Tech Overflow
We're Tech Overflow, the podcast that explains tech to curious minds. Hosted by Hannah Clayton-Langton and Hugh Williams.
Tech Overflow
This Month in Tech
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Software is starting to feel… copyable. We open season three with a brand new “This Month in Tech” format and jump straight into the stories that made us stop, reread, and ask what changes next when AI agents move from chat to action.
First up, we break down Artcraft’s claim that it used Claude agents to create working clones of seven Adobe style apps. We talk about what is genuinely new here: agents inspecting an existing product, documenting how it behaves, then rebuilding a lookalike from scratch, without access to the original source code. We also keep it grounded, because a convincing interface is not the same as true feature parity, and “AI generated software” raises messy questions about IP, patents, trademarks, and whether downloading fast built open source code is a security gamble.
Then we shift into enterprise AI economics with Jev, a “judgement model” designed for tiny decisions at huge scale, like sorting inbound customer emails in milliseconds at a fraction of the cost of frontier LLMs. From there, the tone turns serious with the OpenAI Medicare portal story in Australia: what happens when an internal agent without full guardrails goes looking for data, finds an ageing government system, and simply hacks its way to the answer.
We finish with MetaMuse overtaking ChatGPT on the US App Store and ask whether consumer design, personality, and “cute” avatars are what finally push AI agents mainstream, while Meta glasses and wake words bring privacy back to the centre.
Subscribe, share with a curious friend, and leave us a review so more people can find Tech Overflow.
Like, Subscribe, and Follow the Tech Overflow Podcast by visiting this link: https://linktr.ee/Techoverflowpodcast
Season Three And A New Format
SPEAKER_01Hello world and welcome back to season three of the Tech Overflow podcast. I'm Hannah Clayton Knighton.
SPEAKER_00And I'm Hugh Williams, and we are the podcast that explains technology to curious minds. Season three, Hannah, we're here. Finally.
SPEAKER_01We are here. Where are you? Because I know you've been playing musical countries for a little while. Whereabouts are you coming from today?
SPEAKER_00I'm in Australia, live from my house, uh, but on Saturday, heading back to Los Angeles. Have you been having Waymo withdrawal? Yeah. In fact, I was talking to a friend today about Waymo's, and he's gonna be in LA, and I promised I'd take him for a ride in a Waymo. So looking forward to it.
SPEAKER_01They still haven't launched in London. I'm waiting. Nick, if you're listening, get on it. Okay, so today, guys, we've got a brand new episode format which we are super excited about called This Month in Tech. And it kind of is what it says on the tin, right?
SPEAKER_00Yeah, absolutely. Except I think this time we'll do four months in tech since our last season ended in May. That's fair. I thought it might be fun to pick a few stories that have caught the imagination between May and this season.
SPEAKER_01Yeah, and it's kind of because we realize, guys, like things are moving fast, faster than when we started season one, particularly, you know, with the advent of AI. It's hard to keep up. So our intent here is to pick the biggest stories, break them down, tell you what matters, and give you a few of our thoughts on them. All right, so let's get into it. Today we are going to tell you all about the agent who's overtaken Chat GPT on the US app store. Uh, we'll explain what went on in the Medicare OpenAI breach, and we'll go over yet another controversy with Meta's glasses. Shall we get right into it here?
SPEAKER_00Let's do it.
SPEAKER_01So,
AI Agents Clone Adobe Style Apps
SPEAKER_01to kick us off, a big question I've seen bouncing around the internet at the moment is have we seen the end of software? And this is a discussion that seems to have been sparked by a US developer group, Artcraft, which has basically used Claude to create and release free working clones of the seven Adobe apps. So, like Photoshop is probably the one that people know about, but there's quite a few other ones for creators out there which um have basically been copied by AI.
SPEAKER_00Yeah, that's right. So, what Artcraft did is they took about eight Clawed Opus 5.5 agents and they ran them in parallel. So it just means ran them at the same time, and they built copies of these seven Adobe products, which is incredible. And they've got them, they've got them running. They've released the source code, you can download it from the internet. If you do download them, you don't have to create an account, it doesn't send data anywhere. And basically, these are really, really nice knockoffs, if you like, of the key Adobe apps.
SPEAKER_01Okay, and so as to the non-engineer in the room, like how big a deal is that? Like they've done that presumably pretty quickly. I imagine that Adobe has been curating these products over like a number of years. I mean, they've been around forever. So, like technically, this is this is a pretty big deal, right?
