OpenAI has a Muse problem

Welcome and introductions

0:00Support for the show comes from Atio. There are sales teams out there who never seem to stop moving. Always a step ahead, always working with a level of precision that doesn't seem normal. It's no secret. They're on Atio, the agentic CRM. Atio runs the work behind every win. It's built for humans and agents with the guardrails of human software and the scale of agentic infrastructure. This is the frontier of go to market. Teams like Parallel, Turbo Puffer, and Wardsmith are already using it.

0:30Atio to set the pace. Now it's your turn. Try Atio for free by going to atio.com slash decoder. That's A-T-T-I-O dot com slash decoder.

0:46Support for the show comes from ServiceNow. AI is moving fast across the enterprise, but without visibility, it's just chaos. Different tools, different models, different teams using AI in completely different ways. ServiceNow turns that chaos into control. With the AI control tower, you see all your AI across the business in one place. What it's doing, what it's done, and what it's about to do. So you stay in control. To put AI to work for people, visit servicenow.com.

1:18Low energy, unfocused, foggy? You might be dehydrated. Whether it's hot yoga, over 100 degree weather, or too much sun. Gatorlite, with a specialized blend of five electrolytes, was scientifically designed to help move fluids into the body faster for rapid rehydration. Shop now at retailers nationwide. Gatorlite hydrates faster than water. Is it in you? Symptoms noted are signs of mild to moderate dehydration.

1:48Consult a healthcare professional if symptoms persist.

1:52Hello and welcome to Decoder. I'm Neil A. Patel, Editor-in-Chief of The Verge, and Decoder is my show about big ideas and other problems. Today I'm talking with Hayden Field, The Verge's senior AI reporter, about the new wave of consumer-friendly AI agents. Now, if you've been paying attention to AI, you know that AI enthusiasts have been using agents for a minute now. Homebrew open-cloth setups led to a surge in Mac mini sales earlier this year. With the launch of Meta's Muse, OpenAI's Dots, and XAI's Grockbot has brought easy-to-use agents to millions. Dots, and in particular Muse, have had the highest-profile product launches lately, and they're fascinating to compare and contrast.

2:26Both Meta and OpenAI have decided to pitch their agents to mainstream users and businesses, in the form of cute animated mascots that can do everything from the boring, restaurant reservations, and inbox triage, to more sophisticated tasks. In OpenAI's case, they are even offering specialist Dots for marketing, legal work, and accounting. There are other differences. Muse, notably, is free, while Dots is not. And you'll hear Hayden and I talk about why Meta, which still doesn't have a frontier model of its own, might have a meaningful edge here, because it's so much better at making and distributing consumer products.

2:56There's also a huge decoder-style tension wrapped up in the agent race, a conversation about what AI is good at today, what it still can't do, and the privacy and security implications of handing over your credit card information, your email inbox, and other sensitive data to an AI that might be able to do things for you without ever requiring an app or a phone in your hand. That might be the future of all computing, but it's unclear whether anyone wants to hand over all the data to make it happen. Okay, The Verge is Hayden Field on Muse, Dots, and trusting AI agents. Here we go. Hayden Field, your senior AI reporter here at The Verge.

3:27Welcome back to Decoder. Thanks so much. I'm excited to talk to you. This time, there isn't any wild interpersonal drama. No one's feelings have been hurt. I would say Elon tried to hurt some feelings and didn't get there. Alexander Way from Meta has maybe been taking some shots, but none of the Real Housewives stuff we usually end up talking about when you're on the show. Just some product launches. It's been a bit of a pause in the soap opera antics for now, which I'm really grateful for. Just some good old-fashioned product competition in the marketplace.

3:59Let's see who wins and loses, and maybe everything will kill us all in the end. But for now, cute mascots.

The current state of AI agents

4:05So there's a lot going on. Let's start at the start. Hayden, you were at OpenAI Dev Day in San Francisco. They announced Dots, which is their new agent platform with a cute mascot. That happened just a few weeks after Meta Connect, at which the company was all in on Muse, their agent platform with an adorable mascot. Tell us about the state of the industry right now. Everyone's very excited about agents. What's going on? Yeah, it's funny to me because, you know, I've been covering agents for so long. And I remember that in 2022, it was kind of like the year of ideation.

4:38That's what, you know, tech leaders called it about agents. They were just ideating. They were thinking about what it could look like. There were a lot of Jarvis from the Marvel Universe references. 2023 was, they all called it the year of deployment. Like, let's try things, let's deploy and, like, learn more about what's failing, aka pretty much all of them were failing at the time. Then we had 2024 and 2025. They didn't have any names for those years, but things were still pretty bad, as we saw. And 2026 seems to be, to me, the year of actually, the beginnings of actually useful AI agents for the consumer.

5:11Like, always on. You know, obviously, OpenClaw was the start of all of this. But, yeah, what's interesting to me is that after seeing that one man, Peter Steinberger, was able to create an actually useful AI agent for the consumer, an always-on tool, despite its privacy flaws, it had a lot of privacy and security issues. But it was an agent that was useful enough that people still wanted to use it anyway and try to find ways around these privacy issues that it was having. So that inspired these companies to be like,

5:41oh, if one man can do this over the course of one weekend, we've really been slacking. We've got to get it together. OpenAI hired that guy, and now we have Dots. Meta, of course, was working on its own version of this and put it out sooner, Muse. So, yeah, they're basically just both always-on consumer-facing AI agents that these companies are peddling as a way to essentially be your personal assistant in a lot of ways, like booking flights, booking dinner reservations, buying gifts for people.

