GPT-5.6 Restricted, Claude Tag Lock-In, and Robot vs SWAT | EP106

Could the US government’s AI safety push end up making the whole AI race more dangerous? Frank and Justin dig into the reported GPT-5.6 slowdown and ask whether AI regulation is creating real safety — or simply giving China and open source models time to catch up.

Plus: Claude Tag arrives in Slack as the next AI agent interface, but raises awkward questions about permissions, tribal knowledge, token costs and vendor lock-in. They also get into AI data centre caps, Europe’s AI sovereignty problem, robots interrupting police operations, and whether access to the best AI models is worth giving up more personal privacy.

00:37 Is Claude Tag the AI agent we've waited for?
03:25 Is Claude Tag a permissions nightmare?
04:41 Will convenience beat vendor lock-in?
09:37 Is Claude Tag the next AI interface?
10:55 Did the US gov block the release of GPT-5.6?
14:45 Will AI regulation help China catch up?
20:59 Can Europe depend on US AI models?
25:08 Should we cap AI data centres?
28:44 Did a robot crash a SWAT raid?
32:07 Is AI access worth your privacy?

► SUBSCRIBE
Don't forget to subscribe to our channel for more arguments

► LINKS TO CONTENT WE DISCUSSED

• What is Claude Tag?
• The White House is asking OpenAI to slow roll the release of its new model over safety concerns
• The Digital Sovereignty Trap by U.S. Under Secretary of State for Economic Affairs Jacob Helberg
• Can the Cold War Teach Us How to Slow Down AI?
• DoorDash Delivery Robot Invades SWAT Scene, Won’t Leave as Cops Flashbang Resident

► CONNECT WITH US
For more in-depth discussions, connect Justin and Frank on LinkedIn.
Justin: https://www.linkedin.com/in/justincollery/
Frank: https://www.linkedin.com/in/frankprendergast/

Transcript

Frank: I’m Frank Prendergast. I’m here with Justin Collery to argue over all of the latest AI news. I’m the cautious one. Justin’s the techno-optimist who wants to let AI power ahead at all costs. We’ll see if you still feel like that at the end of this conversation.

Is Claude Tag the AI agent we’ve waited for?

Frank: And so, Justin, interesting news. We talked before, there was that whole thing about Open Claw and the various seven different names it went through, and how you could set it up and you could just text it and tell it to do stuff, and it would go off and do things. And when we covered it, we were like, “Yeah, it’s a bit dangerous right now, wouldn’t necessarily hook it up to everything,” et cetera. And you said, “Look, for general users and for businesses, the AI labs are going to do something like this, and it’s probably better and safer and just more intelligent to wait until there is a business version of this that you can rely on.”

Well, it looks like maybe we have that.

Justin: Yeah, much more than that. Much, much bigger than that. But I see where you’re going with it, right? I see where you can make the inference, but it’s much bigger than that.

Frank: So Anthropic announced Claude Tag, which lives in your company’s Slack, and you can tag Claude and just tell it to do stuff. Is that right?

Justin: Well, it does more than that, you see, and people are up in arms over this. Andrej Karpathy, who is a brilliant educator and stuff, was shilling Claude Tag, and he got a lot of heat on Twitter during the week because he said that this is really cool, and he’s been using it a lot and stuff like that.

So Tag lives in Slack. Slack is an application that a lot of companies use to allow teams to talk to each other, share ideas. It’s okay. In this conversation, it’s important because it’s where the tribal knowledge lives. So it’s where people share their thoughts and ask questions and give answers and all that sort of stuff.

Tag lives inside Slack. So what Tag does is, as these conversations are going on, it’s learning about your organisation. It’s picking up all of that tacit knowledge from all across the organisation, from all the different people, and it’s synthesising and storing that knowledge. You can then ask it questions, and it will answer, which is great.

To me, that sounds like a really good use case, and you can also ask it to do things. So it’s Claude, right? See, it’s Claude Code somewhere in the background, and it can go off and do things. Loads and loads of furore over this for many different reasons.

Frank: And let’s just pause on that for a sec, ’cause I did think that was interesting, and I know Google have something similar and OpenAI are working on something similar, or actually, yeah, OpenAI have something vaguely similar. But Claude, in this instance, works on Anthropic’s servers, so it works in the cloud essentially.

