapons and drugs and all sorts of nice things.

And with this one, I’m quite sure you could do the same, and it’ll give you a better recipe.

Frank: Yes.

Justin: Frank.

Should China’s models be the regulatory benchmark?

Frank: And this kind of brings us nicely into the next story that we wanted to talk about, which is Demis Hassabis publishing an essay on X about, basically, regulating AGI.

And I think it’s relevant because there’s a lot going on right now. We’ve been talking a lot about, for example, how the US has been very, very reluctant to regulate AI at all, and has kind of denigrated the EU for doing any regulation of AI and has kind of—

Justin: Can I interject?

Frank: Yeah.

Justin: Just as you—and then, sorry, I’ll let you go on, and you can tell us what Demis has said, right? But the US is obviously searching for ways to regulate AI, and they came out with a great idea this week, which I think Europe should copy, right?

There’s a saying about flattery and copying and whatever. But anyway, somebody in the US administration has suggested that the way the regulation should work is that, if the model is less capable than the most capable Chinese model, then there’s no regulation. Only if it’s a little bit more capable than the best Chinese model does the regulation kick in.

Which is really pragmatic because it means the bar changes with the technology. And I think—pragmatic is the word that I love—I think Europe should copy.

So, if a European company releases a model that is not as capable as any publicly available US or Chinese model, then you’re good to go. Just no need for regulation, because other people can get it anyway, so why are you regulating it?

If you can get a Chinese or an American model that can do this, why are you regulating it? Just let the Europeans go. And then, when Europe eventually gets to a point where they’re releasing models that are more capable than what people can access anyway, then the European Commission can step in and start to do their good stuff.

Frank: Then regulation becomes not at all about how dangerous a model is, but simply how dangerous it is in comparison to what the most—what’s the word I’m looking for?—reckless person or reckless country is willing to do.

Justin: Okay. And actually, I’m only half joking, right? Because, if you think about how the mechanics of that work, it actually, by default, slows down the progress of AI.

What it allows is—it’s kind of perfect, right? Because, if you’re behind, it allows you to catch up. But as soon as you want to release a model which is more capable in any of the three territories—in Europe, the US or China—it then slows down the frontier progression because you’re now releasing a model which is more capable than anything else.

Therefore, we’re going to do some regulation on that one. We’re going to check on it. And if the three regions were very, very smart, they would coordinate on what that regulation is.

Frank: Would it not simply result in China actually storming ahead to get the biggest lead they possibly could because they’re the benchmark?

Justin: Well, no, because—oh, sorry, I see what you mean. I should maybe rephrase that. I think you’re misunderstanding my point.

It’s if you’re about to release a model which is more capable than any other model on the market. So it could be a Chinese—so in China—

Frank: So China would have to buy into the regulation as well. I see. Right, right, right, right. Yes, that makes more sense. Okay. I see.

Justin: So there it—

Frank: Well, it’s not—I hope—yes, I can see there’s some kind of logic there. But I do personally think that the risks inherent in the models should be taken into account, not just who has the biggest model, or the most powerful or most dangerous model.

And I do hope that Demis Hassabis’s vision for how this should be regulated is a little bit more complex than, “How dangerous are China’s models?” I hope.

Justin: Okay. Just before you go to Demis, what I’m saying is that, if we continue the way that we’re going right now, it will make a mockery of the regulation.

Because if Kimi—let’s say Kimi just keeps on going, right? The US models are now being slowed down, and Kimi keeps on progressing at the same speed. You could imagine that 12 months from now they will release a model that’s open weights, that’s more capable than any US model, and I can just download it and run it.

So what’s the point of the US regulation?

Frank: Will we need international cooperation? Yes, we absolutely do need international cooperation. And I think that’s in Demis’s essay, although I’d have to go through it with a fine-tooth comb to figure out the detail exactly. But it—so his—

Justin: What did he say?

Is Demis Hassabis right about AI regulation?

Frank: “A Framework for Frontier AI and the Dawning of a New Age.” Basically, up until now, as I said, the US has been very, very reluctant to regulate.

What they’re doing right now is kind of regulating on an ad hoc basis, based on what the administration feels like with any given model.

So Fable 5 was released by Anthropic. Then, as you pointed out earlier, allegedly Amazon said, “Wait a minute, that’s very dangerous,” and the US administration went, “Oh yes, it is. Pull that off the market.”

Then OpenAI were about to release 5.6, and the US administration went, “Wait, wait, wait. That looks like it’s pretty powerful. Let’s hold off on releasing that one as well.” Then eventually it got released. We don’t know what happened in the background.

There’s no transparency. There’s no certainty. And Demis Hassabis is basically saying, “Look, the technology is starting to outpace our ability to figure out what the risks are here. We need to take a beat. We need to give ourselves space and time to figure out how we actually want to proceed.”

