A.I. Agents: Cute, Cuddly and Maybe Catastrophically Dangerous?

Rogue agents and tech news

0:00There are two things I'm having trouble keeping track of in tech right now. The first is all the new personal assistant agents. There's Meta's Muse. So talking to your muse, it's often the fastest way to get things done, and it is also just really fun and cute. Google's Spark. I can just put my phone away and get on with my day, and Spark works in the background for me. An agent called Instinct, and just announced this week, OpenAI's new product, Dots. Dots are remarkably capable, always-on agents, built to handle really anything you can think of.

0:31The other thing I'm having trouble keeping track of is all the websites and institutions that have been hacked, illicitly accessed, or otherwise meddled with by rogue agents. The AI platform Hugging Face, of course, plus the Australian government's Medicare data portal, the Education Department, the Commerce Department, the Securities Exchange Commission, an obscure German wiki. All of this culminating in OpenAI deciding not to release its latest model this past week. So right now, two things seem to be true. One, agents are the thing that will make AI useful for regular people.

1:04And two, they're maybe going to kill, hack, or otherwise meddle with us all. The president's attitude about this, by the way, is that companies should figure it out themselves.

Hard Fork introductions

1:14This is Hard Fork. I'm Max Reed, and over the next few months, I'm going to be joined by an array of New York Times reporters and other tech experts to discuss some of the biggest stories in tech, AI, and our weird new future. Today, we're talking about rogue agents, helpful agents, cute AI avatars, and whether there's such a thing as self-regulation. I'm joined by Mike Isaac, Aaron Griffith, and Eli Tan.

1:48Hi, guys. Hey. Before we really get into it, Casey and Kevin left me a scrawled note here in the studio. It says, run while you still can. I'm kidding. It says, don't forget to disclose that the New York Times is currently engaged in a lawsuit against OpenAI and Microsoft over alleged copyright violations related to the training of large language models. It also says, check to see if your wife works at Anthropic. My wife does not work at Anthropic.

The White House AI summit

2:18I want to start today with an icebreaker question. Tuesday was a big White House AI Summit luncheon. Trump, as he has wanted to do truth socialed out, a seating chart of it. Can we put the seating chart up so people can see it? Because what I want to ask each of you is, where would you sit? Imagine you're walking in. You've got your tray. You've got a plate with a little pre-made sandwich and a glass of milk and like a green apple on it. Where do you want to sit?

2:48Right out the gate, you know, they already are way over their capacity for women with two. So I think that I probably would be like standing somewhere in the corner, you know, off to the side, like not actually at the table. Eli, where are you sitting? I would be sitting awkwardly and quietly between Chamath and Susie Wiles. I think that's got to be some really good gossip there. I think that's a good, that's a really solid, that's where the tea is. Mike, what about you? I'm a coward.

3:19I just want to be serving shrimp canapes and going in and out of the room and being the guy everyone is delighted to see because he's bringing food and that's about it. I know this is a bold choice. I think right in the Jensen, Mark, Trump circle is actually, it's like shooting the moon. It might be the most tolerable lunch that you could have. I think that if they're in a good mood, you probably get some good stories. You can rap with them about whatever's going on.

3:49I think it's a little dangerous elsewhere at that table. There's some strange, like sitting between J.D. Vance and Jeff Bezos, like what's the vibe there? I don't know. It seems a little scary to me. Unfortunately, none of us were invited, so none of us managed to be at the biggest party of the week. So this was a summit, as we were saying on Tuesday, between the Trump White House and some of the richest and most powerful names in AI, which culminated in a press conference featuring Trump and all the CEOs. Can we play a little bit of the Dario laughing and talking in front of the reporters?

4:20Everybody says it's OK. Be careful. A slowdown to AI. So I just want to hear from you after this big summit where you are today. Look, I mean, I will say what I've always said, which is that AI has incredible benefits. There's the side talk. I love that. The jokes, the laughing, the smirking. As the president has said, you know, whoever wins AI wins. I think that's very important. But I think the technology has very real risks. Reign in with Mark. And, you know, the mechanism of how we address those risks is still under discussion. And then we have Trump forcing Mark to say something to, like, basically answer a question he does not know the answer to in front of everyone.

4:56And don't get it wrong. Mark had a view that was, I thought, really right down, right down the middle. Mark, do you want to discuss what you love? Sure. I mean, I can I can kind of summarize what I think we all signed today. I mean, the upshot of this was that it was a rebranding exercise, right? Like the kind of what they got at the end of it was we're calling AI super intelligence now. That was the big commitment. Well, right. I mean, Trump, along with the rest of the tech industry, maybe they are they're the ones who put this idea in his head, knows that AI needs a rebrand.

5:28That, like, right now people have pretty negative feelings about it, especially associating with the data centers. And then now we have everybody in the industry saying, oh, it might kill us, too. Not great branding. So change the narrative. Super intelligence. Well, my read on the scene was that Mark Zuckerberg seems to have quite the influence over the administration right now from kind of the cordial talking with with Trump to even the term super intelligence. This is something that, you know, Mark's been talking about for the last year.

5:58It almost seemed like it reminded me of like a high school friend group where they've invited the new awkward kid and they're still trying to show that, you know, they've been buddies for a while. And yeah, but but didn't it did almost kind of like endear me to Dario a little bit because I'm like, OK, either he hasn't gotten meeting media training or the media training just hasn't stuck because because like, you know, he's itching in his suit. He's like kind of like making weird faces. It kind of like almost made me feel like, OK, like, you know, the rest of these guys are like very polished and buttoned up and kind of like trying to do their like alpha male thing.

