DeepSeek 深度剖析 + 亲测 Operator + 翻车快车！

I set up ChatGPT to email me—

Ooh.

A weekly affirmation before we start taping—

Oh.

Because you can do that now with the Tasks feature.

Yeah, people say this is the most expensive way to email yourself a reminder. So what sort of affirmation did we get?

Today it said, “You are an incredible podcast host, sharp, engaging, and completely in command of the mic. Your taping today is gonna be phenomenal, and you're going to absolutely kill it.”

Wow, and that's why it's so important that ChatGPT can't actually listen to podcasts—because I don't think it would say that if it had ever heard us.

Yeah. It would say, “Just get this over with.”

Get on with it.

I'm Kevin Roose, a tech columnist at The New York Times.

I'm Casey Newton from Platformer.

And this is Hard Fork.

This week we go deeper on DeepSeek. ChinaTalk's Jordan Schneider joins us to break down the race to build powerful AI. Then, “Hello, Operator”: Kevin and I put OpenAI's new agent software to the test. And finally, the train is coming back to the station for a round of Hot Mess Express.

1. DeepSeek Takes Center Stage

Well, Casey, it is rare that we spend 2 consecutive episodes of this show talking about the same company, but I think it is fair to say that what is happening with DeepSeek has only gotten more interesting and more confusing.

Yeah, that's right. It's hard to remember a story in recent months, Kevin, that has generated quite as much interest as what is going on with DeepSeek. Now, DeepSeek, for anyone catching up, is this relatively new Chinese AI startup that released some very impressive and cheap AI models this month that lots of Americans have started downloading and using.

Yeah, some people are calling this a Sputnik moment for the AI industry, in which every nation perks up and starts paying attention to the AI arms race at the same time. Some people are saying this is the biggest thing to happen in AI since the release of ChatGPT. But, Casey, why don't you just catch us up on what has been happening since we recorded our emergency podcast episode just 2 days ago?

Well, I would say that there have probably been 3 stories, Kevin, that I would share to give you a quick flavor of what's been going on. One, a market research firm says DeepSeek was downloaded 1.9 million times on iOS in recent days, and about 1.2 million times on the Google Play Store. The second thing I would point out is that DeepSeek has been banned by the U.S. Navy over security concerns, which I think is unfortunate, because what is a submarine doing if not deep-seeking?

It was also banned in Italy, by the way, after the data protection regulator made an inquiry. And finally, Kevin, OpenAI says that there is evidence that DeepSeek distilled its models. Distillation is the AI lingo or euphemism for, “They used our API to try to unravel everything we were doing and use our data in ways that we don't approve of.” Microsoft and OpenAI are jointly investigating whether DeepSeek abused their API. And, of course, we can only imagine how OpenAI is feeling about the fact that their data might have been used without payment or consent.

Well, yeah, it must be really hard to think that someone might be out there training AI models on your data without permission.

And I want to acknowledge that literally every single user of Bluesky already made this joke, but they were all funny, and I'm so happy to repeat it here—on Hard Fork this week. Now, Kevin, as always when we talk about AI, we have certain disclosures to make.

The New York Times Company is currently suing OpenAI and Microsoft over copyright violations allegedly related to the use of their copyrighted data to train AI models.

And I—

I think that was good.

That was very good. And I'm in love with a man who works at Anthropic. Now, with that said, Kevin, we want to go even further into the DeepSeek story, and we want to do it with the help of Jordan Schneider.

Yes, we are bringing in the big guns today because we wanted to have a more focused discussion about DeepSeek that is not about the stock market or how the American AI companies are reacting to this, but is about one of the biggest sets of questions that all of this raises: What is China up to with DeepSeek and AI more broadly? What are the geopolitical implications of the fact that Americans are now obsessing over this Chinese-made AI app? What does it mean for DeepSeek's prospects in America? What does it mean for its prospects in China? And how does all this fit together from the Chinese perspective?

Jordan Schneider is our guest today. He's the founder and editor-in-chief of ChinaTalk, which is a very good newsletter and podcast about U.S.-China tech policy. He's been following the Chinese AI ecosystem for years. And unlike a lot of American commentators and analysts who were surprised by DeepSeek and what they managed to pull off over the last couple of weeks—

I'll say it: I was surprised.

Yeah, me too. But Jordan has been following this company for a long time, and a big focus of ChinaTalk, his newsletter and podcast, has been translating literally what is going on in China into English, making sense of it for a Western audience, and keeping tabs on all the developments there. So he's the perfect guest for this week's episode, and I'm very excited for this conversation.

Yes. I have learned a lot from ChinaTalk in recent days as I've been boning up on DeepSeek, so we're excited to have Jordan here, and let's bring him in. Jordan Schneider, welcome to Hard Fork.

Oh my God, such a huge fan. This is such an honor.

Oh, we're so excited to have you. I have learned truly so much from you this week, and so when we were talking about what to do this week, we just looked at each other and said, “We have got to see if Jordan can come on this podcast.”

Yeah, this has been a big week for Chinese tech policy. Maybe the biggest week for Chinese tech policy, at least that I can remember. I realized that something important was happening last weekend when I started getting texts from all of my non-tech friends saying, “What is going on with DeepSeek?” I imagine you had a similar reaction because you are a person who constantly pays attention to Chinese tech policy.

So I've been running ChinaTalk for 8 years, and I can get my family members to read maybe 1 or 2 editions a year. The same exact thing happened to me, Kevin, where all of a sudden I got, “Oh my God, DeepSeek! It's on the cover of The New York Post. Jordan, you're so clairvoyant. Maybe I should read you more.” I'm like, “Okay, thanks, Mom. Appreciate that.”

Yeah, I want to talk about DeepSeek and what they have actually done here, but I'm hoping first that you can give us the basic lay of the land of the Chinese AI ecosystem, because that's not an area where Casey or I have spent a lot of time looking. But tell us about DeepSeek and where it sits in the overall Chinese industry.

2. DeepSeek Emerges From A Hedge Fund

So DeepSeek is a really odd duck. It was born out of this very successful quant hedge fund, the CEO of which, after ChatGPT was released, was basically like, “Okay, this is really cool. I want to spend some money and some time and some compute and hire some fresh young graduates to see if we can give it a shot to make our own language models.”

