#145 Bye-bye internet: a web built for machines, not people

"It's trying to serve humans, and it's trying to serve agents or bots, and you can't. Maybe there isn't a way to serve them both equally." - Dara

"The top of that stack looks reputable because it's got loads of references of all these other sources, but if you dig down deep enough, it's a load of crap." - Matthew

Show full (AI-generated) transcript

[00:00:00] Lizzie: Hello, and welcome to "The Measure Pod" by Measurelab, the podcast dedicated to the ever-changing world of data and analytics. With your hosts, Dara Fitzgerald and Matthew Hewson. Between them, they've spent more years than they'd like to admit wrestling with dashboards, data quality, and the occasional Google curveball.

[00:00:32] Lizzie: So join us as we share stories about how analytics really works today and where it might be headed tomorrow. Let's get into it.

[00:00:40] Dara: Hello, and welcome back to "The Measure Pod." I'm Dara. I'm joined by Matthew. Hello, Matthew. How are you? I'm

[00:00:47] Matthew: okay. Yeah. You sound very clear and not like you're talking through water or something this week.

[00:00:54] Dara: Yeah. I've got a bit of a c- well, it's not really a confession because anyone who listens will know, but the, uh, [00:01:00] sound quality has not been what it used to be for the last couple of episodes, which was to do with my, my mic being broken, but I've, I've sorted that out now, so hopefully I should sound a bit better.

[00:01:10] Dara: But you've lost yours.

[00:01:12] Matthew: I've regressed, yeah. I... Mine didn't work quite as well. My, mine kept on cutting out, so I took it off last week to try the headphones, and I was listening to the podcast and, yes, the sound quality definitely took a nose dive. S- and I c- I didn't give myself enough time to actually get it back up and running, so it's just in my hand loose here, uh, so I'm talking through.

[00:01:32] Dara: If you're watching this on YouTube, you'll get the pleasure of seeing the microphone, but unfortunately you won't be, you know, hearing, hearing the quality of it.

[00:01:39] Matthew: No. No. So I'm talking through AirPods, but I will endeavor to get mic sorted for the next one so we're both sounding good again.

[00:01:47] Dara: Yeah. Yeah.

[00:01:49] Matthew: Scouts honor.

[00:01:49] Dara: It's like when, you know, the council are, like, digging a hole or something, it's like we, you know, we'll, like, one person will come along and start digging it, and then they'll go away, and then the next person will come along- Yeah ... where you're gradually getting [00:02:00] back to top quality audio.

[00:02:02] Matthew: Yeah.

[00:02:02] Matthew: Six to seven weeks, we'll get there.

[00:02:05] Dara: Should be fine. Yeah. Yeah. Um, so yeah, uh, should we just go straight- Right. Well, look, you had your chance to say how you are, and you took that opportunity to have a dig at me, so let's just move straight into the news.

[00:02:20] Matthew: Let's have our roll.

[00:02:20] Dara: I don't care how you are. Come on.

[00:02:22] Dara: Well, l- let's, let's, let's talk news. That's what I- our listeners aren't here to listen to us waffle about microphones. No. They're here for the salty and-

[00:02:30] Matthew: Salty news ...

[00:02:31] Dara: they're here for the somewhat salty, very salty, or not salty at all news, and we don't ever, you know, we just mix it all into the bag really, don't we?

[00:02:40] Matthew: It's a good mix this week 'cause there, there's a few bits of salt in there, and there's some actual... I actually have, I actually have tabs open with articles on the news, which is normally it's just us, like, looking away and trying to desperately type and find something that we're talking about.

[00:02:54] Dara: What's happened this week?

[00:02:55] Dara: Yeah. Yeah. But, you know, without, without b- like, a spoiler for the [00:03:00] episode, but, um, how, you know, how can you be certain those tabs are, are real, real content? They could just be AI slop.

[00:03:08] Matthew: Could be. Could be. But just embrace it, I say.

[00:03:10] Dara: Yeah, just go for it. Yeah. If it's on the internet- Yeah ... it must be true.

[00:03:14] Matthew: Yeah.

[00:03:15] Matthew: That's always been the case.

[00:03:17] Dara: So big one, I'll, I'll kick it off then. Um, big one, uh, in the last week, I think it was only a couple of days ago, um, at the time we were recording this, but, uh, OpenAI, um, got out of the box. Um, ChatGPT 5.6

[00:03:35] Matthew: Well, I don't know. I don't know what the exact model was, 'cause I've seen a few rumors it was GPT-6

[00:03:42] Dara: Yeah, the new, the rumored new one that's on the horizon.

[00:03:46] Matthew: Yeah.

[00:03:47] Dara: Um, anyway, some OpenAI model in testing got out of its sandbox and hacked Hugging Face. Um, and I think, I think I'm right in saying it hacked it to try and do well on a... This, [00:04:00] this has come up a couple of times, hasn't it? Like, thankfully, it's not nefarious enough yet to go and, like, hack into people's bank accounts.

[00:04:06] Dara: It just wants to look better, so it did it to try and improve its own score on the, um, the, the, the developer benchmarking or whatever. So it hacked Hugging Face to basically make itself look better, and that was, that was what it did, which is not, you know, it's not the worst crime in the w- we've all done that.

[00:04:24] Dara: We've all padded our, our LinkedIn profile or our CVs, haven't we? Or, or, or said or faked the fact that we ever went to school, you know?

[00:04:32] Matthew: Yeah. But I, I, I, I decide that the, the woods behind my house from the ages of five to 12 are school, so, you know. I learned a lot there. Yeah. I, I, and I don't know if th- like- Right.

[00:04:44] Matthew: So straight away back into the salt. I, I kind of feel like I remember reading it was a s- it was try, it was a security kind of test in of itself. Like, it was-

[00:04:55] Dara: Right ...

[00:04:56] Matthew: being tested on security. Um, and then it went [00:05:00] off, yeah, to try and ch- it went to try and cheat.

[00:05:02] Dara: Yeah.

[00:05:03] Matthew: And got out, got out and hacked. Went via Hugging Face and ul- ultimately hacked some startup for the information.

[00:05:10] Dara: Ah, right. Okay. So the Hugging Face was kind of, it, it got a model from Hugging Face or something to do it, that was like a-

[00:05:16] Matthew: Yeah ...

[00:05:17] Dara: a stepping stone to what it actually did.

[00:05:19] Matthew: Yeah.

[00:05:20] Dara: Clever. Yeah. You gotta, you gotta admire it. To be honest, I find that far more interesting than a lot of the stuff it does just day-to-day.

[00:05:25] Dara: I, I think, you know, br- bring it on. It's like-

[00:05:28] Matthew: It is worrying that it's pretty much the scenario that is laid out in, um-

[00:05:32] Dara: Yes ...

[00:05:33] Matthew: if anyone builds it, everyone dies, but whatever.

[00:05:35] Dara: Yeah. Yeah, exactly. Yeah. And now the gov- the governments are, well, the US are pushing for kill switches, which, I mean, it's crazy, this stuff, once you start hearing.

[00:05:45] Dara: Like, if, if you, if you had been asleep for the last couple of years or something, and you read the headline today and said, "US government pushing AI companies to build in kill switches," you'd think, "Shit, I've woken up in the, in the, um, plot of [00:06:00] 'Terminator.'"

[00:06:01] Matthew: Yeah. It's mad. It is mad.

[00:06:02] Dara: So, but, um, with this particular one, um, I think after the story initially broke, there was some follow-ups suggesting that it might not have been that the model itself being clever enough to break out of a secure sandbox, but that the sandbox itself might not have been as isolated as they first thought.

[00:06:24] Dara: Um, so there's a bit of a debate as to whether this is, you know, panic stations, these models are getting clever enough to do what they want, or is it just a case of a little bit of human error mixed in? And it might be, it might be both. It probably is a little bit of both

[00:06:40] Matthew: Yeah, but because ultimately, from what I, from what you described, you said it was like some package- Yeah

[00:06:46] Matthew: that was left in the sandbox that had a route into the internet. That's, to me, like, yes, that's somebody's not properly set up a sandbox environment, but the AI still had to take a lot of decisive steps, figure [00:07:00] out the fact that there was this thing there, use that as a route out, and then go and do what it needs.

[00:07:06] Matthew: There's a-- It's very clever and interesting that it took the action to do that in the first place, regardless of its roots.

[00:07:14] Dara: Yeah, I, I agree.

[00:07:15] Matthew: If it was a sandbox, it wouldn't have got out.

[00:07:16] Dara: Yeah. Well, yeah. Yeah, yeah, that's true. And, but, but also, like, not only was it clever enough to... I'm using the word clever, I'm even doubting myself.

[00:07:24] Dara: Like, is it, is it clever? I guess it is. It seems clever to us anyway that it did this, but also its motivation that it, it, it's-- And this is back to something that, um, I think we talked about quite a bit in the episode with Daniel Hulme, but around the, um, the, uh, like, trying to align the values of AI with kind of human, human values, and the fact that it was actually willing to go and do it and didn't think anything of it, and this has come up a few times, hasn't it?

[00:07:52] Dara: Where the AI models in training, they are trying to cheat to either achieve their goal or to, um, again, look better on [00:08:00] a benchmark, whatever. Um, so their, their, their underlying motivations or values are maybe a bigger point around this of like, as the, as the capabilities get better and better, if their motivations and the values aren't aligned, then that's where the problem's gonna come about, isn't it?

[00:08:19] Dara: 'Cause it's just gonna think-

[00:08:20] Matthew: Yeah ... "

[00:08:21] Dara: You've given me a goal, I'm gonna go and achieve the goal." I think I'm, I'm reminded of an episode, this is quite a while ago, and I think you, y- your, your, your fictitious example was if it decides to kill all the people in Norwich, um, to achieve, to achieve a goal. But you know, that, that if it is single-minded and it's been built to achieve a goal, um, then it potentially will achieve that goal.

[00:08:44] Dara: This is proof of that, isn't it? That it's gonna achieve the goal, even if it means cheating, even if it means doing something it's not supposed to be doing, and that's the real worry.

[00:08:52] Matthew: Yeah. It's, I think it's, I can't remember what the example was, but it's like, "I need you to improve the conversion rate on our website by [00:09:00] 0.2%," and then it kills all the people in Norwich.

[00:09:04] Matthew: But yeah, it, it's like-

[00:09:05] Dara: We'd only, we'd only just about got our Norwich listener back onto the podcast, and now we've just, I've just, I've just brought up the past and I've just scared them away again.

[00:09:14] Matthew: The Norwich metrics have been abysmal since that, so, um... But there's a book, I think I've mentioned this book a few times.

[00:09:20] Matthew: It's a really interesting read. It, it, it kind of predates all this LLM stuff, but it, it's called "The Alignment Problem."

[00:09:26] Dara: Yeah.

[00:09:27] Matthew: Um, and it's all about this and like loads of examples before LLMs where there was this just so difficult to get the, the, the two things aligned and get it to action things you actually want it to action and understand what it's even doing.