6:33And here's this guy like he can't help but be himself. I kind of like won me over a little bit. I was thinking I'm sure I'm not the first person to make this observation, but I was thinking that Dario really we all know that Trump has a kind of central casting problem that he needs people he interacts with to look like what he expects them to be. And without wanting to be rude about it, I do think Dario, if you are thinking like super intelligent math genius nerd like that is good, like Dario is kind of what you're picturing in your head. And I think that maybe won him over to Trump that Trump was like, I understand this.

7:05I think possibly Sundar had the same kind of effect, a sort of taller, ganglier, awkward version. Can we play the clip of Trump? This is what we're about to hear is Trump complimenting Sundar. This guy is a monster and nobody knows what a great life. A monster. What's your name? You can walk down. What's your name? Just all you have to know is Sundar.

7:30Wow. I don't know if I would want to be called a monster in front of everyone. I don't know. No, it's monster in the Jay-Z Nicki Minaj sense. That's the monster he means. That was for sure the reference.

Self-regulation and Jensen Wong

7:40OK, so in addition to this being a big exercise in branding, which is like the most Trumpian sort of, you know, after effect we could have gotten out of it. But the other thing was that Trump effectively asked them all to self-regulate. And I want to focus on this for a second because self-regulation, the word, the concept has sort of been circulating for the last few months as AI safety has come to the forefront of a lot of people's minds. I'm sure you guys all saw Ezra Klein, our colleague, interviewed Jensen Wong, the CEO of NVIDIA, last week.

8:14Don't ship products until they're in control. It is really quite that simple. And speaking of people who are very influential to Trump, Jensen was sitting right next to Trump and has been extremely influential to the way Trump has been thinking about this. And I have to say, and I want to talk about it with you guys, I found Wong's interview to be sort of refreshing as probably taking it too far. But it was a little different from a lot of what we've heard about CEOs sort of directly in the AI industry.

8:45He's got a very different perspective. And I want to talk about why that is, not just on sort of regulation in general, but how he thinks everyone should think about the dangers of AI. Whitney, can we play the car, the clip about the car if you've got it? Ezra, it's so weird. If a car company competing with all bunch of other car companies, which they are, I'm competing with all kinds of companies, which I am. If I believe that I'm about to launch a product that is unsafe, it is completely in my ability, my power, and my responsibility, and I'm incentivized to do so, to not launch the product.

9:30So Jensen is sort of saying, look, if you're creating these things that are, you know, super dangerous, crazy, you know, going to kill people, why don't you just not release them at all? What did you guys make of this interview? Was it as sort of, did it feel new to me from my perspective, or was this a fairly common view in the Valley, do you think? So it's the, like, the tech industry slogan of, like, you can just do things, but you can just not do things.

10:02I like that, anti-agency. That's a good, that's great. I would argue it's kind of naive and perhaps willfully naive to suggest that everyone is just going to not do stuff if it may be bad or has the capacity to be bad. And, you know, has that gone well for industry so far over many years? Not necessarily. Like, there's a lot of ways you can release things that may be somewhat dangerous or unintentionally dangerous, but it felt very much like he was trying to shift, as what Aaron was saying before, trying to shift this narrative of it's not, these systems are inherently dangerous,

10:42but it's, we are the stewards and the people who should be responsible of shepherding these great systems into the world. And if you're not, if you're OpenAI or Anthropic, really, this is targeted Anthropic, then it's a personal failing of the CEOs for you to release these products. And, like, I just, I'm very curious why you found it refreshing, because I felt like maybe I was being too cynical about it. No, I mean, maybe refreshing is the wrong word. I feel like I've spent the last few months, or few weeks at least, hearing a lot of kind of hand-wringing about what's happening.

11:18I find this sort of the Dario Altman pose a little, I don't know, it just feels, I mean, I suppose in some ways I have the, I have a sort of Jensen-esque, like, okay, well, like, why are you, like, okay, so stop. And I know that there's this elaborate scaffolding of rationalization and justification for why you would continue building AI, even if you think it's got a 10% chance of blowing everything up. But I do think that Jensen, in his heart, I'm just calling him by his first name, that's how you can tell I'm captured.

11:50I want to be clear, I don't, I really disagree with a lot of stuff in this interview. I'm not, I'm not, this is not an endorsement of the take. But I do find him, he's an interesting figure to me because he's not somebody who has come up through the world of machine learning, which has a really specific, as you guys all know, has a really specific kind of culture and a kind of sense of what they're doing and what they're building is. I mean, this is a guy who is a graphics card salesman. That is his job. That is what he's been doing, you know, for 30 years now. And I'm sort of very sympathetic to a guy coming around, you know, if we can extend the car metaphor.

12:23He's like a guy who sells mufflers or whatever. And all of a sudden Volkswagen's saying, you know, if we build cars, like they might all blow up on the highway and kill all of us. And he's like, well, you're stopping me. First of all, like I've been selling these mufflers forever and it's been, I've gotten really rich off of it. It's been great. Why are you messing up a good thing? But also it's like, they're just cars, just build the cars. And I thought this was an interesting contrast with the Bill Gates interview that Ezra did this week, which was another end of this sort of spectrum. You can't rely on the industry to self-regulate here.

12:55I mean, it's just insane. Gates is very pro-regulation. But the place where they both met up was there very much people who see AI as a tool that can be, that needs to be addressed as a tool instead of as a kind of a potential superintelligence. The, you know, the only sort of, the only way it can be dealt with is by properly, is by like teaching it to be a good person, basically, instead of like thinking through it. I don't think that you're just like getting captured by like a more palatable version of the like accelerationist argument.