As in, you don’t need your organisation having a Claude app running and make sure that the laptop doesn’t go to sleep for it to work. It’s going off into the cloud and making the changes and interacting with your GitHub and your…

Justin: Yes, I was gonna say we should have a bell here for bing, bing, bing.

Is Claude Tag a permissions nightmare?

Justin: This is the first red marker, the first sort of uh-oh, back up the truck. So one of the very first big issues with it is it takes actions for you, but the actions are permissioned by it. So there’s a lot of concern among the community about how do you, A, protect knowledge.

So if it’s collecting all of this tribal knowledge across the organisation, well, big organisations work on a need-to-know basis. You don’t want everybody knowing every piece of information. You only want the information that you need in order to do your job. So how do you protect that? And apparently there’s ways of setting it up and whatever, but that’s on you, the company, and so there’s big risk in that.

Second thing is, if you tell it to go and do something, it interacts with your systems using its own credentials, not your credentials. And there’s a very complicated setup about how you make sure that it has the right credentials to do the right thing. Loads of scope for it to go wrong. So what happens if you should not have access to a particular SharePoint site or a GitHub repo or a thing?

The teammate that you have now that lives in Slack might have those permissions, and there may be a way for you to basically jump the control barrier to have access to things that you shouldn’t have. Loads of concerns around that as well. But those aren’t the big concerns, Frank.

What are the big concerns?

Will convenience beat vendor lock-in?

Frank: Well, I would imagine that the big concern would be token costs because I’m hearing, I’m reading a lot of articles right now about token costs being a massive concern for companies, and they’re finding that people are burning through tokens doing stuff that would probably be faster and easier for them to have done by hand and certainly cheaper.

And I’m reading a lot of big organisations who are saying, what they are focused on right now is figuring out, “Okay, what’s the workflow here, and how do we figure out the exact right context to give the AI in order to perform the task that we want done efficiently?” And this seems like the opposite.

This seems like, “Yeah, it’s fine. Just give Claude access to everything, give it access to all the information, and let it figure out the right context and the right…” And that sounds to me like it’s gonna burn through tokens.

Justin: It may do, right? And for the thing we’re gonna talk about after this slot, right, I wouldn’t be concerned about the token cost. I think token costs are gonna go through the floor in the next 12 months. So even though people worry about it today, I’m not sure you’re gonna be worried about it in six months.

Frank: We’ve been promised that since the very beginning of AI though, haven’t we? We’ve been promised that the…

Justin: We can thank the American government for ensuring that it’s now gonna happen, so don’t worry about it. So people are concerned about, yes, token cost. You’re totally correct. People do say that, but I think that’s not the biggest concern. The biggest concern is vendor lock-in. So now you have a company which has all of your tribal knowledge, and they have it, you don’t.

So if you use this product and you get a lot of use out of it, and in six or 12 months you decide, “Yeah, this is really cool, but the token costs really, they are, they’re a bit heavy. I’d like to go and switch to another provider,” how do you get your data out? How do you get this tribal knowledge out of that and into another system so that you can switch from one provider to another?

There is no way that we know of at the moment to do it. So the more you use it, the more you are locked into using Anthropic’s products, and that is the thing that is getting people most exercised.

Frank: That’s really interesting, and I will be really fascinated to see how this plays out because we live in a culture that is just obsessed with convenience, and I just wonder how much convenience will win out over kind of thinking this through and being like, “Yeah, we kind of need portability and we need cost-effectiveness.”

I wonder, will people just be like, “This works. You just set it up. Yeah, let’s just use it.”

Justin: You’ve clearly never worked in a large organisation where usability is not the prime factor. Control is the prime factor. Paranoia and control are actually the driving factors, and if something takes 10 tickets to set up in three weeks as opposed to just signing up for it, so be it to ensure that we have control.

So usability for a big org isn’t the prime… I agree with you. For a smaller…

Frank: Let me push back on that. No, no, but let me push back on that even for the big orgs for a second because, for example, right, just a quick detour. There was a Japanese model called, I think, Sakana, Sakana Fugu. Did you come across this?

Justin: This is the new one that was released this week. It’s brilliant, apparently.

Frank: And so what’s fascinating about it is that it’s not really, in terms of the benchmarks that they put out, it’s not really that it’s a new frontier model that is making it powerful.