6:11But not only that, some of them are kind of going further to be your work personal assistant. And, you know, we'll get into that in a minute, but this is kind of how OpenAI and Meta are warring right now. They're trying to kind of present these two AI agents as a little bit different from each other, even though they're essentially the same.

Defining AI agents and open claw lineage

6:31Just broadly speaking, to define some terms, we're going to keep saying agent, and what we mean here is an AI model wrapped in a harness that lets it use a computer. Right. The way that I describe AI agents usually is an AI tool that can complete multi-step complex tasks on your behalf without you hand-holding it the whole time. So, just like, you know, I was a personal assistant in one of my first jobs in New York, so I would have gotten fired if every time my boss asked me to book a flight for her,

7:04I was like, well, what time? And also, what's your SkyMiles number? And also, what's your loyalty number again for this other thing? And also, oh, do you want me to try to get you first class? I mean, no, they want you to have common sense and just book the flight. And so, this is what they're peddling these AI agents as being, things that can work on their own mental kind of scratch pad in the background and not ask you every single time there's a step. They can complete multi-step processes in the background, kind of doing their own reasoning, and then present you with a final completed task.

7:38That's the big idea. I just, just to say it clearly, both Muse and Dots come out of the Open Claw lineage, which is a technical approach to building an agent, right? There's a lot of different ways you might be able to accomplish it. It'll remember your SkyMiles number and book you a flight. But the way everyone's going at it now is the idea of a harness around a model. That was the big innovation in Open Claw. And the computer Open Claw was using was often your own Mac Mini on your own network, using like a wide open browser with your own data, which is where a lot of their security problems came from.

8:09It's the same model with Muse, only it's a little Linux computer that Meta's giving you in the cloud. It's the same model with Dots, although the way Chattapedia is set up with codecs is very complicated, and that computer could be in a lot of different places. But this notion that what you really need is a computer with a browser, and some harness is going to direct a model to do a bunch of stuff for you using that computer, that's what we've landed on today. And it just seems to be the default winning model.

8:41Are all of these things just Open Claw? Are they just rifts on Open Claw? Are they the same as Open Claw? You could say that. I mean, they're different technically. Meta keeps saying, hey, we did build this from scratch, but it was inspired by Open Claw. There are a lot of similarities there, but they did say that they built it from scratch. And then, of course, OpenAI hired the creator of Open Claw, so there's probably a lot of similarities there as well for Dots. But, I mean, technically they're their own things, but they were all heavily inspired by Open Claw, and clearly these companies are kind of trying to get it as close to Open Claw as possible,

9:14without the security issues. The reason I wanted to stay there for one second is that I think it's important to identify the fundamental technical approach that is happening with these agents. Because there's a lot of different ways it could have gone, and the whole industry has picked this approach. What you need is a web browser, a harness, and a model that can go use that web browser.

OpenAI versus Meta strategy

9:34What's really interesting to me is that Meta has done a really good job, on top of that fundamental model, of making a cute consumer product. Meta is good at consumer products. Obviously, Instagram and Facebook and the rest are undeniably good consumer products. People really like them. The Meta quest for all of Meta's failings in VR, I thought, has always been a great consumer product. Like, you take it out of the box, it helps you use it, you can have fun with it right away. Whereas Dots seems like it's kind of made for software engineers,

10:04but you can also plan a wedding in it. And I'm wondering if that's how you see it, too, that one is kind of an enterprise product with a little consumer gloss on it, and then Meta obviously is just dead ahead consumer product. What's different here is OpenAI needs money a lot more than Meta needs money. And so they're honestly pretty similar, the products, it seems like. The difference is that OpenAI is marketing it to be an enterprise-friendly product a lot more than Meta is.

10:36So Meta, yeah, it's just a consumer-facing product. They're happy with that. Their goal is to make it as easy to use as humanly possible. So, you know, one-tap downloading, you know, just like putting it in your face literally everywhere. Even when I went to my Instagram profile the other day, there was a, like, pop-up for using Muse. So, yeah, they're trying to make it free, accessible, and super easy to broaden the audience. They want anyone to be able to just use it with one click and get used to it. They're trying to kind of just flood the market with AI agents and make it normal

11:10and kind of have a chat GPT moment, if you will, for AI agents. They want Muse to be that. Now, with OpenAI, they're still marketing it as a consumer-facing product. I mean, when you're using chat GPT, if you have the right subscription, it's right there. It's like, use your dot to do this, that, and the other. They really want you to, you know, go for it as a consumer as long as you're paying. So this is for their $100 to $200 a month subscription tier, unlike Meta. So that's why there's a huge difference here.

