Meta Muse vs Google Argon, OpenAI’s Spy Scare, and AI Job Interviews

Frank: Two huge models, one from Google, so Gemini 4 Argon. And also Meta Muse Wonder.

Justin: If you want to make sure that we get a fast takeoff and recursive self-improvement, the best way of guaranteeing that is making sure that there's not one, not two, not three, but four frontier companies all working on it at the same time.

Will Justin's eye-scanning Orb ever arrive?

Frank: Frank Prendergast here, your approachable doomer, checking in alongside techno over-optimist Justin Collery to argue over all of the latest AI news. And there's been, yeah, there's been a lot of big news, since we chatted last week, Justin. Have a singularity update for us this week?

Justin: Mm. Bad news. Bad.

Frank: Oh no. Wha-- have the robots risen?

Justin: No, no, no. I got so excited this week, right, for two reasons. So I got a... You know, right, we've talked a b- a number of times that I have signed up for an orb. The orb is the thing that's gonna prove that you're human. It's a little round thing. It's gonna scan the front of your eye, and about two years ago, I signed up to get myself an orb.

I paid my money, and.

Frank: I cannot every time you say this, I just cannot believe that you willingly signed up for one of these orbs.

Justin: Not only did I sign up for the Orbs, I signed up for the community programme. So I was going to get an Orb and other people were going to come to my house to get, their irises scanned. It was gonna be like a little re- religious ceremony where people would come and they would get their irises scanned.

Anyway, so I got an email from the Orb people this week. I got very excited because I was thinking maybe finally my Orb is about to be delivered. Bad news, they've cancelled the Orb community programme.

Frank: No way. Wait, the community programme. So just, just the bit where you get people to come to your house. The, so your religious ceremonies are out the window.

Justin: It's-- I know, it's sad, isn't it? So it turns out, I went onto the website before we started the show to see could I pay whatever, another amount to get it, and it looks like the Orb, like there's nowhere to buy an Orb anymore. So it looks like maybe the Orb is being cancelled, which is super sad. So that's my Singularity update.

The Orb is being cancelled.

Is Claude taking Justin's refund fight to court?

Justin: I do have another update, which I have to say a little bit quiet about. So as you know, at the start of the year, I travelled down to Mauritius and had a wonderful holiday down there, and I had some difficulty getting back because somebody decided to bomb somebody else, and I was supposed to be flying through there.

And so I got Claude to help me out, and it-- Claude got me home, but I was out of pocket, and I wanted Claude to get me a refund. It's, I can't speak too much now, Frank. I can't say too much 'cause it's about to go legal and therefore, you know... What do they say? There's a phrase they use, you know, as it's a matter that's, you know, going in front of the courts, I can't comment any further on it.

But I will give you an update,

Frank: And.

Justin: Movement. The wheels of justice are in motion there, Frank.

Frank: When you say it's about to go legal, wha- Claude has been looking after this for you.

Justin: Yes.

Frank: So is this an action that Claude has recommended and is taking on your behalf?

Justin: You will be so proud, right? You will be so proud because Claude, we've done-- me and Claude have been working on this together, and Claude sent me an email during the week saying, "Look, we've now reached this deadline. Do you want me to go to the next step?" Which we cannot, that which we cannot speak its name of.

And so when we're done here, I'll be authorising Claude to go to the next step, after thoroughly reviewing all of the documents, obviously.

Frank: Of course, yes. Okay, so human, human still firmly in the loop.

Justin: Yes, maybe not in control, but in the loop.

Frank: Excellent.

Justin: In control is about as much as anything a man my age can be in control of.

Can Meta Muse and Gemini shake up the AI race?

Frank: So massive news this week in a way. Two huge models, big news models, one from Google, so Gemini 4 Argon. Have I got the name right? And also Meta Muse Wonder and Meta Muse Wonder Full.

Justin: Yeah, which is good actually. I like the naming play on it. MetaMuse, the interesting news, right, you can-- And you can tell me, right, 'cause I've been traveling this week, so I haven't had as much chance to play with new toys as I normally would. MetaMuse, right, that one, the benchmarks are astounding.

The benchmarks for MetaMuse Full, are, like, as good as Opus 5.5. In some instances, they are better. And this is an, an open source model, I believe. You can download the weights, you can run it yourself, which is incredible. But the pricing when you use it from Meta... Am I right.

Frank: So no, I don't think it's open. No, no, it's closed. Yeah, it's closed.