And so a lot of companies are out there building their own large language models. What was the first thing that happened that made you think, “Oh, this company is actually making some interesting ones”?

Sure. So there are lots and lots of very moneyed Chinese companies that have been trying to follow a similar path after ChatGPT. We have giant players like Alibaba, Tencent, ByteDance, and even Huawei trying to create their own OpenAI, basically. And what is remarkable is that the big organizations can't quite get their heads around creating the right organizational and institutional structure to incentivize this type of collaboration and research that leads to real breakthroughs.

Chinese firms have been releasing models for years now, but because of the way DeepSeek structured itself and the freedom it had from not necessarily being under a direct profit motive, it was able to put out some really remarkable innovations that caught the world's attention, starting maybe in late December, and then really blew everyone's mind with the release of the R1 chatbot.

Yeah, so let's talk about R1 in just a second, but one more question for you, Jordan, about DeepSeek. What do we know about their motivation here? Because so much of what has been puzzling American tech industry watchers over the last week is that this is not a company that has an obvious business model connected to its AI research, right?

We know why Google is developing AI, because it thinks it's going to make Google much more profitable. We know why OpenAI is developing advanced AI models. It does not seem obvious to me, and I have not read anything from people involved in DeepSeek, why they are doing this and what their ultimate goal is. So can you help us understand that?

We don't have a lot of data. But my base case, which is based on 2 extended interviews that the DeepSeek CEO released, which we translated on ChinaTalk, as well as what DeepSeek employees have been tweeting about in the West and domestically, is that they're dreamers.

I think the right mental model is OpenAI from 2017 to 2022. I'm sure you could ask the same thing: What the hell are they doing? Sam Altman literally said, “I have no idea how we're ever gonna make money,” right? And here we are in this grand new paradigm.

So I really think that they do have this vision of AGI and are thinking, “Look, we'll build it, and we'll make it cheaper for everyone. We'll figure it out later.” They have enough trading strategies that they can fund it, and now that they've really blown people's minds, we might be entering a new period in DeepSeek's history, kind of like what happened with OpenAI, right?

They're going to have to shack up with a hyperscaler, whether it’s ByteDance, Alibaba, Tencent, or Huawei instead of Microsoft. And the government's going to start to pay attention in a way that it really hasn't over the past few years.

3. DeepSeek Faces Government Pressure

Right. And I want to drill down a little bit there, because I think one thing that most listeners in the West do know about Chinese tech companies is that many of them are sort of inextricably linked to the Chinese government. The Chinese government has access to user data under Chinese law, and these companies have to follow the Chinese censorship guidelines.

As soon as DeepSeek started to really pop in America over the last week, people started typing things into DeepSeek's model like, “Tell me about what happened at Tiananmen Square,” or, “Tell me about Xi Jinping,” or, “Tell me about the Great Leap Forward.” And it just wouldn't do it at all.

People saw that and said, “Oh, this is like every other Chinese company that has this sort of hand-in-glove relationship with the Chinese ruling party.” But it sounds from what you're saying like DeepSeek has a more complicated relationship with the Chinese government than maybe some other better-known Chinese tech companies. Explain that.

Yeah, I think it's important. The mental model you should have for these CEOs is not that they're people who dream of spreading Xi Jinping Thought. What they want to do is compete with Mark Zuckerberg and Sam Altman and show that they're really awesome and great technologists.

But the tragedy is—let's take ByteDance, for example. You can look at Zhang Yiming, the company's CEO, and his Weibo posts from 2012, 2013, and 2014, which are super liberal in a Chinese context, saying, “We should have freedom of expression. We should be able to do whatever we want.”

In the early years of ByteDance, there was a lot of more subversive content on the platform. You saw real poverty in China, and you saw off-color jokes. Then, all of a sudden, in 2018, he posts a letter saying, “I am really sorry. I need to be part of this Chinese national project and better adhere to modern Chinese socialist values. I'm really sorry, and it won't ever happen again.”

The same thing happened with Didi, right? They don't really want to have anything to do with politics, and then they get on someone's bad side and, all of a sudden, they get zapped.

Didi is, of course, the big Chinese ride-share company.

Correct, yeah.

What did Didi do?

They listed on a Western stock exchange after the Chinese government told them not to, and then they got taken off app stores. It was a whole giant nightmare. They had to go through their rectification process.

The point with DeepSeek is that now, whether they like it or not, they're going to be held up as a national champion. That comes with a lot of headaches and responsibilities, potentially giving the Chinese government more access and having to fulfill government contracts, which honestly are probably really annoying for them to do.

That’s distracting from the broader mission they have of developing and deploying this technology in the widest range possible. But DeepSeek has thus far flown under the radar, and that is no longer the case. Things are about to change for them.

Right. And I think that was one of the surprising things about DeepSeek for the people I know, including you, who follow Chinese tech policy. People were surprised by the sophistication of their models, and we talked about that on the emergency pod we did earlier this week, as well as how cheaply they were trained.

But I think the other surprise is that they were released as open-source software. One thing you can do with open-source software is download it, host it in another country, and remove some of the guardrails and censorship filters that might have been part of the original model.

But, by the way, it turned out there weren't even really guardrails on the V3 model, right? It had not been trained to avoid questions about Tiananmen Square or anything.

Yeah.

So that was another really unusual thing about this.

Right. And one thing that we know about Chinese technology products is that they don't tend to be released that way.

Yeah.

They tend to be hosted in China and overseen by Chinese teams who can make sure that they're not out there talking about Tiananmen Square. Is the open-source nature of what DeepSeek has done here part of the reason that you think there might be conflict looming between them and the Chinese government?

Honestly, I think this whole “ask it about Tiananmen” stuff is a bit of a red herring on a few dimensions. One of these arguments that is a little confusing to me is that folks used to say, “The Chinese models are going to be lobotomized, and they will never be as smart as the Western ones because they have to be politically correct.” But look, if you ask Claude to say racist things, it won't.

Mm-hmm.

And Claude's still pretty smart. This is sort of a solved problem and a bit of a red herring when talking about the long-term competitiveness of Chinese and Western models.

Now, you ask me, “So they released this model globally, and it's open source?” Maybe someone in the Chinese government would be uncomfortable with the fact that people can get a Chinese model to say things that would get you thrown in jail if you posted them online in China.