It’s that they’ve trained one model to be this master orchestrator of all these other models, and apparently, according to their own benchmarks, it’s really good. So if we just imagine for a moment what a company could do, you could work out your own workflows, figure out, use all of your SOPs, use a system like this that uses lots of open source models, have an orchestrator picking the thing that picks the one that’s gonna be most cost-effective for this task, and…

But that’s a huge project in and of itself. Or you could say, “Oh, Anthropic, that’s a big brand name. We know Anthropic. They’re focused on safety. They’re focused on ethics. Let’s give them our money,” and it just works. Do you not think there’s…

Justin: Yes. Yeah, so yes, they’re gonna get customers. I think it’s a great product. I just think that it’s not gonna get as many customers. Big companies won’t go for this. Some of them will, but not all of them. Now, but look, this is the first iteration of the product, so I wouldn’t get too upset about that, right?

If there was another iteration of the product where that thing didn’t run on the Anthropic cloud, it ran within your own cloud, right, and you can host it yourself, then it becomes a much more interesting proposition for the companies that I pointed out before. I still think that for companies that are not in a regulated industry, companies that are not absolutely massive or paranoid, this sounds like a great product.

Is Claude Tag the next AI interface?

Justin: So okay, why is it a groundbreaking product? It’s a groundbreaking product because we’ve gone through a number of iterations of how to interact with AIs. So the first one was the chatbot interface, right? Where you typed stuff and then you copied it, and you wrote code that way and you did stuff.

Second one was in a terminal, right? So you use Claude Code or Codex or something like that in a terminal. This could be the next iteration where it’s a team member, and it sits in your channels that you communicate with your team members with, and it picks up knowledge. It’s obvious to me that this is a really good idea, right?

You could imagine at a meeting, right, that you could even call… I can imagine in the future, there’s gonna be, you could just ask it questions, like in natural language like this. You don’t even have to type, and it’s like, “Here, did we not talk about this before?” Or, “Can you read us out the minutes from the last?”

You know what I mean?

Frank: In a way, it’s a limited version of, I think it was one of our very first podcasts where you said, what companies need to be doing now is they have microphones everywhere and just record everything, and so this is like a limited version of that. It’s using the Slack conversations, not any and all audio conversations.

But if we… Yeah, I think you definitely deserve a prescient badge there for seeing this coming.

Justin: Well, let’s see. We’ve had a… So let’s move on, right?

Did the US gov block the release of GPT-5.6?

Justin: We were promised, we were promised a couple of months ago by OpenAI that every Thursday they would release something big and new and shiny, and it would be a fantastic Thursday. We meet every Thursday to discuss what we’ll talk about on a Friday, and it seems like every Thursday evening something happens to scupper our plans.

And so it was this week, where we should have got GPT-5.5, 5.6? 5.5, 5.6. Thank you. And we didn’t, because the American government has stood in and said, “No, no, no, OpenAI, you cannot release this model. We want to approve every single person, person by person, organisation by organisation. We want to see exactly who they are, and we want to be the ones who approve if they have access or not.”

This is a moment of mendacious folly. It is the most ridiculous thing I have ever heard, and it’s gonna lead to terrible outcomes. But you must be delighted.

Frank: Well, I’ll tell you why I think it’s interesting. First of all, we talked about… So Anthropic basically used this model voluntarily. When they came out with their Mythos level of models, which were really good at cybersecurity, they figured this could be dangerous. Bad actors would be able to hack systems too easily.

So they created Project Last Wing, a group of 50 companies that was later expanded to, I think, 200 or whatever, and they rolled it out slowly just to those companies, and Mythos has not been available to the public. Mythos was then shut down by the US government, who said, “Actually, only Americans should be able to use this.”

But when we were talking originally about Anthropic limiting the release of Mythos, you were saying, “Look, it’s not gonna matter because OpenAI are gonna come along and they’re just gonna release a model.” And my argument was, this is why we need regulation, because you can’t have two companies just taking their own approach to safety.

We need something that says ideally globally, but certainly every country needs to figure out, well, what do we consider safe here? So this is interesting because I think at least it shows that when the US government stepped in and said to Anthropic, “Right, we want you to make these models only available to US citizens,” this now shows that that at least was a genuine safety concern and that the US administration is now taking AI safety seriously, and that it wasn’t part of just the fight that was going on between Anthropic and the US administration.

