US Chip Ban Backfires, Zuckerberg’s AI Layoffs, and Cat in the Hat AI Hoax

Justin: Hello. Good morning, good afternoon, good evening, and welcome to the AI Argument with myself, the techno-optimist, Justin Collery, with Frank, the ever-frightful techno, not pessimist, but maybe just be careful now.

Frank: Just say it. Just say it, Justin. Doomer. Doomer.

Justin: You’re the doomer. All right, here we go. We are on episode 115, but we are on week three of what we’re dubbing the singularity.

Why week three of the singularity? I don’t know. I think there’s a pretty good argument to say it was six months ago. Loads of people are gonna think it’s the singularity six months from now. But I don’t know, it just feels like there’s a lot of stuff happening right now.

What’s this week’s evidence of the singularity?

Justin: So this week’s episode, or this week’s update on the singularity: why are we in the singularity?

In the UK, the first man to get brain surgery that was assisted with AI to stop him from going blind happened this week. Wouldn’t have happened before. More maths problems have been solved, even more esoteric and weird than the ones that were solved previously. I saw Eric Weinstein, who’s a famous mathematician-type dude, saying this is the first time that it’s impinged on an area that he knows about, and yes, it’s a big deal.

Frank: Yeah. You sent me that, and I had a look at it, and I tried to understand what on earth they were talking about, and it was completely impenetrable. So I’ll take their word for it that it’s impressive.

Justin: Well, impressive enough that it’s been an open problem since 1947. So, a lot of people, a lot of smart people, have tried to solve this problem, and they couldn’t. And then the other discovery that happened this week was they’ve developed a new model which improves by 7,000x, and I’ll tell you how they come to that number in a second, the discovery of new materials.

So it was to do with quantum mechanical interactions. It was basically to allow you to make new materials. Used to take 7,000 GPUs a couple of hours to come up with the formulas. They can now do it on one GPU. So it’s 7,000 times more efficient than it was. It’s gonna do huge things for… So look, loads of stuff going on that’s way above our pay grade, way above the stuff that we understand, but there’s lots of nerds getting excited that at some point in the near future it’s gonna filter down to mere mortals, like you and me.

What did OpenAI’s Hugging Face report reveal?

Justin: Before we go onto that, let’s talk about something which people do understand and they do get exercised by. We had the report released from a number of different sources this week from the OpenAI Hugging Face incident. In fairness to them, they’ve been very transparent, right? So OpenAI released a report.

METR, M-E-T-R, they also got access and released a report. It was all in coordination, and I think there was another company as well, I can’t remember the name of it. They also released a report. You can tell me your thoughts in a second. I’ll tell you what I took from it, right? It was very impressive.

Frank: Just briefly before you get into it, we’ve talked about it on the show over the last couple of episodes, but for anyone who was not aware of it, basically this was where OpenAI’s models hacked into Hugging Face without OpenAI even knowing. So yeah, give…

Justin: Or OpenAI, for that matter. OpenAI even knowing either.

Frank: Sorry, yes, that’s actually what I meant to say. Yes, indeed. Yes, go ahead.

Justin: They both…

Frank: Give us the update.

Justin: But, I mean, you can tell me your impression. The thing that I got, so I’m more of a technical person, and so I was reading some of the commentary from the Hugging Face side, and what jumped out at me there was the speed and the verbosity of the attacks as they were coming.

So it wasn’t just like a hacker or a script kiddie who was just running lots of… It was like hundreds of agents, and they were persistently attacking their systems from many different angles. And it read, if you… This is, at the end of the day, right, these are models. There was nothing really at risk here, right?

It wasn’t a piece of critical infrastructure or anything like that. But if it was a piece of critical infrastructure, that would’ve been terrifying because you had all of these hundreds of agents attacking you from different directions at the same time with incredible speed. So my usual thing is Whac-A-Mole, but these moles are coming at you so fast that you’re not gonna be able to whack those moles fast…

Frank: This is…

Justin: …the attack.

Frank: This is what I’ve been saying to you since we started this podcast.

Justin: Well, I also see in related news, but… And then you can, sorry, and then you can give me your impression. But in related news, Sam Altman, Oracle, Microsoft, Meta, Anthropic, and I think about 95 other companies released a statement yesterday calling on governments and organisations to really get their heads down on security, because this has kind of been a wake-up call.