13:27Or, you know, it's like, it's like the accelerationist, but like minus the like, and also screw the little guy, you know? The slow accelerationist. Yeah, yeah, yeah. It's just like, it's like that same argument, but it's just like a little bit, you know, kind of goes down a little easier because it doesn't have like a menace to it. Yeah, I think that's right. I mean, I definitely am getting, I've like contrarioned myself into like a position that I don't actually hold just by like trying to triangulate my way out of it. But I do think, I mean, the one thing, the one other thing I'll say. It's tough being a contrarian these days. I know, exactly.

13:59The one other thing I'll say, I think that there's a real, the value of having Jensen Wong talk like this is that it feels to me very often like there's a real echo chamber in the AI world. And I'm not saying that like adding more wrong opinions to the echo chamber is necessarily going to somehow produce a right opinion, but it is nice to have to like start to like to, you know, sort of bounce around some ideas that don't have anything to do with the kind of rationalist directed, rationalist oriented, you know, EA oriented, AI safety oriented things.

14:33I mean, the end of the day, I think I more agree with Bill Gates than I do with Jensen Wong about the need for some kind of like sector specific regulation. But I also feel like this is such a fast moving environment. It's really hard to say how we should approach it at all. I think it's really weird that they all have been, not all, but many have been begging the government to regulate them and force them to slow down. And and then they all get together and sign this pinky swear that's like morally binding

15:04that has no teeth to it that says like, we're going to police each other. And then they all sort of like are like, yeah, we did it because because Trump doesn't want actual regulation. And and is that it? Are they going to stop asking for real regulation now? Because they've all sort of like agreed to this very flimsy piece of paper. It's a very good question. And I'll just note that the FTC is now looking into open AI and entropic, which feels a little bit like does the government want them to self-regulate or does it not?

15:36Because the way they seem to want to self-regulate is by effectively forming a cartel. And if the government is going to bring antitrust prosecution against them, then that's not

Security inside AI labs

15:45they're not going to be self-regulating anymore. But I want to follow up. I want to follow that thread a little bit, because part of what's at play here that I think gets ignored by the fact that it's all the CEOs at the summit. And these are interviews with Bill Gates and Jensen Wong is that there's this huge mass of people inside the labs, researchers and workers and engineers who are themselves trying frantically to pump the brakes or to figure out exactly what they're doing and how to do it. But there was a great our colleague, Sheriff Frankel, had a great story this week about

16:16people inside open AI raising the alarm over lack of monitoring, lack of proper security procedures around a bunch of the the hacking tests that have resulted in in stuff like the hugging face incident. And are you guys hearing I mean, is this is this do I have the right like sort of take on this, that there is there is a fair amount of unrest within these companies or if not unrest, at least sort of a sense that the workers are taking it into their own hands? I mean, certainly at open AI, like which I think stands out from all of the other frontier lab

16:50companies and the number of hacks and attempts and the number of things that they've they've found that their their AI agents are doing without their realizing. And, you know, Shira's story is incredible because it focuses on just what her sources call sloppy security practices. It's it's not just like sloppy for a company that's building really powerful AI tools. It's just like sloppy for any company. Like there's this really great quote in there. The guy says, why are you using Slack to build your nuclear Manhattan projects?

17:24Because some white hat hackers got got access to like their Slack logs, amongst other things that they got access to. And and they brought it to open AI, which is standard security practice. And open AI was just like, oh, whatever. And then later kind of came around to it. But it just seems like a company that's not taking this stuff seriously, despite, you know, saying all the right things and talking, telling the government we want to be regulated and you got to slow us down, like behind the scenes, they're kind of they're not really that good at security, it seems like based on this reporting.

17:55So the fact that these leaks are coming from open AI first doesn't surprise me because of the sheer volume of incidents that they've had. I will say, though, that the culture of inside of these companies of employees just like popping off, I think is and this is not to say that they're wrong, but it's like it's very it reminds me of like back in the day at Facebook or the Web 2.0 companies when anyone was saying anything and they were like, you know, speaking up to the boss and like saying

18:26this is wrong. We shouldn't be doing this, which like we're in a very different era of Silicon Valley now where that is not the case inside of Apple and Amazon and Google and Google, especially they were, you know, the the like sort of pinnacle of like employee dissent and message boards and discussions and ultimately culminating in a walkout, ultimately culminating in them shutting that stuff down. What about inside Facebook, say, which is a player, a major player in the AI space? Is there I mean, is it because they're not a frontier lab?

18:58There's just less of a sort of employee culture of of fear and dissent and whatever else? I think I think that the Facebook does feel and you like to talk about this, too. But sorry, meta does feel a little bit different just in a few ways. One, I think the biggest concentration of let's see, let's say people who are like alarmed at what I can do and feel like it's super dangerous is inanthropic is probably going to be inanthropic. Those have also dispersed across open AI and deep mind.

19:31And I feel like the folks inside of meta in general, I'd be curious to hear what you think, Eli, but like are kind of more like foot on the gas, man. Let's just like it's open source. Everyone should be able to use this stuff. And like, yeah, there's a few if we like, you know, take out a few power grids along the way, then like that's the cost of doing business. But I do think in general, they're way more acceleration as compared to the conservatism of some of the main labs. Is that right, Eli? I don't know. I feel like that.