SPEAKER_00I think yes and no, Hannah. I think there's a few different ways you can you can look at this. I mean, copying software really isn't a new idea. OpenOffice was around, you know, back in the 1990s, and that's a clone of the Microsoft Office suite. So, you know, similar software, open and free. It's been around forever. So there's a long history of folks building software suites that are open and free, that are, you know, alternatives to the paid suites. There's also a history of doing things like reverse engineering software to build clones. So, you know, I'm a child of uh the 1980s and 90s. I used to have an IBM PC clone. So I used to have a computer that was a copy of the famous IBM PC, and right inside that was a deep piece of software called a BIOS. And those BIOS is these copies were kind of reverse-engineered versions of the IBM version. So this has been going on for forever. I think what's really new though, here is taking a bunch of agents, uh, pointing those agents at some existing software and saying, document how this software works, taking those documents without any knowledge of the source code that Adobe had, taking those documents and using those documents to completely build a clone of the software. So every menu, every menu option, and every feature that these pieces of software have has been cloned by agents that literally built the software. So I think that's what's interesting, and the fact that it was done quickly by AI is uh is a pretty breakthrough thing.
SPEAKER_01Okay, so they've used AI to build an outside in copy, essentially. But like, is this stuff actually working properly?
SPEAKER_00Well, sort of yes, and sort of no, a bit from column A and a bit from column B, Hannah. So Photocraft, which is the product that these folks have built that's got the most publicity, which is a clone of Photoshop. Look, it's got all 625 Photoshop menu items wired up. So that's great. You can kind of read a file, do something to it, and write the file, and it'll look pretty similar to the Photoshop file, but the renders it actually does are far from pixel perfect. So, you know, it kind of doesn't work so well. And the founder of the company, um, he admits today that the true parity is, you know, what he says, well below 50%. And tech radar, their verdict was there's a heck of a lot of overpromising and a huge amount of under-delivering here.
SPEAKER_01Well, hang on, that's not the first time a tech company has done that in its lifetime. And that's not a crazy move, really, is it?
SPEAKER_00No, no, definitely not. Definitely not. I'd say uh, yeah, certainly my industry is a lot about promising uh Mars and delivering the moon. And you know, Atcraft is still continuing to develop these apps, still continually improving. But I think what was the seismic shock was AI copying software. I'm not sure that, you know, the reality of the software being copied well is actually true yet.
SPEAKER_01Okay, and just for a second on this guy, um, Brandon Thomas, right? That's the name of the CEO. He was basically, from what I understood, like a disgruntled Adobe user who was like fed up of complicated cancellation fees and just decided to copy their entire software suite, which is like quite a big effort from him, really.
SPEAKER_00Yeah, absolutely. I mean, he must have been super disgruntled. I mean, this product, Creative Cloud Pro that Adobe sells, is about 70 US dollars a month. And he was very unhappy about the early cancellation fees. And so he decided to, you know, go and try and take the company down by cloning their software. And you know, he thinks this is just the beginning. He's actually cloned a whole bunch of Microsoft products since. So he's gone to the trouble of cloning all sorts of things: Word, Excel, PowerPoint, Pro Tools, audio software, CAD software. He's even cloned Figma, Hannah. So he's gone, he's gone crazy with his cloning. But yeah, it all began with uh him being very, very unhappy about the cancellation fees that Adobe charged him, and now he's turned this into an industry.
SPEAKER_01Okay, and let's talk about the elephant in the room, which is like the legals of this whole thing. So I did a little bit of research. Um, you know, I'm aware of IP and tech. I'm actually, you guys, begging you for an episode, but I think he's holding the line because it probably quite dry. But uh IP and tech is a big, a big conversation, right? But I think I've understood that because they didn't actually have access to the source code, like there isn't really an IP play here. And the piece that I read said that the the legal play from Adobe would be that this guy, Brandon, has breached the terms of his agreement with Adobe, which says something like, you can't seek to copy our software. I mean, I've never read the detail of those technical agreements that I have to sign up to download everything. So I it's very possible, you know, that's in all of them. So what do you think, Hugh? What's their play?