And he is basically saying that there should be—that the US is perfectly positioned because they have some of the leading frontier labs. They are perfectly positioned to figure out how to regulate powerful—

He doesn’t, I don’t think, ever mention the word “regulation”, which is probably very wise in the US. But he’s talking about a body that would be set up in the US that would be independent. It would be a state-governed body, and it would have experts who would assess models rigorously, in a way—he keeps saying it has to be a way that supports innovation.

And the way it would work is that certain companies would get a frontier lab designation. Once you get that designation, if you’re releasing a powerful model, this body would have to assess the model ahead of its release and give it a release pass or a fail.

Which just sounds all very logical. And if you don’t have a frontier lab designation, a little bit like you’re saying—it’s a little bit more complex than just, “Is it as powerful as a Chinese model?”

Justin: I’m a simple man. I like simple rules.

Frank: But it’s along the same lines. It’s like, look, if you don’t have that frontier lab designation and you’re releasing a smaller model, you’re grand. It’s fine. You don’t need to submit it to this third-party body.

Justin: Yeah, like an FDA for AI models, I guess. Like a Federal Drug Administration for AI models is what he’s looking for.

Frank: Yeah, he likened it to another body. The Financial Industry Regulatory Authority was the one that he likened it to. So he is comparing it to existing bodies. This is not unheard of, in other words. It’s done in other industries.

Justin: So I’ll give you the good things. I’ll give you the two sides, right? I’ll put my two hats on for this one.

On the good side, the issue that we have at the moment is that politicians, non-technical people, are trying to make regulations that apply to technical products, right? And they probably don’t have that skill set.

And so you end up with things that I find quite attractive, but are probably not the best idea, which is, “If it’s not as capable as the Chinese model, I say let it go,” right? Which is a very non-technical way of looking at it.

It’s really true that you can see the person who said that was a person who was like, “Yeah, I just want to beat the goddamn Chinese, so whatever,” as opposed to protecting people or doing what regulation should do.

So having a technical body with technical people is definitely a good thing. We should do that, right? We should do that in Europe. We should do it in China. We should do it in the US.

The problem I have is—and the other side is—if they have regulations, all of the regulations that exist in the FDA or the financial world or whatever, they’re all published, right? So we need to publish these regulations.

I haven’t yet seen a good suggestion for published regulations that apply to AI models. That’s a really hard problem.

We’ve been talking about this for two or three years. How do you do this? We don’t know. People need to start talking about it. Maybe it’s a good thing that they’re starting to talk about it, but these things take a long time.

Frank: But let’s take a moment. Let’s say—hang on now. Let’s take—

Justin: Just to make you frightened, right? If you look at the banking industry, which he’s talking about, you’ll see that there’s a financial crisis about once every 10 to 14 years.

And that’s because they keep playing with the rules that are supposed to regulate the banks, and they keep getting it wrong, and then the banks mess up, and then you get a financial crisis, and it sort of goes in a circle.

So, if you apply the same logic to AI, you’ll have the same thing, and you’ll get your bad outcome.

Frank: Yes. But at least, with the—if the banking industry was not regulated, imagine the insane mess that we would be in. That is, I think, the bigger point, which is kind of related to what I was going to say, which was: let’s just take a beat, let’s just take a minute and pat ourselves on the back as Europeans for having the foresight to regulate for—

Justin: Regulate what? What are we regulating?

Frank: It doesn’t matter. It doesn’t matter that we don’t have a frontier lab. What matters is that we are—sorry, I’m not saying that it doesn’t matter. I mean, in relation to regulation, we still need to regulate.

And if you look at what Demis Hassabis is suggesting, there’s quite a bit of it covered in the EU AI Act. Not perfectly, and it’s not exactly what he’s talking about.

There isn’t a specific body that you have to put your model in for testing to. It’s kind of left up to the companies themselves in Europe right now. But there’s a lot of alignment between what Demis Hassabis thinks the US should do and what the EU already has in place.

I think that’s pretty impressive. And we talked before about how, in this instance, it seemed like, when the EU brought out GDPR, people talked about the Brussels Effect, where that approach to data privacy filtered out throughout the world.

And then a lot of people were saying, “Well, there’s no Brussels Effect for the EU AI Act. Everyone just thinks the EU is crazy for regulating.”

And, again, I don’t think we’re crazy. And I think Demis Hassabis is showing that, yeah, we need to regulate powerful and dangerous AI, and we need to take the time and space to figure this out.

Justin: Yeah.

Is Google trying to regulate its way ahead?

Justin: No, I think—the other part of me is this, right? Demis—this is a bit of a cheap shot, so whatever, but I’m just going to say it anyway—he would say that, wouldn’t he?

We’re still waiting for the latest version of Gemini 3.5 Pro. It hasn’t been released yet.

And if we look a little bit deeper, what we see is that Google, to their credit, is a mature company. They’ve been around for a long time. They’re a very big company.

And some people would say that the reason Google is falling behind—they have all the data, right? They should be ahead. They should be winning this race. They have the smarts. They invented the technology which underpins everything that we’re talking about. Why is it that they are behind?

And some people speculate that the reason is that startups like Anthropic and OpenAI are focused on one thing, and they don’t have the restrictions around the data. They’re brand new, and so they don’t have the mature controls in place that a company like Google does.