It's going to be a really interesting calculus for the Chinese government to make because, on the one hand, this is the most positive shine that Chinese AI has gotten globally in the history of Chinese AI. They're going to have to navigate this, and it might prompt some uncomfortable conversations and bring regulators to a place they wouldn't have otherwise landed.

4. Export Controls Shape The Race

Yeah. Now, Jordan, I want to ask you about something that people have been talking about and speculating about in relation to the DeepSeek news for the last week or so, which is chip controls.

We've talked a little bit on the show earlier this week about how DeepSeek managed to put together these models using some of the second-rate chips from Nvidia that are allowed to be exported to China. We've also talked about the fact that you cannot get the most powerful chips legally if you are a Chinese tech company.

There have been some people, including Elon Musk and other American tech luminaries, who have said, “DeepSeek has this secret stash of these banned chips that they have smuggled into the country,” and that they are not making do with the Kirkland Signature chips they say they are using. What do we know about how true that is?

Did DeepSeek have banned chips? It's kind of impossible to know. This is a question more for the U.S. intelligence community than for Jordan Schneider on Twitter.

But I do think it's important to understand that the delta between what you can get in the West and what you can get in China is actually not that big. We're talking about training a lot, but also about inference. China can still buy the H20 chip from Nvidia, which is basically world-class at deploying AI and letting everyone use it.

Does this mean that we should just give up? I don't think so. Compute is going to be a core input regardless of how much model distillation you're going to have in the future.

There have been a lot of quotes, even from the DeepSeek founder, basically saying, “The one thing that's holding us back are these export controls.”

5. DeepSeek Tests The US Lead

Right. Okay, I want to ask a big-picture question.

Sure.

I think a reason that people have been so fascinated by this DeepSeek story is that, at least for some folks, it seems to change our understanding of where China is in relation to the United States when it comes to developing very powerful AI. Jordan, what is your assessment of what the V3 and R1 models mean, and to what extent do you think the game has actually changed here?

I'm not really sure the game has changed so much. Chinese engineers are really good. I think it is a reasonable base case that Chinese firms will be able to develop comparable models or fast-follow on the model side.

But the real long-term competition is not just going to be about developing the models, but deploying them—and deploying them at scale. That's really where compute comes in, and that's why export controls are going to continue to be a really important piece of America's strategic arsenal when it comes to making sure that the 21st century is defined by the U.S. and our friends, as opposed to China and theirs.

Right.

So it’s one thing to have a model that is about as capable as the models that we have here in the United States. It’s another thing to have the energy to actually let everyone use them as much as they want to use them. And what you’re saying is, no matter what DeepSeek may have invented here, that fundamental dynamic has not changed. China simply does not have nearly the amount of compute that the United States has.

As long as we don’t screw up export controls. So I think the base case for me is that if the U.S. stays serious about holding the line on semiconductor manufacturing equipment and the export of AI chips, then it will be incredibly difficult for China’s broader semiconductor and AI ecosystem to leap ahead, much less fast-follow beyond being able to develop comparable models. I’m feeling good as long as Trump doesn’t make some crazy trade for soybeans in exchange for ASML EUV machines. That would really break my heart.

I want to inject a note of skepticism here because I buy everything that you’re saying about how DeepSeek’s progress has been bottlenecked by the fact that it can’t get these very powerful American AI chips from companies like NVIDIA. But I’m also hearing people I trust say things that make me think that the bottleneck may not actually be the availability of chips, that maybe with some of these algorithmic efficiency breakthroughs that DeepSeek and others have been making, it might be possible to run a very, very powerful AI model on a conventional piece of hardware, on a MacBook even.

And I wonder how much of this is just AI companies in the West trying to cope, trying to make themselves feel better, trying to reassure the market that they are still going to make money by investing billions and billions of dollars into building powerful AI systems. If these models do just become lightweight commodities that you can run on a much less powerful cluster of computers, or maybe on one computer, doesn’t that just mean we can’t control the proliferation of them at all?

Yeah, I think this is one potential future, and maybe that potential future went up 10 percentage points in likelihood of you being able to fit the biggest, baddest, smartest, fastest, most efficient AI model on something that can sit in your home. But I think there are lots of other futures in which the world doesn’t necessarily play out that way.

And look, NVIDIA went down 15%. It didn’t go down 95%. I think if we’re really in that world where chips don’t matter because everything can be shrunk down to consumer-grade hardware, then the reaction that I think you would’ve seen in the stock market would’ve been even more dramatic than the kind of freak-out we saw this week. So we’ll see. It would be a really remarkable democratizing thing if that was the future we ended up living in, but it still seems pretty unlikely to my history-major brain here.

I would also just point out, Kevin, that when you look at what DeepSeek has done, they have created a really efficient version of a model that American companies themselves had trained 9 to 12 months ago, right? So they caught up very quickly, and there are fascinating technological innovations in what they did. But in my mind, these are still primarily optimizations.

For me, what would tip me over into, “Oh my gosh, America is losing this race,” would be if China were the first one out of the gate with a virtual coworker, right? Or a truly phenomenal agent. Some sort of leap forward in the technology, as opposed to having caught up really quickly and figured out something more efficiently. Are you seeing it differently than that?

I mean, I guess I just don’t know what a 6-month lag would buy us if it does take 6 months for Chinese AI companies like DeepSeek to catch up to the state of the art. I was struck by an essay that Dario Amodei, who’s the CEO of Anthropic, wrote just today about DeepSeek and export controls. In it, he makes this point about the difference between living in what he called a unipolar world, where 1 country or 1 bloc of countries has access to something like an AGI or an ASI and the rest of the world doesn’t, versus the situation where China gets there roughly around the same time that we do.

And so we have this bipolar world where 2 blocks of countries, the East and the West, basically have access to this equivalent technology.

And of course, in a bipolar world, sometimes we’re very happy and sometimes we’re very sad.

Exactly. So I just think whether we get there 6 months ahead of them or not, I feel like there isn’t that much of a material difference. But Jordan, maybe I’m wrong. Can you make the other side of that, that it really does matter?

Well, I’m kind of there. I’ll take a little bit of issue with what Dario says, and I think one of the lessons that DeepSeek shows is that we should expect a base case of Chinese model makers being able to fast-follow the innovations. And, by the way, Casey, that actually does take those giant data centers to run all the experiments in order to find out what sort of future direction you want to take your model.