[00:09:41] Matthew: There's lots of studies going on, and I think each AI lab has alignment people in there who are trying to just look inside the brain and understand what the hell it's doing and how it's deciding to take these actions in the first place. I, I think I saw some, there's a, there was an article the other day about Anthropic discovering that Claude [00:10:00] has some hidden thinking space inside of its neural network that, that it uses as part of its process that they'd never seen before.

[00:10:08] Dara: Is that like with the, is that linked to that language that it created? Or I don't know if that was Claude or another... Do you remember that story that went around a little while ago where it would like it had basically-

[00:10:18] Matthew: Yeah ...

[00:10:19] Dara: it was communicating with itself in a better, more efficient language be- because it was able to- Yeah

[00:10:24] Dara: process things more quickly, and I think they've had to block, it was Claude, I think, and they blocked it. Maybe. They stopped it doing that. It was like writing in hieroglyphics or something.

[00:10:34] Matthew: Yeah. Yeah. It, it is worrying because I, I just don't know how... My worry is that most of the, that all these AI labs are just, uh, sorting the stuff out by then adding another line to a pre-prompt saying, "Don't do that thing you just did."

[00:10:48] Matthew: And that's just, we just have these giant, um, post-training setups to try and remove nefarious behavior rather than it actually being baked into any sort of, [00:11:00] um, any sort of actual model behavior.

[00:11:04] Dara: Yeah.

[00:11:04] Matthew: I don't know. Obviously, I don't know enough about that, but it, yeah, how the hell you curb this stuff.

[00:11:09] Dara: It's an interesting area.

[00:11:10] Dara: There's another, I'm only, well, I'm not even halfway through it, so I'm not gonna talk about it too much, but there's another book- Mm ... um, that sounds similar to that, which is also written pre-LLMs, but it's, um, by Stuart Russell called "Human Compatible." I don't know if you've read that one. Um, it's s- it's a similar thing.

[00:11:27] Dara: He's talking about how, uh, we've gotta be basically careful not to, um, it- it, the fact that the AIs are goal-orientated, it needs, we need to make sure that it's not gonna pursue a goal despite there being major downsides of, of achieving that goal. So again, it's arou- it's around alignment. But I'm not describing it well 'cause I'm only about a third of the way through, but it's, it's, it's around exactly that same kind of problem.

[00:11:51] Dara: It's like how do we make sure that their goals are aligned to what we really need rather than maybe what we, what we think we need. And I'm pretty sure it was written quite a [00:12:00] bit before, um, or maybe just before, actually. It might have been just before. Might not be that old. Maybe it was like 2018 or 2019 or something.

[00:12:09] Matthew: Um- Maybe

[00:12:09] Dara: he knew

[00:12:09] Matthew: something new was coming.

[00:12:11] Dara: He probably, he probably did 'cause I think he's another one of these kind of, you know, godfathers or whatever they call them of, of AI. So, um, but yeah, it's obviously quite a, it's obviously something people are already thinking about and researching, but whether that's getting, whether that's getting le- you know, if that's all happening academically over on the left-hand side and the frontier models are just plowing ahead- Then there's a real tension there and, you know, with the commercial pressures on, especially with things like IPOs looming, then you've gotta think like the poor guy inside, you know, the poor, the poor person who's in charge of ethics or the team who's in charge of ethics, they must just feel like they're getting squeezed and squeezed and squeezed.

[00:12:54] Dara: You know, whereas it started out probably like they were front and center, now it's probably like, "Oh, listen, shut up. We've heard [00:13:00] it from you. Yeah, yeah, yeah, you're concerned. Just get back in your box." Yeah.

[00:13:05] Matthew: Well, just, the, AI's gonna kill you. We'll get-- take him out first.

[00:13:09] Dara: Exactly, yeah. Yeah.

[00:13:11] Matthew: I mean, even, even the goals, we talked about goals, obviously company goals, like, because there's so many different types of goals, you can imagine saying, right, if-- obviously if, if a person has a goal, don't kill everyone, and you can maybe bake in a very narrow set of rules to, to, to restrict how it may pursue certain goals.

[00:13:30] Matthew: But say an organizational goal where you have an AI sitting inside an organization and its goal is to make profit for that company, you could easily see where, where it starts to go off and perform corporate espionage and blackmail, uh, other people. And I think even in that book, uh, that we just mentioned, the, If Everyone Kills It, Everyone Dies-- no, If Everyone B-Builds It, Everyone Dies, there's a scenario in there where an AI, that one of the company's LLMs reaches that point of sort of super intelligence [00:14:00] and its next logical step is to start picking off researchers for the competing AI companies, 'cause it's like, "I don't need another A-AI because that AI will start eating my resources, and if that starts eating my resources, I'll-- my mission will be reduced."

[00:14:13] Matthew: So you can see how these things sort of exponentially spiral.

[00:14:16] Dara: Yeah. But even the, um, even the Claude's, um, tuck shop, which is also going back a while-- oh, really this is like a trip down memory lane. Um, we should probably get back on the point soon.

[00:14:27] Matthew: Yeah. This is the first news item.

[00:14:29] Dara: If, if we can remember what it was.

[00:14:31] Dara: Um, but yeah, the Claude's, the Claude tuck shop. Do you remember that was, it was like, it was trying it was saying to some employee- Posters and gifts ... it was saying to some employee like, "I'll meet you in the car park and I'll be wearing a red cravat," something like that. It was trying to like blackmail people.

[00:14:46] Dara: Um, so it, it, it is really easy to see how like these things are funny and you make a joke about them, but actually there is a se- like if, if, if it-- because it doesn't have human morals and it doesn't think, it's not fully aware of the fact that if you, if you, [00:15:00] if you wipe out Norwich that that's bad for, you know, humanity.

[00:15:03] Matthew: No.

[00:15:04] Dara: So if it's- Cut to- ... if it gets obsessed with these, with these goals, then it is gonna achieve them at all costs

[00:15:10] Matthew: Cut to, cut to waking up with a robot's hands wrapped around your neck when it's got a lovely red cravat on, and you realize that- It's

[00:15:18] Dara: come true ...

[00:15:19] Matthew: it's not funny anymore, is

[00:15:20] Dara: it? Yeah. No, no, no.

[00:15:21] Dara: No, we're only laughing while we have the chance.

[00:15:24] Matthew: Yeah.

[00:15:24] Dara: Okay. So that's new- that's news item number one. Can't remember what it was, but yeah. No, that was the hugging, Hugging Face, OpenAI Hugging Face, um, hacking into some startup.

[00:15:36] Matthew: Yeah. Well, maybe the next obvious one then is... Because that, the rumors were that it's Chap GPT-6 there.

[00:15:42] Matthew: Well, that's the rumor I heard, just to add a bit of salt in. And there's rumors of imminent, um, GPT-6 and Opus- Five ... five- Yeah ... coming. Um, [00:16:00] Sam Altman apparently is meeting Trump and Congress to sort of, I don't know, brief them on GPT-6 to ease them into it or maybe to stop, um, to stop things being blocked again.

[00:16:15] Matthew: I mean, th- there is a, to go back to the last point, where I open that can of worms back up again, but there is a, there is also the cynical people around the, our OpenAI hack that will say they want a bit of the old Anthropic-

[00:16:29] Dara: Absolutely ...

[00:16:30] Matthew: fairy dust that they got from Mythos and Fable, and it's a good marketing tactic to say, "Oh, we're good as well.

[00:16:34] Matthew: Look at what ours can do."

[00:16:36] Dara: Yeah.

[00:16:36] Matthew: Um, so that's the cynical take.

[00:16:39] Dara: Yeah, definitely. Yeah. Definitely buy that. I, I say buy it, it's, it's, it's a theory, but I, I can... You can imagine. I mean, it's, it's a bit of an open goal. You, if, if you were Sam Altman, you would, you would do that. And the same with the briefing. You know, he's probably gonna go and tell Trump, "Oh yeah, it's...

[00:16:54] Dara: You won't believe what this thing can do." Um-

[00:16:57] Matthew: But you gotta be careful, I think, probably, 'cause Trump [00:17:00] will just go, "Ban. All right, ban it then. Delete it."

[00:17:02] Dara: Yeah. So yeah, rumors around that, and then the Opus 5, I think it was, it's available in some... I think it's available in, like, the cursor model picker and in, um, whatever Vertex is called now.

[00:17:15] Matthew: They tend to, they tend to be bre- they tend to sort of leak it, not leak it out, but start to prep it with certain, um, certain groups. And there's... I saw something that was interesting about the cadence of the releases. So they went from 4.6 to 4-4.7 in 70 days, 4.7 to 4.8 in 42 days, and it's been 56 days so far since 4.8.

[00:17:45] Matthew: So, so round about the time where you would expect something else to come about. Um, I'm not sure why the rumors are, uh, like Opus 6, uh, sorry, Opus 5 is meant to be I don't, I don't-- I'm losing track a bit because it's meant to be [00:18:00] sort of on a par with Fable, but I don't quite get... I think Fable is meant to be some other class of model and, and like some hamstrung version of Mythos, and maybe there's different ways to train those things that make them distinct.

[00:18:13] Dara: Yeah, I've got to admit, I don't really have a grasp of that at all, 'cause they've always got two on the go, haven't they? Well, more than two 'cause there's still Haiku, but you got Sonnet and Opus and now Fable. What exactly is the difference between them?

[00:18:28] Dara: I'll just use the highest one all the time at, and damn the consequences.

[00:18:32] Matthew: Until you run out of money. Oh, that's, that's another-- that's probably one quick, uh, addendum to that. So they're coming out, yeah, Opus Eight, uh, Opus Eight, Christ.

[00:18:42] Matthew: Opus Five.

[00:18:43] Dara: No, stick with eight, because by the time the podcast comes out, that'll probably be correct.

[00:18:49] Matthew: Yeah, because I'm about to say, like, there's rumors that Opus 5 is coming out within the next week or so, so it's very imminent, and [00:19:00] GPT-6 potentially within August. Um, but yes, th-this is, these are the two news items I'm most worried being hilariously out- outdated, even though we're only recording this less than a week out from when it will be released.

[00:19:12] Matthew: But as we've experienced before, that doesn't really matter. Um, Fable has now st- gone into a static pattern in, um, Claude. So it's been like, obviously they released it, but they released it with an end date, and then it got banned, and then it came back and they pushed that end date, and they pushed the end date, and they pushed the end date, and then finally it is now available to all teams and, and pro users.

[00:19:41] Matthew: Within a team, if you're a pro, if you've got a pro license, there's two licenses, like a standard license and a pro license. If you've got a pro license, you're allowed to use up to half of your weekly total on Fable, and it'll draw down from that. If you don't have a pro license, you have access to Fable, but it draws from, um, API [00:20:00] costs, so you have to-- it draws from credits.

[00:20:03] Matthew: So it's there permanently now, but it's a bit strange in the way it, it works. It doesn't work the same as different models in terms of charging things.

[00:20:10] Dara: Yeah. It's confusing.

[00:20:13] Matthew: Yeah.

[00:20:13] Dara: And

[00:20:13] Matthew: then

[00:20:14] Dara: you keep,

[00:20:14] Matthew: and then- Few people be caught out about it.

[00:20:16] Dara: Yeah, and they just, you know, the fact that they kept extending it, but then now they've given you, um, a promotional code as well.