So I think that is why I’m kind of happy. I’m happy that this is a signal that AI safety is being taken seriously at the highest levels in the US administration. What I’m still not happy about is that surely, it’s not scalable to do this model by model and just interject when they see on Twitter that Sam Altman is saying, “We’re getting close to releasing 5.6,” and saying, “Whoa, whoa, wait a minute, wait a minute. This should be regulated.”

Justin: To be fair to Sam Altman, from what I’ve read, they were working with the White House or with the regulators in the US even before Fable was pulled from the market in order to work through this regulation, right, and try and do this. So this is a terrible idea for a number of reasons, right?

So you cannot have regulation without slowing things down, as has been demonstrated now. We don’t have access to Fable, we don’t have access to Mythos, and now we don’t have access to GPT-5.5. And this will slowly be rolled out to people. There’s a couple of implications, right, of the course that’s happening at the moment.

Will AI regulation help China catch up?

Justin: So one is, if you think about the open source models, particularly from China, they are currently six to 12 months behind, and that lead has been maintained because the frontier has been pushed forward by the US closed source companies. By implementing this regulation, which slows the rollout of these frontier models, it gives time for those open source models to catch up with the American ones.

So that’s a problem, right? It’s a problem for mainly OpenAI and Anthropic, because their business model is based on being able to charge a premium for the products that they develop, right? So if the Chinese models catch up, it makes it harder for them to charge a premium for their products. And also this approach then reduces the size of their total addressable market.

So if you have to be a US citizen based in the US, and you have to go through some friction to get approved to use those models, less people are going to use them. So their market size gets smaller, right? So what the logical outcome from the US government’s point of view of doing this is, you ready to be shocked, they’re going to ban the Chinese models.

So the US government doesn’t want to bankrupt OpenAI and Anthropic. They want to make sure that they’re in the lead, but they do want to control them. The open source models, as a result of doing that, will catch up and be nearly as capable, and therefore, to protect their own companies, the US government will ban the open source models, try to stop people from using them.

That’s a terrible idea, but that’s what they’re going to do. So then we come on to the Europeans, right, the mid powers, the Europeans, the Canadians, the Australians, those people. What do they do in this, right? So they’re now caught between should we use the Chinese open source models or should we try and buddy up, right?

The Americans are going to try and foist this regulation upon the mid powers, and it’s the only point of leverage that we have. So if the American government turns around and says, “Well, you must follow this regulatory regime in order to access these models,” the question back is going to be, “That’s fine, but you’ve got to guarantee us access to those models in order to follow your regime.”

And there’ll be loads of Europeans that are really upset that they’re going to have to share their passport details and God knows what else in order to get access to the latest model, right? So again, this just causes friction and reduces the total addressable market for OpenAI and Anthropic. Now, the other thing that I see that comes out of this is OpenAI and Anthropic are funded by two things, right?

One is investor confidence, and two is revenue from the clients that they sell to. And therefore, if you reduce the total addressable market, that means they’ve got less money, both from the VCs and from the money that they receive from clients. That has to slow progress. Has to slow progress.

There’s less money to spend on developing models. And so the outcome of all of this, that I see, is they’ll try and ban the Chinese models, and what will actually happen, this comes back, sorry, to my earlier point about don’t be worried about token costs. What is the incentive for OpenAI and Anthropic in that sort of environment, right?

The Chinese models are catching up. They’re basically free. Their total addressable market is low. Their motive is to make money at the end of the day. So what they will do is they will make models that are about as capable as Fable or just a little bit less capable than Fable, but they’re faster and cheaper.

So you don’t have to have government approval to use them, and everybody can use them, and they’re just gonna get cheaper and faster, and they have to compete against the Chinese models.

Frank: But this is what’s kind of crazy about the situation we find ourselves in right at this moment because, for example, you’re saying create a model just under that capability and you’ll be fine. But there’s no guarantee of that because there’s no regulation. There’s nothing anywhere on paper that says this model’s fine, this isn’t, or this amount of compute or this amount of training.

There’s nothing. So you just have to go off and create a model and then hope for the best that the administration isn’t gonna say, “No, actually, we don’t like that one. Let’s hold that one back.” This is the argument that the EU makes for regulation, that it creates an environment of certainty that actually does allow innovation to flourish.

I think that this level of uncertainty right now in this moment, it will also be bad for AI labs in the US.

