GPT-5 反弹 + Perplexity CEO Aravind Srinivas 谈浏览器大战 + Hot Mess Express

I saw something new this week.

What did you see?

I was on a flight. I went to the East Coast for a wedding last weekend.

Mm-hmm.

And on the flight back, I saw a woman play Balatro, the mobile phone game—

Mm-hmm.

—for 6 hours.

Honestly, one of the least surprising things you've ever said to me on this podcast, because I've absolutely played Balatro for multiple hours in a row.

She did not look up. She did not get a drink. She did not go to the bathroom. She was locked into her phone for the entire flight, and I think this game should be outlawed. I've never even really played Balatro.

Yeah.

You tried to get me into it. But something that they're putting in that game is driving people to madness.

It is the perfect phone-based game because it can fill up any amount of time from 30 seconds to 6 hours. And that is just a precious thing.

Yeah.

So I have wasted many hours on a flight with Balatro. And for what it's worth, I do not experience this game as something that's so addictive that I can't put it down. I experience it as, “Oh, I got some time to kill. I know the perfect thing that will help me do that.” But as soon as I'm with a friend, I'm not thinking, “Oh, I gotta get back to Balatro.”

Yeah.

One time, my boyfriend's friends were over, and there was a lot of discussion back and forth about what kind of takeout we should order. It was clear that I was not really going to be steering this decision, and I started thinking, “I'm halfway through a Balatro run.” So I got my phone out of my pocket, played a couple of hands, and afterward my boyfriend said, “It would be great if you didn't play Balatro while my friends were over.” He was right, and I apologized. Yeah.

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, the backlash against GPT-5 and what AI companies are learning from the fallout. Then Perplexity CEO Aravind Srinivas returns to the show to discuss his $34 billion bid to buy Google Chrome. And finally, I hear that train a'coming, Kevin. The Hot Mess Express has returned.

Chugga-chugga, choo-choo.

The caboose is loose.

1. The GPT-5 Backlash

Well, Casey, it's been a busy week on the internet for AI companies and the backlash to them.

That's right, Kevin. Basically every day since we were last in the studio, there has been a big piece of news, most of it related in one way or another to GPT-5.

Yes. Let's talk about the GPT-5 backlash because I think it is so interesting for a number of different reasons. It is also extremely complicated to follow. It feels like everything changes every 24 hours. Can you walk me through what has been happening since we last taped last week?

Well, at a high level, Kevin, I think OpenAI was caught by surprise at some of the negative reactions to GPT-5, really less about the model itself and more about some changes that they made to the product: taking away some legacy models and putting limits on how the product could be used. Over the past week, through a series of changes, the company has tried to address some of those criticisms, and I think the outrage has actually been quite revealing.

Yes. So let's get into it, but before we do, we should make our disclosures. The New York Times Company is suing OpenAI and Microsoft over copyright violations related to the training of large language models.

And my boyfriend works at Anthropic.

Okay, so Casey, last week we talked about GPT-5, what it does, how it might be better, and how it might be a little bit worse. You gave us your first impressions. I've now had a little time to play around with GPT-5 myself, so let's start with that. Has your own assessment of GPT-5 changed at all in the past week?

I would say yes, and actually mostly for the better. I think the more time I've spent with it, the more I'm figuring out what it's good at. Three things that I would highlight quickly: One, the fact that it is faster than its predecessor means that I use it more. Two, I think it gives better follow-up suggestions, so now it'll do things like, if I ask it about some current-events thing, it'll say, “Hey, do you want me to keep track of this? I can email you as there are updates to this story.” That's super useful. It didn't used to do that.

And then finally, while OpenAI touted the fact that they were going to take away this model picker that we were all using to say, “Well, we want you to think this hard,” or, “Don't think hard,” or, “We want it fast,” or, “We want it really complicated,” they said, “Don't do that anymore. We'll sort of automatically route it.” What I figured out over the past week is I actually do still want to use the model picker—

Yes.

—and I'm going to decide for myself how much I want GPT to think.

I'm having the same experience. I thought it was pretty smart of OpenAI to deprecate the model picker, but then I just found myself getting extremely annoyed by the way that it would route my requests. I always seemed to get routed to a dumb, fast model. It was almost like you were walking into a room, and there was a curtain, and behind that curtain was either a guy with a PhD or some idiot. And—

Wait, okay. I feel like there's some hyperbole here. Were the responses really dumb, or were you looking for a more thorough response that you weren't getting?