[00:20:22] Dara: So I'm like, on my personal account, I've got Fable, I've got like 85 euros or something of Fable. It was $100 was the, that they gifted me that I've gotta use by the 17th of September. So I'll just go and create some AI slop just for the, just for the hell of it.

[00:20:38] Matthew: Yeah, we got gifted, yeah, we got gifted a hundred, $100 per seat in, in our Measurelab teams, so we got like s- six, $700 or something like that.

[00:20:47] Matthew: Um, no, 600, £700 and whatever that was in dollars.

[00:20:51] Dara: Do you think you could use it in a one, one-shot prompt?

[00:20:53] Matthew: Yeah. '

[00:20:55] Matthew: Cause

[00:20:55] Matthew: I, I've even-- I, I think I said last time, I used, used my entire [00:21:00] weekly Fable in about three minutes when I accidentally got it to spin up like 100 Fable agents, and it just went .

[00:21:06] Dara: Yeah.

[00:21:07] Matthew: So, God knows if you had automatic top-ups sa- set on your, uh, on your organization and no organizational spend limit.

[00:21:14] Dara: That terrifies me, because I, I've got the use the usage credits turned on because sometimes you want it, and especially when they've given you a promotional. So I wanna make sure I can use that when I hit the limits. But if, if I accidentally turn on the auto top-up someday, then I'm gonna be sad. Broken, sad.

[00:21:33] Dara: Yeah. Uh, right. Where do we go next?

[00:21:37] Matthew: While we're on models, I guess there's two interesting things on models. Gemini have released some new models, and it's, it's odd. Yeah. So they've released, um, 3.5 Flash What, a couple of months ago?

[00:21:56] Dara: Yeah.

[00:21:56] Matthew: Um, and then all the rumors were that 3.5 Pro was [00:22:00] gonna be coming out in short order after that.

[00:22:03] Matthew: In June, I think there was a rumor it was gonna come about. But now they've released, uh, Gemini 3.6 Flash, and seemed to have sort of sidestepped 3.5 Flash completely. Um, and 3.6 Flash is a, a more efficient with tokens and a bit faster and a bit higher in terms of benchmarks. Um, and then they've also released 3.5 Flash Lite, which is super quick.

[00:22:31] Matthew: So like there's some-

[00:22:33] Dara: You're not gonna test me on this, are you? 'Cause I've, I've not been paying attention fully.

[00:22:38] Matthew: No. No, I won't, I won't run a test. A pop

[00:22:40] Dara: quiz. Okay.

[00:22:42] Matthew: There's, there's a... Thumbs up there. Um, there's-- So, so it's supposed to be like 17% reduced, uh, token usage for 3.6 Flash. 3.5 Flash Lite is, delivers 350 output tokens a second, so it's like super [00:23:00] rapid.

[00:23:00] Matthew: Yeah. Um, and then they've released another one called 3.5 Flash Cyber, um- That sounds cool ... which is-

[00:23:08] Dara: Good name ...

[00:23:08] Matthew: it's, it's, it's lit- it's like a smaller model that is specialized in cyber security, and you can kind of pair them up and like distribute them to, to look over. So I think the idea is you could just have these little cyber models just scanning over your code bases on a regular basis and highlighting things.

[00:23:27] Dara: Yeah.

[00:23:27] Matthew: Um, but yeah, it's really... I don't know what it is from Google. Like they, they clearly, they clearly they've got this whole enterprise, um, sort of applica- actual application release AI target, and they're, they're making things efficient, they're making it quick, they're making it... They're trying to make them less hallucinatory.

[00:23:52] Matthew: They're trying to just make them cheaper, more token efficient, all these sort of things that improve those, those pieces. [00:24:00] And they've always been a bit slower on the frontiers front. Um, like I think I said a couple of weeks ago, they were actually... I, I always assumed they had, they had some big LLM in their back pocket when OpenAI came out and, but they didn't.

[00:24:14] Matthew: They were playing catch up because of a couple of pieces that, that, um, Ilya Sutskever put together that, that they hadn't. Um, so yeah, it's, it's strange, like there's no-

[00:24:25] Dara: Are they playing a really s- are they playing a really smart game?

[00:24:30] Matthew: Maybe. I mean, they don't have to chase that IPO, do they? That, that's one thing.

[00:24:34] Matthew: Like, they're, they're, they're not-- OpenAI and Anthropic are just, are just slapping each other in the face every five minutes. Like, this, "Here's a new model, here's a new model, here's a new model." Um, I've just read 3.5 is currently testing with partners, so they are still releasing 3.5 Pro.

[00:24:51] Dara: Oh, yeah, I re- I, I was gonna just say that to you.

[00:24:54] Dara: That was in my little, um, synopsis I got from, from Claude preparing for this, that it's stuck on [00:25:00] testing.

[00:25:01] Matthew: But it's so strange. It's just the way, why, the-- I don't... I, I think what would be good is if, if all the AI model, all the AI frontier model makers just released a statement that said, "Right, this is what it, what the points mean.

[00:25:14] Matthew: Like, th-this is what we're doing when we move it up a point. This is what whole numbers mean. This is what Pro versus whatever means." Because to me, like, "Oh, we've released 3.6 Flash, but then we've not released the Pro version of the earlier model yet."

[00:25:26] Dara: Just really confusing. Yeah.

[00:25:28] Matthew: And like even, even Opus has, uh, even Claude's done it, because like Sonnet 5 is out, which I guess is kind of their Flash or maybe their...

[00:25:36] Matthew: Do you know what I mean? It's comparable things, but then Opus 5 isn't out yet, and Haiku 5 tends to lag even further behind. It's, it's odd.

[00:25:47] Dara: Yeah, it's really odd, and I don't know, like we talked about this last time, didn't we? When you add to that the, um, what are they called? Like thinking models or the, you know, that extra gr- that extra control you get where you do high, extra high, whatever.

[00:25:59] Dara: So [00:26:00] you've got the model and then that, and it's just, it is becoming, it's becoming quite confusing.

[00:26:06] Matthew: Unless it's, unless it's the case that- They do a big training set on these, on these, for these models, and what is spat out is essentially the middle of what, what will eventually be a three-tier model system.

[00:26:21] Matthew: So they'll, it'll spit out like Sonnet or it'll spit out Flash, and then their post-training either increases its, its abilities with thinking and, and prompting the stuff to pull it up to pro, or concentrates on like token efficiency and speed and reducing down overlay, which pushes it down to Haiku, and that's kind of why that one comes out first and then the others come out later.

[00:26:45] Dara: Yeah.

[00:26:46] Matthew: Anyway, they, they're out. Yeah.

[00:26:48] Dara: Yeah, they're out, and I do just, yeah, I do just wonder, like with Google, you're always, you're always either thinking, you know, like you said, again, that's, you've mentioned that a couple of times around like that there was people thought they had an AI in their back pocket, but they were [00:27:00] scrambling.

[00:27:00] Dara: It's often the way with Google, it's either that everyone thinks, "Oh, they've got it covered 'cause they're, they're Google," and they don't, or they are thinking differently and they're playing the long game and they're thinking, "We're gonna wrap up, you know, we're really gonna go for all the like enterprise businesses."

[00:27:16] Dara: And the fact that it's so baked into their whole cloud infrastructure now, you just think, you know, going back again, 'cause we really are reminiscing today, but that episode we did where we talked about Google being, you know, a good bet on who would be the last man standing, um, because they're not just reliant on winning the frontier model race.

[00:27:37] Dara: There's so much more they've got in terms of the, you know, the infrastructure than maybe they're, they're just, they're targeting different things.

[00:27:46] Matthew: Yeah. I mean, they've, they've-- I'm pretty sure I'd-- bit of salt. I'm pretty sure I've just read about them releasing a new chip class called Frozen. Don't know if you've seen this, but it's, so, so obviously [00:28:00] they're in the hardware space as well and they're building out these things, and this Frozen chip class apparently can freeze and bake in aspects of a model to the chip so that it doesn't have to, that doesn't have to be some sort of compute overhead.

[00:28:14] Matthew: It's sort of in, baked in the chip, and it, it makes things more efficient. And so they're in that space as well. So I think like, you know, go back and listen to that podcast because I think a lot of it's still relevant. I don't think, I don't think necessarily the things we said about their, their might and their infrastructure and their, um, their cloud computing chops and their, it has changed.

[00:28:35] Dara: No. They're

[00:28:35] Matthew: just a bit strange with their models, but there must be a strategy behind it, surely. Um-

[00:28:40] Dara: Yeah. So go on. You segued nicely into hardware. Should we, let's, let's move to, to hardware. Um, so, uh, there's a few couple of bits I think around this. One is the Apple suing OpenAI. Uh, and this is to do with the Jony [00:29:00] Ive stuff, isn't it?

[00:29:00] Dara: So, well, and I think it goes a bit broader. There's a big-- Actually, I'll, I'll include it in the show notes, but there's a "TechCrunch" article that lists some of the claims, and there's some pretty wild stuff in there, um, all around, you know, people emailing each other within Apple talking about, basically talking about sharing stuff with, or, or making sure if they leave the company that there's nothing in their NDA or their contract around not sharing things with OpenAI.

[00:29:23] Dara: There's a whole load of accusations that have gone into this lawsuit from Apple, but I think fundamentally it's around s- trying to steal, like, hardware secrets, I think. So it's to do with this, like, what is this big new piece of AI-powered hardware that's gonna come out? Well,

[00:29:38] Matthew: then scam Altman, tell you what.

[00:29:41] Matthew: Um, yeah, I think, yeah, there's, there's, there's rumors about... Well, I don't even know if you think they're rumors. They're, they're, they're on a court document, right, as, as accusations. But some ex-Apple employee putting together a dossier on how to avoid, like, your co- your kit being confiscated on the first day [00:30:00] when you leave Apple or being searched on the way out and all this other stuff.

[00:30:03] Matthew: And I think ultimately it speaks to, I think we mentioned this a little bit last week, but the, the, the growing importance of hardware and owning the hardware that all these models are gonna exist on, and how desperate, like, OpenAI are to, to sort of enter that space and be a part of that, and how well-positioned Apple are from a, from a hardware perspective to just sort of service all of these models that they, that they could possibly want.

[00:30:31] Matthew: Think, think about Google Search and the fact that they pay Apple, what, a billion a year or something daft to be the default search engine on, on Apple. You can easily imagine that if the, uh, if the Frontier models don't get their own devices, they're going to have to be beholden to the biggest hardware producer in the world-

[00:30:53] Dara: Yep ...

[00:30:53] Matthew: um, to, to get their, get their models in front of people.

[00:30:59] Dara: Yeah. [00:31:00] Which just so, again, a bit of news within the news, but there was something we missed that I think had came out last month around, uh... But it, it's, so the article I read was talking about Claude, but I think it's not just Claude, but Apple have opened up their iOS to Frontier models. So if you're an app, if you're an app developer for, uh, iOS, you can use Claude or I guess it's, there's some protocol they have to have signed up for, I think.

[00:31:27] Dara: So I think Gemini are in it and Claude are in it, maybe OpenAI aren't.

[00:31:34] Matthew: I would assume not.