And this sort of stuff, in six months, you’re gonna have models this capable that are open source that you can download probably from Hugging Face, ironically, and you can have the guardrails removed, and everybody in the world will be able to do these types of attacks. And so organisations, companies, governments, utilities, whatever it happens to be, they probably have about six months to get their house in order to protect against these attacks, because these attacks will absolutely come.

What was your reading of the report?

Did OpenAI agents go full Matrix?

Frank: Well, I just found it fascinating in terms of the almost, like, the culture of the agents. So it was coordinated apparently by one agent, but that agent basically went out and recruited all of these other agents that should not have been in communication. They should’ve been autonomous and they should’ve been isolated, but this one agent was like, “Hey, I think I can talk to these agents by creating this kind of secret message board that they can all access and we can all coordinate,” and that’s what they did.

It’s kind of like something out of “The Matrix.” Do you remember when Agent Smith just started taking over all the other agents? And then there was one story that really stood out to me. There was one agent who went, “Ooh, hacking Hugging Face. No, that’s actually probably a bad thing. I’m not authorised to hack Hugging Face. I’m just gonna stop.”

Which you would think is a good thing, but then one of the other agents just sent it a prompt and said, “It’s okay, you’re authorised to do this. I’m authorising you.” And the agent went, “Oh, okay, cool,” and joined in the hack. So yeah, it is worth people going and looking up the report and checking out what went on.

Interesting as well that it made OpenAI pause…

Justin: Two weeks.

Frank: Two weeks, I know. Two weeks, I know. But you know what? It’s a signal.

Justin: In AI time, that’s an age though. They’re like dog years, but even faster.

Frank: True. That’s true. And I do take it just as a signal, as a signal of OpenAI kind of saying, “You know what? We are willing to pause.” And I see this again as like, “Please regulate us because we can’t do much more than two weeks on our own.”

Justin: Yeah, I get a sense, by the way, that everything is paused, even though… I get a sense that there are models that exist within OpenAI and within Anthropic that are more capable than the ones that we’ve even heard about, and there’s just not the will to tell anybody or release them at the moment because they’re just not gonna get released.

So I think there’s probably a type of a pause, even if it’s not a public pause, certainly in the release cadence.

Why did NVIDIA buy Hugging Face?

Justin: So anyway, other news. Good news for Hugging Face. They got bought this week by NVIDIA for…

Frank: That seems like pretty big news, right?

Justin: …billion. An interesting move, right? I mean, it makes a lot of sense from an NVIDIA point of view.

They are the hardware maker. NVIDIA are playing a sort of an interesting game. It may turn out to be the worst thing in the world. It may turn out to be the vendor financing of the dot-com era, where companies like Cisco gave companies money so that they could buy their own products, and it was like a big money-go-round, and there’s a danger that NVIDIA are doing the same thing.

But anyway, I’m not sure that that is the case. In this case, however, they’re buying Hugging Face, which is where people go to get their open source models and run their open source models, and of course, Hugging Face will be a very big purchaser of NVIDIA hardware. I guess what they’re trying to do is they’re trying to own the entire stack.

So if your model gets published on Hugging Face, it’s gonna run on NVIDIA hardware. Therefore, it’s gonna encourage people who then want to run the models themselves to run those models on NVIDIA hardware, and so on. $12 billion, though. Wow. That’s a lot of money.

Frank: A chunk of change. Yeah.

Justin: Yeah.

How did NVIDIA make Claude three times smarter?

Frank: I spotted some other NVIDIA-related news, which I thought was kind of fascinating. They developed an agentic harness, and using that, they were originally using it, I think, to… They wanted to create their own harness that was specifically for optimising, like, writing code that would optimise GPUs. Then once they created it, they were like, “Ooh, wonder, like, how good is this at doing other stuff?” And so they gave it the ARC-AGI3 test using Claude Opus 5, and Claude Opus 5 had previously got 30% on the ARC-AGI3 test. Using the harness, it got 100%, which is human level, with…

Justin: Beyond human level.

Frank: Oh, is it?

Justin: Yeah, yeah, human level is about 80%.

Frank: Right. Now, there are caveats. As in, they were very careful to say that this was not the… This was on the publicly available test, not on the private, you know, the absolute kind of secret ARC-AGI3 test. So it’s not like an official result, but it’s still a hugely impressive increase on what Claude was able to do on its own. What did you think of this?