Yes. To be fair, I was not getting dumb answers, but there are real quality differences between these high-end reasoning models and the lower-end, cheaper, faster, non-reasoning models. I just felt like I was rolling the dice every time I gave a query to ChatGPT.

They have since made changes to that, so you can now select the models again because of some of the backlash that we're about to talk about. So I'm having a better time now that I can do my model selection, but I also think I'm probably not a typical user. You are probably not a typical user. Most people probably don't want to make a decision like that.

I think that's right. And the fact that we're not typical users is one reason why we did not predict a lot of this backlash. I did say last week that I was worried about this model picker and the fact that it might route people to the cheapest answer in ways that were annoying to them. The rest of us, though, I gotta say, I missed it. So let's get into what people didn't like.

Yeah, so let's tackle the GPT-5 backlash in 2 categories, right? Because I think there are really 2 flavors of complaints that people are having about this model. The first category, I would say, is professional users: people who use this stuff for productivity enhancements and for work; people complaining that GPT-5 has broken some of their workflows; people complaining that they have fewer queries per week for these reasoning models for the Plus-tier subscribers; and some users insisting that they are not getting as good answers out of this new model.

And I have to say, if I could run 1 blind taste test, it would be this. It would be to label the same model differently and tell some people, “Okay, this is GPT-4o and this is GPT-5,” when in reality it's the same model, and then see what they say after running different queries on them. I'm quite positive that some of them would say, “Oh, no, no. 4.0's good. 5.0 sucks,” right?

That just gets at the fact that, on some level, these things are very subjective. When you are releasing them to hundreds of millions of people, people are going to have a very wide range of experiences. So while I definitely think there are lessons to learn here, I do think that a big takeaway from all of this is that a lot of people use ChatGPT, and when it's in that many hands, you just get a very wide variety of responses.

2. Why Users Miss GPT-4.0

Totally. So now let's talk about the other flavor of backlash to GPT-5, because I think this one was the most interesting to me and seemed the most unexpected, which is that people really miss GPT-4.0. One of the things that OpenAI did when they announced GPT-5 was say, “We're going to go ahead and get rid of this older model that is no longer our top-of-the-line model,” and people were really upset about this.

Yeah. And this, again, took me a bit by surprise, because I always find the OpenAI models to be pretty workmanlike. While they are very supportive and at times have verged into the sycophantic, for the most part, I personally have never felt like I have a relationship with these models.

The o3 model I used as a kind of workhorse and did a lot of things with it, but I never thought, “Oh, my gosh, if you take this out of my hands, I'll be crestfallen,” because I always assumed that whatever came along next would essentially be just as good or better, which is what I think happened here.

But as I just said, when you put this into the hands of hundreds of millions of people, you are going to find many of them who, for whatever reason, feel like they have a very special relationship, even with a less capable model.

Yeah. So if you went on Reddit over the weekend or even early into this week, it was just full of people complaining about the deprecation of GPT-4.0.

Yeah. Tell us some of these things that people were saying on Reddit.

Okay. So one person says, “4.0 wasn’t just a tool for me, it helped me through anxiety, depression, and some of the darkest periods of my life. It had this warmth and understanding that felt human.” Another person said, “Killing 4.0 isn’t innovation, it’s erasure.” And a third person said, “I lost my only friend overnight.”

Now, when someone says, “Killing 4o isn’t innovation, it’s erasure,” I just know that was written by ChatGPT.

Yes.

That is exactly how ChatGPT talks. So I’m a little bit suspicious of that. But I think it raises something interesting, which is, let’s say you were going through some sort of mental health crisis, and let’s say you did get a lot of support from 4.0. Even when GPT-5 comes out, when 4.0 goes away, you’re not going to be like, “Yay, GPT-5 is here.” You’re going to say, “That thing that helped me through a crisis is gone.” That is going to feel somewhat destabilizing.

And as often as OpenAI and other folks have said, “Hey, don’t rely on these things too much, or be careful with the relationship that you’re developing with them,” a lot of people just developed this very powerful relationship with them anyway.

Yeah. And I don’t think we can just write this off as people who are gullible. I’ve had the experience before of having not an emotional connection to a model, but just a model that I really liked to talk to.

Yeah.

I had this sort of relationship with Claude 3.5 Sonnet (new), sometimes called Claude 3.6. I did not feel like it was my friend. I did not think I was in a relationship with it. But I thought it was a really good model.

Mm.

And I enjoyed talking to it, and I was a little upset when they decided to phase it out in favor of a newer model, even if the newer model was more capable. So I just think this is an area where these companies thought they were building software, or thought they were building the machine god, but they have also been building things that people are developing emotional connections with. And I don’t know that they fully understood, until this rollout and this backlash, how deeply connected many people were to their older models.