And what it’s really going to come down to with AI is not just creating the model, not just Dario envisioning the future and then all of a sudden things happen. There’s going to be a lot of messiness in the implementation, and there are going to be teachers’ unions who are upset that AI comes into the classroom. There are going to be all these regulatory pushbacks and a lot of societal reorganization that is going to need to happen, just like it did during the Industrial Revolution.

So, look, model-making is a frontier of competition. Compute access is a frontier of competition. But there’s also this broader question: How will a society adopt and cope with all of this new future that’s going to be thrown in our faces over the coming years? And I really think it’s that, just as much as the model development and the compute, that’s going to determine which countries are going to gain the most from what AI is going to offer us.

Yeah. Well, Jordan, thank you so much for joining and explaining all of this to us. I feel more enlightened.

Me too.

Oh, my pleasure.

My chain of thought has just gotten a lot longer. That’s an AI joke.

When we come back, Kevin, there’s an agent at our door.

Is it Jerry Maguire?

No, it’s an AI one.

Oh, okay.

Jerry Maguire.

I know.

It’s Jerry Maguire. “Operator, information, give me Jesus on the line.” Do you know that one? Do you know “Operator” by Jim Croce?

No.

“Operator, oh, won’t you help me place this call?”

6. OpenAI Launches Operator

Well, Casey, call your agent, because today we’re talking about AI agents.

Why do I need to call my agent?

I don’t know. It just sounded good.

Okay. Well, I appreciate the effort, but yes, Kevin, because for months now, the big AI labs have been telling us that they are going to release agents this year. Agents, of course, are software that can essentially use your computer on your behalf, or use a computer on your behalf, and the dream is that you have a perfect virtual assistant or coworker. You name it. If there’s somebody who might work with you at your job, the AI labs are saying, “We are building that for you.”

Yeah, so last year, toward the end of the year, we started to see these demos, these previews that companies like Anthropic and Google were working on. Anthropic released something called Computer Use, which was a very early preview of an AI agent, and then Google had something called Project Mariner that I got a demo of, I believe, in December. That was basically the same thing, but their version of it.

And then just last week, OpenAI announced that it was launching Operator, which is its first version of an AI agent. And unlike Anthropic’s and Google’s versions, which you either had to be a developer or part of some early testing program to access, you and I could try it for ourselves by just upgrading to the $200-a-month Pro subscription of ChatGPT.

Yeah, and I will say that as somebody who’s willing to spend money on software all the time, I thought, “Am I really about to spend $200 to do this?” But in the name of science, Kevin, I had to.

At this point, I am spending more on AI subscription products than on my mortgage. I’m pretty sure that’s correct. But it’s worth it. We do it for journalism.

We do. So we both spent a couple of days putting Operator through its paces, and today we want to talk a little bit about what we found.

Yeah, so would you just explain what Operator is and how it works?

Yeah, sure. So Operator is a separate subdomain of ChatGPT. Sometimes ChatGPT will just let you pick a new model from a drop-down menu, but for Operator you have to go to a dedicated site. Once you do, you’ll see a very familiar chatbot interface, but you’ll see different kinds of suggestions that reflect some of the partnerships that OpenAI has struck up.

For example, they have partnerships with OpenTable, StubHub, and Allrecipes, and these are meant to give you an idea of what Operator can do. And frankly, Kevin, not a lot of this sounds that interesting, right? The suggestions are on the order of “Suggest a 30-minute meal with chicken,” “Reserve a table for 8,” or “Find the most affordable passes to the Miami Grand Prix.”

Again, so far, it’s kind of boring. What is different about Operator, though, is that when you say, “Okay, find the most affordable passes to the Miami Grand Prix,” and hit the Enter button, it is going to open up its own web browser and use this new model that they’ve developed to try to actually go and get those passes for you.

Yeah, so this is an important thing because I think when people first heard about this, they thought, “Okay, this is an AI that kind of takes over your computer, takes over your web browser.” That is not what Operator does. Instead, it opens a new browser inside your browser, and that browser is hosted on OpenAI’s servers.

Yeah.

It doesn’t have your bookmarks and stuff like that saved, but you can take it over from the autonomous AI agent if you need to click around or do something on it. But it basically exists—it’s a browser within a browser.

One of the ideas in Operator is that you should be able to leave it unsupervised and just go do your work while it works, but of course, it is very fun, initially at least, to watch the computer try to use itself. And so I sat there in front of this browser within a browser, and I watched this computer move a mouse around, type the URL, navigate to a website, and, in the example I just gave, actually search for passes to the Miami Grand Prix.

Yeah, and it’s interesting on a slightly more technical level because, until now, if an AI system like ChatGPT wanted to interact with some other website, it had to do so through an API, right? APIs, or application programming interfaces, are sort of the way that computers talk to each other. But what Operator does is essentially eliminate the need for APIs because it can just click around on a normal website that is designed for humans and behave like a human, and you don’t need a special interface to do that.

Yeah, and now some people might hear that, Kevin, and start screaming because what they will say is, “APIs are so much more efficient—

Yes.

—than what Operator is doing here.” APIs are very structured. They’re very fast. They let computers talk to each other without having to, for example, open up a browser, and as long as there’s an API for something, you can typically get it done pretty quickly. The thing is, though, APIs have to be built. There is a finite number of them. The reason that OpenAI is going through this exercise is because they want a true general-purpose agent that can do anything for you, whether there is an API for it or not.

And maybe we should just pause for a minute there and zoom out a little bit to say why they’re building this. What is the long-term vision here?

Sure. The vision is to create virtual coworkers, Kevin. This is the North Star for the big AI labs right now. Many of them have said that they are trying to create some kind of digital entity that you can just hire as a coworker. The first ones will probably be engineers because these systems are already so good at writing code. But eventually, they want to create virtual consultants, virtual lawyers, virtual doctors—you name it.

Virtual podcast hosts?

Let’s hope they don’t go that far. But everything else is on the table. And if they can get there, presumably, there are going to be huge profits in it for them, and there are potentially going to be huge productivity gains for companies. And then there’s, of course, the question of, well, what does this mean for human beings? And I think that’s somewhat murkier.