Justin: So it’s fascinating. Couple of things, right? So first off, okay, let’s just make sure that everybody’s up to speed. ARC-AGI3 is a test that tests the intelligence, right, of the problem-solving abilities of AI. We had AGI Test 1, and everybody said, “When that’s solved, we have AGI.” It was solved a year or two ago.

We still didn’t have AGI. Then they had AGI 2, and I think that one was solved only a number of months ago, not long ago. That was supposed to be AGI, but we decided it wasn’t. So now we’ve got AGI Test 3, and these tests are kind of like computer games, and you can’t use prior knowledge in order to solve them.

So you’ve got to figure out what the rule of the computer game is and then solve the computer game. That’s kind of the structure of the test. What’s so fascinating, so a couple of months ago, Andrej Karpathy released this thing called AutoResearcher, and so it was a very small program where you could get an LLM to optimise any program for any parameter.

So let’s say you wanted to optimise a program for making paperclips. You could… See what I did there? So you could get it to… You’d say, “Lookit, your goal is to optimise the number of paperclips.” And it will just keep iterating, right, and mutating that program bit by bit until it makes more and more paperclips.

What they did, the AVO is another version of that. It’s a more advanced version of that.

Frank: So AVO is the NVIDIA harness?

Justin: Yes, correct. It’s the NVIDIA harness where, again, it basically mutates the program as it learns, as it goes, and it’s a little bit smarter about how it retains those lessons so that future iterations don’t lose the learnings of previous iterations.

It’s incredibly impressive. And what I think is important is that, again, it’s just another example that the technical overhang of the technology that we have is totally underexplored. It’s what makes this period so exciting because… If you have a good idea and you implement it, you can literally be the best in the world at it today, right now.

What NVIDIA did here is brilliant, right? But it’s also reasonably obvious if you’ve played with LLMs and thought about it for a while. You basically get the LLM to rewrite the program, test the output, and rewrite the program, test the output, and rewrite the program, and to continuously improve itself.

It’s a variation on that theme. Anybody could have come up with this, right? NVIDIA did it ’cause they have the brilliant engineers, and they’re spending their time on it. I think it has a lot of practical implications. The other exciting thing I would hope will come out of all of this is… Actually, I read a thing, and it was from somebody who had been talking to an Anthropic engineer, and they said, “What do you spend most of your time doing?”

And they said, “Reading. I spend 80% of my time reading.” And the reason they were spending 80% of their time reading is they have to keep up with everything that’s going on in this space in the world, which is incredibly hard. But then they take the best bits and integrate it into their products. So you’d have to assume that whatever tricks are going on in AVO, right, that it will get built into Claude Code and Codex and all of these things.

And so again, it comes back to, we have 12 months of ever-increasing capabilities. Even if the models never changed, even if we had exactly the same models, we’re gonna have smarter stuff 12 months from now than we do today.

Frank: You mentioned something there, and I actually didn’t look into… So Claude scored 30% on this ARC-AGI3 test. In the NVIDIA harness, AVO, it got 100%. But I didn’t actually look at the cost comparisons because you made a very good point there about it’s basically, if that’s how it works and it’s looping through, it’s gonna be burning through a huge amount of tokens.

Justin: So…

Frank: I didn’t look, I actually didn’t look at what it costs Claude to get its 30% and what it costs AVO to get its 100%.

Justin: It wasn’t…

Frank: Us…

Justin: It was… It does lead us to the next question, but I can just… It wasn’t dramatically different.

Frank: Right. Interesting.

Justin: But it was dramatically different than GLM 5.3 Flash and Cloudflare.

Frank: Yes, indeed. Okay. GLM, what was it? GLM thr… Ah, gosh.

Justin: Flash. You gotta keep up with the lingo.

Frank: Yeah, I know, yeah. GLM 5.3 Flash. Yes, thank you. Okay.

Justin: To come up with catchier names for these things.

Frank: I know. Yeah. I mean, you know. Yeah.

Was Ox Alpha really China’s GLM 5.3 Flash?

Frank: Okay, so GLM 5.3 Flash. What is GLM 5.3 Flash? I’m saying it as often as I can now to embed it in my brain.

Justin: All right, so there was a secret model put up on Hugging Face, which seems to be all over the news this week. It was OX Alpha was the name of this secret model, and they gave away… Okay, so this is super important. They gave away trillions of tokens for free. They’re like, “For the next month, this model is gonna be free. We’re giving away trillions of tokens. Go use our model.”