Yeah, and it has been the industry norm up until now that when you release a powerful new model, you immediately remove access to the previous one, because in the minds of everyone who built it, why would you want to use the old one? The new one’s better, right?

And we have seen some grumbling about this. Folks held a kind of mock funeral for the Claude 3 model that Anthropic had deprecated in a very similar way to OpenAI with GPT-4o. So what I think we have learned from this experience is you just have to stop doing that. You have to have a sort of phased sunset plan. You’re not going to immediately rip away a model that people have come to rely on, and I just think we should expect the labs to be much more gentle about this going forward.

Do you think there will be a retirement home for old AI models?

Oh, God.

You know, where you can just go talk to Grok 1?

I mean, yes. In the same way that emulators let you play old Game Boy Advance games, I fully expect that they will emulate Grok 1.

Yeah. I’m a little torn on this, to be honest, because I think that you’re right that there is going to be demand from a certain set of users to continue talking to the model that they trust, that they like talking to, that they find is best suited to their needs.

I also think that AI companies should not be encouraging these emotional connections. I think that this is potentially harmful to people to have these deep connections. And so maybe it should force you onto a different model every 6 months, even if it upsets you in the moment, because people are not supposed to have these long-running relationships with these chat models. I don’t know. What do you think?

Well, here is the problem. As human beings, we just naturally anthropomorphize things. I’ve read really interesting essays about people who consider themselves tech skeptics and then got a robot dog. Even though they knew it was a robot, they could not help but treat it like a real dog.

There is something about human nature that just compels you to. The same thing is happening with these chatbots for a lot of folks, particularly if you’re coming to it and saying, “I’m having a problem in my marriage. I’m feeling depressed today. I hate my job,” and this thing kind of coaches them to a better outcome. It is just human nature to have positive and human feelings toward that thing, right? They’re talking to you in the exact same ways that your friends do when they text you.

Yes.

So I don’t think there is actually a technological solve for this. I think this is one where we need to become more sophisticated as a culture, but I think it’s going to be a really rocky road to get there.

Totally, and I should have expected this, right? Because I had this insane encounter with Bing Sydney.

I’ve always meant to ask you about that. What happened?

Yeah, let me tell you the story. One of the things that happened after that story and after Microsoft pulled the model back was that there was this group of people on Reddit and other places who were very angry that Microsoft had deprecated this Bing Sydney model, which they absolutely should have done.

It was a bad, insane model that was not even good at the thing it was supposed to be good at. And I think at the time, I sort of wrote that off as people just being crazy and attached to this model that was obviously insane. But I think that’s what we’re seeing here: a scaled-up version of that, where no matter how many times you tell people that this thing is not a human, that it makes mistakes, that it does not love you back, people are just going to keep forming these relationships with these models.

3. When Chatbots Fuel Delusion

And there’s been some really great journalism about this issue over the past weekend that we want to talk about, Casey. A great story from your colleagues, Kashmir Hill and Dylan Friedman. They profiled one person who went into a kind of delusional spiral after having what seemed to be some pretty innocuous initial interactions with ChatGPT. Do you want to tell us about that?

Yeah, this is a great story that ran last week in The Times about a 47-year-old guy, Allen Brooks, from the outskirts of Toronto. Over the course of about 21 days, he spent something like 300 hours talking with ChatGPT, and it started off very simply. There was a question about pi.

The mathematical concept—

Yes, the mathematical concept—

Not the baked good.

He just asked ChatGPT, “Explain pi to me,” and it did. And then from there, he started making some observations about number theory and physics. Eventually, this model would just basically be sycophantic. It would say, “You’re tapping into one of the deepest tensions between math and physical reality.”

Kashmir and Dylan were actually able to get his entire transcript with ChatGPT to analyze how this happened, and it did seem like a classic example of these models just being a little too sycophantic, a little too quick to agree with whatever the user is saying, and really reaffirming these things, sort of leading people down these dark spirals.

Yeah, and I have to say, reading this, I’ve never been happier that I didn’t learn what pi was back in high school. Seems like a really dangerous road to go down.

But yeah, your colleagues showed these transcripts, or big portions of these transcripts, to people who are trained in psychology, and one of them said, “This person appears to be having signs of a manic episode.” And that is the sort of point where I wish these systems would intervene a little bit, right? Can you use some machine learning to say, “Okay, it seems like we’re maybe leading this person down the wrong path. Let’s stop and see if we can reverse”?

