呼叫我的AI智能体：Meta Muse与全面接管生活的个人助理时代

Natalie Kitroof: From the New York Times, I'm Natalie Kitroof. This is the daily. Over the last few weeks, millions of people have rushed to download a new class of powerful AI apps called agents. They're incredibly useful. They're honestly kind of cute, and all they need to work is your most sensitive personal information. Today, my colleague Eli Tan on why he and so many Americans are opening up their entire lives to technology that we're just beginning to understand. It's Tuesday, October 6th. Eli Tan, first time on the Daily and we are very happy to have you on the show.

Eli Tan: Happy to be here.

Natalie Kitroof: Okay, so we've been talking a lot here recently on the daily about artificial intelligence gone ary, AIs going rogue, doing things they shouldn't, hacking into places they shouldn't, aka AI behaving badly, doing things that feel scary to a lot of people. But you, Eli, have been focusing on a very different side of AI, a potentially very helpful version of this technology. and we want to talk about what exactly that has looked like. So, just tell us what you've been up to.

Eli Tan: For the past few weeks, I've given my life over to this app called Muse, which is an AI agent created by the company Meta. And it's really the first agent of its kind that's been released to the public. Think of an AI agent as a chatbot that has access to its own computer where it can use a monitor and a keyboard and a mouse. So instead of just conversing with you like a person, it can actually go online and do tasks for you and work on its own autonomously in the background for hours at a time.

Natalie Kitroof: Basically, it can be your personal assistant, right? Powered by AI and not in this case an AI chatbot, but an AI agent.

Eli Tan: Exactly. A personal assistant is what Meta is calling it. And in the world of AI, these agents are a real leap forward in the kind of technology that right now mostly powers chat bots. Agents can operate software, they can operate apps, they can sign into your accounts. And the hope here is that in the future, you can just say, "Hey, book me a flight to Atlanta next week for the cheapest price," and it'll just go into Expedia or Delta on a browser and it will book that flight for you. So it fundamentally takes out the leg work of you having to use a computer.

Natalie Kitroof: Wow, that sounds incredible. And I want to see this in action. Do you have your agent with you right now? Can we take this for a spin?

Eli Tan: Yeah. I mean, what do we wanna what do we want to do? Like,

Natalie Kitroof: uh, I have a desire for lunch, a sort of salad like thing with protein. Um, can you do that? Can it order food?

Eli Tan: It can. Let me ask it right now.

Natalie Kitroof: Okay.

Eli Tan: Okay. I'm saying my colleague Natalie is hungry for lunch. She wants some kind of salad. Can you order a salad to our office?

Natalie Kitroof: Okay, I'll pay you back.

Eli Tan: And if you don't pay me back, I'll have it uh send you an email reminder uh in one week and it'll charge your Venmo.

Natalie Kitroof: Wow. Okay, so this is a real accountability thing. Got it. And so wait, like just so I'm clear, so it has your information to a delivery service that you have as well?

Eli Tan: Yes, it's also signed into my Door Dash account.

Natalie Kitroof: Got it.

Eli Tan: And it says happy to order the lunch. I'm being asked just to approve that this is the right lunch. And it suggested a harvest bowl from Sweet Green.

Natalie Kitroof: Okay. Unfortunately, that is actually my favorite order from Sweet Green. So, I'm getting a little freaked out.

Eli Tan: It's my favorite order, too. Maybe that's why.

Natalie Kitroof: Oh, interesting. Okay. So, it assumes we have the same taste.

Eli Tan: Yes. And then it says placing the order. And right now I can actually watch it do this, moving the cursor across the screen clicking on the different buttons. Just

Natalie Kitroof: Are you watching it right now?

Eli Tan: I am. Yeah, I have it.

Natalie Kitroof: Can I see it?

Eli Tan: So here it is. It's typing the address of the office.

Natalie Kitroof: Oh my lord.

Eli Tan: And and putting in the order.

Natalie Kitroof: This is nuts.

Eli Tan: Lunch is on the web. All right. So this thing, this service, this has long been the holy grail of Silicon Valley, right? Being able to have a computer run your errands for you. So how does Meta get here to building this? Because frankly, they haven't always been at the front of the pack in terms of the AI race.