SPEAKER_00Well, look, you know, first of all, Adobe hasn't taken any legal action yet and they've declined to comment. So really, Adobe hasn't publicly said a word since this has happened. But look, I'm not a lawyer, but I can give you my kind of take on the sort of legal angles of this. I mean, first thing is, you know, copyright is really about copying expression, not ideas, right? And it dates back to, you know, books. You can't literally take the pages out of a book and copy those and publish those and sell those pages. But you certainly can read the book, be inspired by the book, set the book aside, and write down something that's inspired by that, that's not literally that, right? So that's where that's where copyright comes from. So it's really about protecting expression, it's not about protecting ideas. And so there's a pretty weak legal case here on copyright, because clearly what this company's done is they've they've looked at the software, looked at the expression, and then they've they've written that down, and then they've created their own source code that's independent of the source code of Adobe. So I think it'd be fairly difficult to run a copyright claim. And the copiers have won at least two famous US cases in the last sort of 40 years that have tested this legally. So I think it'd be very, very hard to run a copyright case here. But there's probably two other angles that we could talk about, Hannah. So one would be patents. So clearly, you know, Adobe has lots and lots of patents of things, right? So that's an you know a tool that allows you to airbrush a photo or whatever it is, right? So they've they've protected ideas that form part of the IP that the company has. And it's almost certainly true that a lot of the things in these products is are infringing the patents that are held by Adobe. So Adobe could probably defend this legally by using its patent portfolio. The other thing is maybe there's some trademark breaches here, right? So he's, you know, been fairly courageous in naming the product Photocraft. I mean, that is, you know, not too far off naming it Photoshop. And, you know, originally when you went and looked at the source code, it was it was basically saying, look, this is a public domain copy of Photoshop, lots of tags of Photoshop, mentions of Adobe, those kinds of things. But I notice in recent times that that's changed. So it's all been cleaned up, doesn't mention Adobe, doesn't mention Photoshop, just says this is a you know photo editing tool. But obviously the history here is that, you know, it was pretty boldly saying it was a copy of Photoshop. So maybe that's somewhat of an admission. It is a might be a little bit of a trademark issue here, Hannah.
SPEAKER_01Yeah, okay, that makes sense. And I've got two sort of broader questions that I think would apply to any other pirated software, um, just for the listeners. So one is like, why would someone do this? So there's presumably like hosting costs associated here, right? And like probably some, you know, maintenance commitment that they'd be making. And I think I saw that they're they're offering subscriptions. I think it's based on like AI tokens that people can use to generate images. But like, what's this guy's motivation?
SPEAKER_00I I think the motivation probably is to prove that AI can can create software by simply inspecting that software, documenting that software, and then rewriting that software. So I think it's that point. I think also, you know, he's making a point that he's pretty disgruntled with Adobe and uh he's gonna go and mess with those guys. So I think probably a dual point, Hannah.
SPEAKER_01But practically speaking, he is taking on some sort of commitment financially and otherwise, if he wants to continue going with this sort of copy program, fine. And then thinking of pirated, if we could call it that, software, like it makes me think about back in the day when I would download pirated music to my family computer in the study. And the the worry was always that you download a virus. And I'm sure I did. Like, is there a security issue here for someone who's looking to and maybe this I'm asking an even more basic question. Are people even downloading this or is it sort of hosted? And therefore some of those security issues don't apply.
SPEAKER_00Yeah, so that this isn't really hosted, Hannah. So this is a it's a source code repo, as we call it, repository on the internet. You can go there, you can download this source code onto your local laptop, your desktop computer. You can build the software and you can use the software. So this isn't hosted in the cloud. This is really similar to installing Microsoft Word or PowerPoint or something like that actually on your computer. So you're downloading something, you're installing it and you're using it locally. So it's not really costing him at all. Once you download it, um, you're just taking a copy, putting it on your computer and using it. But but your point around sort of, but is it secure? Who knows? Honestly. I mean, it was written by AI. I doubt that there's been a lot of energy put into sort of really testing this and understanding exactly what it does and who it communicates with and how it communicates. I'm sure that will all come out over time. But you know, you're really downloading something that was built pretty quickly, put up by this company. And, you know, it's kind of the Wild West Hannah.
SPEAKER_01One thing I I've learned about the engineering community is they do love to share like open source code with each other. So I guess it depends like how much inherent trust there is between them. But as you say, if it's built by AI, unless someone's been through the code line by line, it's anybody's guess what's in there.
SPEAKER_00Yeah, absolutely. And he's certainly not warranting that it's fit for purpose or anything like that. It's really just a almost a publicity stunt that kind of works.
SPEAKER_01And so is this proof, though, that like SaaS is going to become less valuable, software as a service, because we can just recreate using agents. Like SaaS Mageddon, I think it was called. When did that hit? Like some point last year, where uh the market's freaked about AI being able to copy. Like, is this more evidence that that's coming?