11:42For OpenAI, they don't have the money, it seems like, to, you know, just be front-loading all of this and say, hey, use it for free, no problem, we'll catch up with you later. They want this to be a money-making product, and that's why they're marketing it to the enterprise. They talked a lot about privacy to try to, you know, make this set apart from Meta's Muse. They talked a lot about what you could do with this as a knowledge worker, whether you worked in marketing, graphic design, software engineering, product, tons of different industries

12:14that kind of gave examples of how you could use this as a assistant, as a co-worker. Sam Allman even put it like a chief of staff. He didn't want it to be just an assistant, he wanted it to be your chief of staff. And something else that I thought was interesting is that Sam Allman seemed to kind of subtly dig at Meta on stage a few times during Dev Day, but one of the times he said, look, I don't want this to be just something that's used to book flights or, you know, plan things with friends and look at your calendar. I want this to go beyond that and to really do knowledge worker tasks, things that, you

12:47know, you and your industry would need a niche assistant to perform. And that's why they also introduced specialist dots that are good at marketing, legal analysis, things like that. So they're really going for the enterprise angle here, but I think it's because they need money ahead of their IPO, obviously. The revenue piece is really interesting that it's only for paid customers that they're higher tiers. OpenAI's tiers are very confusing, but it's $100 and up. And that's just going to keep a lot of people from using dots at first.

13:18The flip side of that is OpenAI does have a lot of customers who are paying them a lot of money. If they can get those customers to start using dots, then those customers will be on their agent platforms. And maybe you get some lock-in, right? You've given those agents a lot of data, they're part of your workflows, and that might solve the problem I think we've seen all the Frontier Labs have, where a new model comes out and it's slightly better than the competitor model. And everyone switches for six months, and the next new model comes out and everyone switches back. Is there some thought that these agent platforms create lock-in, create actual loyalty on top

13:53of just the models being better? Yes. They're basically hoping that they can build a moat with their users for these things. So, you know, if Meta and Google already have super easy integrations with all of each company's existing tools and services, you know, they can easily keep you there. If you're, for example, I have an Apple iPhone, but I use almost exclusively Google apps. So if Google had an AI agent that I was using every day, it would be a little tough for me

14:24to switch. Meta is hoping that they can get you so into the Muse ecosystem that it would be tough for you to switch once maybe one day you do have to pay. With OpenAI, they're trying to get you locked into their ecosystem right away. Even though you have to pay up front, they're thinking, okay, well, you know, we have integrations with almost every tool and service. Maybe it's not as good as the company that owns that tool and service, but we have the variety and the breadth. They're kind of going for quantity here. One of their employees told me that.

14:55They said, you know, look, yeah, maybe our Gmail integration, for example, isn't going to be as good as Google's. But, you know, the fact that we have a Gmail integration and an Outlook integration and X, Y, and Z, they're going for just the power of choice and the fact that power users will hopefully stay in this moat.

15:15Support for the show comes from CrowdStrike. AI is already fundamentally changing how companies build products and run their businesses. If you listen to the show, you probably already know this. But while AI drives innovation, every new model and agent also creates new cybersecurity vulnerabilities. That's why CrowdStrike built AI Detection and Response, or AIDR. AIDR helps organizations discover where AI is being used, monitors AI activity at runtime,

15:50and helps teams detect, investigate, and respond to threats involving AI systems. As companies adopt AI, security has to be built in from the start. More than 70% of the Fortune 100 trust CrowdStrike to protect their businesses. Now CrowdStrike is helping secure the AI era. You can learn more about AI detection and response at CrowdStrike.com slash decoder.

16:18You're either all in on this model, or maybe you're building with another. You're either speed, or is it security? You're either custom, or are you ready to use? Or, you're AWS AI. With hundreds of models, ready-to-use agents, speed, and security built in, you don't have to pick a side. You can have them all. How will AI revolutionize your business?

16:46AWS AI is how. Support for the show comes from Pipedrive. Sales teams spend up to 50% of their time on admin work, rather than selling, building relationships, and closing deals. They could really use a tool that covers all the admin stuff. That way, sales teams could actually focus on the work they were hired to do. That's where Pipedrive comes in. It's an intelligent, AI-powered sales CRM loved by growing sales teams. Their new meeting intelligence features, like an AI note-taker, are built right into the CRM.

17:19Pipedrive automatically pulls deal history, email records, and previous conversations. You can walk into a call already briefed. During meetings, Pipedrive's AI note-taker records and takes notes. Then, it turns those notes into accurate, auto-drafted CRM updates, ready for you to review and approve. The end result? Less admin, more selling. Switch to a CRM built by salespeople, for salespeople, and join the over 100,000 companies already using Pipedrive. Our link gets you an exclusive 30 days free instead of the usual 14-day trial.

17:51No credit card or payment needed. Just head to pipedrive.com slash decoder to get started. That's pipedrive.com slash decoder, and you could be up and running in minutes.

18:09It's funny, you mentioned Google. Google announced an agent, Gemini Spark. We should mention that SpaceX AI, the very coherent company that everyone understands, SpaceX AI, has an agent called GrokBots, which it is promoting heavily. People seem to like it. That was built by the Cursor team, which they acquired. There's a lot of this going around, but it seems like Meta really got there first with a consumer product that people could understand. Obviously, GrokBot and Dots are much more enterprise-focused, coding-focused.