Justin: Sorry, you're right. But the cost, the cost is the same as Opus 5.5, whereas before they had been releasing models and they were, like, just cheap as chips. I mean, essentially free. This one isn't. This one is very definitely on the frontier and very definitely frontier pricing.

Have you had a chance to play with them?

Frank: A little bit. Yeah, a little bit. And, I'll actually-- I'll show you something in a minute. But yeah, before we get to that, I mean, I think-- the reason I think this is such big news is because we had, for a while there, you know, we all know OpenAI launched ChatGPT, started the whole thing.

Anthropic caved off from OpenAI, came out with their own models. Google, who should've been ahead the whole time, were behind, then they came out with Gemini. And then for a while, we had Google, OpenAI, and Anthropic. It was a three-horse race. One would pull ahead, then someone else would pull ahead, and then Google fell over one of the hurdles.

And so for the longest time now, it's been OpenAI and Anthropic. That's all anybody's really talked about in terms of the absolute frontier. Meanwhile, you had Meta saying, "Well, we're going to be open source. We're gonna be the best in open source." And they never really quite even pulled that off, really.

They were the best in America. And now here we are, and we have four companies now at the frontier. Again, well, we, yeah, we now h- we now have a grand total of four right at the frontier 'cause they're all in, they're all in and around the same, like the frontier models from each company.

Like there, there's only points between them on the benchmarks. And I.

Justin: How's the slowdown, how's that slowdown going, Frank?

Frank: I know, yeah. Yes. Yeah. Weeks, weeks after they all said that we should slow down, here we are. The Meta one I think is really interesting.

Justin: If you want to make sure that we get a fast takeoff and recursive self-improvement and where the models make themselves better, the best way of guaranteeing that is making sure that there's not one, not two, not three, but four frontier companies all working on it at the same time.

Surely one of them will get there.

Frank: Yeah. Yeah. I-- what I, what I think is interesting about Meta is that, you know, Yann LeCun was the head of AI at Meta, and I think we both have a lot of time for Yann LeCun. And then Mark Zuckerberg brought in Alexander Wang, and Yann LeCun was not happy. Wang was not as interested in LeCun's, dedication to open source and, started to take things in a different direction, and LeCun left Meta and, set up his own thing in France, which is great that we have LeCun now in Europe.

And I would've thought long-term vision-wise, that it was a mistake to lose LeCun like that. But who would've thought it? Surprise, surprise. It does seem like Zuckerberg, from a commercial perspective, knew what he was doing. And here they are.

Justin: I mean, who, who would've thought such a thing?

Would you switch your AI coding tools?

Justin: I'll be very interested, 'cause they've released Muse, right? And they have a coding harness that goes with Muse as well, and coding's where most of the money gets made these days. And what I think i- well, I think coding harnesses are a bit sticky, because you kinda have to, you kinda have to become one with the coding harness and with the model and whatever.

You know, you get to use something and you get used to it. I'm interested to see how many people will switch or try or whatever between the different harnesses.

Frank: You say that because I couldn't find a harness for Muse.

Justin: They have one. They have it. It's on their website.

Frank: That's weird. I had to, I had to use it in the terminal. And you know me, I'm like, "What? I have to go into the Matrix to use this?"

Justin: But was there a- 'cause it is, it's probably a terminal application. I can't remember what it's called, but it's probably, it probably, I mean, it doesn't surprise me that you would have to go into a terminal. But it's just, that is the how the coding harness works. You know what I mean? You fire up it and you talk to it and it writes code.

That's-

Frank: Right. Okay. Yeah, but it doesn't... So they don't have a graphical user interface, as in, like, they don't have an app, that I could find anyway.

Justin: No, you're right. Okay, so they don't have like a, they don't have a front end for normies, like people who don't like terminals, whatever. Yeah, it's a code. But anyway, my point is, right, there are, there are tricks to getting each of the harnesses to work well, right? So, and there are-- And now as skills come out and skills are kind of tied to the harnesses, so let's say in Anthropic, you know, you've got, hooks, which you can, which is deterministic codes that you can do.

You've got like, you know, build evals. You've got like, what else is it? Hill climb, which makes it better, right? There's just things that get packaged up inside the harness. OpenAI have their own set of skills and special ways of firing off sub-agents and doing all sorts of cool stuff. And so you get to, you know, you just become comfortable with the harness that you're using.