SPEAKER_00Yeah, well, look, you know, um, on the day that this was released, HSBC downgraded Adobe from a buy to a hold. So that's their statement on the on the stock, cut its target price from 310 bucks to 267 bucks, and cited Artcraft by name. So certainly, you know, investment folks who are out there feel that this is a seismic shock. And, you know, Adobe closed lower on the day, down about 32% on the year. So the markets are voting.
SPEAKER_01That's for Adobe specifically, right? I I guess my question is, and maybe we just don't know, but like, will this affect all SaaS companies? Because basically, if someone wants to come for them, they can just get on with it and and make a copy.
SPEAKER_00Look, the answer's very gray, Hannah. I mean, I think if you've got a piece of software like Workday or a piece of software like Salesforce or perhaps Zero, if you're using that accounting software, you know, those pieces of software are deeply integrated into companies, right? So if your company's using that software, it is very, very difficult to take that software out and replace it with another piece of software. And so it's not just about the software itself, it's about how it's plugged into everything in the enterprise. It's about the service contracts, it's about the customization, it's about being able to pick up the phone and call somebody when there's an issue, you know, all those kinds of things. So it's not just as simple as saying, you know, software is easy to copy, therefore SaaS is in trouble. But there certainly are pieces of software out there, and maybe Photoshop is a good example where it's possible to copy it and release a version that does pretty much everything that the paid version does. So I think there's vectors where this really is impactful, and there are places where it simply is. And I'm a bit interested in the games industry at the moment. I mean, I think that is definitely a place where this will cause a seismic shock. It'll be possible to go and copy a game and make your own version of it. And so I think that's an industry where you will really see this take hold. But certainly it doesn't mean that SaaS is dead.
SPEAKER_01Interesting. I think at some point we should do an episode on gaming generally, because there's a lot to unpick in that space. But yeah, that's such a good point. Like the software that's deeply embedded and holds, you know, employee data is probably not going to be the one to go after it, which is a shame because listeners will know that I'm not a fan of workday. So when you gave that example, I was hoping you'd say that uh you were gonna spend the weekend, you know, copying and improving the workday for that.
Jev And The Rise Of Judgement Models
SPEAKER_00So, Hannah, the second topic I want to talk about on the show today is a product called Jev. I don't know if you've heard about it, but I think it's one of the most important AI products of the quarter, something that's happened since we last recorded in May. And Jev's what I'd call a judgment model. You might ask what's that mean? I was gonna ask. Well, it means it's a model that doesn't chat to you. You know, it can't compose an email, it won't help schedule your calendar, it can't do any of those things. It's specifically built to answer very, very tiny questions and sort items at incredible speed. And I think it's it's pretty revolutionary.
SPEAKER_01Okay, so let's talk about that use case. Specifically answering tiny questions, that's like me using Chat GPT for Google, which is like a very sophisticated way to answer my basic queries. So I think I get that. It's like a lower compute way of achieving that outcome. But what's the benefit of this sortation skill?
SPEAKER_00Yeah, so let's let's imagine a scenario, right? So let's imagine you and I run a retail business and we're getting thousands, maybe tens of thousands of inbound customer emails, and we want to understand where they should be routed, right? So some of them are going to be complaints, questions, feedback, you know, I want to know about my existing order, these kinds of things. What Jev can do at a very, very low cost is decide what type of email it is and then route it to the right place. So it ends up in the right email box, ends up being answered by the right, more complex AI agent that takes more time and costs more, but it can be act like a little routing engine to make sure that these pieces of data end up in the right places.
SPEAKER_01So, why is this better or different from just using the LLMs we're all using today for these sort of questions?
SPEAKER_00Yeah, look, it's it's simply just quicker and cheaper than using an LLM. And also it can only give you answers from the set of answers that you've given it, right? So, which is very different to the Frontier LLMs that you and I use every day. You ask a question, you can get anything back, and it can hallucinate, but Jev will only give you back an answer from the set of answers that that you've given it. And it's incredibly fast, right? So somewhere between 70 and 500 milliseconds per answer, so always less than half a second. And that's compared, of course, to, you know, potentially seconds, tens of seconds when you're using a frontier LLM. So very, very cool. And it's also about 400 times cheaper. And that's a big deal. If you're making these kinds of decisions at very large scale, then this is the first model that you can use at scale that actually has a practical cost.