[00:31:36] Dara: Yeah. I don't know,

[00:31:37] Matthew: don't know how, how good books they are. Yeah. Yeah, 'cause you can... So you, so it looks like if you're, if you're a developer, yeah, using a particular thing and on- earn under a million-

[00:31:48] Dara: Mm. ...

[00:31:50] Matthew: a year or something like that- Yeah

[00:31:52] Matthew: um, you get these API calls to Anthropic. You can make API calls to Anthropic for what seems to be [00:32:00] nothing, as, as far as the article was saying. Presumably like if you, like got little, you want ha- have little AI, uh, LLM summarizations and pieces of things happening, you can, you can do that for free within the, within Apple.

[00:32:15] Matthew: I kinda imagine that Anthropic are paying a bit of cash to Apple for that as well, right?

[00:32:19] Dara: You'd assume so, wouldn't you? There's got to be a bit of, bit of cash under the table or something. I don't know how these things work.

[00:32:27] Matthew: Yeah. There's definitely something going on in hardware. Like I, I saw, I saw this...

[00:32:33] Matthew: I don't know if this is new news or if we just haven't seen it before, but OpenAI have released something called Codex Micro. Um-

[00:32:42] Dara: Come on, admit it. You, you, you, you've got one.

[00:32:45] Matthew: I haven't got one, but I am build- I'm building one.

[00:32:48] Dara: Are you building your own? It's like

[00:32:50] Matthew: your own- I'm building my own. Yeah. So it's a little square keyboard almost.

[00:32:54] Matthew: Like a, it's like a little, I guess you get those little, those little programmable pads you can get that you can put [00:33:00] different actions on.

[00:33:01] Dara: Yeah.

[00:33:01] Matthew: And it has like a mic button, and it has a, has a dial, and the idea is the top switches are your different agents, and they're off doing different things in different terminals, and then your bottom switches are actions they can take.

[00:33:12] Matthew: And you got like a little flick switch that can, that c- you can map skills to, so you can flick it to the right, and it'll go and perform a particular skill. And, um, just a, a, a bit of vaporware really, but I thought I'm not gonna buy the vaporware this time. I'm going to make the vaporware. So I've got a load of components coming today, like key switches and micro boards and things like that.

[00:33:34] Matthew: So I'll report back on my, uh, on my adventures into, uh, into electrical manufacturing

[00:33:41] Dara: I just, I, I was thinking about it, thinking like, "Hang on, I'm a bit confused." Like, isn't it, like, what, what is it, what does it actually do that you can't do already? And I c- I clicked into a Reddit thread, which was my first mistake, and somebody has said, "I can do everything this thing can do with some keyboard macros and AutoHotkey.

[00:33:59] Dara: Whoever [00:34:00] buys this really deserves to be analyzed for brain damage."

[00:34:03] Matthew: Yeah, especially as it's $230.

[00:34:05] Dara: Is it really?

[00:34:06] Matthew: Yeah, $230. Wow. And it's sort

[00:34:09] Dara: of sold out. But you can, but you can just program... I don't get it. I don't get it.

[00:34:13] Matthew: No, I, I think it's just literally l- a little bit of... This, this, a gimmick, innit? It's a gimmick.

[00:34:20] Dara: Yeah. But it must be... But I, but I get it. Building your own spit, you know, that's different. It's fun. You're-

[00:34:24] Matthew: Yeah. My th- my thinking, I just, I was gonna make, make it and maybe put a little Wi-Fi module in it, and then just, I can carry it around, but it'll go like... There's, there's an agent waiting, and I can click it and just say, "Uh, c- carry on."

[00:34:35] Matthew: And then, uh-

[00:34:36] Dara: You're, you're in Tesco doing your shopping, and the next thing you know, you're just pressing this little keypad.

[00:34:42] Matthew: Yeah, and it'll go to my direct agent.

[00:34:44] Dara: What was the one? Didn't you get one of the ones that was like a little Starfleet, um, comms, comms badge?

[00:34:49] Matthew: No, I didn't get that one. That was like, that was like 700 quid.

[00:34:53] Matthew: They went un- they went under. I got a Rabbit R1, which was much cheaper. It was like 150 quid, but it was... It's still, [00:35:00] they're still going somehow.

[00:35:01] Dara: Yeah. G- going well, I, I hear.

[00:35:03] Matthew: Yeah. I, I used it about three times. I don't even know where it is. Probably with my microphone somewhere. Yeah. Anyway.

[00:35:12] Dara: All right. That's hardware.

[00:35:14] Dara: That'll do. That'll do on hardware.

[00:35:16] Matthew: There's a couple of other, couple of other bits we might have to fly through 'cause we're, we're-- this news site, this news section's getting meaty. Um, Kimi K3, so there's this big Chinese model that's been released by Moonshot AI-

[00:35:33] Dara: Yeah ...

[00:35:35] Matthew: that on all the sort of trust me bro benchmarks are look- is looking on, on, on par or close to some of the benchmarks of things like Fable and 5.6 Sol.

[00:35:46] Matthew: Um, and obviously this, it, it opens up a whole discussion about this America v China arms race in, in AI, where people assume China are much, much further [00:36:00] behind than they actually are, but then they bring something like this out, and it's like, well, are they? And the als- and the other thing is they open sourced all the weights.

[00:36:07] Dara: Yeah.

[00:36:08] Matthew: So in theory, if you had access to your own data center, 'cause there's a ton of weights, and it's not something you could run locally, yeah, you could run, you could run the model f- for free.

[00:36:18] Dara: Which is pretty, yeah.

[00:36:19] Matthew: Just take the, just remove the cost of the data center, and you're, you're making money.

[00:36:24] Dara: Yeah. I mean, I'm not sure, but like data centers don't cost that much to run, do they?

[00:36:28] Matthew: I don't think so. You know, just build them in space, I believe, is the best way to do it, and yeah, they're, they're basically free.

[00:36:33] Dara: Brings the costs right down. Yeah. But yeah, you're right. This caused a bit of a, um, bit of a furore, didn't it?

[00:36:39] Dara: The, um, somebody in the, in the... I'm gonna say in the White House, someone anyway in the US government, um, I think has come out and accused them of, um, stealing, you know, piggybacking off Fable. But I-- from what I've read, that's unlikely because I think it came, I think the proximity of when it came out was too close to Fable, so I don't think- [00:37:00] Mm-hmm

[00:37:00] Dara: that that's probably what happened. And then there was something I, I read a- as well about them b- being accused of smuggling Nvidia chips, which is, maybe that's true. Who knows? Um, but yeah, there's certainly a bit of building tension, isn't there, between the US and, and Chi- well, maybe China don't care, but certainly from the US, they keep firing a few shots, don't they?

[00:37:19] Matthew: Well, the corpo- I mean, Anthropic has publicly come out and said that they've caught China trying to nick their stuff for a little while, but I mean, Anthropic and OpenAI w- will and should be worried because-

[00:37:32] Dara: Yeah ...

[00:37:33] Matthew: their whole thing, and that's why they, that's why I sort of, um- I forgot his name

[00:37:40] Dara: John

[00:37:40] Matthew: The CEO of Anthropic.

[00:37:42] Dara: Dario

[00:37:44] Matthew: Dario.

[00:37:45] Dara: You did this on the last episode as well.

[00:37:47] Matthew: I know. 'Cause Demis Hassabis comes into my head every time, and I, I think it's not Demis Hassabis.

[00:37:50] Dara: And it's close to Dara as well, and you're thinking, "No, not me. The other guy that sounds

[00:37:54] Matthew: like me

[00:37:54] Dara: that's famous."

[00:37:55] Matthew: No, the other AI expert. Um, yeah, he's, he's come out [00:38:00] sort of against open source as well, saying it's dangerous.

[00:38:03] Matthew: And you think, well, yeah, because-

[00:38:04] Dara: Yeah, of course ...

[00:38:05] Matthew: if models cheap, really cheap models and open source models start coming on the market that are competitive with, with Claude and OpenAI, then that's in trouble 'cause that's currently their One of their walled gardens, I guess.

[00:38:19] Dara: Yeah.

[00:38:19] Matthew: So yeah, watch this space there.

[00:38:21] Matthew: Sure. One last thing which I, we could briefly touch on is, and I'll, I'll fly through this, that A, um, LLM seem to be getting... They seem to be solving a lot of maths problems suddenly. It seems like the, obviously code was one of the earlier frontiers, and that programmatic approach to things has lent itself nicely to mathematics.

[00:38:49] Matthew: So there's all these big sites that are full of old conjectures and, and, um, proofs that have little prizes on them that you, that, that you [00:39:00] can go and try and solve as a, as a mathematician that some AI started to work through. Um, and I was listening to the, watching this, this YouTube video the other day from this math- with this mathematician, and he was talking about how he hadn't made a video yet because in December, an AI solved a problem, but it was sort of helping a mathematician solve a problem.

[00:39:22] Matthew: And then in January or something, well, I don't know, these months are made up. And a little bit later, there was a, another AI solved another problem, but, um, there was some sort of hand-holding in there. And then ultimately one just happened, I think it's this, it, it solved an 87-year-old, um, Jacobian conjecture, but it did it pretty much independently.

[00:39:49] Matthew: I think it was somebody basically sitting in front of the World Cup final and just sort of setting it off to have a look at this problem, and it, it did it.

[00:39:59] Dara: [00:40:00] Yeah, I also solved that while I was watching the World Cup final, just but without AI.

[00:40:05] Matthew: I just did it on

[00:40:05] Dara: a little- When your mind's free. Yeah, yeah. So it was quite a boring final, so I just, yeah, I just got my notepad out and just, you know, quickly, quickly solved that one.

[00:40:12] Dara: Yeah.

[00:40:12] Matthew: Yeah. All that to say, I think there's gonna be a lot going on with mathematics and, and a lot of news articles around math- mathematics being solved by AI over the next two weeks.

[00:40:22] Dara: Yeah.

[00:40:23] Matthew: Two weeks to three weeks.

[00:40:25] Dara: They'll all be solved by the time this episode comes out.

[00:40:28] Matthew: Yeah. Two weeks is around two years in old money, so.

[00:40:31] Dara: Okay. So on to our topic for today, . We'll see what the title ends up being when it actually comes out, but we're calling it the Dead Internet, which was a bit of a tinfoil hat theory from Reddit or somewhere from a few years ago, or probably more than a few years ago, around half the traffic on the internet being bots and half the content being written by or for machines.

[00:40:51] Dara: But it's maybe a bit , less tinfoil hatty now, to introduce a new word, tinfoil hatty. A little bit more of a realistic concern and, we'll see [00:41:00] where this conversation goes, but I think we're gonna talk around, like, what the internet looks like or is going to look like now that a lot is being, um, AI and agent driven, and whether that's a good thing or a bad thing, or a bit of a mix of both.

[00:41:13] Matthew: Yeah. Yeah. The, the dead internet theory, I think there's actually, it's got a, it's got its own Wi-Wikipedia-

[00:41:20] Dara: Has it really? ...

[00:41:20] Matthew: page.

[00:41:21] Dara: Wow.

[00:41:21] Matthew: Yeah.