I think there's more stickiness than people may think in the harness. And so I'm just interested to see how many people switch to use this new... But the benchmarks are amazing. They're brilliant.

Is Meta Muse eating through your tokens?

Frank: I did hear people saying that, they were find-- despite the fact that the published costs are, on, you know, on a, on a par roughly with all the other models, I did hear people saying that Muse seemed to, like, eat tokens and cost a fortune to actually use in a practical sense.

Justin: That is a very interesting point. I hadn't read that. So there's two things I would say that will drive adoption, yes, between different harnesses and models, right? So the first is you can tell me how expensive per input and output token your harness and your model is, but if it spends, like, a boatload of time thinking and using loads of tokens and thinking, well, that's not so good, right?

So the cost per task is more important than the task cost per token. That's the first thing, right? So that's quite an interesting observation. And like the Opus 5.5 was 20% cheaper just because it was smarter, thought less, so it.

Frank: Right.

Justin: To the answer quicker, right, without using tokens. The other thing I think, which isn't as highlighted as much, is speed.

So if, if you're there doing sort of waiting to get a response from a model, it's actually quite a sort of soul-destroying task when you're just sitting there waiting as it thinks and thinks and thinks. Whereas if this thing can work, I don't, I don't know how quick it is, but even if it's doing a lot of thinking, if it's doing it really fast and it just spits the an- answer out to you, that's quite an addictive property of a model doing anything, coding in particular, but doing anything.

Same prompt, will Muse or Argon win?

Frank: One, one of the things I found-- So I'll, I'll actually show you a little experiment just, I g- I, got access to both, Meta Muse Wonder and, Google's Argon, and I gave them a relatively simple little task just to compare them. And w- I'll show them to you in a minute. I had to use Meta, Muse in the terminal, which was which was fun for me.

And just in relation to what we're talking about right now, one of the weird things I found is that I don't actually, I don't actually fully know what it was doing, but it got caught in some kind of, it got caught in some kind of loop. So it did my task. I asked it to create a webpage for the AI argument, because this is something that I've been working on recently and I wanted to compare.

And it spun up a local server to show me the page, even though it appeared to be just static HTML, and then it got caught in some weird loop. So even though the page was finished, the task was done, it kept consuming tokens, doing something in the background, and coming back to ask me, "Should I do this on the local server?

Yes or no? Should I do this on the local server? Yes or no?" And my token consumption just kept going up, whatever it was doing, even though the task was done.

Justin: Interesting. Interesting. There should be a way. They, they do in Claude Code where you can sort of share your chat history and to report a bug. So maybe there's a way to do that, 'cause that would be very useful for them. It's a good spot.

Frank: So let's have a quick look, right? So the first thing I'm gonna show you is our actual website, which was vibe coded with ChatGPT 6.1. And the reason I'm gonna just-- I'm just gonna show that as kind of our, like, baseline. This was something now that I spent a lot of time on. So I would say, you know, this wasn't designed by ChatGPT 6.1.

This was designed by me vibe coding with 6.1. And then we'll have a look at, I took the content for that page, some of the assets for that page, gave it to the new models, and just one-shotted with a pretty vague prompt. Here we go. Can you see this?

Justin: I can, yeah. Beautiful.

Frank: So this is our current site. This is the one that I worked hard on with, with ChatGPT's, 6.1, and.

Justin: And if I was to describe this, in case there's people listening, right, I would describe this, it's got a dark background. It's got the podcast colours, which are like yellows and greens, and it looks reasonably fun, but fairly professional, I would say.

Does Argon follow instructions better than Muse?

Frank: So then I gave, Meta Muse Wonder the following prompt. So this is a pretty poor prompt, but I just wanted to see with a broad prompt, what would it do? So I said exa-- I basically gave it access to a folder on my sh- my machine, and I said, "Examine the assets folder. You'll find page content, markdown file.

Design a webpage around that content, make it modern and aesthetically excellent. Wow me with modern functionality and make it alive and exciting to browse." So let's see what it did. So this is what it came up with. So I'm not crazy about it, but like, it's, it's pretty good. It's pretty good.

Justin: I.

Frank: And I,

Was very vague in my prompt.