18:43Why was Meta able to get there first?

Meta getting to market first

18:46I think it's because from all my reporting on Meta, they have been behind in AI for a while. And Mark Zuckerberg was like, look, we've got to lock in and do a product that people are going to use on our platforms. Alex Wang apparently was more looking to create a frontier model that it was going to compete with the other leading frontier models. He's like, we got to think long. We got to think super intelligence. Zuckerberg was like, no, we got to think cold, hard cash, a.k.a. getting people to use stuff on our own platform.

19:24And so I think that where they met in the middle was this, an AI agent that people would use on the Meta platforms, but that still represented a huge step forward, they hoped, for AI systems and what they could do in the everyday world. This is a fascinating dynamic in that Meta still doesn't have a frontier model. I don't think anyone at Meta is claiming that the Muse Spark model competes at the frontier with GPT-6, but their product is better. And certainly they have a massive distribution advantage in that they don't have to pay for advertising on Meta platforms,

19:59and OpenAI does. And certainly they have a consumer design advantage. They are just better at consumer products. Does that change how you feel about the dynamics of competition in AI at all? For so long, it's just been about the capabilities of the frontier. But now it's a little bit about marketing. It's a little bit about product design. It's you don't have the best model, but you made a consumer agent product, and you connected it to giving everybody a computer in the cloud, so now it can actually do things for you. And its frontier capability is actually not the important thing.

20:29It's its capability capability. I think the more money and power you have at one of these companies, you get a little bit mentally farther away from what the average consumer wants. So let's say the average consumer does want an AI agent, which many of them don't, as we know. But let's say they do. If they do, they're going to want it to be free or cheap, and they are going to want it to be safe to use without creeping them out too much. You know, they're willing to deal with a little bit of creepiness in order for it to be free,

21:01but not too much. Meta's muse, the headlines about some of the privacy issues with it have gotten a little bit above that comfort level, I think, for some users. That's probably why OpenAI at DevDay made a lot of, you know, remarks about privacy and the importance of that, especially if they're trying to, you know, court enterprise users with their AI agents.

Consumer versus enterprise adoption

21:25Right now, these companies are trying to, especially Meta, just kind of trying to flood the market and create a mass market appeal for this stuff. Similar to how ChatGPT became the default AI chatbot, Meta wants to be that for AI agents, and they feel very hungry about it. They want to create a ChatGPT moment for this technology by making it the default use for the general public muse. I just think there's something really interesting about the idea that a less-than-frontier model wrapped in a better product

21:56is actually going to be the success here versus a bleeding-edge frontier model that is a little bit harder to use. And there's some dynamic in there that is the history of all consumer tech products, right? The worst product that's easier to use often wins, but we haven't seen that yet in AI. Right. I think that's really important to note here, because not everyone needs the bleeding-edge frontier model. You know, software engineers do. They see a huge difference with coding prowess from one model to another, and that's why they're always switching.

22:27But the general public, if they're using an AI agent, as long as it's useful, they're good, you know? However, not many of them are useful yet. We're still testing Muse and Dots ourselves. Muse, apparently, I saw a couple anecdotes yesterday that it had a big flop for a day or so. People felt like the capabilities got a lot worse. Dots, you know, the jury's still out. We're still testing it. But it depends on what industry you're in and what you're really using this stuff for. That's why OpenAI is going for enterprise,

22:57because it makes money and they're confident that their agent is going to be more and more helpful in these niche industries. That's also why they launched Specialist Dots, because as we know, the more specialist, like, niche info you feed into an agent or model, the better it's going to do, obviously. And they don't want to have that imbued into every single agent or general purpose agents, just too pricey and too compute heavy. So they're like, you know, let's launch these specialists. I feel like you and I talk about consumer versus enterprise adoption of AI

23:29every time you're on this show. And here, it seems very important to note that OpenAI, asking people to pay a lot of money and asking people to use it primarily in enterprise contexts, is going to overcome what I think of as the brittleness problem, just because enterprise customers, business customers, have a lot of economic incentives to figure out how to make something work. So if I'm paying OpenAI 100 bucks a month and I'm using the marketing dot to solve some marketing automation and it doesn't work the first time,

23:59it is very likely I will try again. Because if I get it right, then maybe I've saved some money or some time in my business. That is really good. And I'll do it again. And eventually that will lead to more revenue for me, the business owner. What I know about consumers is that the first time it doesn't work, they just walk away. They have no incentive to try again, to overcome frustration or brokenness. And we've seen this happen with Muse now a few times. I've seen it myself with Muse, where it pops open a browser and tries to do something and it runs into a brick wall or a CAPTCHA

24:31or Amazon blocking it. And it says, hey, I can't do it that way. Do you want me to try another way? And I'm like, ah, whatever. I was just screwing around and I'm never going to try it again. These are very different markets. OpenAI has vacillated between wanting to be the big consumer company and knowing that the revenue is an enterprise. Do you see any focus in dots or do you see that personality shift back and forth in the product itself? So far, I'm a little bit more impressed with it than Muse just because Muse, everyone was impressed right away.