Natalie Kitroof: Exactly. About a year ago, Meta had fallen behind in the AI race. They were developing what were called models, essentially the underlying technology that powers AI chatbots. And Mark Zuckerberg had apparently grown tired of falling behind. And so at an all hands meeting, he told the company that they needed to shift their focus entirely to building consumer AI products. And something he was saying inside the company at the time was that they should focus on building AI products that even his mom could use.

Eli Tan: And what did that mean? Like how did he envision that actually occurring? So Mark Zuckerberg's mom isn't a tech person. She's a psychiatrist. She's not really techsavvy or actually know much about AI.

Natalie Kitroof: AI for dummies basically.

Eli Tan: AI for dummies. Exactly. And around this time, this app comes out that becomes a viral hit. It's called Open Claw. And this is essentially an open source version of an agent. It catches fire in the developer community and everyone starts using it.

Natalie Kitroof: People can give over their computers to it and it can do all sorts of tasks that previously required human beings. And what were people in tech and the app actually doing with it at this point?

Eli Tan: A lot of them are using it to write code. So instead of having to sit at your computer coding all day, you could just tell the agent, "Hey, go make an app for me," and it would do it. But another developer I talked to, he used it to run his entire day. It would look at his calendar, it would tell him it was time to go to a meeting. He even rigged a camera up in his house so the agent was watching him and every time he drank water, it could see him drinking a glass of water and it would say, "Good job."

Natalie Kitroof: Wow.

Eli Tan: He even connected it to his Tesla and at one point it rerouted him to Whole Foods and it said, "You got a blood test that said you're low on magnesium. You need to be taking this magnesium." And he did it.

Natalie Kitroof: Okay.

Eli Tan: Now, the thing about OpenClaw is that it was the first tool to really show what was possible with these agents. But it was very geeky. It was very hard to use. You had to know how to use the terminal. You had to know how to set it up. It wasn't something that was going to work for normal people. But Meta was looking at this and thinking, this could be the thing that if we package it into an app that's really simple that anyone can just download on their phone, this could really be the first big AI product hit.

Natalie Kitroof: They're basically looking at this and thinking, we can turn this into a consumer app, something that normal people would actually use in their day-to-day life.

Eli Tan: Yes. And they were thinking we can be the first company because of all the resources that we have to bring an agent like this into the masses.

Natalie Kitroof: And when we're talking about the resources that Meta has, we're talking about all this money, but also their expertise in making things that people use every single day.

Eli Tan: That is right. While they don't have as long of a history of making the most cutting edge, experimental models, they know how to make software that people like using. That's why there are 3.5 billion users that are on one of Meta's products.

Natalie Kitroof: Yes. Three and a half billion people that already use Meta's products for hours and hours every day.

Eli Tan: Mhm.

Natalie Kitroof: So by April, they had a version of this that was pretty good at all the things that OpenClaw was good at, coding and scheduling and ordering food. But it looked like something anyone with a smartphone could use. And then last month, they released it to the public and they make it free to use.

Eli Tan: Okay. And talk to me about the decision to make it free. Obviously that gets a lot of downloads. But that costs a crazy amount of computing power for Meta. So why did they do that? It's clearly about getting this in the hands of as many people as they can.

Natalie Kitroof: And how's that going?

Eli Tan: So far, it's been a bit of an early success. Muse shot up the App Store rankings, millions of people have downloaded it. And in a lot of ways, it's delivered on what Mark Zuckerberg originally envisioned. But in my experience of using it, in order for it to be useful, you have to take this really big step, which is you have to give your life over to this AI and fully let it in.

Natalie Kitroof: Yes. In order for this app to actually be useful, you have to give over all aspects of your life, not just your Door Dash login. And what did you give over? What did that look like? Walk me through it.

Eli Tan: One of the first things I did was I signed into my credit cards, my bank accounts. I gave it my Social Security password, my health insurance password. I gave it access to my contacts and I gave it my girlfriend's email and her address.

Natalie Kitroof: Oh my.

Eli Tan: In case I wanted to send her something. Yeah. I thought of all the possible personal sensitive pieces of information that I have and I I gave it all to Muse.

Natalie Kitroof: Wow. Okay. And did that give you pause? Were you nervous about doing that?