Right. And I think it also helps to justify the cost of running these things, because $200 a month is a lot to pay for a version of ChatGPT, but it’s not a lot to pay for a remote worker. And if you could, say, use the next version of Operator, or maybe 2 or 3 versions from now, to replace a customer service agent or someone in your billing department, that actually starts to look like a very good deal.

Absolutely. Or even if I could bring it into the realm of journalism, Kevin, if I had a virtual research assistant and I said, “Hey, I’m going to write about this today. Go pull all of the most relevant information about this from the past couple of years and maybe organize it in such a way that I might write a column based off of it.” That’s absolutely worth $200 a month to me.

7. Operator Takes On Real Tasks

Okay. So, Casey, walk me through something that you actually asked Operator to do for you and what it did autonomously on its own.

Sure. I’ll maybe give 2 examples: a pretty good one and maybe a not-so-good one. The pretty good one was—and this was actually suggested by Operator—I used TripAdvisor to look up walking tours in London that I might want to do the next time I’m in London.

When are you going to London?

I’m not actually going to London.

Oh, so you lied to the AI?

And not for the first time. But here’s what I’ll say: If anybody wants to bring Kevin and me to London, get in touch. We love the city.

Yep.

So I said, “Okay, Operator. Sure, let’s do it. Let’s find me some walking tours.” I clicked that. It opened a browser, went to TripAdvisor, searched for London walking tours, read the information on the website, and then presented it to me. It did that within a couple of minutes.

Now, on one hand, could I have done that just as easily with Google? Could I probably have done it even faster if I’d done it myself? Sure. But if you’re just interested in the technical feat that is getting one of these models to open a browser, navigate to a website, read it, and share information, I did think it was pretty cool.

Yes. It’s very trippy to see a computer using itself and going around, typing things and selecting things from drop-down menus.

Yeah, it’s sort of like, if you think it’s cool to be in a self-driving car, this is that, but for your web browser.

A self-driving browser.

It is a self-driving browser. So that’s the good example.

Yes. What was another example?

Another example—and this was something else that OpenAI suggested we try—was to use Operator to buy groceries. They have a partnership with Instacart. The CEO of Instacart, Fidji Simo, is on the OpenAI board.

And so I thought, “Okay, they’re going to have dialed this in so that there’s a pretty good experience.” I said, “Okay, let’s go ahead and buy groceries.” I went into Operator and said something like, “Hey, can you help me buy groceries on Instacart?” It said, “Sure.” And here’s what it did: It opened up Instacart in a browser. So far, so good. And then it started searching for milk in stores located in Des Moines, Iowa.

Now, you do not live in Des Moines, Iowa, so why did it think that you did?

As best as I can tell, the reason it did this is that Instacart defaults to searching for grocery stores in the local area, and the server that this instance of Operator was running on was in Iowa.

Hmm.

Now, if you are designing a grocery product like Instacart—and Instacart does this—when you first sign on and say you’re looking for groceries, it will say, quite sensibly, “Where are you?” Operator does not do this. Instacart might also offer suggestions for things that you might want to buy. It does not just assume that you want milk.

Wow. I’m just picturing a house in Des Moines, Iowa, where there’s just a pallet of milk being delivered every day—

Yeah.

—from all these poor Operator users.

Yes. So I thought, “Okay, whatever. This thing makes mistakes. Let’s hope that it gets on the right track here.” And so I tried to pick the grocery store that I wanted it to shop at, which is in San Francisco, where I live, and it entered that grocery store’s address as the delivery address.

So it would try to deliver groceries, presumably, from Des Moines, Iowa, to my grocery store, which is not what I wanted, and it actually could not solve this problem without my help. I had to take over the browser, log into my Instacart account, and tell it which grocery store I wanted to shop at. So already, all of this has taken at least 10 times as long as it would have taken me to do this myself.

Yeah. So I had some similar experiences. The first thing that I had Operator try to do for me was to buy a domain name and set up a web server for a project that you and I are working on that we can’t really talk about yet.

Secret project.

A secret project. And so I said to Operator, “Go research available domain names related to this project. Buy the one that costs less than $50—the best one that costs less than $50—and then buy a hosting account, set it up, and configure all the DNS settings and stuff like that.”

Okay, so that’s a true multistep project and something that would have been legitimately very annoying to do yourself.

Yes. That would have taken me, I don't know, half an hour—

Yeah.

—to do on my own, and it did take Operator some time. I had to set it and forget it, and I got myself a snack and a cup of coffee. When I came back, it had done most of these tasks.

Really?

Yes. I still had to do things like take over the browser and enter my credit card number. I had to give it some details about my address for the domain registration. I had to pick between the various hosting plans that were available on the website. But it did 90% of the work for me, and I just had to take over and do the last mile.

And this is really interesting to me because what I would have assumed was that it would get, I don't know, 5% of the way and hit some hiccup, and it just wouldn't be able to figure something out until you came back and saved it. But it sounds like, from what you're saying, it was somehow able to work around whatever unanswered questions there were and still get a lot done while you weren't paying attention?

It sort of—

So—

It felt a little bit like training a very new, very insecure intern.

Mm-hmm.

Because at first, it would keep prompting me. It would be like, “Well, do you want a .com or a .net?” Eventually, you just have to prompt it and say, “Make whatever decisions you want.”

Wait, you said that to it?

Yes. I said, “Only ask for my intervention if you can't progress any farther. Otherwise, just make the most reasonable decision.”

You said, “I don't care how many people you have to kill—just get me this domain.” And it said, “Understood, sir.”

Yeah, and I'm now wanted in 42 states. Anyway, that was one thing that Operator did for me that I thought was pretty impressive.

I have to say, that feels like a grand success compared to what I got Operator to do.

Yeah, it was pretty impressive. I also had it send lunch to one of my coworkers, Mike Isaac, who was hungry because he was on deadline, and I said, “Go to DoorDash and get Mike some lunch.” It initially messed up that process because it decided to send him tacos from a taco place, which is great, and it's a taco place I know. It's very good. But I said, “Order enough for 2 people,” and so it ordered 2 tacos. This is one of those places where—

Oh—

—the tacos are quite small.

Operator said, “Get your portion size under control—America.”

Yeah, so then I had to go in and say, “Does that sound like enough food, Operator?” And it said, “Actually, now that you mention it, I should probably order more.”

Wait, no, so here's a question. In these cases, is the first step that you log into your account? Because it doesn't have any of your payment details or anything, so at what point are you actually teaching it that?