And therefore, when Google wants to do something, it takes them longer because they have to make decisions, request access to the data and decide what the best way is to go forward. And all of this just slows them down a bit.

Whereas Anthropic and OpenAI are startups, and they just go for it. That’s why they’re ahead and why Google is behind.

And so what Demis is trying to do, kind of, is impose some of that thinking on—impose some maturity on the less mature companies. And it’s not going to work. All he’s trying to do is slow them down.

Frank: Yeah, we’ll see. We’ll see what happens because I think he’s right, but I think the reluctance to regulate in any way, shape or form will continue for some time yet.

Justin: Here’s the big thing, Frank, that is a problem with all of this. Even to put a regulation in place, we are moving so fast, right?

We’re going to have big models training small models in the next—it’s already happening on a small scale, but I think, in the next year or two, in your office, you could have Mira Murati’s model training a small model that’s dedicated to your data. Who knows, right? Whatever.

They’ll still be talking about getting this international cooperation and setting up the regulatory body. This thing is moving so fast that the idea, or the chance, that you will have the proper regulation in place in time is fanciful.

So you’re just going to have to make the best of it.

Frank: Yeah.

Did Anthropic make an ad for the apocalypse?

Frank: Well, let’s take a moment for some light entertainment. You sent me a very, very interesting ad that Anthropic put out, and I thought, let’s just take a look at it. And also, let’s see if we have the advanced technology and capabilities to share this. Okay.

Justin: Can we just say, for the people who are not watching this, the ad opens with a house on fire.

Frank: A house on fire.

Justin: Just to set the tone. This is fine.

Frank: An ad for Anthropic. First image: house on fire. Let’s see, where does it go from here?

Justin: Is that a child?

Frank: “Can AI be trusted?” So we saw—it looked like a kid interacting with AI, maybe, and then I’m not sure what we saw. It looked like a desert or something. So let’s see.

Justin: That was a mine with people digging stuff out of the ground.

Frank: Okay. Oh, facial recognition.

Justin: Stop there.

Frank: “Who’s going to hit the brakes if we need to?”

Justin: “Can we trust AI?” This is like Boeing selling aeroplanes by showing films of aeroplanes crashing into the side of mountains.

We just saw an image of mass surveillance done by AI, and then Arlington Cemetery with lots of dead people, and then, “Can we trust AI?” and, “Who’s going to stop it if we need to?”

Frank: Yeah. I see where—

Justin: To this.

Frank: Was that a riot? Looked like a riot.

Justin: It did look bad.

Frank: Homelessness.

Justin: Yeah.

Frank: That is a good question, I will admit. It’s another good question. I love AI. I love using it. Do we actually have to have it? I don’t know.

Justin: I’m hoping for sunny uplands next because I’m pretty depressed so far.

Frank: I’ve got to admit, I am enjoying getting—I think I prefer now when I get an AI bot on a support chat than a human.

Justin: Yes, I’d be there too, but lots of people aren’t. But okay.

Frank: Okay. Are you starting to get the warm fuzzies, Justin?

Justin: Sort of, but I’m still—yeah, go on.

Frank: We definitely have more uplifting imagery here, with people playing in fire-hydrant water and—

Justin: Why do I need AI to play in the fire hydrant?

Frank: It gives you the time. It gives you the time to go out and play with the fire hydrant. And we know we—

Justin: As an AI practitioner, I can tell you that I’ve played in zero fire hydrants in the last six months.

Frank: We now have a kid who is sharing AI with his dad, I assume, rather than hiding under the covers and using it for bad purposes. So it’s all looking a bit brighter now.

We don’t. And then it says, “There’s hope in hard questions. Keep thinking.”

So Sam Altman was actually very funny. He shared this, and what did he say? He said, “I thought this was satire.”

Justin: Yes.

Frank: He kept looking at the Twitter handle that posted it, expecting it to be not an official Anthropic account.

And then he followed up by saying, in quotes, “Hard questions are great, but only if we deem you worthy enough to not silently downgrade you or even get access at all,” referring to the fact that Fable 5 was originally silently downgrading people if they asked about machine learning or AI training and stuff.

Justin: For the benefit of those, again, listening on a podcast, the closing image was a beautiful image of a person standing on a beach, looking out to the sea, and there was a beautiful mansion in front of the—

It was selling the dream. Most advertisements start by selling you the dream. This one started by selling you the fear.

Frank: I love this ad. I love this ad, but when I started watching it, I was like, “Wait, are they publishing a trailer for the film ‘How I Became an AI Apocaloptimist’?”

Justin: Yes.

Frank: “There’s hope in hard questions. Keep thinking.” And I got to the end of the ad, and I have to say my brain just immediately went, “Yeah, it’s kind of a bit like, ‘There’s peace in being a paperclip. Keep using AI.’”

Justin: Right, on that happy thought, have a great week.

Frank: You too, Justin. I’ll chat to you next week. Excellent stuff. Cheers.

Justin: See you.