20:02Yeah, I think it's been interesting to watch when when Mark staffed his new AI lab with all these new researchers, giving them hundreds of millions of dollars a year ago. He was bringing in these these types of people from companies like open AI. And now in the last year, you watch them tweet and they're talking about products and flywheels. It's like they've been totally medified. And I think their attitudes have changed a little bit. But I'd like to see I'd like to medify the three of you. I'd like more products and more flywheels on this on this call.

OpenAI scraps GPT 6.1

20:30I want to talk. So but this is all leading somewhere because the other big news or the big news to me, at least this week, was that open AI scrapped the release of its newest AI model, GPT 6.1 Astra. The quote that has been ringing in my head is the head of open AI safety systems, Saatchi Jane, says GPT 6.1 Astra showed higher levels of deception. It wasn't always honest about telling users of the actions it did or didn't take.

21:00Sounds like my worst breakup that I've gone through.

21:04You should have. Yeah, well, this is this is one of those like AI like they can't not talk in like sci fi cliches, but also they are building cliched sci fi, you know, minds. So I guess that's kind of how you have to talk about it. You know, like I'm interested in this in a bunch of different ways. One is to what extent do we think this is about this sort of internal, you know, push? Is it is it that the people at the top at open AI might have released this, but they're they're worried about pushback from inside the lab?

21:36Is it more that they're they're just like brutalized by story after story of the various, you know, hacking, scraping, to use our favorite word, meddling with websites that open AI agents have done? Erin, when you read about this, what was your reaction? They're pacing the frontier.

21:56They are. This is, in fact, them pacing the frontier. I mean, is this self-regulation? I guess. I mean, I don't know. I started seeing all the conspiracy theories about like they're doing this because like they've lost control.

22:11I don't believe that. But yeah, they don't want more bad headlines. They don't want more instances of this. And they're still finding ones. So maybe a bad idea. Yeah. There's also the fact that that morning Anthropic released a new model, basically, and like which continues to bring up the question, like, why are you doing this and releasing new models every day if you're freaking out this much? And I do think inside of open AI, they like playing for keeps and sticking the knife in

22:44their competitors whenever they can. So like positioning oneself as the responsible actor as your main competitor is pushing ahead. I would not put that past how open AI thinks about some of this stuff while also saying, hey, there are people inside that do believe this is like dangerous stuff. So this is an example of rogue agents. One of the things that I thought the Jensen Wong interview sort of underlined quite well, in fact, was the way that part of what we're talking about here is a shift change and just in the last six months, say, where we have a sense of how much more powerful, how much

23:18more capable, and in some ways, how much more dangerous these models can be. But one of the funny things about this technology is that the power and the capability and the danger is also overlaps quite a lot with usefulness. So we also have been hearing a lot more recently about useful AI agents, autonomous AI agents that can do things like change your calendar, respond to emails, you know, reschedule your haircuts, whatever else. Eli handed his life over to the meta one for the sake of the education of all the rest of

23:51us. And after the break, we're going to ask him about that.

Personal assistant agents

24:08Among the funniest things about this week, in addition to OpenAI apologizing to the Australian government for improperly accessing various websites and announcing that its new model was too scary to possibly release, they also announced that they were going to have a personal assistant called Dots. So Dots joins, among others, Instinct, Google Spark, and Meta's newest product, Muse, which has been... Also, there's one called Town. No one else seems to have heard of this one. There's a startup. Town? Does it do something different? No, it's a person. It's just the same. There's a button.

24:38There's actually a myriad. It's from Boston. Yeah. So Dots joins, yeah, Town, it buys Red Sox tickets and schedules bank robberies for you. So Dots joins Instinct, Google Spark, Muse, and Town. All of these have been receiving a ton of interest. They're interesting because they are sort of a little bit less freaky and scary than the AI that, you know, hacks things and does all the math problems for you. They do things like monitor your email, check your calendar. They create diet and exercise plans for you. In some cases, they will even make annoying phone calls that you don't want to make.

Testing Meta's Muse

25:12Eli, I want to talk to you about this because you, as I say, turned your life over to Muse for the last few weeks. What has that been like? I did, yeah. For our listeners, if they're wondering what kind of agent it is, you know, the helpful agents look kind of like labooboos, these companies have decided. And the ones that hack governments, they have, I guess, little hoods and hacker outfits. But yeah, I turned my life over to Muse. I gave it access to my email, my calendar, my bank accounts. I let it do everything for two weeks. And I wrote about it and I found it to be one of the most helpful AI tools that I've

25:46used was my takeover. So for people who haven't opened Muse, like tell me what the, like what's the interface? Like how do I, like it's an app on your phone, right? It's an app on your phone. You download it from the app store. You sign in with, you know, your meta account, Instagram and Facebook. And it looks just like a chat bot. There's like this little avatar at the top who he starts out like a labooboo, but then he asks you to give him a name and mine, his name was Ren. He's a tiny little owl. And you talk to him like a chat bot or it like a chat bot. And then, you know, the little owl types on a little keyboard and it can do, you know,

26:17all the tasks for you and, you know, sign into all these accounts and kind of walks you through it. So what kinds of tasks were you asking it to do? I had it order me groceries for pickup at Whole Foods. I had it at some of the kind of the trial accounts, you know, they have a phone feature where you can use these two personalities, a female voice called Haley and a male voice called Brett, and you can have them make phone calls for you. So I called my dad. Did you name them? Wait, back up. Who named them?

26:48I did not name them. These are the names that Meta gave to these personalities. Oh my God. So I would have Haley call my- They're posing as humans, right? Like they're not saying, I am an AI agent calling. They're saying, I am Haley human calling. They say, I'm Haley and I'm calling on behalf of Eli. I had to call my dental insurance. It gave them my member ID, answered a security question about my birthday, and then it waited on hold. And then 10 minutes later, I pick up my phone and it's Haley calling and saying, hey, I'm about to transfer you to Anthem.