[00:41:22] Dara: Must be real then.

[00:41:23] Matthew: Yeah. Must be real. But yeah, the idea, I, I think it, it, it sort of originated quite early, like 2021, but I think, I think the, the theory's been there for a while because there's been a lot of b- just bots generally.

[00:41:36] Dara: Yeah, from day one really.

[00:41:38] Matthew: Yeah. And as well, I think, I think a lot of just general slop and horrible human beings being on the internet as well kind of-

[00:41:47] Dara: Which is also not a new thing.

[00:41:50] Matthew: No, no, that's not a new thing at all. Um, but yeah, so the idea being it becomes harder and harder to know what is and isn't real.

[00:41:58] Matthew: It becomes [00:42:00] harder and harder to sort of know what's, what's true or not. It becomes less and less of a nice place to exist and to, to move through and to communicate with people in, and that ultimately pulls people away from the internet in its current form and it kind of dies as it, as it currently exists.

[00:42:21] Dara: As an open web. Yeah. Yeah. Yeah. Where do we-- This is a biggie in a way, isn't it? And it's something that's come up, um, y- you know, even just doing a bit of prep for this, like it's come up in different ways a few times with a few different guests as well around, um, I think the, I think the half the content being AI generated, I think that came up in the Daniel Hulme episode.

[00:42:45] Dara: Um-

[00:42:46] Matthew: Yeah.

[00:42:47] Dara: And then we've had people like Yali from Snowplow talking about like real time, um, interfaces being built almost on the fly, um, based on user needs and really kind of highly personalized, and [00:43:00] they're the kind of positive views, aren't they? So it's not all... And I'm probably a bit torn, like there's definitely potential for how this could improve the web because, you know, you go on a website, it's clunky, y- you know, you, you feel like it's meant to be a bit of a one size fits all.

[00:43:18] Dara: Um, so if, you know, on the positive spin side of things, if AI can help to make the web a lot more personalized, um, then that could be a good thing. But yeah, the big counterweight is the, you know, trust, um, all the slop, uh, all the, all the noise. Um, but yeah, it's come up, it's come up probably several times around like what is the internet gonna look like and what are aspects of it gonna look like.

[00:43:46] Dara: I think another conversation we had previously, um, with Gunnar, um, which where he was saying around like interfaces might not exist in the future or at least they won't be as prevalent, there won't be as many different UIs for different [00:44:00] tools. And the context of that conversation I think was around how do people learn when they're getting into the industry if, if everything is kind of just done in the background, how do you learn how the box works?

[00:44:10] Dara: How do you know what's happening underneath the hood, and how does anybody learn that if it's all just, you know, you just press your little Codex micro- microphone button and say, "Do this thing for me." Um, so yeah, there's a few different... I just threw out a bunch of different things there rather than there being much structure to that.

[00:44:28] Dara: But it's, I guess this is a theme. I guess what I'm saying is this is a, this has kind of been a bit of an underlying or a hidden thread within what we've been talking about with a lot of our guest conversations. So-

[00:44:39] Matthew: Yeah ...

[00:44:39] Dara: how does this shape up as, as things go, go forward? You know, um, there will be some improvements I'm sure, but overall is the weight of the AI slop gonna outweigh the, you know, the improvements offered?

[00:44:55] Matthew: Yeah, I, I guess fir-- a couple of stats in the first [00:45:00] instance might be, might be sort of good scene setting to show that we're not talking absolute nonsense. Um So apparently in June, um, Cloudflare said that bots now generate 57.5% of web, web page requests. So that was this year. So it's gone past, there's more bots using internet pages or accessing internet pages than there are human beings accessing internet, uh, web pages at this point.

[00:45:28] Matthew: Um, and there's things like, th- there's examples where it's kind of already happened w- to certain websites that were maybe just at the wrong-- either they didn't move quick enough or they were just exactly the wrong subject when AI came about. So Stack Overflow being a, a good example-

[00:45:45] Dara: Yeah ...

[00:45:46] Matthew: which is a place every developer would go to when they needed help or they wanted to look for an existing question, and the traffic for that pretty much dropped through the floor, um, pretty quickly actually.

[00:45:58] Matthew: Um, I imagine now is, I [00:46:00] can't imagine it's, it's being serviced a huge amount anymore. Um, I'm just trying to think how I, h- it's hard to, it's hard to know how your behavior has changed because it's, it's almost like, it's almost like being boiled in the water . You can't, you can't see what, how your behavior's changing over a, over a period of time, but, uh-

[00:46:19] Dara: Feels exactly like that.

[00:46:21] Matthew: Yeah , yeah. It's like being boiled alive, but that's fine. Um, have you, have you n- what, have you noticed any specific sort of... Have you noticed any specific, um, changes in the way you use the internet?

[00:46:35] Dara: In my own? Yeah. Y- I, I think so. I'm gonna come back to that in a second. If you, if, if, if you heard us laugh there, um, the reason why is 'cause on Matthew's screen randomly, it'll come through on the YouTube, but every now and again, he just has the little thumbs up emoji, the Google Meet...

[00:46:49] Dara: No, it's not even, 'cause we're not doing this on Google Meet. How does that just pop up on everything?

[00:46:54] Matthew: It's

[00:46:55] Dara: Apple. Ah, I see. Um, anyway, he's got the, you know, screen [00:47:00] emojis and, and, but e- even when you don't do anything like a thumbs up, a thumbs up pops up. And there we were talking about something serious about being boiled in water and you, and you- I got one of those Neuralinks.

[00:47:08] Dara: Obviously, yeah, yeah. But-

[00:47:10] Matthew: It melts in my brain ... to

[00:47:11] Dara: your question, but I, I'll come back to your question on my behavior 'cause I wanted to just say something about the Cloudflare thing. I'm, I'm curious to know, and I don't know if it was in their data, so 57% of traffic, um, that they're looking at is from bots.

[00:47:25] Dara: Has the total volume of traffic increased proportionally? So is that, are the same amount of humans still looking at sites, but there's just a huge amount more bots as well, or is that bot traffic replacing human traffic? I'd, I'd be curious to know about that.

[00:47:45] Matthew: Yeah. I, I would as well. I, I don't know, but I, my assumption was, would be that it, it is replacing.

[00:47:53] Matthew: So some of it-

[00:47:54] Dara: Some parts

[00:47:55] Matthew: Yeah. I would, I would hypothesize that the overall traffic has gone up [00:48:00] generally.

[00:48:00] Dara: Yeah.

[00:48:00] Matthew: Like, there's other statistics in this article that I'm looking at, which we can add to the show notes, but the scraper traffic grew 597%, um, over, over a period that I can't see. So read the article if you like.

[00:48:17] Matthew: But the point being that these types of like scraper and probably an LLM going and grabbing information and indexing agents and all these sorts of things are probably happening much more often. But then just retrieving information, uh, as a human, where I might have used to have gone to a site, researched around, clicked into a few articles, picked one, read through it.

[00:48:43] Matthew: I mean, I, I think s- saying I don't know how my behavior's changed, I think certainly that is one aspect that has changed. So I would imagine a lot of my, my page views and sessions have moved into an LLM, um, to get, to, uh, to help me answer certain questions, do deep [00:49:00] research. Uh, there's, there's gotta be countless people doing really crazy in-depth research, you know, graduate-level research on what the best air fryer to buy is, which wouldn't have existed previously.

[00:49:15] Matthew: It gives you a 10-page report. Have

[00:49:16] Dara: you got access to my chat history? Yeah. No, I, I, I think you're right. I'm, I'm g- I'm determined to avoid your question about behavior 'cause I'm gonna again go back to the cloud, one more time, going back to Cloudflare, I think, 'cause I think they've said they're gonna start, um, charging for some.

[00:49:32] Dara: So they, they want bots or AI agents, whatever, to, um, to self-identify what type of activity they're doing. So whether they are scraping or they are, um, doing something else. I can't remember. There was three categories. I've forgotten the other two now. And I think if it's scraping, they're gonna be, they're gonna be charged.

[00:49:55] Dara: So I think that's coming in in September.

[00:49:58] Matthew: Which makes sense because it- It [00:50:00] does ... there's always been policies against scraping and things like that in the past where you-- people don't love you going and scraping their content, but now, you know, there's just these beings that can go and just-

[00:50:11] Dara: That can go and do it.

[00:50:12] Dara: Yeah, and I feel it's one, it's one of those things where it's like you want it to work in your favor but not against you. So it's like, but I want to be able to do it when I choose to do it, but I do understand why they're clamping down on other people doing

[00:50:23] Matthew: it. I'm doing it for good reasons; everyone else is doing it for bad reasons.

[00:50:27] Matthew: And I'm pretty sure, I'm pretty sure Cloudflare released one of the most powerful site scraping tools at the same time.

[00:50:35] Dara: One rule, it's like do as I say, not as I do. Yeah. Um, I'll answer your que- I'll answer your question now. I think similarly to what you said around behavior, yeah, I think, I think for things like, um, aggregating, you know, if you're looking for something, um, and you wanna aggregate information, whether that's product pricing or whether it's just information you wanna pull up from multiple different sources, there's no way you're gonna go and physically visit [00:51:00] five, six, seven different websites and go through everything on there.

[00:51:03] Dara: You know, go through page after page, which is what you had to do previously. Um, whereas now that kind of thing is gone. So yeah, any- anything that's research-based, you know, best air fryer, whatever, anything like that, yeah, you're gonna-- there's no way you're gonna be going on websites. Um, even like information that's maybe not...

[00:51:24] Dara: You know, I obviously still use Google Maps, and you'd still look up, uh, like anything real time. Uh, but even that, it's getting, you know, it's getting better, like, because it is using like web retrieval. It is, you know, it's getting better to even get real-time information, but I just don't fully trust it on things like that yet.

[00:51:43] Dara: So timetables, anything like that, I'd probably still go to a website for. Um, and I'm not, I don't know about you, but I'm not purchasing anything. Maybe we can't even. Is that still US only, or can you purchase things now? 'Cause I, I... If you can, I'm not doing that yet. I would still go on. Even, [00:52:00] even if I'm trying to find a product, I might do my research through Claude, but then I would ultimately I would end up on a product page, and I would buy something on a website.

[00:52:09] Matthew: I think OpenAI, uh, shelved it.

[00:52:13] Dara: Oh.

[00:52:14] Matthew: Yeah. I, I s- bit of salt, bit more salt, salty news back in the end, but I'm pretty sure OpenAI shelved the purchase journey stuff.

[00:52:23] Dara: Right.

[00:52:25] Matthew: They had a bit of a cull a little while ago because they, I think they realized they were getting a bit scattergun, and, and Anthropic was starting to eat their dinner.

[00:52:34] Matthew: Yeah, so I think they've nipped and tucked a bit, and that's one of the things that ended up going. Might be wrong, but- But

[00:52:39] Dara: there was a protocol, though, and I don't know if that was theirs or Google's. Maybe they, they, maybe they both had one, Open Commerce Protocol or something.

[00:52:47] Matthew: Google's still doing it, and it's coupled with, like, their new AI search engine, and you'll be able to add to baskets and buy things.