25:05And then, yeah, since then, there's been a lot of up and downs. Plus the privacy stuff has been really interesting. Even building profiles on your family and closest friends, there's just a lot of, you know, Meta has never been a super trusted company. None of these companies are very trusted, but Meta has made a lot of missteps with user privacy. So, yeah, I think dots seems to be a little bit more helpful to me right now, but I'm still testing. So I'm not fully sure yet. I need to run through all my typical things

25:38that I do to like kind of compare. Everyone has their own short list of tasks. Right now, for example, I'm trying to create an album of my bachelorette photos and it's been unbelievably hard to get any agent to try to do that. So I'm like, I'll just do it myself. But that's a pretty easy thing to do. Once again, I mean, I'm like, how useful are these things for our everyday lives? Really? We'll see. I mean, they've gotten a lot better, I will say, over the past few years. I've been shocked at how slow they've been moving, but they have gotten better. So, you know, I think dots is a little bit

26:08more impressive to me right now, but we'll see. I've Muse plugged into my Instagram because I don't worry about sharing my metadata with Meta in the way that I worry about maybe everything else, but they already have my Instagram. So I've Muse plugged into my Instagram and it's just doing all kinds of stuff. It surfaces interesting comments that I should reply to on Instagram posts. It tracks follower counts. It suggests things I should make videos about. It has horrible ideas. Muse is absolutely convinced I should make a video almost every single day about obscure Australian tech regulation.

26:39I don't know why it has this idea. I mean, I'm up on it now. I'm deep in the weeds on Australian tech regulation just because I'd see it every morning. But there's something about all that that still to me is work, right? I've just assigned it work and then it tells me what work I should do. And I haven't found some great consumer use for it yet. Or organize my photos as a consumer use. It kind of still rhymes with work, but it at least does not actually work. Have you found a great consumer use for these yet?

27:10Not that works. That's the problem. You know, it's like, I'll try to have it do something for me for my wedding or for my personal life and they haven't been good yet. But that's why I'm still kind of running through and testing these because it's like, so far, it's always interesting factually when they can do something, but I haven't been able to trust them end to end with something that I really, really need done. It kind of reminds me of the analogy for AI agents.

27:40People always say, right now they're a bad intern and what you want them to be is a really good assistant. Or in Sam Allman's words, a really good chief of staff. I think they're no longer a bad intern, but I think maybe they're an okay or eh assistant. I don't think they've reached the point of being a really, really good assistant yet, but I still need to run through the rest of my tests to see. I mean, you know, they're at least not super consistent, it seems like yet, which is something that you need

Trust and privacy concerns

28:10in a really good assistant. That ties into trust. You've mentioned trust several times now. I'm only willing to give Muse access to, again, the meta platforms that Meta already has access to. I can't worry too much about that. I've given it access to my spam Gmail, so it is collecting a lot of ideas about my spam Gmail. But Meta really wants my credit cards, right? It wants to go through all of my credit card receipts. This is one of the suggestions in Muse. But Muse, you know, one of its suggestions is, let me log into your credit cards and I'll cancel streaming services you're not using,

28:41or I'll negotiate your Verizon bill for you. And I am absolutely not ready to give Meta one ounce of data more than it already has. This feels like a big issue for every agent. And there's an argument that you should give Meta all that data, one, because they already have it, and two, because they're a big company that you could sue if they get something wrong. Whereas if I gave all of my data to OpenClaw, it's just my own fault if it gets leaked. If I gave all my data to one of the hot startups, like Instinct, and they go out of business, well, that's just the end of that.

29:11How do you feel about this? Do you feel comfortable giving your data to Meta? Do you think people should do that? No, I don't feel comfortable. And I also don't think that the general person does either. If I've seen one thing in a lot of the comments on our Meta story that we published in the last couple of weeks, it's that people don't trust Meta at all. And I have to say, I mean, why would they? But I don't think that, you know, you should trust OpenAI a lot more either. It's kind of like, it's everyone's own kind of comfort level

29:42that they have to abide by. I think that some people are going to get a lot out of this stuff. Some people may say, hey, like, I don't care if they use my data as long as they are negotiating down my bill accurately or canceling streaming services that I wouldn't be canceling on my own. So they're saving me money. It's fine. But one, are they actually going to be able to do that accurately? Some people have said no. And two, yeah, I mean, like, as we know, when you release data like this, it's kind of, you can't get it back. So I don't know. I always think about that old example of the insurance company

30:13that theoretically buys data on what's in your fridge and maybe charges you a higher premium if you're buying too much beer and not enough, like, bananas. That's just a theoretical example. But, I mean, that's the type of thing that can happen. So I think that people should just kind of operate with caution, abide by their own comfort levels. But I'm certainly not comfortable yet giving my personal data to Muse. I don't even know if I'm comfortable putting Muse on my Mac. Apple just recently changed how apps can access the disk on Macs

30:44because Muse was so aggressive about reading data. In particular, Jason Aiton is a columnist at Inc. It read his notifications off the disk because he had just given it a permission and he didn't realize Muse would take that permission and go do something else with it. This is a lot of spiraling concerns, right? You're going to have a little robot that can use your computer for you, whether that computer is in the cloud or whether it's on your desktop, and you might not know what it's doing. And that ties directly into, well, OpenAI often doesn't know