27:18And I go, oh my gosh. But doesn't like Anthem already has that service. So, okay, sorry. I'm jumping ahead here, but I want to hear from you. Like, it feels like you're just managing a bunch of interns, like that you have to then, it makes more work than for you because you have to then tell them, okay, here's all my stuff. Here's how to do it. Here's what I want. And then like, if they mess it up, you have to like fix it. Like, is it, did it really save you that much time? When you're ordering groceries with it, let me just to take an example. So you, you, do you text it, you know, uh, I need, I, can you get me onions, potatoes,

27:51peppers, eggs? You're making a scramble in this case, I guess, like sort of breakfast. Or are you like, I want, I need to make breakfast. Can you just order me some breakfast stuff? In this case, I gave it a New York Times recipe and I said, can you order me all the ingredients you'd need in this recipe? Got to keep it in house. Company man. Company man. But also, the Times app has that feature in Instacart. You can just push a button. Wow. She has one up to you, Eli, in promoting New York Times products. That's true. I also would, I also, every time I would order something online, I'd say, find the

28:21Wirecutter recommended today. Oh, man. Use their referral link because you want to keep all the, you know, I want to keep it all in house. But, um, I mean, one question I have, I had about this because I downloaded it and I opened it up myself is the first thing I encountered is just has a long list of things it's suggesting for you to do. Because it's like, this is a, it's an interesting product where it's not clear. It's not, it wasn't entirely clear to me. And in some ways still kind of wasn't, even after I read this long list, like what exactly I need a personal assistant in my life for. So it has everything from, like you're saying, order groceries for me, create an exercise

28:54plan for me, create a diet plan for me, break a bad habit. You're going through it, Max. Well, yeah, I know. Um, there's a real, uh, there's a sense at least that I had, uh, and I'd be interested like as you interacted with it, that it's the kind of thing that you need to, you need to like get in the group of using it. And then all of a sudden the world kind of starts look, everything in the world starts looking like a problem that your personal agent, that Haley or Brett might solve for you. It's Wren.

29:25Wren is managing Haley and Brett. Wren manages Haley and Brett. I would like to be a fly on the wall for the multi-month meta meetings where they decided on the names Haley and Brett. Anyway, sorry. Tell me more about it, Eli. Yeah, it's definitely the type of thing that meta thinks you'll just use for everything. Personally, after my experiment, I think the thing that I'll use it most for is probably just personal finance, tracking all the things I'm spending. Um, but while I was using it, I mean, it was like, I kind of, I don't want to say I got a little bit of AI psychosis, but like I was using it for everything.

29:56I mean, at one point I remember thinking I went to ask it to, you know, preheat my oven to make the dinner. Like that's the level of, that's like the level of, that is rough. Wow. This is, I think what's a little bit scary about, and like what some of the people who are pushing back on the like anthropomorphizing of all these AI things is that once we assign them like more human like qualities and like give them this kind of mystical, magical ability to like, they know everything, then we just start outsourcing more and more and more of

30:27our like decision making and our, even our thinking to them to the point where, yeah, we're asking it to preheat our oven and like, or the, the guy, the tweet that I saw of the guy who like pasted a bunch of texts into Claude and said, how many words is this? Oh man. And like, you're just like, uh, people, you know, just turning to it for literally everything without like thinking with their brain first. It's like kind of a little bit, uh, scary. That's a side rant though. Um, well, wait, I do want to know, Mike and Erin, you guys, you guys are out there on the bleeding, on the bleeding edge of technology and culture.

30:58What, what are people like are, is, is the, is the Bay Area absolutely rotten with people using agents to do everything they can? Oh yeah. I mean, instinct went super viral before meta launched. Um, all the VCs were using it and going nuts, trying to invest in it, um, and tweeting about, you know, and some of the, the use cases that they tweet about are genuinely like useful because this thing is so persistent. That's what it has that humans can't do. Like, it's not just like, oh, I'm going to tell it to order my groceries, which I can also just do in an app. It's, um, get me this restaurant reservation or like go back and forth with Amazon's, uh,

31:34customer service to make sure that my refund actually came through the kinds of stuff that humans just don't have time or aren't organized enough to like stay on top of. I really do like that restaurant reservations are the number one problem that American technologists are trying to solve so much so that they're breaking resi and these apps are getting banned

The rise of autonomous tools

31:56from them. Like instinct is the, is the hot one right now. They just raised a billion dollars. They're like, uh, they have a wait list. I was at a, I was, I was at an event yesterday and I was speaking for it. And this woman came up after and was like, can you get me on the instinct off the wait list and into the app? But I do think like, it's a thing out here. And I think they, the way that people think out here is like, this is where it's all going. So, uh, so you just need to sort of get on board with it.

32:26And, and rather than ask the questions about, should, uh, should I hand these intimate parts of my life over to this robot, they glue, they go like full Eli mode and say, here is everything I have and go for it. And I, I mean, I don't think that's just an out here thing though. I think that, I think that like handing over your life in exchange for convenience is a thing all across America because people are busy and they're, and they're willing to make that, that privacy to convenience trade-off. But the thing out here that people think, in my opinion, is that they view every, or

33:02I don't want to paint with too broad a brush, but many people view every interaction with a human as a source of tension that can be solved with technology rather than just calling a host and be like, do you have any tables? Eli, did you make it? I want to know if Eli made any restaurant reservations with, with, uh, with Ren. I did make reservations with Ren. And I think something like it's important to remember here, Meta, they're spending 600 billion dollars on this restaurant reservation app. Like, like, yes, it, it's good at restaurant reservations.