Eli Tan: I was. Yeah, even though Instagram I feel like it it knows so much about me and it knows all my preferences, to actually give your passwords to a bot, it feels like a really big step. But it was also the only way it was gonna work. If it can't sign into my accounts, it can't buy things for me. It can't do the things that I want it to do.

Natalie Kitroof: Absolutely. Yeah, Meta has had a ton of issues with privacy in the past. Obviously Cambridge Analytica and leaking the personal information of tens of millions of users in the past. Were you thinking about that?

Eli Tan: Absolutely. Yeah, Meta has had a ton of issues. It's been an important part of the conversation around privacy and data. But when I was doing this, I actually set up an isolated sandbox. I wasn't using my real main bank accounts, I used dummy accounts with small amounts of money. Because for me, giving over my real life savings was still a line that I wasn't willing to cross.

Natalie Kitroof: Okay. So once you openly embrace Muse, how did it go for you? I mean just describe the experience of it.

Eli Tan: It went surprisingly well in the beginning. I started giving it the most basic tasks, like "Send a text to this person" or "Check my calendar and see if I'm free this afternoon." And it would pretty much do them all as planned. Hm.

Natalie Kitroof: Mhm.

Eli Tan: Then I gave it instructions to do more complex tasks. And this was a time when I started to notice things getting a bit hairy, where it would get tripped up with the more complicated stuff.

Natalie Kitroof: And what was that more complicated stuff?

Eli Tan: Telling it to call my dental insurance and ask for the status of a reimbursement check. With dental insurance, you're usually waiting on hold for over an hour. And the agent actually made the call, navigated the phone trees, talked to the voice bot on the other end, and waited on hold until it had the exact right person on the line.

Natalie Kitroof: What? This is crazy.

Eli Tan: So for this insurance case, I looked at the transcript and it had found my member ID, and when the phone tree said, "Answer these questions to confirm your identity," it got through all of the CAPTCHA questions that are meant to make sure that I'm human.

Natalie Kitroof: Oh my god. Waiting on hold is the exact use case that I want AI for. Can I just say like that's good.

Eli Tan: Me too. of all the things that it can do. If all it could do is just wait on hold for me, I would pay for that.

Natalie Kitroof: And was it always that good at these more complex tasks?

Eli Tan: No, it wasn't always that smooth. That was probably the best example. A lot of times it was pretty awful. Like I tried to have it make a custom morning podcast for me about the news that I care about in tech and sports, and I listened to it on the way to work. No, Eli,

Natalie Kitroof: No, Eli,

Eli Tan: it was pretty awful. The host, you know, they were monotone and it was boring and I had to turn it off after 5 minutes. I haven't listened to an episode since ending the experiment.

Natalie Kitroof: Yeah, I I was under the impression you could not replace the indisputable charm of a human host and that podcast hosts were immune to AI disruption.

Eli Tan: They are immune to AI disruption, and any company trying should just give up because it's just so bad. And then there were other times where I asked it to manage my fantasy football team. And it totally blew it. It made these terrible roster decisions, benching players that were healthy and starting players that were injured. It basically cost me my matchup that week. It's been the worst loss of my fantasy football league so far this year.

Natalie Kitroof: Okay. Just setting aside the quality of the advice that the agent was giving you, hearing this makes me think, why did you outsource that stuff anyway? Like fantasy football, isn't the whole point of that the fun of doing it yourself and making the decisions? Explain that to me, Eli. What is the point?

Eli Tan: Yes. This is actually my biggest criticism of Muse and AI agents, which is that you lose the serendipity and you lose the human connection. Like I had it order a book for me, and it did, but I love going to my local bookstore. There's a guy at the bookstore who I enjoy talking to

Natalie Kitroof: and I would have rather just either picked up the phone or god forbid go to the store and talk to him in person and and browse the shelves.

Eli Tan: And why did you outsource that stuff to a robot if you actually enjoy those human interactions? Well, when I first started using Muse, I thought I might just use it for certain things, like just calling my insurance or just ordering food. But what I found was that I ended up relying on it for everything. And instead of a lot of times just thinking for myself, my first instinct was to go to Muse and ask it instead. At one point, I pulled out my phone to ask Muse to preheat the oven to make the dinner that I was making with all the ingredients and bought me

Natalie Kitroof: Eli. It's like I had just given over all my cognitive ability to this thing and it was concerning like I really felt like you know it had taken over my brain and I think that takeover is really what the companies imagine all of us to be using these kind of agents for. That's the vision that they have.