31:15what its agents are doing, and that has led to catastrophic security problems across this industry. Can you connect those dots? Is your own little agent doing your travel booking the same kind of threat as OpenAI's agents hacking Hugging Face? I don't think it's the same type of threat, but it's definitely something that you should keep in mind when you're using this stuff. I mean, I think that back to the kind of Facebook marketplace debacle that we saw with Meta's Muse. To recap, basically, a Muse agent

31:48without someone's knowledge invited potential buyers of this item that this guy was selling to his house, said, hey, he's available now. He wasn't home. And then the potential buyer was like, hey, are you here? Like, I'm trying to get the item. And then Muse responded on his behalf. Yeah, I'm here. And then he just never came down because obviously he wasn't there. So there was a lot of drama going on there. Some people said, hey, well, he did give Muse access to his Facebook marketplace. You know, he should have known this could happen. A lot of people don't read

32:18all the fine print when they check yes to one of these things. So there's a lot there. There's a lot to unpack in that gray area. But the point is, I think that people should just be a little wary before, you know, checking yes to every possible thing that these agents want to access. Like you said about the credit cards or, you know, it's just, it's easy to just check yes and be like, you know, whatever. I'll figure it out later. But people should be reading the fine print or even use AI to read the fine print. You know, like this is one great use of AI.

32:49Get it, get it to read the terms and be like, what are the worst possible things that could happen here? Lay it out for me. Those are things that people should keep in mind before they move forward. And then as for, you know, the, you know, so-called rogue AI agents that have been like hacking other companies and stuff from these labs. Yeah, I think it's just important to notice that if they can't control some of their most powerful unreleased AI models, how well are they going to be able to control your data or your agent?

33:20Sam Allman said during Dev Day that you have a ton of control over your dot. He said that there were a lot of like, if this then that rules you can create. There are a lot of specific stipulations that you could make in terms of permissions of what your dot could and couldn't do. So that did give people some peace. But I mean, as we've seen, yeah, these labs and these large tech companies have had a lot of missteps about hacking other companies

33:50and more are coming out. It feels like every single week, more news items about these cyber attacks. You know, we've seen Meta have one. We've seen OpenAI have a ton. We've seen Anthropic have a ton. So I think that this is all just stuff to keep in mind. And one cool thing that I saw someone do was name their AI agent after the CEO that runs that company and then see how much data you feel comfortable giving it. So, you know, if you want to do that, I think that's a great idea. And then it just kind of makes you think twice. If you're saying,

34:21hey, Sam Allman, can you, you know, go through my email and see what needs to be deleted may make you, maybe you still feel comfortable with that, but maybe it just makes you think twice and think, oh, you know what, maybe I'll give it access to my spam email only, like you said. At T-Mobile, the new iPhone 18 Pro is here with a big leap in battery life and you can get it for nothing up front with so many reasons to choose T-Mobile,

34:52you'll want to talk about it. I heard at T-Mobile, you can get iPhone 18 Pro for nothing up front and get $1,000 off. Pass it on. Pay nothing today. No taxes, no fees, nothing down. Pass it on. Hey Mickey, pass it on. iPhone 18 Pro has the most advanced iPhone camera ever. Pass it on. At T-Mobile, you can get iPhone 18 Pro for nothing up front. Pass it on. At T-Mobile, get $1,000 off the amazing iPhone 18 Pro

35:23with no trade in needed. The 36 monthly bill credits when you switch to T-Mobile on a qualifying plan. If you cancel entire account before receiving all bill credits, credit stop and balance and required finance agreement is due. Example, 119999 iPhone 18 Pro 256 gigabytes. Bill credits in if you pay off device early. Qualifying credit, new line, port in, and service. $85 more per month with auto pay plus taxes and fees required. $35 device connection charged due at sale. Visit T-Mobile.com. Low energy, unfocused, foggy, you might be dehydrated. Gatorade Zero Powders have been scientifically designed to help improve

35:53hydration balance in the body. With a clinically proven electrolyte blend and no artificial colors or flavors. Great tasting hydration no matter where you're going. Shop now online or at retailers nationwide. Gatorade hydrates better than water. Is it in you? Symptoms noted are signs of mild to moderate dehydration. Consult a healthcare professional if symptoms persist. Support for the show comes from ServiceNow. AI was supposed to handle the parts of the job you hate. Instead, it just describes them, suggests what to do about them, and then leaves you to do it.

36:25That's not help. That's homework. ServiceNow's AI specialists are different. They're not a tool. Think of them as digital teammates who actually do the work from start to finish. Cases get resolved. Requests get processed. Loops get closed. And most importantly, no extra work for you. Because when you can truly delegate to AI, you can get back to the work only you can do. The work that requires a person with ideas and judgment. And you know, a pulse.

36:55To learn how to put AI to work for people, visit ServiceNow.com.

Mascots and anthropomorphizing agents

37:00So this ties into the fact that all these things are really cute. Muse is cute. The mascot Jolly is very cute. Alexander Wang has been waging a relentless campaign on X to make it not only cute but sexy, which you can feel any way you want about that. But Alexander Wang keeps tweeting memes of the mascot with a butt. Dots are obviously very cute. They come in shapes and sizes.