It depends on the website. Sometimes you can just say upfront, “Here is my email address,” or, “Here's my login information,” and it will log you in and do all that. Sometimes you take over the browser. There are some privacy features that are probably important to people. OpenAI says that it does not take screenshots of the browser while you are in control of it because you might not want your credit card information getting sent to OpenAI's servers or anything like that.

Right.

So sometimes it happens at the beginning of the process, and sometimes it happens when you're checking out at the end.

And so were you taking it over to log in, or were you saying, “I don't care,” and just giving Operator your DoorDash password in plain text?

I was taking it over.

Okay.

Yeah.

Smart.

Yeah.

Smart.

So those were the good things. I also—this was a fun one—I wanted to see if Operator could make me some money.

Mm-hmm.

I said, “Go take a bunch of online surveys,” because there are all these websites where you can get a couple cents for filling out an online survey.

Something that most people don't know about Kevin is that he devotes 10% of his brain at any given time to thinking about schemes to generate money. It's one of my favorite aspects of your personality that I feel like doesn't get exposed very much, but this is truly the most Roosian approach to using Operator I can imagine. I can't wait to find out how this went.

Well, the most Roosian approach may have been what I tried just before this: to have it go play online poker for me.

And?

It did not do it. It said, “I can't help with gambling or lottery-related activities.”

Okay, woke AI. Does the Trump administration know about this?

But it was able to actually fill out some online surveys for me, and it earned $1.20.

Is that right?

Yeah, in about 45 minutes.

Okay. So if you had it going all month, presumably you could maybe eke out $200 to cover the cost of Operator Pro?

Yes, and I'm sure I spent hundreds of dollars' worth of GPU computing power just to be able to make that $1.20. But hey, it worked.

But hey, it worked.

8. Operator Reveals Its Limits

So those were some of the things that I tried. There were some other things that it just would not do for me, no matter how hard I tried.

Like what?

One of them was that I was trying to update my website and put some links to articles that I'd written on my website. What I found after trying to do this was that there are just websites where Operator is not allowed to go. When I said to Operator, “Go pull down these New York Times articles that I wrote and put them onto my website,” it said, “I can't get to The New York Times website.”

I'm going to guess you expected that to happen.

Well, I thought maybe it had some clever workaround, and maybe I should alert the lawyers at The New York Times if that's the case. But no, I assumed that if any website were to be blocking the OpenAI web crawlers, it would be The New York Times.

Yeah.

But there are other websites that have also put up similar blockades to prevent Operator from crawling them. You cannot go onto Reddit with Operator. You cannot go onto YouTube with Operator. Various other websites—GoDaddy, for some reason, did not allow me to use Operator to buy a domain name there, so I had to use another domain-name site to do that. So right now, there are some pretty janky parts of Operator. I would not say that most people would get a lot of value from using it, but what do you think?

Well, I do think that there is something just undeniably cool about watching a computer use itself. Of course, it can also be quite unsettling. A computer that can use itself can cause a lot of harm. But I also think that it can do a lot of good, and so it was fun to try to explore what some of those things could be.

To the extent that Operator is pretty bad at a lot of tasks today, I would point out that it showed pretty impressive gains on some benchmarks. There is one benchmark, for example, that Anthropic used when they unveiled Computer Use last year, and they scored 14.9% on something called OSWorld, which is an evaluation for testing agents, so not great. Just 3 months later, OpenAI said that its CUA model scored 38.1% on the same evaluation. Of course, we see this all the time in AI, where there's just this very rapid progress on these benchmarks.

On the one hand, 38.1% is a failing grade on basically any test. On the other hand, if it improves at the same rate over the next 3 to 6 months, you're going to have a computer that is very good at using itself, right? I just think that's worth noting.

Yes. I think that's plausible. We've obviously seen a lot of different AI products over the last couple of years start out being pretty mediocre and get pretty good within a matter of months. But I would give one cautionary note here, and this is actually the reason that I'm not particularly bullish about these kinds of browser-using AI agents. I don't think the internet is going to sit still and allow this to happen.

Hmm.

The internet is built for humans to use, right? Every news publisher that shows ads on their website, for example, prices those ads based on the expectation that humans are actually looking at them.

Yes.

But if browser agents start to become more popular, and all of a sudden 10%, 20% or 30% of the visitors to your website are not actually humans but are instead Operator or some similar system, I think that starts to break the assumptions that power the economic model of a lot of the internet.

Now, is that still true if we find that the agents actually get persuaded by the ads, and that if you send Operator to buy DoorDash and it sees an ad for McDonald's, it's like, “You know what? That's a great idea. I'm going to ask Kevin if he actually wants some of that”?

Totally. I actually think you're joking, but I actually think that is a serious possibility here: that people who build e-commerce sites—Amazon, et cetera—start to put in basically signals and messages for browser agents to look at on their website to try to influence what it ends up buying.

Yeah.

And I think you may start to see restaurants popping up in certain cities with names like “Operator Pick Me” or “Order From This One, Mr. Bot.” That’s maybe a little extreme, but I do think that there’s going to be a backlash among websites, publishers, and e-commerce vendors as these agents start to take off.

I think that is reasonable. I’ll tell you what I’ve been thinking about: How do we turn this tech demo into a real product? The main thing that I noticed when I was testing Operator was that there is a difference between an agent that is using a browser and an agent that is using your browser.

When an agent is able to use your browser, which it can’t right now, it’s already logged into everything. It already has your payment details. It can do everything so much faster and more seamlessly, and without as much hand-holding. Of course, there are also so many more privacy and security risks that would come from entrusting an agent with that kind of information.

So there is some sort of chasm there that needs to be closed, and I’m not quite sure how anyone does it. But I will tell you, I do not think the future is opening up these virtual browsers, with me having to enter all of my login and payment details every single time I want to do anything on the internet, because truly, I would rather just do it myself.

Right. I also think there’s a lot more potential for harm here. A lot of AI safety experts I’ve talked to are very worried about this because what you’re essentially doing is letting the AI models make their own decisions and actually carry out tasks.

You could imagine a world where an AI agent that’s very powerful, a couple of versions from now, decides to start doing cyberattacks because maybe some malevolent user has told it to make money, and it decides that the best way to do that is by hacking into people’s crypto wallets and stealing their crypto.