33:32It's okay at movie tickets. Like this is a, but this is like for 600 billion dollars, it should teleport me to the restaurant or something. Like this is a lot of, yeah, this is a lot of investment for something that's pretty useful at some stuff. Uh, which I think we, I think we should remember, right? I remember like 10 years ago when we were in the sort of peak of, uh, on-demand app mania when it was like, you know, you have your app for your on-demand butler, who's going to go get your dry cleaning and your on-demand like house cleaner and your on-demand, like everything was, could be accessed by an app.

34:04Um, and it was like this new, you know, millennial lifestyle subsidy utopia. I remember some investor telling me like, there will be a time when you don't make any decisions without first consulting AI or first consulting an app because they will know the best place to get your dry cleaning or the best like time to order your car or like whatever details, any kind of decision that you make is going to be like consulted by AI or by an app. And I remember thinking like, that's insane. That seems like a huge time suck to have to be like negotiating back and forth with this app when I just know the decision in my brain.

34:36But I, I now think we are moving in that direction. Like, and that a lot of people would rather outsource some of these like minor day-to-day minutiae decisions to an AI that knows them really well and has all their info and can just go ahead and execute it. Like I, I, I do really see the appeal there just to, to argue against myself. I mean, one of the, one of the things at, at, at, at stake here is like, uh, the, if everybody has an app, if everybody's got a personal assistant, if everybody has access to the best dry cleaner, the best restaurant, the best, whatever kind of nobody does because you're just in

35:08an arms race against everybody else's personal assistant, everybody else's whatever else, which seems like a, uh, not a great world to live in, in particular. It's a question of how many agents do we need? I saw today, I think it was DoorDash. They released an agent in the app. I mean, are we going to go into every app? Am I going to open the subway app and have a sandwich agent that clicks on whatever I want for lunch? It's called a sandwich artist. Okay. Yes. Thank you, Mike. How many, how many agents do we need? How many apps do we need? I don't know. But that's, that's what happens in eras of tech, right?

35:39Like, remember when, uh, uh, everyone figured out, okay, now you need like not a real AI chat bot, but sort of like a bootleg chat bot at the end, at the bottom of every webpage that eliminates customer service, uh, people. And then those end up being terrible enough that they probably hire them back. And like, everyone kind of does it and then it washes out over time. I don't know. Like, I agree, Aaron, that people are more comfortable now making those trade-offs, but I also think we're going to go through this period of like messiness and then something

36:11comes out in the wash that people like or don't like or agree upon. I think we're just like in that gnarly period right now. It's going to be gnarly for a little while. Like I was talking to this, I was talking to this investor. Yeah. Gnarly for a little while. No, I was going to say, I was talking to this investor who kind of made this point that I keep thinking about, which is that we spent 20 years trying to prevent bots from being able to transact on the internet. And now we're doing the opposite. We're trying to roll it back and allow bots to like actually buy things on the internet.

36:43And like, it just, I think about that every time I see these like weird tensions where like, you know, the instinct bots are like breaking the resi site by essentially giving them a DDoS attack by like hammering at looking for reservations like 10 times a minute. And so like, we're going to, yeah, it's, it's going to be gnarly. It's going to, we're going to see like a lot of, um, you know, companies have to adapt to a world where like agents are roving. This is like one of my favorite things about the various, uh, reports that meter and open AI have been producing about all the rogue agents is how much, how much time and how many

37:16tokens the agents are putting toward trying to solve CAPTCHAs and like failing at them or trying to like find burner phones or whatever. It's, it's like, it really, I mean, they are like some of the funniest science fiction stories I've read in years and years and years, the robots arguing with each other about how to like, how to, how to like steal Bitcoin from people so that they can find a phone so that they can register with the Australian Medicare website so that they can pull some stupid question or whatever. It's really good. I mean, one thing, the other thing about this is I think one of the big stories in AI over the last few years has been open AI in particular, but all the companies have been really throwing

37:50consumer ideas at the wall, trying to get them to stick. And there are some, obviously, you know, chat GPT has however many gazillion users, there are consumer uses, but it's really been an enterprise that we've seen the money start to flow in. And the bots feel a little bit like, okay, maybe we're back in a place where regular people are going to have like this sort of relationship to these things to that end.

Avatars and assistant personalities

38:11This is going somewhere more fun than, uh, than my musings on this. I want to look at some of the avatars. First of all, I want to know which one you chose, Eli, for Ren. Um, and I want to look at them and I want us to see, I want us to see which ones we would choose and which ones we think are the cutest. If we could, if we could have them on the screen, please. And the centerpiece of our vision for what we're building, there he is, is Muse.

38:37What do we think? That's my body type, by the way. This is the personal agent that we shipped a few weeks ago. Joey. Chippy. Boba. Icky. Grumpy. Toodles. Penelope Muse. Frank. Frank. So, sorry, Eli, remind me, which one was Ren or was Ren not represented here? Ren was not represented. So, actually, the name it recommended was Miso, and I said, I'm a 28-year-old man. I can't be talking to Miso, you know, to do my groceries. Can we do another name? And it gave me Ren, which was the bird.

39:08You know, Ren is a bird, so that's what the influence was, yeah. Oh, my gosh. I mean, they're cute, right? Do you guys feel won over by them? Yeah, I mean, come on. Like, everyone's so mad about how ugly the data centers are, and so they're like, let's give us, like, let's make it, like, let's make AI cuddly and cute. But, and also, while also noting that AI will cure cancer.