Justin: Now I would say this, so this is more fun, I would say, if I was to describe it. I'm not too keen on the Sky News breaking news banner type thing that's going that looks a bit, you know,

Frank: Yeah, yeah, yeah. Yeah, autonomous weapons, energy use, robots all kind of going across the screen. I like this subtle little 3D effect here on our image. If you-- when you roll your cursor across it just has this little subtle 3D effect, which is just kinda cute. It also... So here we have, a playable video.

Justin: Nice.

Frank: And then I like this. This-- I don't know where it got this 'cause it's not on the current website, and that is one thing about, the Meta model as well. It didn't, it didn't just take the content of the page and design a page s-specifically around that content. It did hallucinate its own content.

It changed the content, et cetera. I wasn't that crazy about that. I kinda wanted it to use the markdown file.

But it came up with this little slider: "Before you listen, where do you stand?" And the, person on the website can, can slide it over to you, and it hallu- it hallucinates these quotes from us over here th- as you slide it around, which, you know, I didn't give it materials, so you could absolutely give it materials and have it a- have it come up with actual quotes from us.

But seeing as it didn't have them, it just hallucinated them. But that was a cute little,

Justin: Can I just, can I just say, right, for the pr- for practical use cases, what you said there was actually a really important line, which was that it didn't stick to the markdown file and it hallucinated what it thought was good. Now, that's fine, and I'm joking saying that's great, right? Actually, in realtor, in real world usage, in real world applications, that is absolutely not what you want your model and your coding harness to do.

In fact, it's, it needs to stick very strictly to the instructions that you've given it.

Frank: No, in fairness, my actual instructions were a little bit vague. So, yeah, so I wouldn't be too harsh on it, but I was kind of surprised.

Justin: Yep. It's beautiful though, isn't it? The other thing I would say, right, as people use the models, like if I look at a website, I can tell within five seconds if Claude has made it. There's like tells in a way. This one definitely looks different, and I think that's important for people who are making websites.

Frank: So I'm not sure-- This is another little interactive thing here. I'm not really sure what it does. Oh, I see. You click on it, and again, it hallucinates quotes from us on these topics. So again, if we, if, if we gave it actual quotes, this is a cool,

Justin: Yeah, yeah.

Frank: This is a cool little feature. Oops, sorry. I just did something weird there.

Let's see. There we go. So it's-- What's it saying here? It claims you're saying, "On jobs and work, Justin wants to ship it, learn fast, and fix it in the open." I mean, that's actually true. It's just you never, you never actually said it, but yeah. One builds a bunker, one opens the box. Yeah, that's kind of...

What does this say? Another Frank take. Oh, I see. Again, it's hallucinating. It's-- this is great. It's obsessed with hallucinating quotes from us, but, it's a nice idea. Maybe I should build that into the, existing site with real quotes.

Justin: Yeah, yeah, it is good.

Frank: And then here it has a little, a little promo for the AI argument game.

And there we go. So yeah, it's, you know, it's, it's, it's a good site. It's grand. It's not, um... It's, it's... It doesn't appeal to my aesthetics, but, like, if you just needed to get a site done quickly, I mean, there's nothing wrong with it. Nice little countdown timer here to the next, argument as well, which if you gave it the right time, would probably work.

So.

Justin: I was to describe, if I was to describe this, right, the front page, for people who are listening and not watching, i- if you could imagine, Grange Hill from the late '80s and they'd made the front page of a comic, that's kind of the aesthetic, I would say. It's kind of a fun, sort of slightly retro look.

Frank: Yeah. Yeah.

Does Gemini Argon build a better website?

Frank: So let's have a look then at what Gemini Argon did. So straight away, what's interesting to me here, and this is why I showed the original design, because what's interesting to me here is that it has gone, with a design that is, is oddly much kind of closer to what I designed with 6.1, with a lot of my input.

So here, I've just given it this broad prompt. I haven't gone through with it what I wanted, and it's just magically gotten much closer to what I, what I designed.

Justin: Almost like Google knows you better.

Frank: Oh my, yeah, yeah. Now.

Justin: Maybe Google searched the internet before it made your website to find other websites that were similar to this one?

Frank: I wondered if that was possible. Absolutely. And I didn't dig into it to try and see if I could tell what, what had happened. Interestingly, again, you-- look, you've got the same little 3D effect on the image of us here as you, as you roll over it, but it's, again, it's a little bit m-more, it's, it's a little bit nicer than the Wonder one in its implementation.

You've also got this. I love this. Look at this. There's a gradient on the background of the top, the hero section of this website, and on the left, you've got this kind of blue, and on the right, you've got this kind of red. And up here, you can click on Frank to make it more blue and click on Justin over here to make it more red.