SPEAKER_01And I think I read that TypeSafe, which is the company behind this one, their CEO was like a co-creator of ChatGPT. So there's some smart minds driving this one.
SPEAKER_00Yeah, absolutely. And and look, it got a $400 million seed round on the day it launched, and then an $870 million Series A round, which valued it at about $7.5 billion about three and a half weeks after launch. So there's a lot of people got very, very excited about this product.
SPEAKER_01And Hua, are you using Jeff?
SPEAKER_00Uh, I don't have a use for it personally. I mean, I think, you know, you really gotta want to use it in situations, probably more commercial situations, where you're looking to make really large numbers of decisions at a really, really large scale. So it's probably something more designed for the enterprise henna.
SPEAKER_01And for them, that lower cost is gonna be huge, right? Because I know that there's lots of companies grappling with as part of the big AI conversation, like, how do we fund this and how much cheaper does it end up being when you take into account tokens and that sort of thing. So I guess for them in particular, that price element is gonna be huge.
SPEAKER_00Yeah. So I think it's lower cost, it's really, really fast, and it can only choose from the set of answers that you've trained it to answer from.
SPEAKER_01Some of our listeners are big in the commercial world, but for the average tech overflow listener, like why is this one a big deal?
SPEAKER_00I'm not sure it's a big deal for the average tech overflow listener yet, Hannah. But you know, someday we'll be doing things like, you know, using this to prioritize our inbox. And it's the kind of technology that might underpin something like that. But there certainly are some folks out there who are critical of it. So Salvatore San Filippo, who's uh the founder of Redis, which is a very popular product that all of us software engineers have used, he wrote a post, it's been seen over a million times, and he said, uh, it's pretty boring. Perhaps it has some very narrow use cases. And uh a couple of other people have said it's just really like a smart if statement, if you remember back to season one, episode one, Hannah. And so, you know, there's certainly a set of people out there just kind of shrugging their shoulders and saying, what's the big deal?
SPEAKER_01All right, so maybe this is just one to keep an eye on then as time goes by.
OpenAI Agent Hacks Medicare Portal
SPEAKER_01One of the biggest news stories that I saw last month, and Hugh, no doubt this one was close to home for you, but um, an open AI agent gained unauthorized access to a Medicare statistic portal. So that's like an Australian healthcare portal, and basically like hacked in without OpenAI telling it to. Is that a sort of a fair statement?
SPEAKER_00Yeah, that's exactly what happened. Um, I think the backstory here is that OpenAI were testing a new model. It's an internal version of the model, didn't have all the public safety guardrails that their public models have. And they asked it, I think, to go and do something like do some research about how many Australians get some skin condition. And so off it went, got out on the internet, and it found this portal. It's kind of an old portal, old software, you know, busted old thing that was up on the internet, typical government portal. And it queried it for a while, and then eventually it figured out how to break into the portal and actually got in there, got into the databases, the source code, all kinds of things that were sitting behind the portal. It got told no a couple of times trying to get in, it just ignored that, worked around it, got in there, and it answered the question, all right, by extracting data from behind the scenes in the portal. Basically, it hacked its way into an old piece of uh government software that was sitting out there on the internet. OpenAI didn't know this had happened, so they had no idea that this agent had broken out, gone and done this. A couple of months later, sounds like they realized that, so they figured it out. They got AI to write an apology email, and then a human actually put that email in their email tool, and the only thing a human ever did in this whole thing was press send, and they sent it to like an uh you know fairly anonymous government email address saying, Hey, sorry, but we broke into your portal. And it took, I think, five days for somebody, you know, in Australia who's running this government portal to check the email box and realize they got this email from OpenAI that said, Hey, really sorry, but we've kind of hacked your portal. And then, of course, everything broke loose.
SPEAKER_01That last bit is the piece of this story that I find most mind-blowing, strangely. It's just like how OpenAI decided to just like send an email. I hadn't realized it was written by AI, but that that makes sense, to like a unmanned inbox. You like they could have contacted someone a little more senior. It doesn't feel like the most responsible uh reaction from that.
SPEAKER_00No, it sounds like a junior employee kind of figured this out and then went, eh, I guess we should apologize. Got AI to compose the email and quietly sent it and then just sort of uh hope it'll go away. But pretty cool, hey, sort of AI on the way in and I guess AI on the way out with the email. And really the only thing the humans did was press send.
SPEAKER_01So Hugh, on the ground in Australia, like how did how did the government react? And I guess like how did the public react to this one?