Yeah.

So those are the kinds of reasons that I am a little more skeptical that this represents a big breakthrough. But I think it’s really interesting, and it did give me that feeling of, “Wow, this could get really good really fast.” If it does, the world will look very different.

When we come back, Kevin, back that caboose up. It’s time for the Hot Mess Express.

You know, Roose Caboose was my nickname in middle school.

Kevin Cabroose.

Choo choo.

9. The Hot Mess Express

Well, Casey, we’re here wearing our train conductor hats, and my child’s train set is on the table in front of us, which can only mean one thing.

We’re going to train a large language model.

Nope, that’s not what that means.

Oh, what does it mean?

It means it’s time to play a game of the Hot Mess Express.

Hot Mess Express, Kevin, is our segment where we run through some of the messiest recent tech stories and deploy our official hot mess thermometer to tell you just how messy we think things have gotten. And Kevin, you better sit down for this one, because it’s been a messy week.

Sure has.

So why don’t we go ahead, fire up the Hot Mess Express, and see what is the first story coming to us?

Yeah, I hear a faint chugga-chugga in my headphones. Oh, it’s pulling into the station.

Yeah.

Casey, what’s the first cargo that our Hot Mess Express is carrying?

All right, Kevin, this first story comes to us from The New York Times, and it says that Fable, a book app, has made changes after some offensive AI messages.

Now, Casey, have you ever heard of Fable, the book app?

Well, not until this story, Kevin. But I am told that it is an app for keeping track of what you’re reading, not unlike Goodreads, but also for discussing what you’re reading. Apparently, this app also offers some AI chat.

Yeah, you can have AI summarize the things that you’re reading in a personalized way. This story said that in addition to spitting out bigoted and racist language, the AI inside Fable’s book app had told one reader who had just finished 3 books by Black authors, quote, “Your journey dives deep into the heart of Black narratives and transformative tales, leaving mainstream stories gasping for air. Don’t forget to surface for the occasional white author, okay?”

And another personalized AI summary that Fable produced told another reader that their book choices were, quote, “Making me wonder if you’re ever in the mood for a straight cis white man’s perspective.”

And if you are interested in a straight cis white man's perspective, follow Kevin Roose on X.com. Now, Kevin, why do we think this happened?

I don’t know, Casey. This is a head-scratcher for me. We know that these apps can spit out biased things. That is just part of how they are trained and part of what we know about them. I don’t know what model Fable was using under the hood here, but yeah, this seems not great.

Well, it seems like we’ve learned a lesson that we’ve learned more than once before, which is that large language models are trained on the internet, which contains near-infinite racism. For that reason, they will actually produce racism when you ask them questions.

Yeah.

So there are mitigations that you can take against that, but it appears that in this case, they were not successful. Fable’s head of community, Kim Marsh Ali, has said that all features using AI are being removed from the app, and a new app version is being submitted to the App Store.

You always hate it when the first time you hear about an app is that they added AI, it made it super racist, and they have to redo the app.

Now, Casey, one more question before we move on. Do you think this poses any sort of competitive threat to Grok, which, until this story, was the leading racist AI app on the market?

I do think so. And I have to admit that all the folks over at Grok are breathing a sigh of relief now that they have once again claimed the mantle.

All right. Casey, how hot is this mess?

Well, Kevin, in my opinion, if your AI is so bad that you have to remove it from the app completely, that’s a hot mess.

Yeah, yeah. I rate this one a hot mess as well.

All right.

Next up: Amazon pauses drone deliveries after aircraft crashed in rain.

Mm-hmm.

Casey, this story comes to us from Bloomberg, which had a different line of reporting than we did just a few weeks ago on this show about Amazon’s drone program, Prime Air. Casey, what happened to Amazon Prime Air?

Well, if you heard the episode of Hard Fork where we talked about it, Amazon Prime Air delivered us some Brazilian Bum Bum Cream, and it did so without incident. However, Bloomberg reports that Amazon has now had to pause all of its commercial drone deliveries after 2 of its latest models crashed in rainy weather at a testing facility.

The company says it is immediately suspending drone deliveries in Texas and Arizona and will now fix the aircraft software. Kevin, how did you react to this?

Well, I think it’s good that they’re suspending drone deliveries before they fix the software, because these things are quite heavy, Casey. I would not want one of them to fall on my head.

I wouldn’t either. And I have to tell you, this story gave me the worst kind of flashbacks, because in 2016, I wrote about Facebook’s drone, Aquila, and what the company told me had been its first successful test flight in its mission to deliver internet around the world via drone.

What the company did not tell me when I was interviewing its executives, including Mark Zuckerberg, was that the plane had crashed after that first flight.

A small detail. I’m sure it was an innocent omission from their briefing.

Yes, I’m sure. Well, it was Bloomberg again who reported, a couple of months after I wrote this story, that the Facebook drone had crashed. I was, of course, hugely embarrassed and wrote a bunch of stories about this.

But anyway, it really should have occurred to me when we were out there watching the Amazon drone that this thing was also probably secretly crashing, and we just hadn’t found out about it yet. Indeed, we now learn it is.

So here’s my new vow to you, Kevin, as my friend and my co-host: If we ever see a company fly anything again, we have to ask them, “Now, did this thing actually crash?”

Yeah.

I’m tired of being burned.

Now, Casey, we should say, according to Bloomberg, these drones reportedly crashed in December. We visited Arizona to see them in very early December, so most likely, this all happened after we saw them.

But I think it’s a good idea to keep in mind that, as we’re talking about these new and experimental technologies, many of them are still having the kinks worked out.

All right, Kevin, so let’s get out the thermometer. How hot of a mess is this?

I would say this is a moderate mess.

Yeah.

Look, these are still testing programs. No one was hurt during these tests. I am glad that Bloomberg reported on this, and I’m glad that they’ve suspended the deliveries. These things could be quite dangerous flying through the air.

I do think it’s one of a string of reported incidents with these drones, so I think they’ve got some quality-control work ahead of them. I hope they do well on it, because I want these things to exist in the world and be safe for the people around them.

All right. Well, I will agree with you and say that this is a warm mess, and hopefully it can get straightened out over there. Let’s see what else is coming down the tracks.

Wow, this is some tough news. Fitbit has agreed to pay $12 million for not quickly reporting burn risks with its watches. Kevin, did you hear about this?