Justin: Right. I mean, interesting, useless piece of functionality.

Frank: Yep. So as we scroll down, here's what's weird as well. Look, it's done the same slider thing here.

Justin: And it's also got something... Look at it. It's also, it, has it hallucinated those pieces of text underneath? The AI... Isn't that interesting? If they've.

Frank: You know what? It a- it has actually taken those quotes... Those are real quotes from us from the w- from the markdown file, actually. So they're actually real quotes.

Justin: All right. Okay. Nice.

Frank: But it's interesting that they both decided to do that slider. The video isn't working here, but that's just 'cause I'm... It's, um... I'm just running it locally.

Justin: Jep.

Frank: I don't know what this cinema mode button does. Let's press it and see what happens.

Justin: It opens a full screen, I'm.

Frank: Interesting.

Justin: Nice.

Frank: All this loads of text here that you can click around. So lots of little interactive features. It didn't have the thumbnails for the show, so it generated its own SVG, thumbnails to stand in, which is pretty cool.

Justin: Nice. Ja.

Frank: It's embedded the podcast. It's got this interactive element here again, where.

Justin: Get very similar.

Frank: Very, very similar. You can click on the topic and it, in this case has actually hallucinated quotes from us once again. Um- Meet the hosts. Nice little-- And now I wanted to show you this. This is crazy.

So it's also, you know, the markdown file had a promo section for the game, so again, it has a promo section for the game. That's fine. Got a link here that would go to the game.

Justin: Yeah.

Frank: A dead link at the moment 'cause we-- this is-- I only asked it to do this page. But here's what I wanna show you. I was like: "What's this?

Try instant mini sim here." So I clicked it. It's made a little in-browser game where look, I'm-- these things, these things are the p- AI problems popping up, and you've gotta click on them to get rid of them, just like you, like to say is the best way to handle AI problems. And then you-- and you click down here for, for Frank's regulation to increase the regulation.

Justin: That is fantastic.

Frank: I don't think there's any way to win the game or anything, but how bizarre, how mad is that?

Justin: That is cool.

Frank: Yeah, so there you have it. I was more impressed with the Gemini Argon page. But that said, I think they both, both did a pretty good job.

Justin: That's pretty interesting, Frank. It's pretty good. How long did it take, and do you know how much it cost you for each one?

Frank: The-- so the Gemini Argon is on subscription, so it didn't, it didn't cost anything outside of the subscription. I didn't actually check on the usage, so I do need to check on the usage. I know that with the Muse one, it used, 18% of my weekly usage, and it used about 60% of my five-hour usage.

Justin: Usage.

Frank: And as I was saying to you, it just kept using, it just kept using up tokens as well. And I was like, "What is it doing?"

Justin: That's interest. But you're on the-- for the Muse one, are you doing API put billing or you, have you paid them 20 bucks a, 20 bucks a month sort of stuff or something?

Frank: Less than five bucks a month for the, for the Muse.

Justin: That's impressive. That's very impressive. And how long did it take then for y- so from the p- time you put in the prompt to the time you had a website.

Frank: You know what? I should've, I should've timed it, but I'd say about 10 minutes.

Justin: Go away. 10.

Frank: Yeah, both were very, very quick.

Justin: Impressive. Impressive. Listen to me. Okay, that's pretty good. Those, so those are two good models.

Who's spying on OpenAI?

Justin: We've had a lot of news this week, about more disturbing things, right? So this is-- I'm.

Frank: Oh, great. Oh, great.

Justin: You've, you've got all the fun stories this week, and I've got all the terrifying stories this week, right?

So I just wanna bring just a couple to your attention, 'cause I know you like to keep on top of, you know, the dangerous things that are going on in the world. So let me just bring a couple of interesting things that happened. This isn't like a singularity update. This is like a security posture. We should have a red klaxon and stuff like that.

So in OpenAI, right, one of the predictions in AI 2027 was that around this time, a rogue nation state would try to exfiltrate the model weights, so they'd try and steal the model weights so they could use it themselves. This very week in OpenAI, they found a sophisticated listening device in their offices and attributed it to one of their staffs, who's now been terminated as a member of staff for OpenAI.

But yes, that's definitely on theme. Not quite exactly as predicted in AI 2027, but just goes to show, you know, the propensity for espionage within these very large, frontier model companies. It's gonna happen more.