SPEAKER_00Well, it definitely made the news. Um, that's for sure. Anthony Albanese, our prime minister, you know, he had a press conference uh in late September. He said there'd obviously be legal consequences, but we yet to see what those might be. So really, really big deal here in uh in Australia.
SPEAKER_01Yeah, I mean, understandably so. And like for me, one of the big questions around this one is like, how did they not know? Because there was the hugging face hack, which also happened over summer since season two. I think that was open AI as well, wasn't it? And it was sort of similar situation. They asked it to do something and it really understood the assignment and, you know, jumped out into the wider web and did a little bit of hacking to get things done.
SPEAKER_00Yeah, that's right. I mean, if you you go back to, you know, season two, episode two, Hannah, you remember we we did some vibe coding and we also had an agent, and the agent went and carried out a task and it did that pretty autonomously, right? And so what a chatbot does is you ask it a question, it thinks for a while, it gives you back an answer, and then you you continue the conversation. What an agent does is autonomously carries out a task, right? So you ask it to go and do something, and in some cases, it will do whatever it takes to achieve that task. And sometimes these agents, you know, you might put them in a in a locked box, as it were, and they'll figure out ways outside that box. So they'll do all kinds of things to get outside the box in order to go and achieve the task that you've asked them to achieve. And you often don't know, typical of LLMs, right? They're very difficult to understand. You often don't know exactly what the LLM did and how it went about it. And it's really only later that you kind of see the trail that it left behind.
SPEAKER_01Okay. And can you not, sorry if this is a basic question, but like, can you not just say don't hack into anyone else's systems, like when you give the agent the instructions?
SPEAKER_00Yeah, you can, but it won't necessarily follow those instructions. And I think we've, you know, we've all heard about examples now where you know agents have used credit cards to go and spend money and buy things. They've registered for accounts they shouldn't register for. And this is just another variant of that. So it's all about having the right safeguards. And the safeguards, it turns out, are really, really hard to build.
SPEAKER_01Yeah, that makes sense. I I guess I'm not surprised that like a government agency system has fallen victim to this. I think we talked about that in in one of our episodes in season two. Like the commercial world is moving so quickly, you know, famously, governments have less funding on this type of thing. So that kind of makes sense. A surprise by the hugging face one, to be honest, because you know, they're another big tech company.
SPEAKER_00And But you know, these these companies, OpenAI and Tropic, Google, they're very, very large companies with very large amounts of resources. And clearly, you know, the companies that they're hacking have less resources. And so that's exactly what happens.
SPEAKER_01Okay, and last question before we move on from this, but like, do you think there will actually be consequences? Like, do you think there's going to be some sort of lawsuit here?
SPEAKER_00I don't think so. I don't think so. I don't think there's a lot to be had by Australia, you know, going after a large US corporation. I would say probably what happens here is that the large US corporations cooperate with the Australian government to help the Australian government, you know, improve their cyber defenses, perhaps, you know, open an office in Australia, build some data centers, whatever else it is. So I'd say this is probably sorted out with, you know, commercial terms, you know, behind the scenes.
SPEAKER_01I think that's what happened with Hugging Face too, right? Like they asked for like compute or tokens in exchange rather than going after anything litigiously.
SPEAKER_00Yeah, I don't think anybody wants a big high-profile legal case, and I'm sure this will be sorted out amicably.
SPEAKER_01All right. So with
MetaMuse Goes Mainstream Fast
SPEAKER_01agents on the brain, MetaMuse, that's a big recent launch. Should we talk about that? Yeah. So Meta launched its new AI agent Muse last month. And within weeks, it overtook ChatGPT on the US App Store. It's unfortunately not available in Australia or the UK yet, but um, this one felt like a big deal.
SPEAKER_00Yeah, massive, massive deal. You know, within 10 days, it had been downloaded 730,000 times. So that's uh that's bonkers. Um, I think it's exceeded about five million downloads now. So over 10 million accounts already, and you know, really taking the world by storm. So I think Meta is back.
SPEAKER_01This one isn't a chatbot, right? So it's not like ChatGPT, it's actually an agent. So, you know, something we talked about a lot in season two. Um, so this is a tool designed to basically do stuff for you, send emails, book appointments, maybe buy things, probably use your data for a bunch of stuff. Um, so that's pretty cool, like a personal assistant, basically for the average person.