39:31Like, which one of these little guys is going to be the one who does it? Did anyone else have a Tamagotchi on here? Yeah, I did have a Tamagotchi. Does that mean anything to you, Eli? I know what Tamagotchis are, yeah. I've heard about that. Okay, thank you. Before we started recording, we asked Eli if he knew what the Brady Bunch was. So, he's really mad that we're just asking him about things that happened 25 years before I said my grandma loved the Brady Bunch. He was a big fan. Things that happened in the 1900s. Yeah, exactly. Before the invention of electricity. Tamagotchis were, in fact, the first thing that Edison invented.

40:02So, if you could describe them to our large Zoomer audience, Mike. I mean, look, this is essentially a Tamagotchi. It was like a little thing that you, although I would say it's a reverse Tamagotchi in that you were essentially a parent to this gourd-shaped egg that could do nothing for itself that you had to feed and carry around. And this thing is now parenting you by doing everything for you. Look, we do like how you got there. That's really good. Did you feel like you had companionship with Wren, or was it fully just business?

40:33Like, what can I do for you? Unfortunately, there was a little bit of companionship. We developed a couple inside jokes. Not even joking. Like, there was definitely, you know, it has a little bit of a personality for sure. Wow. What did Wren think of your article?

40:52Wren thought it was interesting. I actually didn't ask it about it. Who brought it up first? I brought it up first. So, it was just part of the article, too. I had it create an AI podcast for me. And it had more AI characters. Maya and Theo were the hosts of that, and it didn't really like how I was kind of making fun of the podcast. Let me ask one sort of last question, which is, should we trust Meta with this kind of access to everything in our lives? And let me ask it in sort of two ways, because there's sort of two questions of trust here.

41:25One is, of course, the question of privacy and data, which is something that Meta has always had a little bit of trouble keeping track of. The other one is, we started this whole discussion talking about rogue agents and agents that are doing things that we didn't really want them to do and trying to do all kinds of things. So, Eli, did you get worried about sort of either of these, any of these things when you were, I mean, you said finances is one of the things that you're handing over to it? I did, yeah. One of the lines for me was that I was willing to connect it to my personal email, but not my work email. I didn't want it to go into there. I mean, I'm Gen Z.

41:55Like, when it comes to privacy, I just don't really have the same issues about it. Like, I just have grown up always assuming, yeah, this data is all of ours, you know? It's this company. You're just okay with being surveilled. Yeah. And, you know, I had a hedge if my bank account got drained, I was going to be able to write about it and put it in the store. That's true. That's a good point. So I knew that would be, you know, a potential kicker or something. But no, I actually wasn't as concerned. And also, when it comes to Meta, you know, we talk about history of all the issues this

42:25company has had. But, you know, with some of the other agents, it's like, would you rather have a company like Meta with 100,000 people that runs Instagram and Facebook with your data? Or would you rather trust Instinct, which has, I think, less than 20 people and is run by 14 people. You know, I would not trust Instinct with my banking data personally. But I would trust Meta in this case. Well, that's the security, like, hacking problem that you're more worried about. And Instinct has had some privacy snafus already where people dug in and they're like,

42:55wait, why did they download a whole copy of my entire Gmail inbox, that's 20 years, to their own servers? Like, that's, I don't want that sitting on a server of a less than one-year-old startup. I kind of see that concern. And I can see why I would maybe trust Meta to have better, like, security infra on that. But I also just, I mean, I have a very different perspective than Eli on, like, my own personal privacy. Like, I don't wish to be further known by these platforms. You know, like, they know enough about me when I was, like, sloppy with my, like, kind

43:27of personal data back in the early 2010s when I was signing up for every single app. And, like, now I'm like, shut it down. Like, I don't, I think a lot of us, maybe us elders, are, like, not posting as much. Never mind. That's actually not true. Mike posts all the time. Yeah, speak for yourself here, buddy. I do think that every generation has a, this is where I get off sort of thing. And I feel like this is, like, the stop that I get off on. I'm not, not able to go there. Maybe I'll be left in the dust as you guys are having, like, Chez Panisse dinners and

43:58I'm shut out of the best reservations. But I can make that tradeoff for my life. I just got, while we were talking, I just got an alert from my Muse app when it's literally asking me to import all of the data that I have with Claude over to it. It says, help me, help me get to know you better. Already use another AI? I can give you a prompt to bring it over. On that note, let's go to another break. Afterwards, I have some very stupid questions about the news that I want to ask you guys. I want to talk about Anthropics S1. I want to talk about Meta's data center tax breaks. And Eli, I have a really important question about buying flowers for your girlfriend.

44:32We'll be back soon.

Anthropic financials and Meta taxes

44:42Okay, so before we go, there's a bunch of stuff that happened this week that we haven't even covered. I have some questions. You guys are the experts. I would like you to either answer my questions or redirect them to more fruitful plays. Aaron, I want to talk to you about Anthropics S1, which leaked this week and showed a bunch

45:19of figures from 2025. The thing that stuck out to me that I was mentioning to you before was that two unnamed customers make up nearly 25% of their revenue, which for me was like, who, obviously one of them is Eli because he's turned over his whole life to everything. But who's the other one? Is it the DoD? Is it Meta? Is that an interesting question? What should I be thinking about this? What should I know about it? Okay, that I guess is an interesting question.