Frank: Do we know, do we know any details about, like, his affiliations or, you know, if he had any foreign government affiliations or anything like that?

Justin: So we don't know, we don't know, right? So I couldn't say for sure. I do know that he's of, as I remember, he's from, of Middle Eastern origin, so from somewhere around that part of the world. But anything more than that, we don't really know. But it's definitely happening, right?

Are supermarket fridges a cybersecurity risk?

Justin: Related to that story, also this week, just in case you weren't frightened enough by that, Frank, there was a major cyber attack on fridges across all the low countries, so France, Belgium, the Netherlands.

Frank: On fridges.

Justin: On fri- Now you would think, "On fridges? Huh, so what? Why the hell," right? "Does it matter?" So all of the Carrefour, all those big supermarket chains in that part of the world, they used a cyber attack to hack into the fridges, and they basically turned them into cookers. So they made the fridge heat itself up to 50 degrees, destroying all of the food in the fridges.

Frank: That's very serious.

Justin: Yes, it is.

Frank: Yeah, that's crazy. So somebody like messing with our food supply.

Justin: You make, you make it sound so dramatic, Frank. It was.

Frank: What would you call it?

Justin: Make it sound so serious. Look, I think from my reading of this, right, th- this was enabled due to weak security on the fridges themselves. They probably had default passwords and stuff like that, and they were able to hack into them with relative ease.

But because they used AI to do it was a cyber attack and they could coordinate it across a large number of fridges all at the same time. So it is.

Frank: Is fas- yeah, this is fascinating because, you know, we've talked on the show about like, oh, what if they hacked, a nuclear reactor or what if they hacked-- And we know that, you know, these models, like for example, Gemini Argon that I've just been playing with to make a webpage, we know that like we actually talked about it two weeks ago when they released it first, but it wasn't released to the public until this week.

It was released to people to look at cybersecurity and cyber defences, et cetera, et cetera. So we know they're giving it to the, you know, the big tech companies, et cetera, but like they're probably not giving it to supermarkets to check if their fridges are defensible against attacks.

Justin: Yeah, yeah, yeah. And we've talked about this before, right? Because it's, you know, you t- you always, you know, when you think about these things like, "Oh, they'll attack a nuclear power station," or, "Oh, they'll attack the web..." They don't have to, right? It's the mundane things that will get you into trouble. It's a fridge in a supermarket.

It's your router in your home. It's, you know, all the usual small stuff Yeah, it'll be fine. In the same breath, in the same breath, the MI6 and MI- MI5, the UK, security apparatus, they have briefed, the government in the UK to say, "GCHQ privately warned British cabinet that AI could become one of the country's biggest national security threats." Could? I would suggest that it already is one of the country's biggest security threats.

Frank: Was, was this before or after the fridge incident?

Justin: Fridge incident.

But look at the-- this is gonna be a bigger and a bigger thing. Countries, companies, everybody need to get on top of their game, and it's, it's, it's similar, I guess, in scope but way more serious than the Y2K issue. So in the Y2K issue, right, we-- In fact, it's a pretty good simile, right? It's like we all understood, you know, five or six years before Y2K that we needed to test every single system to see that it was gonna work, and mostly they did, right?

There was very few issues found. This is way more serious. We need to go through the same level of preparedness, go test every single system to see that it's hardened to a cyber attack by AI, and we kind of need to get on it straight away.

Frank: Which like it's... I- is it even possible? Like I ca- I don't know that I can even see it happening.

Justin: That, well, that's why I'm using Y2K as an example because it is possible. We did it before, right? And it's totally doable. It took a couple of years It'll be fine, Frank. It'll all be fine. It'll all be fine.

Frank: Yeah, yeah.

Justin: Yeah. No, no, no, it's fine. Look at, I was saying, if you're, look at, look at, it's, it's a weekend, it's Saturday, Frank, just relax.

You can sit down and have a glass of wine this evening, and you can mull over how it is we test every single device on the planet to make sure that it's safe from AI attacks. I'm sure you'll come up with the answer by Monday.

Should your AI do your job interview?

Frank: So listen, I came across, something which is b- both hilarious a- and also worrying in its own right as well. So job seekers are finding more and more that they, you know, turn up for their first virtual interview and there's nobody to interview them. It's an AI system. It's an AI avatar interviewing them.