37:32There's something here where you're not quite anthropomorphizing the agent, but you're making it a character. You're giving it a personality. That connects to the idea that you've called AI Tamagotchis, where eventually we're going to have a little hardware device with your little cute mascot on it. And you'll talk to that instead of your phone. And this is the big promise this whole industry has been working towards forever. And they're going to replace your phone and the entire paradigm of the app store with AI agents that are going and using computers for you.

38:02Do you think that any of these products are a meaningful step towards that? I think they're just one building block along the journey of the goal that they're chasing here. I mean, I haven't heard of anyone that's using Muse or Dots so far for everything. It's not a Jarvis from Iron Man, despite how much these CEOs want both these products to be. I think that they are a meaningful step. But as for the cuteness, I think one of the most important points here is that they are intentionally disarming.

38:39These tools are made to be cute because you don't find them threatening. Or you find them decidedly non-threatening. They're cute. They're there to help you. It's kind of a response, I think, to the AI populism backlash. It's saying, hey, this isn't going to hurt you or do anything bad. It's just a cute little thing with eyes or a sexy little thing with eyes, whatever. But yeah, I think... To be clear, just to be clear, I do not think it's sexy. I think Alexander Wang has been relentlessly trying to make it sexy.

39:13No, I also do not think they're sexy. But I will say Elon Musk has been trying so hard to make all agents and chatbots sexy. So I'm sure it's not far off for someone to think that. But yeah, I think that whether the AI agents of the future are cute or sexy, it's going to be equally disarming to people. That's the whole spectrum. That's the spectrum, the cute, sexy graph. Yeah, it's going to be really disturbing to see when we go more along that graph.

39:44But yeah, I think that the point is there's supposed to be disarming and that people should really think twice before giving too much of their data. Or just pick and choose what you're comfortable giving, right? Like how you did with the spam email versus your regular email or saying no to your credit card. These are both new products. So kind of just see how the cookie crumbles before you decide to just, you know, give everything over. We're already seeing a bunch of changes week to week with how people are using Muse or what they're upset about. So, you know, just kind of see how it shakes out.

Monetization and business models

40:14The last piece I want to talk about is just how any of this makes money. You and I have talked many times about the bubbly nature of the AI economy. Zuckerberg has said his plan is for Muse to take a cut of transactions on the back end. So Zuck has given everybody a little Linux computer in a cloud with eight gigs of RAM. Obviously, he's given away tons of free tokens. He does not have to pay advertising costs. He can just put Muse across all of his platforms. And to the extent that meta is a well-oiled machine, the amount of Muse promotion that is happening across meta platforms instantly is kind of incredible to see.

40:48They just turned it on. I saw an ad during NFL games this weekend for Muse. They are spending money to promote this. They're giving it away for free. They're giving the tokens away for free. And then somewhere at the end of this, they're going to say to merchants, hey, this person was shopping for shoes. Muse brought them to your doorstep. You're going to pay for this customer to buy shoes from you. I don't know any merchants that are excited about this. And it certainly feels like merchants that pay meta more money would be preferred merchants. And that means that not only do I have an assistant or chief of staff, I have a bribeable chief of staff.

41:25Like, I have a corrupt butler. Is that what you want? That is a great term for it, a corrupt butler. Yeah, that is not what I want. And I don't think many people do. It's just crazy to me that, like, we've gotten so far from what people really want, which is they want to trust a recommendation and they don't want to be scammed. The leaders in AI don't think about how this has been, how certain things have been mirrored in other industries. Like, think about when people were paying for reviews for their businesses or, like, paying to delete reviews for their businesses.

41:59That was a huge uproar. Or influencers getting paid to recommend things without disclosing that it was an ad. Then we saw the rise of micro-influencers who were more valuable for a while because they were seen as more trustworthy. They didn't have that many followers, so the things they recommended must be really good, right? And then we saw the, okay, well, their hashtag ad. But also, they only said they were going to recommend things they really believe in, so it's okay that it's an ad. It's like these mental gymnastics that people have to do to figure out what is really a trustworthy recommendation these days.

42:33What is going to convince people that an AI recommendation is trustworthy if it can be bought? So I think that's something that these companies are really going to have to reckon with. You know, the ChatGPT ads blowback was huge, and OpenAI had to say, look, it's going to be a clearly disclosed ad underneath your query, not influencing your query. So, I mean, if you're using Muse to shop, and it's an undisclosed, like, potential, you know, meta customer that you're being redirected to to buy those shoes instead of these other ones, I feel like so many people aren't going to use it to shop.

43:14So I get it on the meta side. I understand why they would think about monetization that way. They're a consumer company. This is a consumer product. They know that the second they charge even a dollar for it, the uptake will drop precipitously. So they've got to keep it free, and then a bunch of people are going to want to shop. And go buy me a new iPhone cable. You know, at the end of the day, it's all just the same Amazon garbage anyway. So whatever. And then they can run an auction on the back end to be like, who wants to pay us meta the most for the privilege of selling an unbranded USB-C cable to this customer?