Eli Tan: It's like I had just given over all my cognitive ability to this thing and it was concerning like I really felt like you know it had taken over my brain and I think that takeover is really what the companies imagine all of us to be using these kind of agents for. That's the vision that they have. And the design of it, they are little Labubu characters. So when you download the app, everyone has an avatar named Muse, but then the first thing it does is it asks you to name it and give it its own personality. My avatar, it name is Ren. Ren's a bird, so it's a little a little bird with these little brown feathers. Okay,

Natalie Kitroof: so these avatars, they're cuddly, they're furry, and the thinking I think is that why would you be worried? It's just a little guy. It's just a little fuzzy guy. It's totally okay if he has your information because he's he's not nefarious. He's not going to go rogue. Oh, this the sweet green order is is ready for you in the lobby. That's what the buzzing was.

Eli Tan: Oh my god, my order is here. I need a humanoid robot to go pick it up because I'm still the one who's going to have to physically go and pick something up.

Natalie Kitroof: Okay.

Eli Tan: Hi. Are you here for Natalie? Oh, for Eli. Yes. Thank you so much. Wow.

Natalie Kitroof: A harvest full. Thank you. Oh my goodness. All right, we got it. I mean, it worked.

Eli Tan: Yeah, it did.

Natalie Kitroof: I didn't doubt you, but there's something about just receiving a package that was completely ordered by an AI that feels different. Like I saw the video of you picking it up, so I knew.

Eli Tan: No,

Natalie Kitroof: I knew it had worked.

Eli Tan: Oh my god. Okay. I mean, I checked. It is in fact a harvest bowl.

Natalie Kitroof: Does it have protein in it? Did it Did it meet all the requirements?

Eli Tan: It does. And it's crazy. The experience is crazy as advertised. Eli, let's just say there is a universe in which everyone starts to have these kinds of experiences all the time where there is truly widespread adoption of these AI assistants. Just to step back, how do you think that would change the way we interact with technology and the role that these tech companies play in our lives? In some ways, I think it makes us even more comfortable with the convenience that technology can offer. But I think the big risk is that we are outsourcing what makes us human. If AI is taking over all the chores, it's also taking over all the decisions. We lose the ability to make decisions for ourselves, and we lose those serendipitous moments with other people in the physical world. And when we rely on a company like Meta for every facet of our daily life, we are essentially giving our entire lives over to big tech. We've already seen how their products, like Instagram reels, for instance, can make us feel like we're detaching from reality instead of connecting more with it.

Natalie Kitroof: Okay. Well, for now, I'm just going to be happy to have my salad, Eli. And don't worry, I'm going to Venmo you. You don't have to send your AI agent after me.

Eli Tan: Well, I already told it to to look out for your Venmo. So, if if it doesn't come through, you'll be hearing from it.

Natalie Kitroof: Perfect. Eli, thanks for coming on the show.

Eli Tan: Thanks so much for having me. You can hear more from our tech journalists on the Times podcast Hard Fork, which just returned with a new guest host, Max Reed. Episodes drop every Friday, so check them out wherever you get your podcasts or on the New York Times app. We'll be right back.

Natalie Kitroof: Here's what else you need to know today. On Monday, President Trump reversed course and announced that his super PAC would now pay for nationwide TV ads that have been promoting him at taxpayer expense. More than $10 million in taxpayer funded ads have run since September, paid for through a contract that used funds from the Department of Homeland Security. At least one of the ads featured footage that originally appeared in a Trump campaign video. Democrats and some Republicans had criticized the ads as a possible violation of federal laws for using public funds for propaganda. Trump said in a social media post on Monday that the ads were quote positive promotion for our great USA and defended them as quote a rather standard thing to do. Today's episode was produced by Stella Tan and Carlos Prito. It was edited by Mark George with help from Michael Benois, contains music by Dan Pal and Marian Lozano and was engineered by Chris Wood. Our theme music is by Wonderland. That's it for the daily. I'm Natalie Kitroth. See you tomorrow.