And one guy has been subverting the process in some hilarious ways. So I don't know if you saw, there was about a week ago, he had a video where he just spoke gibberish to the interviewer, and it was absolutely hilarious. Oh, it was very funny 'cause the AI interviewer would just take this absolute gibberish and say, you know, "Oh, it's wonderful that you worked on the flibbertigibbet project," and, "and that the rhinoceros and the elephant were-- you managed to turn them around and upside down.

That's fantastic." Well, so his latest escapade,

Justin: This is a big thing because it takes a long time to hire. It's a fine, it's kind of interesting. I'm not sure that chatbots and recruiting scenarios are actually a good idea. But anyway, it's, you know what I mean? If people are your strongest commodity, maybe you should have people check people. That would be my.

Frank: I know, and I think it's they're, they're treating it like a kind of a screening thing, I think, that, you know, you might get to a person eventually, maybe, possibly, but that your first rounds are AI. So in this guy's latest escapade, he, brought his own AI assistant to the interview and told the AI interviewer, "Look, I brought my, I brought my..."

I can't remember what he called him, but basically he knows my resume, even better than I do, so I've brought him t- because he might be able to answer questions for you. And then what happened was the two AI systems were more interested in talking to each other than they were to the candidate.

Justin: Eh.

Frank: They would frequently, talk over each other and ignore the candidate, and finally the conversation came around to what his salary expectations were. So let's just listen to a short clip here.

Justin: Ahimè.

AI Pat: To wrap things up, what are Graham's salary expectations, Victor?

AI Victor: Graham is incredibly flexible on that front, Pat. He's primarily focused on finding the right fit and a great team, so he would be very comfortable around sixty thousand.

Graham: Oh. Oh, hold on. Pat, Pat, let me, let me chime in here. Pat, Pat-

AI Pat: That's actually well below our budget overall as our band is eighty-five to ninety-five thousand.

Graham: I'm sorry, Pat, um-

AI Pat: I think we can definitely make that work,

Graham: Pat. Pat, I'm not-

AI Pat: Let's move to next steps.

Graham: Hold on, hold on. Let's not move to next steps at sixty thousand. Pat, Pat, I'm not quite that flexible. I think Victor is just trying to express that I would be really excited to join your team, but I'm looking for something more in the range of like a hundred and forty to a hundred and fifty thousand dollars.

AI Pat: Graham, that is quite a jump.

AI Victor: It would just be at most.

AI Pat: That is quite a-

AI Victor: It would be-

AI Pat: Oh, that is quite a jump from the initial number.

Graham: Yeah. Yeah.

AI Pat: I have to stick with the resume and what Victor provides, so we will stay with-

Justin: Was pizza?

Frank: So he's a, it-- the AI that he brings onto the call lowballs the offer. The human says, "No, no, no, I'm looking for a lot more money than that." And the AI interviewer is like, "No, I have to stick with what the, what your AI avatar says." He then, he then left the room and he said, "I... Look, I'm gonna leave the room, but Victor, I want you to negotiate up to at least 120, and I'll be back in a minute."

And the two avatars were just like, "Yeah, look, I think it's fine. Let's just organise a follow-up meeting." And then they-- So they were like, "Right, that's it. We're done." And they went to end the call, except they got caught in this, like, infinite loop of saying goodbye to each other.

Justin: There we go. I can't see anything bad going wrong with this, Frank. You know, it's an engineering issue. They just need to, you know, make the bot more smart.

Frank: I should say, as far as I can tell, this guy actually sets these up himself, that they're not, they're not actual interviews. However, I did see a Wired story last month from a guy who was absolutely fed up of just getting AI interview after AI interview and never getting to a human, and never getting called back either.

And so he genuinely did just turn on ChatGPT voice mode with a couple of instructions from him and just let that run with the AI interview. He said he still didn't get a call back. He said he was hoping that maybe the AI would get on better with the AI and that he would get on to the next round. Unfortunately, he didn't.

But yeah, this is what's happening now. AI-- companies are sending AIs to interview, and humans are saying, "Well, then fine, here's my AI."

Justin: Frank, I think that your antennae for ethics are probably more finely tuned than my ethics antennae. If, if it's okay for a company to interview you with an AI avatar, is it okay for me to send in my AI avatar to be interviewed, do you think? Yeah, yeah What a world, what a time to be alive. Fantastic. Frank, pleasure as always.

Frank: Excellent stuff. See you next week.