43:47And they'll run some auction, and that will, in its way, just sort of maximize the money that gets turned around. And that is broadly how advertising works today. That's how the Google search results page worked until recently when it was full of AI. And they're probably going to do the same thing. Meta can do that because they already have a giant advertising business that prints money. They can give away a consumer product. Mark Zuckerberg has been shown willing to blow money trying to win a market like he did with Reality Labs for years and years and years. And they can just run it until it works.

44:19OpenAI, even at $100 a month, is probably not making money on those customers. In fact, I think we know they're not making money on those customers. They don't have any monetization. And when they do try to turn on ads, people get mad at them. How can they keep up with Zuck? Only via enterprise. I think that's why they are pushing this so hard. You know, we saw a ton of enterprise discussion, demo videos, customer feedback, tons of stuff at DevDay.

44:49And that's why they're pushing this so hard, because they can't keep up with Zuck on the consumer front. They just can't. This is the part to me where when people ask if it's a bubble, I find myself wondering how much of the economy does Zuck have to win in order to give away millions upon millions of free tokens and 8 gigs of RAM to everyone in the cloud? Because that is just an expensive thing to give to everyone. And how many percentage points of every shoe sale in America do you need to take to fund it?

45:20And the answer feels like a lot. I think that's huge here and something to remember, because we already saw Zuck kind of crash and burn with Metaverse stuff, and he is not wanting to be embarrassed again. So, you know, how long is this going to go on, and how are they going to fund it in a real way on the back end, like what you were saying? It's just a little bit tough to imagine. Luckily, Meta has a ton of money, and they can redirect stuff from other departments, which we've already seen them do a lot.

45:52But it will be interesting to see how this fares on the next couple of quarterly earnings calls, you know, to see how much money is being funneled into this and how much it is making, if any. And how many people are using it? You know, they've announced it at the top of the App Store, but our own David Pierce has written an entire story about how being number one in the App Store doesn't mean much, right? It's sustained success that actually means something. Right. It reminds me of when I was in college, Uber and Lyft, or maybe it was just Uber, honestly, were new and VC-backed, so it was like $4 to call an Uber anytime you wanted to go out or down to the bars or anything.

46:29So they got everyone using it, and then they got everyone hooked, and then they had to, you know, up the prices. That happens, that's a tale as old as time with like every VC-backed thing. So what I'm wondering is if this is really going to work to become that, are they going to get enough users and enough like shoe sales on the back end that this is going to be a really viable business model? Or is it just going to kind of be a potentially, you know, big bet that they switch away from later?

47:00That's what some Meta employees that I spoke with in our Meta piece told me. They said they expect this to kind of be something that they, you know, are putting out, trying to make a big deal up, and then they may kind of revolve away from later, or at least partly. We'll see how that goes. I mean, if your bet is that the future of computing is talking to a little agent, it certainly seems like Meta has gotten farther along that path than anyone else. But we'll see. OpenAI has hardware designed by Johnny Ive that will theoretically come out.

47:31It is certain that Google and Apple and all the rest will respond at some point.

Closing remarks

47:34Hayden, this was great. Thank you so much for being on Decoder once again. Thanks so much.

47:40I'd like to thank Hayden for taking the time to join me on Decoder, and thank you for listening. I hope you enjoyed it. To get new episodes every Monday and Thursday, subscribe to our YouTube channel, at DecoderPod, and find us on TikTok and Instagram under at DecoderPod for more fun stuff we put out almost every day. If you enjoyed this episode, please send a link to it to someone you think you might like it. It really helps us grow the show. You can also subscribe wherever you get your podcasts. Decoder is a production of The Verge and part of the Boxing Beauty Podcast Network. The show's producers are Greg Ott, Kate Cox, and Nick Statt. This episode was edited by Ursa Wright. Our editorial director is Kevin McShane. The Decoder music is by Breakmaster Cylinder.

48:11We'll see you next time. Everyone's got a hard question about AI. Your job, your kids, where it's going. Anthropic was built to surface those questions and share what it finds along the way, because there's hope in hard questions. Ask yours at claude.ai.spotify and keep thinking. That's claude.ai.spotify.

48:36What's going on? I'm Arch Manning, Viore athlete and college quarterback. Whether I'm running, training, traveling, or just unwinding at home, I love doing it in my core shorts from Viore. With a breathable boxer brief liner, they're quick to dry, super versatile, and stand up to even my most intense training sessions. Plus, they come in three inseams and a ton of colors. Ready to try a pair? Go to viore.com slash arch and get 20% off at checkout. I think you're going to love them as much as I do. That's viore.com slash A-R-C-H and get 20% off your first order.

49:09Exclusions apply. Visit the website for full terms and conditions. Not only will you receive 20% off your first purchase, but enjoy free shipping on any U.S. orders over $75 and free returns. Have a great day. Your heart can tell you a lot about your health. Apple Watch Series 12 measures your heart rate every five seconds with the most accurate heart rate sensing in a wearable. So your Vitals app, now with heart rate variability, can tell you when something is off. And your readiness score can let you know when to rest and when to push. Hear the story in every heartbeat with Apple Watch Series 12.

49:41The features described are for wellness purposes only and not for medical use. iPhone 11 or later required. Based on Apple conducted study of heart rate accuracy, August 2026. Visit apple.com slash Apple Watch Series 12.