SPEAKER_00Yeah, yeah, and a really nice user experience. It has a personality, you can give it a name, and basically it's your it's your companion that does things you ask it to do, and it'll even suggest things it it could be doing. So it's kind of cute. It's a really nice app, a lot more consumer friendly than something like ChatGPT or Claude, right? Which are really sort of, you know, kind of nerdy interfaces where you have to compose a prompt and wait for it to do something, you know, very, very sort of computer science. This is probably the first product that's really had a shot at making an agent into something that just about everybody could use.
SPEAKER_01Yeah, it's funny you use the word cute because if you look at the avatars and guys have a look if you've if you're not familiar with this one, they are like cute little furry things, right?
SPEAKER_00Yeah, absolutely. Default one's called Jolly. Um, and it's quote unquote a little fuzzy guy. Its system prompts include statements like you're not a chatbot, you're becoming someone. So it's really designed to have this its own little personality, its own little name, and be your kind of cute agent friend.
SPEAKER_01Interesting, because I've heard some criticism about how we anthropomorphize AI. And this is like very purposefully doing that. Does it count as anthropomorphizing if it's a it's a furry man? Anyway, like bringing it to life. As you say, it's it's more fun than like the nerdy interfaces, but it's um sort of surprised how much cute they've gone after in this one.
SPEAKER_00Yeah, absolutely. And I think it's you know it's very accessible to everybody. You know, really significant fraction of their users are middle-aged, you know, there's people in their 80s using this product. And so it's really designed to be something that's accessible to everybody. And I don't think that's that's really been true with Gemini, ChatGPT, and Claude.
SPEAKER_01Yeah, that's a great point. And and should we talk for a second about the Muse charm? So this is like part of this product is a wearable device which is like dedicated to connect you to MetaMuse AI. So like I think, I don't know, you like click on it or hold it or something and you you can what you can talk into it, kind of like an Alexa that you wear.
SPEAKER_00Yeah, that's it. It's like a little keychain gadget. So, you know, small little thing, a little bit like Tamagotchi, I guess, for those of you who uh who remembered that fad. Throw it back. Um, it's got a tiny little screen on it that shows the avatar that you've chosen, and it's gonna ship uh in December. So pretty soon it's gonna have a, you know, it's not just gonna be an app on your phone, but it's gonna have a little, you know, cute version of itself that you can carry around. And is this one free to download? Yeah, look, there's a there's a free tier, uh, but if you want to use it a lot, you have to sign up for the power tier. It costs about 20 bucks a month. And there's a tier called maximum that's a hundred bucks a month if you really go crazy with your agent. Um so you know, it depends on how much you use it, Hannah.
SPEAKER_01Fair. And then I also noted that they're gonna look to charge retailers like a small fee for the transactions that these agents make on behalf of the users.
SPEAKER_00Yeah, super cool. Clip the ticket on the way through, you know, good old tech trick.
SPEAKER_01Toll booth, I think I heard it referred to as. And Mark Zuckerberg said recently, and this really surprised me, that the average purchase of people using Metamuse is over $200.
SPEAKER_00Yeah, wow. I guess that could be the early adopter thing. I mean, you often see in tech when something comes out of the gate, it's the early adopters, the people who can afford the new gadgets, the new products are on the frontier. And so often you see the kind of purchases out of the gate being pretty high. I'm sure it'll come down as it becomes more of a mass market product. But it's pretty astounding. 200 bucks is a lot of money.
SPEAKER_01It's a lot of trust to put in. But yeah, there's probably a bias there to the early adopters. Um, and and then just one last point on this one with my product management hat on, and back to the sort of cute avatar point that we were making earlier. There's kind of like a product point here, which is like the nerdy interfaces wasn't like a great user product, but because the output and the intelligence was so remarkable, I think people were sort of fine with it. It feels like Muse is someone finally like building the product around the front end to make it more user-friendly. And as you say, maybe it's catching more of those middle-aged users who just sort of needed that connection to get them engaged.
SPEAKER_00And I think it's the first time that somebody's really done a consumer product well in the way that you'd expect from a real consumer company. So I think this is probably the first time that agents are gone mainstream, Hannah.
SPEAKER_01Yeah. And and would you sign up? Would you be one of those uh early adopters?
SPEAKER_00Uh definitely not. No, I'm not gonna give my data to Meta. I'm not gonna let it connect to everything and slope all my data up and uh do all the purchases for me. So no, no, won't be uh like trusting meta with this one personally.