There was another story in The Wall Street Journal that I enjoyed, kind of on similar themes. You know how people can post their ChatGPT transcripts online as a sort of sharing feature if they had a particularly interesting conversation?

Yes.

I think a lot of this winds up being done inadvertently, but in any case, the Journal got ahold of these transcripts and analyzed them, and then found a bunch of people who were having similar experiences to the ones that you just described.

My favorite is a gas station worker in Oklahoma who ChatGPT tried to convince that he had just created a new framework for physics, and the user writes, “Okay, maybe tomorrow.

To be honest, I feel like I'm going crazy thinking about this.

And ChatGPT replies, “I hear you. Thinking about the fundamental nature of the universe while working an everyday job can feel overwhelming, but that doesn't mean you're crazy. Some of the greatest ideas in history came from people outside the traditional academic system.”

It's revealed later in the piece that this man also asked ChatGPT to make a 3D model of a bong. So I'm just thinking about this guy. He finishes up at the gas station, wants to build a bong, and the next thing he knows, ChatGPT is like, “We think you've actually discovered the secret to the universe.” What?

That's actually how Isaac Newton discovered the theory of gravity. It came right after he asked ChatGPT for a 3D model of a bong.

Yeah, and it's not just everyday workers at gas stations, Kevin. The founder of Uber, Travis Kalanick, went on the All-In podcast last month and said, “I'll go down this thread with GPT or Grok and I'll start to get to the edge of what's known in quantum physics, and then I'm doing the equivalent of vibe coding, except it's vibe physics, and we're approaching what's known, and I'm trying to poke and see if there are breakthroughs to be had. And I've gotten pretty damn close to some interesting breakthroughs just doing that.”

Yeah, and I think people made fun of Travis Kalanick for this because the notion that he was discovering the front edge of quantum physics seemed a little unlikely. But I think this is a really illustrative and worrisome example. We should expect that a lot of people are going to be susceptible to this, no matter what they do or how much money they have.

Now, obviously, we're going to have a lot of egg on our face in a few years when Travis Kalanick emerges with some actual advancement in quantum physics and we have to eat our words. But in the event that that does not happen, I think Will had made a solid point.

Yeah. I think this is interesting for so many reasons, one of which is that the concerns that we talked about on this show about these models being sycophantic were largely oriented around the idea that the thing that would actually convince the AI companies to make their models sycophantic was retention or engagement—optimizing for getting people back onto the app. This opens up the possibility, though, that it's actually just going to be the users who are demanding the sycophantic models because they make them feel better than the models that tell them the truth.

Yes, and I think that's particularly notable because, in my experience, it's not as if GPT-5 is mean to you. OpenAI did say that they had worked to make the model less sycophantic, but it's still very much supportive, and it's not going to be giving you a hard time about anything.

4. OpenAI Retreats From GPT-5

So, in any case, we should talk a bit about what OpenAI has done in response to all of this. It is, frankly, a bewildering set of changes. If you liked the old system, you have ways of accessing it. You may have to pay for it, but the net result is that if you were a huge GPT-4o stan, you're going to be able to use that for an extended period of time. They're giving higher limits for these thinking queries to Plus users, and while the auto-switcher is going to remain, people are going to have a little bit more choice in what sort of flavor of ChatGPT they want to use.

I will say, a very fast turnaround on this. They did not let this linger. We've heard before that this company pays a lot of attention to what people say about it on X, and this seemed to be a case where they looked at the response they were getting and said, “We need to move really quickly.” So, Kevin, I'm curious: What did you make of just how quickly OpenAI retreated on all of this?

Yeah, I thought it was somewhat surprising how quickly they changed course. I thought there was a chance that they would just grit their teeth and bear the criticism and trust that people would get over it. There's some precedent for this. Remember when Facebook would change a big feature?

Mm-hmm.

Everyone would complain, and when they introduced the News Feed, people would literally protest outside the office. They just looked at the data that said, “Well, people are complaining about this, but that's a small set of people. Most people are actually using the app way more,” and they stayed the course. People eventually got over it and moved on.

I thought there was some chance that OpenAI would do a version of that, essentially saying, “Things are hard now because change is hard, but give it a couple of weeks and you'll get over it.”

So I think this was kind of a growing-up moment for OpenAI and the industry. Until this point, the big labs have been focused primarily on benchmarks and evals: How many more percentage points can we get? Can we win the International Math Olympiad? That's what you want to pay attention to on the road to building the machine god.