Justin: Hmm. On this we are agreed. I think it’s terrible for the AI labs. I think it’s terrible for the AI labs in the US, and for those, when you think about it, for those really capable models. So you could reasonably assume that there’s a Mythos 2 already developed or very close to being developed, even more capable than the previous Mythos.

As it stands, that model has a customer of one, the US government. That’s the only person they can sell it to right now. They won’t allow you to have it to anybody else. There’s also, I would worry, an incentive for the US government to keep it that way. So if you have all these cheap models coming in from China and so on, and they’re gonna be as capable as Mythos, in order to protect yourself, you would want to be one step ahead, so you want to have the next model that nobody else has access to.

That’s not a good idea either. And the other thing about this regulation is that it’s good for about 12 months, right? In 12 months, you’re gonna have a model which you can run on a couple of Macs in your house that will be as capable as Mythos. There’s no way to stop that.

So what are you achieving by having everybody give their personal private details to some regulatory authority? There has to be a better way to regulate these, and I don’t know what that better way is. To your point, I do agree that there should be a way to test the models or to come up with a thing that says, “If it does more than this, then it requires a higher scrutiny.”

I don’t think that compute is… ’cause the thing, you know that Japanese model you just described…

Frank: Mm-hmm.

Justin: I’ll just create five models that are all really good. One is really good at chemistry, one’s really good at psalms, one’s really good at something else. You know what I mean?

And I’ll just use a router to talk between them, and I won’t break any of your rules. With everything you do, you’re creating incentives for people to get around the regulation or to do stuff. And so it seems to me at the moment, the upper bound of intelligence is not a technical limitation anymore, it’s a regulatory limitation.

Can Europe depend on US AI models?

Frank: You also brought my attention to an article by the Under Secretary of State, Jacob S. Helberg, which is relevant here ’cause he was saying that this idea of digital sovereignty is not a good idea. And it’s relevant here because, of course, digital sovereignty and AI sovereignty is now a huge topic in the EU because we can’t access the latest models from Anthropic or OpenAI, two of the biggest labs, two of the labs creating the most capable models.

And yeah, he was saying AI sovereignty was not a good idea, and I think as you put it to me when you sent it to me, they’re saying, “You don’t need AI sovereignty, just trust us. It’s grand. It’ll all be…”

Justin: Yeah, just us. But not with the latest model. We’ll look after that for you, ’cause you’re not old enough to have that yet. When you’re old enough and responsible enough, we might give you access to the latest model, which we have and you don’t.

Frank: And there were kind of two big assumptions in his essay, I thought. Because, look, he made some valid points, obviously, but the assumptions were, first of all, that if you are chasing sovereignty, you are never going to innovate. And I just think, I don’t know where that assumption comes from because, for example, the EU is chasing, is going after AI sovereignty and, as part of their AI plans we talked about in a previous show, we do have these plans for an incubator for three companies to look for alternative ways of approaching AI.

So it’s not that AI sovereignty completely precludes any form of innovation and that all you do is try to match the latest model. And then the second thing was that he made the point that basically he said that when the world was moving to 5G, the US didn’t have a way, they didn’t have companies that could build it, and so they relied on their allies to do that, and then they focused on the infrastructure and the things that would be all of the apps and interesting things that 5G would be used for.

So his point was, you buy the stuff from your trusted allies. The problem is, do your allies cut you off from all of the latest, most powerful models? And I think there’s a lack of awareness there that the actions of the US lately are having a major impact on how they’re perceived in terms of being an ally.

There was a study that really shocked me recently that said that only one in 10 Europeans currently saw the US as an ally. One in 10. So while I agree with him in spirit that maybe… In fact, originally, way back when, I don’t know when we started this podcast, you asked me about, shouldn’t the EU have our own frontier models?

And I was like, “I don’t know. We don’t have Facebook. We don’t have a Microsoft. We don’t have a Google. Do we need a frontier AI lab?” You know, maybe we can just use the States. And now I’m like, “Yeah, okay. I’m not so sure. I think we need our own frontier model.”

Justin: Yeah, and hopefully we’re not too late. You had a great idea during the week, by the way, I have to say, ’cause the question… What he said, by the way, that secretary, what he said in normal times made total sense, right? Just trust us, work together as allies. If all the other stuff that was going on, what he would’ve said would make total sense.