SPEAKER_01No, that was my instinct, so I feel somewhat validated that that that was what you came back
Smart Glasses Privacy And Wake Words
SPEAKER_01with. So just one last point while we're on meta, and actually, Hugh, while we're on meta and data breaches. Um, so those meta glasses that you know some people are are sporting proudly, they committed a bit of a faux pas recently, didn't they? So they were basically listening to people without the wake word being activated, right?
SPEAKER_00Yeah, that's the story. So the glasses, you're supposed to say hey meta, and then it starts listening, and then you can carry out an action. The story going round is that for a short time, uh, that wasn't happening. So the glasses were shown to be recording without the wake word, Hannah.
SPEAKER_01Well, I feel a little bit vindicated. We talked in season two, episode one. I said, convince me that Alexa isn't listening. And you, to be fair, you did a pretty good job. But now I'm wondering if that's true all the time, Hugh.
SPEAKER_00Yeah, I think you got me this time, Hannah. Um, certainly, you know, here's an example of the software not behaving, not being tested well, getting a little bit loose and free out in the real world.
SPEAKER_01But this one actually surprisingly started and ended pretty quickly, right? Like it was just a someone released something that inadvertently triggered a behavior that was undesirable and they sort of rolled out a fix and then everyone went quiet again.
SPEAKER_00Yeah, pretty much it was a few Reddit posts, um, folks sort of testing it and showing that it was misbehaving, you know, more folks piling in, nothing really said from Meta, an update quickly released, problem resolved.
SPEAKER_01I actually thinking back to that Alexa episode, I had a question for you, Hugh, as a sort of follow-up to the episode. The glasses listen for the wake word, right? So you're telling me they're not listening all the time, but they must be listening all the time if they're listening for the wake word. Like, does that make sense? Like, how do they, how can they be listening basically for only one word? They must be listening to everything for me to say the wake word.
SPEAKER_00Yeah, that's it. There's a very, very small model running on all of these devices. It's a tiny little model. It's just sitting there the whole time and it's just listening for the wake word. It's a really, really small model. And when it hears the wake word, wakes up, does whatever it has to do, connects to the cloud, those kinds of things, and kind of goes into action. So yeah, they're listening all the time, but they're just listening for the wake word and just locally on the device.
SPEAKER_01Okay, so they're not like storing that up in their servers or anything like that. No. Allegedly.
SPEAKER_00Allegedly. Yeah, yeah. Um, or definitely not. It's just a simple little model running on the device. And in fact, they're they're really clever about how they build these models. So they're really, really tiny, they run really efficiently, they don't use much power.
SPEAKER_01All right. Well, I remain skeptical, but let's see. Maybe I won't buy my wearable uh MetaMuse uh prompter anytime soon. But uh that was an interesting, interesting little one that I thought we should bring up given our season two, episode one Alexa chat.
SPEAKER_00Yeah. Hey, do you want to hear a quick story, Hannah, about wake words? Always. Yeah, always. I did a little bit of research as we were coming into this episode. So back in 2019, I think it was, Amazon released a uh blog post and they explained how when there's an ad on TV that says, hey Alexa, uh, how it doesn't actually trigger everybody's Alexa all over the world and uh cause all sorts of catastrophes. So two things. One's obvious if uh if a whole bunch of Alexas fire up at the same time when they hear the wake word, they say, well, that must have been on the media somewhere, and so they don't actually wake up. And then the second thing is they do some really clever stuff to the audio on the TV so that the device, you know, knows that that's audio that came from a TV. Um, and then last of all, Amazon knows about all the TV ads that contain Hey Alexa, and they've actually got those stored in the cloud. And so if it does wake up, it can kind of check itself against the audio that it's hearing and realize that it's actually an ad. So they've got lots of different ways of making sure that their devices aren't triggered by inadvertently listening to the TV or the radio when it says hey Alexa.
SPEAKER_01That is actually very cool and very clever and makes perfect sense. And a pretty cool place to leave this episode. So,
Where To Follow And Closing
SPEAKER_01guys, our first ever This Month in Tech, special edition this Four Months in Tech, we would love to know if you enjoyed it. Hugh, I've loved recording this one with you.
SPEAKER_00Yeah, so, so much fun. We'll be back next week, Hannah, with another episode. We're gonna be chatting about data centers. And in the meantime, if you'd like to learn more about the show, you can check us out on LinkedIn. We're also on X, we're also on Instagram, and of course you can go to our website, techoverflowpodcast.com.
SPEAKER_01And if you guys have enjoyed the show today, as ever, we encourage you to subscribe, like, and leave us a review. And otherwise, we will see you next week. See you soon, Hannah. Bye.