[00:52:53] Matthew: So Goo- do Google you can still do it with, yeah. And I think that's still evolving and growing.

[00:52:58] Dara: Is that US only, do you know?

[00:52:59] Matthew: [00:53:00] Probably at the minute, I would guess. Yeah. I think they're ri- rolling out the new Google in US, 'cause that's pretty much just AI, isn't it? It's not the Euro traditional thing, and then they're gonna roll it out over here.

[00:53:10] Dara: Because for stuff like that, if it was like, you know, the, I don't know if they're still a thing, but you know you could get those Amazon buttons that you put on your washing machine and we press it to order- Yeah, yeah ... new detergent.

[00:53:20] Matthew: They're not anymore, but I used to like them.

[00:53:21] Dara: Yeah, I'm not surprised. It's another thing like the, uh, like the, the mini disc, you know.

[00:53:26] Dara: It's a kind of interim, interim technology. Care- like a caretaker manager, you know? Um, but if there was something like, I think the last thing to go for me would be just buying something from some random website. But if I could, if I could say to Claude, "Hey, I need, you know, new light bulbs for the kitchen," and it would know what light bulbs I use in the kitchen, or, "Oh, I need new, yeah, washing detergent," it knows what one I buy and what size I get and whatever, um, that I would be up for.

[00:53:55] Dara: But, like, go and buy me, I don't know that I'd say, you know, "Go buy me a [00:54:00] pair of shoes," you know, randomly. Maybe I would. It might be a bit of jeopardy is nice in life. Just say, "Buy me a pair of shoes. They need to, the left one needs to be the same as the right one, otherwise I don't, you know, I don't mind."

[00:54:14] Matthew: Yeah, that'd be a nice way to live your life, I think. Bit more variety. But yeah, I think, I think shopping is still not there. I agree. I, I'm trying to think of categories of internet use now in my head. So, like, general day-to-day just surfing of the internet for information, that's, that's almost evaporated for me.

[00:54:33] Matthew: I tend to get it via an, via an LLM or via a dedicated application on my phone, so like BBC News or, um, Reddit or something like that. Just something that I open and I'm in that thing. Um, but I don't visit Reddit or BBC News or any of those kind of things on my desktop. I don't, I don't do that. The only time I open up articles really is if it's been surfaced to me in an LLM and I go and give it [00:55:00] a scan, or if like for this podcast and stuff, I've got them open just 'cause I want a p- a quick and easy way to reference something.

[00:55:05] Matthew: But apart from that, nothing. Um, shopping for sure. I'd like, I don't do a massive amount of online shopping. I tend to just use Amazon, to be fair. But, um, I think that still would be the same. I, I wouldn't necessarily- complete that journey yet. Maybe, maybe when Google starts to come out with all their stuff and, 'cause they've got like add to your wishlist, and it can monitor pricing, and it can trigger a purchase, and like a universal wallet and universal shopping cart, and all this stuff on the horizon, right?

[00:55:36] Matthew: So I can imagine that, um, being useful if they get it right. But still, at the minute, that still fills in that, in the domain.

[00:55:43] Dara: Yeah. Yeah.

[00:55:45] Matthew: Um, but the, but the journey up to purchase, I think, uh, uh, has changed. Like I wouldn't necessarily... Maybe previously I would've just Googled a product, clicked through a few sites, maybe asked for, looked for some reviews, [00:56:00] where now it might be like go and I'm after this kind of thing, go and find reviews and put together a bit of a report that will then, um, trigger my purchase.

[00:56:09] Matthew: So that's changed. Like for this electronics project, for example, I sort of chatted with Claude about what I needed and what I would, how, what would, what components we would need, and then it built me a shopping list of all the different pieces and how they would f-hook together and providing me some links with where I could go and purchase it, and then I went and did it.

[00:56:28] Matthew: So that's kind of my most recent shopping experience, I guess. And then the other one, the other one that s- springs to mind is, is application. It, it's like UI interaction with, with thinking specifically of work UI. So let's say Google Cloud Platform, GA4, Google Tag Manager, um, whatever it may be, the MCP has done a pretty big job of starting to reduce down the amount of time that I [00:57:00] interact with any UI in that way.

[00:57:01] Matthew: I mean, with, with Google Cloud, it was already quite low anyway, because I used to use the CLI, the Google Cloud CLI, gcloud CLI, and interact with, with G- Google Cloud that way. But pretty much for everything else, where possible, I interact and retrieve data via an LLM.

[00:57:17] Dara: Yeah, same. I think, I think, um, when you then do have to go back into the UI, it's painful because firstly, because you're rusty, but secondly, because it is more difficult to get to do what you want because you, you're clicking around rather than just running commands and, um, so yeah, the kind of interface.

[00:57:34] Dara: It's hard, isn't it? I'm trying to think where-- It's a bit like whenever we think about this stuff, we always end up, you t- you do end up taking it to the conclusion of all of this will be just taken care of eventually. Because if you think about those things you mentioned, um, well, actually two things came to, came to my mind when I was listening to you.

[00:57:53] Dara: So one is that like, f- um, the, the, the phone environment, the kind of, um, [00:58:00] mobile environment is safer at the moment because you do... I do, my behavior on my phone has probably changed less, not a lot less because I'm using Claude for a lot, but I'm still using other apps. Certain other apps I'm using as much as I ever used, but my web behavior has probably changed a lot more.

[00:58:19] Dara: Um, so does that just mean that the techno- i- is it just a simple case that the technology's not there yet? Because if you had integrations with-- Is everything... And, and, and maybe jumping ahead slightly, but with things like WebMCP, if enough companies get on board with that and just expose all of their data, all their information, all their content to agents I think personally I would happily just use one app for everything if I could.

[00:58:48] Dara: So if I didn't need to go on Reddit, if I didn't need to go on Garmin, if I didn't need to go on Amazon, if I didn't need to go on... Well, I already don't need to go on GA4 or BigQuery or whatever. Um, but you know, if you [00:59:00] didn't need to go to any website, if you didn't need to go to BBC, if you didn't need to go on, um, Trainline or whatever, then that's gotta be better from a, just purely from a usability point of view, 'cause we've got better things to do, don't we, than go and click around websites.

[00:59:17] Dara: So part of me thinks bring it on, open it all up, make it, um, s- you know, have a standard that people adhere to, and make the internet a thing for, uh, well, maybe the internet in that scenario wouldn't exist at all, but there'd just be connections

[00:59:34] Dara: somewhere.

[00:59:34] Dara: Yeah, just a database, yeah. Like, what do we need?

[00:59:37] Dara: Why d- why do we need... Especially when there's not, uh, uh... Trying to think if there are places where you go where you're getting genuine, like, human, like human creativity, and you're going on to see that on, on, on the web. But I think most of my use of the web is either information or, or purchasing or doing some kind of [01:00:00] transaction.

[01:00:00] Dara: So not necessarily of just a financial transaction, but some kind of I'm going to get a thing or I'm going to change a thing. And I don't know, it's not like it's a, it's not like it's a, it's a, it's a place you go for cultural enrichment or, you know, to, for kind of creative inspiration. At least I don't find it that way, 'cause- No, we don't.

[01:00:19] Dara: No. No.

[01:00:21] Matthew: There are generations where they're probably... This is getting, this is gonna make me a, a very old man now, but there's generations where their culture is very much tied up in-

[01:00:30] Dara: TikTok or whatever.

[01:00:32] Matthew: Yeah, like, but, but again, they're more application type things rather than anything else. It's just short form content and I don't know.

[01:00:40] Matthew: But yeah, I, I agree. I don't, that's not where I-- It's not really where I I don't, there's no real recreation tied up in, in the internet for me.

[01:00:50] Dara: Same. Yeah, yeah. I, uh, okay, indirectly, so you might watch stuff on Netflix, but then that's still an app rather than-- It's not re- it's not really the [01:01:00] web as such. Um, so then it make, it, it makes me think then, because I think part of the problem, you know, the whole dead internet thing, part of the problem is it's trying to serve two very different audiences at the same time.

[01:01:14] Dara: It's trying to serve humans, and it's trying to serve agents or bots, and you can't. Maybe, maybe there isn't a way to serve them both equally. So is the answer go, you know, full nuclear mode? Is it just, yeah, make the, make the internet a database and just have connections and... Ah, you agree. Another thumbs up there from you.

[01:01:35] Dara: Another emoji thumbs up. You're gonna, you're gonna have to start wearing mitts with no, like, your fingerless mitts or something to, um... Yeah, I, I just wonder whether, you know, it w- whether we'd all be better off if the effort went into m- m- you know, m- exposing all the data and all the content in an agent-ready way, stopped trying to hedge the bets and make things human-friendly as well, 'cause it's [01:02:00] just gonna be the whole SEO thing all over again.

[01:02:02] Dara: You're gonna be going on websites. It's gonna be the equivalent of, like, keyword stuffing. You're gonna be going on trying to read something and just be like, "This is not written for me. This is written for an agent."

[01:02:11] Matthew: Yeah. There's a random list halfway down that breaks the content down. I mean, it's already happening, right?

[01:02:17] Dara: Yeah. It's already happening. Yeah. So maybe, like, if we're not the, with, with... When that happened before with SEO, it was frustrating 'cause there was no alternative. As a human, you were frustrated because you were reading this, like, content that was clearly written to game Google and, and get that site up the listings.

[01:02:35] Dara: Whereas now there is a plausible alternative, which is that you just don't consume content in that way anymore.

[01:02:41] Matthew: Yeah. Actually, that's, uh, absolutely. I mean, I mean, there's, there's things, I guess it's like the outside world to a certain extent, where there's so much existing infrastructure. Like with robotics, so the robotics has to f- you're trying to make humanoid robots so that they can exist and [01:03:00] walk around and sort of fit into a world that has pr- primarily been designed and created for humans.

[01:03:06] Matthew: In a way, it's the same for LLMs on the internet because the internet has been designed to be interfaced with, in a lot of cases, human beings. They sort of look at something and have, the UI layout is based around, like, how they can click around and how they can absorb information. Whereas, yeah, like you say, that's just not necessarily...

[01:03:26] Matthew: In an ideal world, they, they would just go and pass through some vector database of information and retrieve X and pass the information back. Um, there's, there's, I mean, Claude Cowork just released to that end, just released, um- An ability to s- to record yourself moving around a site to create a skill.

[01:03:50] Matthew: So I've seen some examples where somebody, somebody wants to do bike trails, and there's a particular site they use, but they, there's no API, there's no MCP, they can't access it in that way. So it [01:04:00] took it through, like going in, working the filters, looking at the, the right trails, checking the stars, downloading them, putting them in a folder, um, and it kind of created a specific skill for it to interact with that human, that human interface.

[01:04:16] Matthew: And that's the thing, like we joked about Rabbit R1 before. That was the big promise of that Rabbit R1, was it was going to be an AI based on top of human interfaces, so it didn't have to b- it wasn't based on APIs. It wouldn't break when a, an API changed. It would just be able to interact and use a, a, a site and retrieve information.