And what I guess it highlights is that we’re moving at such a fast pace that it’s impossible for one arm of the government to know what the other arm is doing. And I mean, I don’t know how to manage that. It seems to me like you need to come up with some fairly simple rules that we can all agree on that we know are easily implementable in order to affect some sort of a control.

Should we cap AI data centres?

Justin: You had a pretty good idea during the week.

Frank: Well, yeah, when we were chatting, we were just talking about how the race mentality is consistently the problem. So the problem is nobody wants to slow down, nobody wants to pause because the other person won’t pause. And we need to have a way to tackle that. And the only way that I could see to tackle it is to get together, cooperate, and agree on data centre caps, because I just don’t see any other way to tackle this.

Justin: I think that’s actually a pretty good idea, right? So there was another article during the week and it was to do with how do we monitor. So if we did have an agreement that we were gonna pause or slow down, and I don’t think anybody’s gonna pause. I’m not really sure anybody’s gonna slow down, but we have to agree somehow to manage this thing, right?

And if you think about when we had agreements in the past, let’s say, to do with nuclear weapons. Well, it was easy to monitor if somebody was blowing up a nuclear weapon or not, right? You had seismometers and they caused great big explosions that you could measure and stuff. How do you do the same with AI to say, “Lookit, we’re gonna slow down”?

And so measuring power or data centre capacity is actually a pretty good way to do it and say, “Lookit, we’re gonna have a cap.” And that is another limiting factor. Will it work? I don’t know. I don’t think so. We’re into the weird time where, and this is pretty foreseeable, right, that a government would start to step in and say, “This is now a national security issue. We’re not gonna allow you to release this model.”

But that has all sorts of implications that the people who are making the decisions haven’t really thought about. And I still say the biggest implication of the thing that’s happened today is that 12 months from now, you will have a Mythos-level model that you can run locally on your machine, because there will be a great demand to do that.

And that is a world which is less safe than the one that you have today, where those big models are housed in data centres.

Frank: Yeah. Yeah. One of the other benefits of the cap on data centres would be that we would be in less danger of either China or the US doing something, because of this race mentality, that endangered the world or the human species. But it would also mean that they’d be up here at their data caps and unable to make an AI that turns us all into paperclips, and meanwhile, the EU could work its way slowly but surely up…

Justin: Give them a target.

Frank: The cap.

Justin: Yeah. And because we’ve got nobody who’s actually doing any AI stuff here at the moment, we could sell that spare capacity to the Americans or the Chinese, and then we’d have some leverage in order to make sure that we have access to the latest models. I’m not sure that’s gonna happen, but we can see.

But I think people should really think carefully about what happened this week with 5.6 and what’s happening with Fable. It’s a very… this is an inflection point that we will look back on. This is a thing, they have to get this thing right, and it’s not been done right at the moment, and it has weird unintended consequences that it’s really hard to predict, except that models will get cheaper over the next 12 to 24 months because that’s the incentive, and there’s gonna be a lot more access to models that you can run locally to get around these restrictions.

That’s what you’re incentivising at the moment.

Frank: Yeah.

Did a robot crash a SWAT raid?

Frank: Well, final story, and I know we usually try and end on a funny story, and this is not a funny story. This is not a funny story at all. And basically, the question I think that this story brings up is, should robots have to obey orders from law enforcement? So what happened was in Arizona…

Justin: If they’re delivering donuts, absolutely.

Frank: So in Arizona, basically a SWAT team or something similar to a SWAT team was on a mission outside this house, trying to figure out, you know, was it time to go in, like you see in the movies, are we going in or is he coming out, and all that jazz, armed to the teeth presumably, and up rolls a DoorDash robot, presumably, as you say, bringing the man inside his donuts or his pizza or his Starbucks or whatever it was.

And the cop said, “Get out of here. This is a police operation. We’re about to go in. Turn around. We don’t want you here.” And the robot had no real way of going, “Oh, this is, I should do what the police say.” Now DoorDash say it did what it’s meant to do, and that it did actually realise that there was something going on and that it was a police activity, and so it stopped where it was.

It didn’t roll up and make a delivery to the guy, which that’s the bit that I was thinking about, and I was like, oh yeah, well, what if it wasn’t a pizza? What if it was an AK-47? What…

Justin: Oh, that’s a good point. I hadn’t thought about that. That’s a great way to deliver all sorts of nefarious things, not just AK-47s. But also it’s a distraction for the guy. Surely the guy comes and he opens the door to get his pizza, then the SWAT team could move in.