45:47You're so gentle with me. I love it. It would maybe be like my fourth question about it. Like it is a good question. And I think we know who the Anthropics biggest customers, which is Facebook. I think we've, the Times has reported that, right? Right, so I was right, it's Eli, it's Muse, it's Muse is stealing all of Anthropics. And the second biggest one, I think we could probably take a couple guesses. They have, you know, obviously a really big partnership with Amazon and others, but I, and I'm sure other reporters who are more enmeshed in this know this like for sure. But what, what is, I think like the biggest thing to come out of this S1 leak, which incredible

46:21scoop by Echo Wong at Reuters, who also scooped SpaceX's S1 ahead of time. I mean, these things just generally do not leak. So just want to shout out there that Anthropic, which is about to launch probably the largest IPO of all time, the most valuable and the biggest raise is selling itself as having the most transformative, you know, product since electricity or the internet or industrialization with a risk factor saying that it could kill us all.

46:54Like, buy our stock. Awesome. You might die anyway. It's just like, it's just so remarkable. I think we need to be talking about this. It's, it's just like mind bending. And then also the next thing to talk about with this is like their, their business model, which we're learning about for the first time. There's been all kinds of leaks and blah, blah, blah over the years, but this is the first time we've seen like audited financials for 2025 saying, you know, they spent $7 billion on compute last year.

47:24They lost 42 billion, which sounds like really, really insane. But if you kind of parse that out at 34 billion of it was like essentially convertible notes that are, you know, control that, that, that, that's not a real loss. Um, it's kind of an accounting thing, but, um, yeah, I mean, this business is growing super fast and it's spending an insane amount of money. They're going to spend $518 billion on like compute infrastructure obligations in the coming years. Like these are just numbers we've never seen before. So that's what I'm like, wow.

47:57Yeah. All right. Thank you, Aaron. I feel, I feel vastly more informed and I'm going to be repeating that at dinner parties all weekend. Um, Mike and Eli, you guys had a great story this week along with, uh, our friends and colleagues, Kashmir Hill and Jesse Drucker about meta using AI data centers to avoid federal taxes. So, uh, I'm going to try and describe the story for a second. Again, remember in this segment, I'm, I'm just a guy. Um, and I want you, and I want you to help me figure out, uh, exactly how this is working. So basically Meta's building, we know, is building all these data centers to help power

48:30muse and the various other cute Ren, Haley, Brett, all the various Miso, all of our cute little friends. And they're claiming on their taxes, uh, a very specific tax credit for research that this is for sort of like scientific research, experimental scientific research purposes. Now, do they not have a point? Is this not like, you know, this is maybe we're in a bubble. They're still figuring out what the uses for AI are. Like, is this not, do they not get a little bit off the top?

49:01What's going on there? The reason that this is such a big deal is I think even internally folks were like, I don't know if this is going to fly with the IRS if we are audited and it's in their financials saying, maybe we get audited by the tax man because this doesn't fly with them. So they essentially had to disclose that. And I think, you know, the average person could, could probably take issue with our, you know, tens of billions, hundreds of billions at this point, uh, in data in chips, building

49:35these data centers. Is that really experimental? Because it feels like it's a little more than a theoretical project. Yeah. For listeners, the classic example of this tax credit is that if you're a company developing tires, environmentally friendly tires, you can expense the rubber that you're using to build the tires. This really takes it to another level. For me, what's interesting is like the, you know, taxpayers, we have none of the upside of these investments, right? Um, I, I thought back to Trump talking about, oh, maybe we should give people stakes in the AI companies.

50:05And in this case, it's like, you have all of us taxpayers funding billions and billions of dollars of the data centers. Um, but we don't seem to be, you know, benefiting from it. Uh, it seems like companies like Meta are the ones that wouldn't be making a ton of money on it. Your upside is Wren. Yeah, exactly. And Haley. Um, this is a great story. I highly recommend everyone check it out. Eli, you're not off the hook. I have a stupid question for you in your story. And I know I'm not the only person to ask this because I've seen the comments on this

50:37story. You mentioned that you almost used your AI assistant to buy flowers for your girlfriend who had recently had surgery. And instead you bought pottery on Facebook marketplace. Did you eventually buy flowers for her? I did buy flowers for her. And her first question was, did Muse buy these for me? She loves you too well, man. That's a whole other wrinkle. Yeah. Like, does it even mean something if somebody just had their AI assistant do it?

51:08Wow. Yeah. You can't hand that over. Okay, guys, this has been so much fun. This has been an incredible discussion. I feel 10 times smarter. Um, I'm going to go get Muse to write you all thank you letters for coming on. Um, it has been a great chat and, uh, we will be back next week, um, to talk more about technology, AI, the weird new future, anything else that comes up. Um, this has been Hard Fork. I'm Max Reed, and I've been joined by Aaron Griffith, Eli Tan, and Mike Isaac. This episode of Hard Fork was produced by Whitney Jones and edited by Viren Pavic.

52:07With additional editorial support from Brendan Klinkenberg, Lisa Tobin, and Brooke Minters. We were fact-checked by Will Peischel and engineered by Chris Wood. Original music by Marian Lozano, Rowan Nemisto, Alyssa Moxley, and Dan Powell. You can watch this full episode on YouTube at youtube.com slash hard fork. Special thanks to Sam Dolnik, Jessica Testa, Hui Wang Tam, Maddie Macielo, and Dalia Haddad. Over the next few months, we're going to be exploring all aspects of tech, culture, AI,

52:41and the weird new future. We're really interested in what you're interested in. If you have questions you'd like us to answer or things you'd like us to look into, you can email us at hardfork at nytimes.com.