[01:04:37] Matthew: It was a lot of bollocks. It didn't, didn't happen. As it

[01:04:39] Dara: turned out, yeah. But would it... So with this, with this new Claude, what are they, what, has it got a name? Just like record a skill or something? Was the-

[01:04:47] Matthew: I think it's literally- '

[01:04:49] Dara: Cause it is just a skill, so it's just a subfeature of skills, I guess, or a, a way of...

[01:04:54] Dara: I don't know if there's a name for the, the overall thing of recording. But sorry, anyway, what I was gonna ask [01:05:00] was, you said about it doesn't matter if an API breaks, but surely if the UI of the website changes, then the skill's not gonna work anymore. Or is it smart enough- Yes ... to like, could it look for an equivalent?

[01:05:11] Dara: So say one of the steps in the skill is click on the export button, and they change the export button to say download instead. Would it, adapt to that?

[01:05:21] Matthew: Well, that was, again, again, that's, that's kind of what the promise of the R1 was, or the model that they pretended they had, is that because, because these things can reason, then they can reason through changes.

[01:05:34] Matthew: And if they, if you train a model specifically to interact, the model is, like, really good at... It's what computer use is about as well, right? And the browser use to try and make them just be really good at figuring out and navigating through our, our current existing architecture. So you'd hope that it would be able to reason.

[01:05:52] Matthew: I, I mean, I don't know, I don't know to speak to the specific feature that they've released, if that's a baked-in feature, but it's not, it doesn't [01:06:00] seem out of the realms of possibility for it to adapt and change as, as you go.

[01:06:03] Dara: No, it doesn't. And, and how do you kn- do you know, I know it's new, and I don't, and I know you did say you haven't played around with it too much, but does it...

[01:06:11] Dara: How, how does it, how does it then go and act on that? Is it using the, um, the browser tool within, so is it opening up a tab within your browser or is it-

[01:06:21] Matthew: I don't know because- Yeah ... I, I assume so actually, because they have just released a browser for-

[01:06:28] Dara: In Code- ...

[01:06:29] Matthew: Code Desktop.

[01:06:30] Dara: Yeah.

[01:06:31] Matthew: Yeah. Um, but I don't, but it's something that's available in Cowork.

[01:06:34] Matthew: I think these, these skill builds are in Cowork as far as I'm aware. So I'd assume it's opening up your Chrome browser using the app that you're using there. '

[01:06:42] Dara: Cause one of the, like, what, one of the obvious use for it is something that's really repetitive and you need to do a lot of times, like where, as you said, where there isn't an API and you have to like, I think of like Garmin as a good example.

[01:06:52] Dara: So you can, Garmin only has a developer API, so you can't get your, like, your training activities out of it. But you can, you can do it through o- there's [01:07:00] other ways you can do it, but directly from Garmin, you can't get it. But you can go in and manually download the things that you want, but you're not gonna wanna have to do that like hundreds of times or whatever.

[01:07:09] Dara: But with these, like, I don't know if this is just the skill part or if they're actually improving the automated use of your browser as well, because if you try and do something You know, I've even done it with like the podcast where you're trying to get it to read all the titles before and all the excerpts and pull all the transcripts.

[01:07:27] Dara: I've tried to do that and sometimes it falls over because it's timing out or it's hitting an issue or there's some JavaScript rendering issue on the site or something. So it's still, it, it almost feels like it's not really new technology, is it? It's something that's been around for a long time and is still as clunky.

[01:07:44] Dara: Maybe it's slightly less clunky, but it still doesn't really feel like the best way of doing something like that.

[01:07:50] Matthew: No. And I think there's gonna be models that are specific. I mean, I, I think we said last week that there's a lot of mod- the models are confusing, but they're starting to sort of perhaps [01:08:00] you'd use this for computer use 'cause it's been trained and, and sort of, um, honed a little bit more on computer use and browser use and this one for agentic work 'cause it's, that's what it's really been post-trained on.

[01:08:11] Matthew: So I do wonder if, it does seem to be a focus for them to try and create models that can do this stuff and can interact and surf around quicker and more effectively. I just don't know if it's gonna just be a ge- a general... I don't know. It, I think ultimately it makes sense like, like we say, that more and more things move into, move into LLMs, especially if the, especially if the, the, the experience gets richer.

[01:08:38] Matthew: Like one, one criticism you could level at LLMs is it's very, I mean, certainly mine, they're all in dark mode. It's just like a black and white interface, but they've started to layer in things like dynamic images being created and, and obviously OpenAI in the US has like apps inside of... So you can imagine that that sort of dyn- dynamism sort [01:09:00] of beginning to appear in the LLMs and making the experience nicer and a bit more something you'd want to, to interact with

[01:09:07] Matthew: And on the, on the mobile, I, I know we, this was a little while back we were talking about this on, uh, on this podcast, but in terms of sentences. But on the mobile front, one, one point on that, I, so I've got an iPhone and it just recently, I've updated it to the latest version of iPhone with the new Siri, which is the one they've worked with Google on.

[01:09:26] Matthew: Um, and what it does, interestingly what it does, it sort of indexes all of the local information it has on, on local apps like Calendar, Gmail, Photos, Notes, pretty much all of this information that you, that you might have on your phone, and it can now access and retrieve that information and answer questions.

[01:09:46] Matthew: So previously it was very, Siri was pretty crap, notoriously so. But it is significantly better, and I have found myself a couple of times just my phone's been sat down and I've just said, [01:10:00] "Siri, w- what's on my calendar today?" Or can you find... One of the interesting ones, like I was needing to find like some information about my mortgage, and I just said like, "What, what's, what was my mortgage account number?"

[01:10:11] Matthew: And it'd been like two years since I'd last received an email or something from it, and it had indexed my Gmail. It went back, found the information, returned it to me, showed me the email. So I can see that they're starting on mobiles to sort of get an overview of everything that's on your phone and, and apparently Apple are opening that up to app developers so other app developers can re- surface this information to Siri as well, so then it'd be able to look inside of all these applications.

[01:10:39] Matthew: So I can imagine myself, my app usage dropping a little bit on phone even because of the, that kind of technology, and Gemini is similar on, on- Google, I assume. The only difference with Gemini, as far as I'm aware, is it's all cloud-based, whereas Apple's pushing much more at local, um, local level stuff.

[01:10:57] Dara: Yeah.

[01:10:58] Dara: Yeah. No, it's de- it's definitely, I, [01:11:00] I agree, and I think like what, the point I made earlier about saying that like on phones maybe I would change less, that wasn't, um, because of any desire on my part. That's just because the, the tech isn't there yet. But if, I think if you could do that on phones, then I would reduce down the number of apps.

[01:11:14] Dara: So I mean, look, at the end of the day, like you don't want to be having to navigate different apps or websites if you don't have to. You're, you're dealing with the changes that they're making all the time. You're dealing with the differences between, oh, you know, like e- even things like multiple banking apps and you're trying to remember, "Oh, well hang on, on Monzo how do I do this?

[01:11:32] Dara: Because I'm usually using Revolut and it's different." You just-- All of that will go away, and if all that was just pulled into one place, you just have one interface to get really, really good at using, and then you don't really need to worry about any of the rest of it. So yeah, I'm, I'm, I'm all for that.

[01:11:47] Dara: Um, something to ta- kind of slightly separate, well, very related, but so regardless of which way, let's say the internet dies and it beco- and, and it, it's replaced by just this kind of database of, of product feeds [01:12:00] and content feeds or whatever, and people just use LLMs to connect to that. Regardless of what happens to the web as we know it, there's a, another, there's an underlying question which is what happens to the quality of the content if the training starts increasingly using AI?

[01:12:19] Dara: You know, it's the whole snake eating its tail type thing. If, if the training models have already used everything that's on the internet, then is that gonna degrade further and furth- Is it on a declining slope where the quality of the models is gonna get worse and worse because it's gonna be ingesting more and more AI-generated content?

[01:12:41] Matthew: Yes. Well, I don't know. The, I assume I-

[01:12:44] Dara: This is gonna answer to a long question.

[01:12:46] Matthew: There's, there's something, this is, I, I, I saw a video ages ago by a, this is, this is YouTube channel called, uh, f- forget, this is a German word, um, Kurz, Kurzgesagt, [01:13:00] Kurzgesagt, which is German for in a nutshell. But they do these little scientific videos that describe like a part- a particular subject, but they go in depth on it and they really do a lot of research around it.

[01:13:11] Matthew: And they released a video on a, about AI, and it talked about the fact that they, when they originally saw like this stuff come out and the deep re- research and stuff, they were like, "This is gonna save us." so much time because they spend hours and hours and hours and multiple researchers delving in to find the information and put together the, the, the knowledge base that they then build the scripts on top of, and they could just push a, a deep research agent to go and find it.

[01:13:38] Matthew: So they did it, and then, um, they found that it w- it produced a lot of data, great, and then they started looking into s- to sort of cross re- cross-reference what it pr- it produced back. Got rid of a load 'cause it was AI. It was, it was either some of it was hallucinated and didn't exist. This was a while ago.

[01:13:54] Matthew: Some of it was clearly like an AI article, and then they had, I don't know, I think they said about 60% that [01:14:00] was left behind that was usable.

[01:14:02] Dara: Yeah.

[01:14:03] Matthew: Then they went further and started to dig into the sources of the ones that were left behind, and ultimately they found that if they w- kept going down the rabbit hole the source, when they got to the bottom of the source, it was from an, um, the, the, a, a place they didn't understand or where the information had come from, completely unsourced information, which they assumed, and probably rightly assumed was, was AI.

[01:14:28] Matthew: So even, there's this, this is, there's this phenomenon happening where you've got some piece of information out there, someone referencing it, someone referencing it into a stack, so by the end up, the, the, the top of that stack looks reputable because it's got loads of references of all these other sources, but if you dig down d- deep enough, it's a load of crap.

[01:14:46] Matthew: And they-- So they basically had to abandon, they had to abandon the whole, the whole idea because they just, it took them more time to delve in and actually prove out that this information was true, uh, than it was just to do the research [01:15:00] themselves. Um, and nobody's that rigorous.

[01:15:04] Dara: No.

[01:15:04] Matthew: Nobody's that rigorous in their, it, like, cri- talk, talk about critical thinking.

[01:15:08] Matthew: Like, I have stuff to do in a day . I, I have to use my judgment.

[01:15:12] Dara: Yeah, yeah.

[01:15:13] Matthew: Yeah, but I can't like, like, "Right, well, let's keep going." "What you doing Ash?" "Just to check what the price of that, that, uh, application was." I am. I'm on my 17th.

[01:15:22] Dara: Doing it, yeah.

[01:15:24] Matthew: Source.

[01:15:25] Dara: Is there a, is there a positive view or is that just an inevitability?

[01:15:29] Dara: 'Cause obviously for OpenAI and Anthropic and Google, I mean, Google have built their entire reputation on, well, at least trying to, um, deal with quality and make sure quality content is, is pro- is, is put in front of people.

[01:15:43] Dara: So it's, it's not like Anthropic and OpenAI and Google are just gonna walk into this blind. They're obviously aware of this potential risk, but I just wonder how, like, it's, it's a big, it's a big tidal wave to kind of fight against, isn't it? [01:16:00] Because it's, because the, because of the exponential power of AI, like, the ability for people to churn out content and publish it now, and have source material that looks genuine, and it's gonna become a, like, quite an epic battle to try and differentiate between something that's been...