Frank: It could, it could have gone either way, I guess.

Justin: That’s brilliant, Sodas.

Frank: But the mere… This caught my eye just because of the mere fact that we now live in a world where a robot blunders its way into the middle of a SWAT team activity, and they can’t get rid of it, and they just have to work around it.

Justin: Do you know what it reminds me of? Do you remember that film, it was The Naked Gun, I think 33 1/3, and they were redoing a famous scene from some other film. Frank Drebin was in some train station, and they were waiting to do basically a similar operation. And just as the paddies are about to come out, there’s a woman comes out with a baby and she’s trying to go down the stairs, and then there’s another, there’s an old lady that needs to be helped down the stairs, and he goes to help her.

And then everything starts going wrong, and then shots start getting fired. He’s trying to catch babies and do all sorts of things. And now we have DoorDashes in there as well, and he sort of stops and picks up a donut. Has a donut, gets a pizza. I think it’s great. Look, it’s a brave new world that hasn’t been explored either in writing or in real life yet as to what you do when you got…

Just think about it, Frank. In another year or two, or three years, we will have loads, we’ll have self-driving cars driving stuff around the place. There’s no need for delivery drivers anymore. We’ll have these DoorDash things coming, delivering pizza. What a time to be alive, eh? Isn’t it great?

Frank: Robots coming in doing your chores for you on a Tuesday.

Justin: Fantastic. Brilliant. Look, and I hope that we have access to Fable.

Is AI access worth your privacy?

Justin: By the way, question for you, Frank. If, in order to get access to GPT-5.6 or Fable, you have to send a biometric copy of your passport to the authorities in this jurisdiction or any other jurisdiction, will you do it?

Frank: I mean, I will.

Justin: And in…

Frank: I will, but am I happy about it? No. I was actually listening to Cindy Cohn from the Electronic Frontier Foundation on Hard Fork recently, and she was talking about… I’m afraid, basically, my mindset is kind of like, well, look, I’ve scanned my face for Revolut and I’ve given my passport to LinkedIn and the horse has bolted.

I’m at that point where I’m like, “Yeah, it’s too late.”

Justin: I don’t think so.

Frank: No, Cindy Cohn was trying to snap me out of it and was making the point that surveillance data goes stale very quickly, was her point, and that what we always need to be worried about is future data, not current, not old data.

And so she was saying you can’t have that mindset of like, “Oh, the horse has bolted, everyone knows everything about me.” You have to start from now and kind of say, “Right, what am I actually willing to part with?” And I think it raises really serious questions.

The reality is, if OpenAI ask me for my passport for 5.6, I probably will give it ’cause I want to play with 5.6.

Justin: Well, let me just say…

Frank: We’re entering into a new era of a level of digital surveillance that I’m not sure I’m comfortable with.

Justin: I was looking at this through a sort of economic lens, right, and a company lens, right? But as I think about it through the Frank lens, it’s horrific, right? So because you’ve got your passport associated with it, that means every chat that you ever have is now associated with you and your passport. And we all know that AIs, we have our deepest, darkest thoughts, like we tell them everything because we think it’s a safe space.

Now it’s not a safe space anymore because somebody’s tracking your conversation. And then as you’re going in from one country to another, and they take your passport and they put it into that machine, they now have a new check they can do on you, which is your deepest, darkest secrets that you’ve discussed with your AI.

Is that an okay thing to do?

Frank: Yeah. So I mean, we are in a situation where the US government are saying to AI labs, “We need to know who X, Y, and Z is if they’re gonna use this.” For example, with Anthropic, they said, “Okay, US citizens only.” That then of course means Anthropic have to be able to check your citizenship.

That means checking your passport. When you couple that with the fact that one of the hard lines in the sand that the Pentagon didn’t want drawn for them by Anthropic was Anthropic saying, “Look, we do not want our systems used for mass surveillance of the domestic population.” When you put those two things together, it’s a little bit scary.

Justin: There you go. Well, on that happy note.

Frank: Yes, indeed.

Justin: On that happy note, Frank, let’s see what happens next week. No doubt there’ll be some other mad thing that will happen on Thursday.

Frank: Of course there will. I’ll chat to you then. Cheers, Justin.

Justin: Take it easy.