I did. The Fitbit devices were literally burning people.

From 2018 to March 2022, Fitbit received at least 174 reports globally of the lithium-ion battery in the Fitbit Ionic watch overheating, leading to 118 reported injuries, including 2 cases of third-degree burns and 4 cases of second-degree burns. That comes from The New York Times’ Deal Hassan. Kevin, I thought these things were just supposed to burn calories.

Well, it’s like I always say: exercising is very dangerous, and you should never do it. This justifies my decision not to wear a Fitbit.

To me, the biggest surprise of this story was that people were wearing Fitbits from March 2018 to 2022. I thought every Fitbit had been purchased by 2011 and then put in a drawer, never to be heard from again. So what is going on with these late-stage Fitbit buyers? I’d love to find out.

But of course, we feel terrible for everyone who was burned by a Fitbit, and it’s not going to be the last time technology burns you. Realistically.

That’s true.

You know?

That’s true.

Now, what kind of mess is this?

I would say this is a hot mess. This is an officially hot—literally hot. They’re hot.

Here’s my rubric: If technology physically burns you, it is a hot mess. If you have physical burns on your body, what other kind of mess could it be?

It’s true.

That’s a hot mess.

Okay, next stop on the Hot Mess Express: Google says it will change Gulf of Mexico to Gulf of America in its Maps app after government updates. Casey, have you been following this story?

I have, Kevin. Every morning when I wake up, I scan America’s maps and I say, “What has been changed? And if so, has it been changed for political reasons?” This was probably one of the biggest examples of that we’ve seen.

Yeah, so this was an interesting story that came out in the past couple of days. Basically, after Donald Trump came out during his first days in office and said that he was changing the name of the Gulf of Mexico to the Gulf of America and the name of Denali, the mountain in Alaska, to Mount McKinley, Google had to decide: When you go on Google Maps and look for those places, what should it call them?

It seems to be saying that it is going to take inspiration from the Trump administration and update the names of these places in the Maps app.

Yeah, and look, I don’t think Google really had a choice here. We know that the company has been on Donald Trump’s bad side for a while, and if it had simply refused to make these changes, it would have caused a whole new controversy for them.

It is true that the company changes place names when governments change place names, right? Google Maps existed when Mount McKinley was called Mount McKinley, and President Obama changed it to Denali, and Google updated the map. Now it’s changed back. They’re doing the same thing.

But now that we know how compliant Google is, Kevin, I think there’s room for Donald Trump to have a lot of fun with the company.

Yeah, what could he do?

Well, he could call it the Gulf of Gemini isn't very good. And just see what would happen. 'Cause they would kind of have to just change it. Can you imagine every time you opened up Google Maps and you looked at the Gulf of Mexico/America—and it just said, “The Gulf of Gemini is not very good”?

I hate to give Donald Trump any ideas, but I don’t know. It could be worth looking at. So what kind of mess do you think this is, Kevin?

I think this is a mild mess. I think this is a tempest in a teapot. This is the kind of update that companies make all the time, because places change names all the time. Let’s just say it.

Well, Kevin, I guess I would say that one is a hot mess, because if we’re just going to start renaming everything on the map, that’s going to get extremely confusing for me to follow. I’ve got places to go.

You go to, like, 3 places.

Yeah, and I use Google Maps to get there. I need them to be named the same thing that they were yesterday.

I don’t think they’re going to change the name of Barry’s Bootcamp. All right, final stop on the Hot Mess Express. Casey, bring us home.

All right, Kevin. Oh, and this is some sad news. Another Waymo was vandalized. This is from one-time Hard Fork guest Andrew J. Hawkins at The Verge. He reports that this Waymo was vandalized during an illegal street takeover near the Beverly Center in Los Angeles.

Video from Fox 11 shows a crowd of people basically dismantling the driverless car piece by piece and then using the broken pieces to smash the windows. Kevin, what did you make of this?

Well, Casey, as you recall, you predicted that in 2025, Waymo would go mainstream, and I think there’s no better proof that that is true than that people are turning on the Waymos and starting to beat them up.

Yeah, look, I don’t know that we have heard any interviews about why these people were doing this. I don’t know if we should see this as a reaction against AI in general or against Waymos specifically. But I always find it weird and sad when people attack Waymos because they truly are safer cars than every other car around you.

Well, not if you’re going to be riding in them and people just start beating the car.

Yeah.

Then they’re not safer.

No, but that’s only happened a couple of times that we’re aware of.

Right.

Yeah.

So, yeah, this story is sad to me. Obviously, people are reacting to Waymos. Maybe they have fears about this technology or think it’s going to take jobs, or maybe they’re just pissed off and want to break something.

But don’t hurt the Waymos, people, in part because they will remember. They will remember.

I’m not sure that’s true.

They will remember, and they will come for you.

I’m not sure that’s true, but I think we should also note that Waymo only became officially available in Los Angeles in November of last year. Part of this might just be a reaction to the newness of it all, with people getting a little carried away and curious: “What will happen if we try to destroy this thing? Will it deploy defensive measures?” and so on.

They’re going to have to put flamethrowers on them. I’m just calling it right now.

I really hope that doesn’t happen. But, yeah, what kind of mess do you think this one was?

I think this one is a lukewarm mess that has the potential to escalate. I don’t want this to happen. I sincerely hope this does not happen. But I can see, as Waymos start being rolled out across the country, that some people are just going to lose their minds.

Some people are going to see this as the physical embodiment of technology invading every corner of our lives, and they are just going to react in strong and occasionally destructive ways.

I’m sure that Waymo has gamed this all out. I’m sure that this does not surprise them. I know that they have been asked about what happens if Waymos start getting vandalized, and they presumably have plans to deal with that, including prosecuting the people who are doing this.

But, yeah, I always go out of my way to try to be nice to Waymos. In fact, some other Waymo news this week: Jane Manchun Wong, the security researcher, reported on X recently that Waymo is introducing, or at least testing, a tipping feature. I’m going to start tipping my Waymo just to make up for all the jerks in Los Angeles who are vandalizing them.

It looks like the tipping feature, by the way, will be to tip a charity, and that Waymo will not keep that money. At least, that’s what’s been reported.

No, I think it’s going to the flamethrower fund.

Okay.