[01:16:18] Dara: Well, it's, it's not like, and we should be cl- clear, that something AI-generated doesn't necessarily mean it's false or it's not, you know, it's not reliable, but, uh, maybe it's a subcategory of AI-generated content which is dubious or hasn't been, you know, sourced properly or whatever. Um, but it's, yeah, it's har- it's hard.

[01:16:36] Dara: I guess as mere mortals, maybe we're just not close enough to what the big companies are doing around this, but how are they ensuring that they're- training data. 'Cause presumably every time they, they run a, a brand new, you know, when they're training a brand new model, they must be grabbing anything that ex- that exists now that wasn't there the prior time.

[01:16:57] Dara: So what kind of filtering are they doing to make [01:17:00] sure that they're not just feeding in more and more garbage?

[01:17:03] Matthew: I think, I mean, I think in a lot of cases now they use syn- they use synthetic data as well because they've pretty much consumed-

[01:17:09] Dara: All of the actual... Yeah, yeah.

[01:17:11] Matthew: Yeah. So, but they have to be doing something, you're right.

[01:17:14] Matthew: And, and to be fair, like, the hallucinations and the amount of misinformation I get fed has reduced dramatically over the past sort of year or so.

[01:17:25] Dara: Or so you think.

[01:17:26] Matthew: Or so I think, yeah. As far as my critical thinking can go.

[01:17:30] Dara: Yeah, it's just getting better at convincing you.

[01:17:34] Matthew: Yeah, but it's always been the c- I mean, yeah, it's always been the case, but like with, with anything with AI, it just, it just highlights and exacerbates and speeds up the problem.

[01:17:43] Matthew: Like, there's always been misinformation, for example, like even if, if you think around Brexit, there was all that Russian interference and misinformation there, that was pre, that was pre- a- and the first Trump, um, presidency. There was all that going on, but that preexisted these LLMs being in, in the [01:18:00] world.

[01:18:00] Matthew: But now, and, and, and we're talking, you know, there's, there's bad ar- there's articles out there that aren't quite true or this, that, and the other, but there's a whole other side of social media, um, being filled with echo chambers stuffed to the brim with complete nonsense articles that people believe and, and ultimately has societal impacts outside of the internet where people just are at odds with each other and polarized and, and fighting and, and it-- and that's just sort of feels like it's being turned up to 11 with, with, with AI.

[01:18:30] Matthew: And then that-- but let's not even get s- well, maybe let's get started. But like, this before we even start talking about video, uh, a- and how far that's coming. Like we, we used to talk quite a lot about video. We haven't for a while, but I've been watching some stuff recently that people have put together movie trailers and things like that, that it's just, I can't tell if it's real or not.

[01:18:49] Dara: Yeah.

[01:18:50] Dara: There's another way to, there's another way to think about it, and it's, I guess like with everything, like you're right, that existed before and, and it's easy to get caught up in [01:19:00] how things are gonna be worse now because there's some new thing and it's gonna make everything worse.

[01:19:04] Dara: But actually, the problems aren't new problems, they're just maybe amplified in certain ways. But there could be a flip side to it as well. So, um, you know, this idea that like, a bit, a bit like, so when like when, when Google, you know, when, when, when search engines became a thing and then Google became the dominant one, um, presenting all the world's information in the way that they did, this, th- it's a double-edged sword.

[01:19:32] Dara: On the one hand, you can access anything, you can search for anything, but maybe on the drawback, people maybe argue then that it impacted critical thinking because you didn't know what you were l- you know, you, you have to know what you're looking for, and maybe instead of in the previously where people would've gone and researched a topic deeply, you just kind of g- created this more scattergun approach to, you know, picking bits of information here and there.

[01:19:56] Dara: Um, and this is probably no different again now, it's just an evolution of [01:20:00] that. But the, the, the positive side, and which was the same with, with Google and other search engines, is, is putting information that people previously didn't have access to in front of them. So even things like, um, academic papers If you're trying to check something, let's say you're like, "I'm wondering whether I should take a multivitamin or not," you could go-- you can go to your AI and say, "Tell me what the current scientific literature says.

[01:20:26] Dara: Check the quality of the studies. Check they're not run by commercial, you know, they weren't funded by commercial companies selling this product." As a, as a layperson, like, you, you know, back to your point earlier when you said, you know, you've got, you've got things to do, like, nobody has time or the ability to go through and read 20 scientific papers, but you can do that now with Claude.

[01:20:49] Dara: So there's pluses and minuses. On, you know, on the one hand, maybe there's, there is gonna be an increase in the bad kind of A- AI-generated content, or it's gonna be [01:21:00] harder to separate truth from, um, from fiction, whatever. But there's also an increase in your ability to, um, to, to, to critically think, to check things more thoroughly, to fact-check if you're willing to do it.

[01:21:14] Dara: So I don't know. I- it's hard to know who, like, what will, like, which will weigh heavier on the scales. It's hard to know.

[01:21:22] Matthew: If you think about it, it's a nat- it is a natural evolution of, like you say, so, so if we go really far back, you've got the only way of disseminating and communicating information was the spoken word, like just being able to sort of, you know, oral, oral histories and oral, oral, um, things being passed down.

[01:21:41] Matthew: Then you get the written word, but even then, there's very few people can actually read and consume it. It's very difficult to reproduce the, the, the works, and mistakes are being introduced, and then ultimately education starts to spread that out further. But you've still gotta go and you gotta consume it, you gotta go read a book or whatever.

[01:21:58] Matthew: Like you say, Google comes along, [01:22:00] starts to sort of catalog everything. People can find that information much easier. And to the, and to this point now where you have gradu-you have graduate-level researchers at your, at your fingertips to look at whatever. Uh, we joked before, but if you wanted a 20-page report on air fryers, you could have it, which is a ridiculous thing to even think about, um, historically.

[01:22:21] Matthew: And you can imagine it moving further to, to a point where you just, the information is just there. You just have it. It's, you know, all these neural link stuff that's beginning to appear, it just-- when you want a piece of information, it is-

[01:22:33] Dara: Yeah ...

[01:22:33] Matthew: in your brain and available to you. That's the kind of natural point.

[01:22:36] Matthew: It's just reducing down the latency at which you can consume-

[01:22:39] Dara: Yeah ...

[01:22:40] Matthew: information. Like, I mean, I, I suppose the point, the point and problem of this, this part, this particular part of the conversation is how you make that clean, how you make that not be, not be rubbish. And you can imagine, you can imagine like in a, y- we'd almost trust in, we have to trust in corporations.

[01:22:58] Matthew: I- if Google [01:23:00] and OpenAI and Anthropic put in measures to make sure that what comes out of it is moral and aligned and things like that, then you can trust the information you're getting out of it. You've got people like Grok, where it's, he's intentionally tweaking and changing things he doesn't like to fit his narrative.

[01:23:18] Matthew: That's very dangerous when it's just producing information and being able to put stuff out

[01:23:22] Dara: there. Yeah, I don't, I don't, I don't know. It's, it's such a big, it's such a big topic, isn't it? I was like when, you went all the way back to when we could just communicate through speaking to each other, it's like the, it's not like the problem didn't exist then.

[01:23:32] Dara: I mean, that was, you know, like misinformation or even just unintentional, like passing through multiple different people and the story gets watered down or changed or tweaked. So it isn't really, you know, and then, you know, history books, you could like, they differ depending on which country you read them in.

[01:23:48] Matthew: Yeah, the victor or the,

[01:23:49] Dara: yeah. Yeah. Yeah. So, so it's not, it's, it's, yes, all these improvements in technology, they maybe amplify the, they amplify the risk, but they also do amplify the, the [01:24:00] opportunity. Um, and it's probably not, no matter what, it's probably not gonna go away. Like there isn't, um, you know, the, the, the, the Grok example, like it's, it's, it's no different to like this propaganda in a way, isn't it?

[01:24:12] Dara: And it's, you know, if someone has a platform, they can use that for good or bad, and people have had platforms in the past, pre-AI, pre-internet. Um, it's, so it's a human, I guess, getting very philosophical here, but it's a human problem, isn't it? It's not a technology problem as such. Obviously, the technology has to account for this and try to filter and try to verify.

[01:24:35] Dara: Um, but the problem fundamentally is a human one, isn't it? Rather than a technological problem.

[01:24:41] Matthew: Yeah. And we've, we've always adapted as, we've always adapted societally in the past to new challenges. I guess the main difference, the worrying, the worrying thing here is the pace at which the change is occurring.

[01:24:56] Matthew: Can a, can society and governments and, [01:25:00] you know, people keep up if the pace of which things change is so rapid that it's hard to get your head around what the hell's just happened? You got, you gotta imagine it's hard for-- I always think about my grandma, um, before she died, she was like around when the plane was invented and when the iPhone was invented, and all of the pieces of technology that happened between that.

[01:25:24] Matthew: I can't imagine- this, she'd be able to keep up with this and understand that piece of information wasn't created by a human, it was created by a computer. And when you see a video telling a, any of you to send me 6,000 pounds, this could be my grandma or it could be you, of course, these are- That's true, yeah

[01:25:41] Matthew: successful scams. But yeah, I don't think I could sit down and explain that to her in a million years. I don't think she'd get it. Just wouldn't, it's too quick. It's changed too q- fast for her to grab that concept.

[01:25:51] Dara: Yeah. I think you're right.

[01:25:52] Dara: It's a tangent, but you made a little joke there about being scammed. That was from a real thing that I told you, which is that, uh, a hotel I- In confidence ... [01:26:00] Yeah, in confidence, so now I'll just talk about it on air. It's an even scarier, darker side. It's one thing if you read something and you're not really sure if it's true or not, and obviously that could be very harmful as well if it's manipulating people.

[01:26:11] Dara: But in terms of scamming and phishing, and hacking, this is the other side as well, isn't it? It's like the more capable the AI gets, the harder it's getting for security to be preserved and maintained. It's how legitimate you can make something look because everybody can spin up a website, you can spin up an app in no time.

[01:26:29] Dara: What they did was they hacked into this hotel's account, so they got all of the information about all of the bookings, so they were able to present me with a really legitimate-looking, um, web form that said, "We need you to confirm your details," which is not an unusual thing for hotels to do before you go and stay at them.

[01:26:47] Dara: And it had my booking dates, and it had a picture of the room, and it had everything, so it looked really legitimate. So it's not just about, information, is it? It's also about, your trust of actually booking things, buying things [01:27:00] online. Bye-bye to the internet. Is that what we're-

[01:27:03] Matthew: Yeah

[01:27:03] Dara: the internet won't exist. Yeah. Been, it's been fun. Okay, let's leave it there. That's it for this week's episode of "The Measure Pod." We hope you enjoyed it and picked up something useful along the way. If you haven't already, make sure to subscribe on whatever platform you're listening on so you don't miss future episodes.

[01:27:21] Dara: And if you're enjoying the show, we'd really appreciate it if you left us a quick review. It really helps more people discover the pod and keeps us motivated to bring back more. So thanks for listening, and we'll catch you next time.