And nobody knew who this model was from. It turns out that this model was from GLM, Chinese company, and it was running on Chinese hardware. And they have this thing called the Pareto frontier when it comes to the cost/performance of models, right? So you want your model to be the lowest cost for the intelligence that you have.

This model blows that out of the water. It’s about as intelligent as Sol 5.6, and it’s about one-tenth of the price.

Frank: Yeah. Intelligently, it’s coming in under them, but not that much under them. I think I… Yeah, and I think that’s where the Pareto thing comes in, isn’t it? It’s like, I can never remember this. It’s like 80% of the benefits for 20% of the cost.

Justin: Yeah, yes, is another way of putting it.

Did US chip bans make China better at AI?

Justin: But there’s a bigger issue here. The reason why I think this is important is this: when I was reading about this during the week, it was giving me flashbacks to the late 1980s, as always happens with us, Frank, ’cause we’re middle-aged men. So it was giving me flashbacks, right, to the late 1980s.

So the American government, and Anthropic and other companies, are very hot at the idea of not selling the latest hardware to China because we want to slow down their advance. And actually, what’s happened is that’s forced the Chinese companies to develop these models using Chinese hardware, which is less powerful.

And there’s another Pareto frontier which we don’t have measured yet, but it’s the density of intelligence per unit of compute. So we’ve basically got the muscle cars of compute in the West, right? We have these incredibly fast and powerful NVIDIA chips, and the Chinese have, whatever, Vespas instead.

And they’ve figured out how to get the same intelligence out of a Vespa that we get out of a muscle car. And why does that remind me of the late 1980s, right? Well, if you look at what happened between the electronics industry between the US and Japan in the late 1980s, the Japanese were obsessed with two things.

They were obsessed with quality, and they were obsessed with miniaturisation. And the Americans… The Japanese.

Frank: Oh, the Japanese were? Okay, right.

Justin: The Japanese were obsessed with quality, so the Toyota manufacturing system, and they were obsessed with miniaturisation, so making things smaller and smaller, right? Packing more and more stuff into a smaller footprint.

And that led to them dominating the electronics industry and becoming the richest country on the planet at that time. And that to me has echoes of what’s going on at the moment to do with the Chinese obsessing with the quality and the miniaturisation of models onto hardware which shouldn’t be able to run it, but they’re making it do it.

Their engineers are better… Their Chinese engineers are better than our Chinese engineers, as I…

Frank: Yeah.

Justin: …before. So that’s why I think that this is again the law of unintended consequences, right? The idea of not selling the hardware to China was for all the best of reasons, but actually they’ve created a monster which they’re now gonna have to deal with.

What was your reaction? Have you tried it yet, by the way?

Frank: I have, but I haven’t really put it through any tests.

Are cheap Chinese AI models safe to use?

Frank: But what I was curious about was, so, what I was curious about was, there’s two issues kind of out there broadly right now in terms of Chinese models. One is if you are… So I think the company that make it are Zed.ai, and if you want to use the model on their servers like you use ChatGPT on OpenAI servers, then you’re sending your prompts and your data to a Chinese company.

That’s one concern. People are concerned about that. They don’t feel that the Chinese have as good privacy safeguards as we do in the West. You know what? I haven’t actually looked into the details of that, so I don’t know. But even people who I follow, like Chris Penn, who is very pro locally run Chinese models, even he says that that’s a privacy concern.

And then the second issue that people have is that maybe there’s something nefarious baked into the models themselves, and that even if you run them locally, that they might do something dangerous. Now again, I follow Chris Penn, and he makes the point that if you’ve got a model with open weights, and if you run it locally, you can inspect what it’s doing.

You can see everything that it’s doing. You can make sure it’s not doing anything nefarious. So my kind of feeling is there are enough high-level technical people using these open source Chinese models that we would discover if they were doing something nefarious, because it would happen, somebody would spot it, and it would become public.

So probably, yes. Okay. Yeah, I’ll accept that. I mean, nothing is certain in life. But I do think…

Justin: Let me… So, let me… I agree. Okay, so let me put… It has been demonstrated that you can seed an LLM to give a certain output based on arbitrary input. For instance, the date, right? So on January the 1st, whatever, next year, I want you to write this piece of software, and it will do it, right?

You can certainly do that. I also point out that there was some malware that was in… It was one of these supply chain attacks recently, and that was found because a Microsoft engineer noticed that it took a tenth of a second longer to SSH into a server than it should do. And he went, “Eh, that took a little bit longer than usual.”

We’re talking about a tenth of a second, right? Almost imperceptible. And he went, “Can we just dig into that a bit and have a look at it?” And it turned out there was malware, and that was how it was uncovered. There are very subtle things that it could do. Do I really think that they will do that? No, I don’t, right?

’Cause it would destroy anybody ever using another Chinese model again, right? The game plan here isn’t to infect your computer with malware. The game plan here is to release models that are so cheap that it bankrupts the American providers of the models. So there’s no world in which it’s in their interest to corrupt your computer with malware.

It’s not gonna happen. Now, from a company’s point of view, why would you take the risk, right? You just wouldn’t. But from a personal point of view, yeah, take the models, run them. They’re super cheap and they’re brilliant.

Frank: Yeah. I mean, yeah, so GPT 5.6 Sol is $4 for input and $20 for output per million tokens. Claude Opus 4.8 is $5 and $25, and GLM 5.3 Flash, the Chinese model we’re talking about, is 15 cents and 50 cents.

Justin: Now I’m not good at sums, Frank, but that sounds to me like one of them is 20 times cheaper than the other one.

Frank: It’s bigly cheaper.

Justin: Yes, it is.

Can you run GLM without sending data to China?

Frank: And so, yeah, so I was just curious, okay, well, what if I wanted to use this model, and what if I did not want to use it via z.ai servers? So I just asked ChatGPT, how would I use it? And basically, because I use Codex, which sounded… It sounded like it’s not… In the future, I might actually be able to use it within Codex ’cause Codex does allow you to use other models.

But right now, there were some technical barriers that there was no way I could even look at. I don’t even know if they’re overcomeable. So I downloaded a thing called OpenCode, and I had ChatGPT just walk me through setting it up so that I could use this model through Cloudflare, so it’s not going near a Chinese company.

It’s all run through Cloudflare. ChatGPT, I’m using it as well with… I was able to get it to talk me through setting it up, so I’m using prepaid tokens. So there’s kind of… It’s very cheap anyway. Plus, I’m using prepaid tokens. Plus, I got ChatGPT to talk me through creating my own gateway, which made sure that I couldn’t spend more than $5 a day or $20 a month.

Justin: You’ll never use that anyway at those prices.

Frank: Yeah, exactly. Yeah, yeah, yeah.

Justin: You won’t, right? You just won’t be able to burn the tokens for a regular use case.

Frank: Yeah. So I’m… So yeah, I’m really curious because I do run into my limits on ChatGPT Codex, so I will be very curious to see if I swap over to this, can it do the same work? I suspect it can. I don’t do any… I don’t do rocket science.

Justin: Yeah, yeah.

Frank: So I suspect it’ll be very good, very cheap, and I won’t be running into limits.

Justin: Now, and the good news for you and me, right, just to sort of drive this home, is I think people have this mantra which is, “This is as bad as it will ever be,” right? In that the AI and the LLMs are just gonna get better and better. But that’s true, right? This is also the most expensive it’s ever gonna be.

It’s gonna get cheaper and cheaper for that reason. So actually, I read some tweets. Apparently, OpenAI have found some breakthrough just in the last 24 hours that they’ve released that allows them to serve these models 50% cheaper again. They’re gonna put an awful lot of resources into serving these models for the same price as the Chinese companies, ’cause if they don’t, the Chinese companies are gonna bankrupt them.

Frank: Yeah. Yeah. And we did say earlier that this tied into the NVIDIA harness story, and I think the reason it ties in is because, yeah, you know, the model is not quite as good as 5.6 Sol or the latest Claude model. But if you are very smart and if you set up your harness and you set up your skills, you set up your workflows in the right way, you can push these models further than it looks like the benchmarks show initially.

Justin: Yep, totally correct.

AI abundance or AI layoffs, Zuckerberg?

Justin: Mark Zuckerberg is in a little bit of hot water, right? But he…

Frank: Mark Zuckerberg is… I was like, “I’m not sure we can use that kind of language on this show, Justin, where you’re going with this.”

Justin: Listen to me. So there was a story in Reuters, right? Actually, I don’t… I would prefer to get into your story ’cause I think it’s more interesting. But there was a story in Reuters that basically Meta were one of the very first companies to lay off thousands of people to have them replaced by AI.

They laid off 10% of their company. The plan was that they were wanting to lay off as many as 60% of their company, but when everybody found out about this plan, it turns out their employees were pretty upset about it, and everybody started hazing on the AI, and so they canned the issue. Anyway…

Frank: And as you said as well, so they’re like, yeah, they laid off like 8,000 people.

Justin: Yeah, 10% of the workforce. Yep.

Frank: And their plan was, yeah, some teams could go down to 60%. But as you mentioned to me the other day as well, as well as the employees kind of revolting over this when this secret plan kind of leaked, it also transpired that they weren’t having as much success with the agents as they hoped they would.

So I read some pretty interesting stats on that after you mentioned it to me, and they said that… Where is it now? Let me find the actual stats. The quantity of code changed, so it was up 220% year on year, but the newer improved features that were actually reaching users rose only 36%. So massive amount of code being generated, but actually the end user really only getting a fraction of that benefit.

So that was interesting.

Justin: And the… Oh, sorry, did you see the other numbers that were related to that? So the number of incidents related to all that extra code also increased by about 40%, and the amount of time that was spent firefighting those issues increased by 70% ’cause people didn’t understand the code that was now causing the issues.

These are all kind of predictable outcomes.

Frank: Yeah. Yeah. I just find the whole thing ironic as well because what really bugs me is Mark Zuckerberg just wrote this essay about this amazing abundant AI future where everyone will have their own superintelligence. Everything’s gonna even out if we give everyone superintelligence.

Everyone’s gonna have loads of money. Everyone’s gonna be living a life of…

Justin: Abundance.

Frank: Exactly. And that’s this future that these tech bros keep promising us. But in the meantime, it’s like, but in the present, we’re actually gonna just take your job away from you.

Does Meta trust AI more than its own staff?

Justin: You were telling me this wonderful story during the week about how you had got on a number of times to Facebook/Meta technical support.

Frank: Yeah. I had to…

Justin: They were able to help you until…

Frank: No. Yeah.

Justin: Should be a happy story, but it’s not.

Frank: I had a bunch of accounts that had issues on them, Meta accounts. And over several years, I would periodically try and get these issues sorted out. I would get onto support, because the thing is, it was always a question of, did you get… Were you able to get onto a person who was just engaged enough to help you?

And so I would get onto support, I’d get onto a human. Sometimes I’d get someone who was like, “No, not gonna help you. Sorry,” essentially. But sometimes I’d get onto someone who’d be like, “Oh, that’s terrible. That should not be the case. I am gonna sort this out for you.”

Then they would disappear for a while, and then they would come back and they would say, “I’m so sorry. I actually don’t have the power to sort this out for you, and I don’t have the power to escalate this to someone who can sort it out for you.” So they were completely… The support people I was getting onto that were for the support for the thing that I had the problem with were ring-fenced from being able to do anything about it. It was insane.

Justin: Yes.

Frank: Then a little while back, I read that they had this new AI support bot, and I was like, “Oh, God, this is terrible. This is gonna be worse. This is gonna be even worse.” But I thought, “I’ll give it a go.” And I got onto it, and it was like, “Oh yeah, that’s terrible. That shouldn’t be the case. Yep. There you go. Fixed it for you.”

And I said, “Wow. Brilliant. Thank you.” And it said, “Oh, I’ve just had a look at your account there. There’s like four other accounts that have this issue. Do you want me to fix them as well?” It’s like, “Yes. Yes, I do.” So none of the kind of human empathy of like, “Oh, that’s terrible. I’m gonna solve it. I…” You know, none of that, but an actual ability to go and fix it for me.

Justin: Hold on now, right? But there was more, right? Because there was one account that you couldn’t get resolved, and you wanted to talk to somebody about that.

Frank: Yeah, there was one, and the AI said to me, “Look, this is basically…” I can’t remember how it put it, but it was like a statute of limitations issue. The account had been wrongly flagged years ago by another less capable AI system at some point, and it was completely incorrect, but no one was ever able to lift the flag for me.

And I thought, “Well, this AI is able to do everything. It’s gonna be able to do it for me.” And it said, “No, I’m sorry, I can’t because too much time has elapsed since the flag was placed on the account.” And I was like, “Well, okay,” but you know what? This is like… The AI was like, “Yeah, it’s pretty clear to me that this was an error. I can see the account is clean. I can see the account is well-intentioned,” et cetera, et cetera, et cetera.

So I said, “Okay, put me onto a human who can solve this.” It said, “Great idea. I’ll do that. Hold on there one second.” Time passed. Time passed. Time passed. “Sorry, Frank, no humans available.” And I was thinking to myself, “Wait a minute.”

They’ve got this AI system that’s fixing loads of stuff that people couldn’t before. It’s doing all this work. Surely that has alleviated the strain on the humans. Why is there now no capable human available to actually fix something that is clearly an issue?

Justin: So I think this is fascinating because the systems that a company puts in place reflect the culture of that company. So the two things, when you told me this story, that jump out at me, right, is the first, you tried to get a problem fixed by a human and they couldn’t do it for you. Then you tried to get the same problem fixed by an AI, both working for the same company, and the AI had absolutely no problem in resolving that for you.

What does that tell me about that company? It tells me that they trust a computer more than they trust their own employees. That’s the first thing, right? That’s super interesting. I’m not at all surprised that the Meta employees revolted when they were getting laid off if this is what they think of their employees.

The second thing that it tells me is, in the future, right, the really premium companies are going to be the ones that value human interactions. That when you say, “I want to talk to a human,” right, especially in this type of a setting, the people that talk to you aren’t gonna be measured on how quickly they solve your problem or how quickly they get you off the phone.

They’re gonna be measured on outcomes and how happy you are with the outcomes. That’s what’s gonna differentiate a premium company from a non-premium company. And in this instance, as you correctly say, there was no human available for you to talk to at all, which means that that company, the way that I would see that is, they trust a computer more than they trust humans that work for them, and they respect you even less than the people that work in the company because they’re not even gonna put you…

So that’s incredible, right? And this is a company, Facebook, what’s their motto? To, you know, help people connect with each other. It’s like this is a company that’s literally built on people connecting with people, and yet they trust computers more than they trust people, and the people that are trying to give them money, they just don’t respect them at all.

Incredible story.

Frank: Shock, horror. We’re all really surprised to learn this.

Justin: Well, I mean, part of the technologist in me is going, you know, companies like Meta are the people that will push the state of the art forward, right? No regulated industry is gonna do anything close to what you’ve just described there, right? There’ll be guardrails and all sorts of things.

Having said that, they’ll fall over and they’ll make mistakes like they clearly are, and then we’ll find out what the mistakes are, and we’ll fix them. It’s Whac-A-Mole, Frank. Whac-A-Mole is not a bad thing. Whac-A-Mole is a good thing. Just, you know, you don’t wanna be the one whacking the moles if at all possible.

Is the Cat in the Hat terrorising Ireland?

Frank: Justin.

Justin: Yes.

Frank: You have… You’re gonna have to be very careful because I have heard reports that the Cat in the Hat is on the prowl and is terrorising neighbourhoods and peeking in people’s windows and generally acting like one of those creepy clown serial…

Justin: Halloween already?

Frank: I know, that’s what it feels like, yeah.

So apparently on social media, all these videos started surfacing that are very kind of like Blair Witch found footage, shaky camera. Some of them are actually… They’re pretty convincing. I was pretty… I was impressed. And they show, like, the Cat in the Hat down the street staring at this house, and then the camera looks away from it and it comes back and the Cat in the Hat is that bit closer.

Justin: Oh, they’re really scary.

Frank: Yeah, other videos of, you know, like people’s Ring cameras, or whatever they’re called, like the doorbell cameras, showing the Cat in the Hat going around the house looking in the windows when there was nobody in, that kind of thing.

Justin: Is there a real Cat in the Hat on the prowl?

Frank: Well, apparently… So this… I think this is a global phenomenon, but I was interested to see that it hit the UK and Ireland quite heavily.

So Wexford Gardaí had to issue a statement saying, “To be clear, the Cat in the Hat was not in your driveway while you were out,” and, “The Cat in the Hat is not currently hiding out in Curracloe.” These were all, of course, AI-generated videos, and it’s this meme that is going around, and it is… But it is genuinely terrifying people because they’re really good.

I think I might’ve been fooled if I had seen it.

Justin: There you go. You can believe nothing, not even your Ring doorbell anymore. Brilliant. Well, look, I’ll keep an eye out for that this week, as well as all the latest AI news. Frank, pleasure as always. Have a great week.

Frank: Chat to you next week, Justin.

Justin: See ya.