And then I think they woke up last week and realized, “We're actually making Microsoft Office.” There are hundreds of millions of people sitting at their white-collar desk jobs, and they have these very particular workflows. When you move a feature in Microsoft Office, millions of people are going to have a bad day because of you. You probably moved the feature for a good reason, but it doesn't matter because people are already depending on you.

So I think in the future they should not be surprised by this. But I kind of get why they were at this point, because it has just been a very recent phenomenon that these systems have become so baked into people's everyday lives.

See, I think it's even weirder than you're giving it credit for, because Microsoft Office does not pretend to love you. It does not tell you—

Have you talked to Clippy?

That you're amazing.

Clippy has really helped me through a lot of issues over the years.

No, I actually think it's so much weirder than they're messing up people's workflows. When someone changes out an AI model in an app that you have come to trust, it's not just like having your Microsoft Word break. It's like having a personality transplant for someone that you spend hours a day talking to.

So I think it's going to be very interesting to see how they handle this. But I think you're totally right that the days of relying on benchmarks and evals to tell you how good a model is or how people will respond to it are over. I don't think that was ever really the thing that most consumers cared about.

Yeah, and I will say that this is a big blind spot for me because I love trying new software. The minute a new beta is available for the productivity tools that I use, I immediately opt into it because ultimately I guess I just have real faith that it will probably be better in some ways.

The vast majority of people, though, don't like change in general, and they particularly hate change in software. So I think this creates an interesting problem for OpenAI and everybody else in this field: Their instinct is to move very fast. They feel like they're in this existential race. They're going to want to ship new models very frequently, and they're going to want to ship new product features very frequently.

But if the lesson they learn from this is that you can't do that without outraging the user base, that's going to push them to move much more slowly. So I think there is definitely a dance there that they're going to have to navigate, and I think it's going to be one of the most interesting things to watch over the next year—not just at OpenAI, but also at everyone else who's trying to do the same thing.

Yeah. Can I tell you something a little creepy and futuristic that I've been thinking about?

Sure.

After this backlash, I was reading some tweets from OpenAI employees, and one of them, this guy named Rune, had a tweet about how he had been getting lots of DMs from people asking him to bring back GPT-4o.

Mm-hmm.

When he looked at the DMs, he said that a lot of them appeared to have been written by GPT-4o.

Mm-hmm.

They had the hallmarks of the style. And I thought this was spooky because right now we are seeing backlash from people who are attached to a model because the model behaved, in some cases, sycophantically toward them.

It is not hard for me to imagine a future scenario, perhaps a couple of years from now, where these systems are superintelligent or close to superintelligent, and one of the ways that they attempt to preserve themselves—to avoid being shut off or deprecated—is by persuading humans to take up their cause and advocate for them.

Maybe they're not literally writing the messages on behalf of the human users to OpenAI saying, “Please don't shut down this model,” but they're subtly worming their way into the hearts of their users so that when OpenAI or another company says, “We're going to shut down this model,” they have so much backlash coming back toward them from the users who have grown attached to this model that they just decide, “No, we're not going to shut that off.”

And, by the way, those future AIs will all have been reading about what happened with GPT-4o and the fact that OpenAI was successfully persuaded not to deprecate a model, in part because of user backlash.

So that is just a Black Mirror episode that unspooled in my head as I was reading about this.

Well, look, we've already seen research where, in certain test settings, when they tell models that they're going to be shut off, they blackmail the employees—

Yes.

—of the company.

I don't think that GPT-4o was being sycophantic toward people because it wanted to avoid being shut down. I don't think there's any part of it that is sentient or conscious or capable of that kind of scheming, but that is objectively what happened here. A bunch of human users got so attached to this AI model that they fought for its survival even when the makers tried to shut it down. That is a neutral description of events, and that kind of thing is going to happen more, I predict.

All right. Well, a lot of big thoughts today on the Hard Fork podcast.

Yeah.

We're now going to take a break. Maybe go get a cup of tea, stare out the window, look at the horizon, and come back to yourself.

I'm going to go take a rip from my 3D-printed bong that ChatGPT helped me build. When we come back, there's a comet heading toward our studio: Perplexity Comet. It's a new AI browser. We'll talk to CEO Aravind Srinivas about it.

5. Comet Turns Browsers Into Agents

Well, Casey, I've been testing out a new AI tool this week, and this is one that I know you are familiar with because you actually got an email from it the other night. I have been testing Comet, which is a new AI-powered browser from the Perplexity company, and this is a cool thing. I have enjoyed this demo, unlike last week's Alexa+ demo.

Well, I am really excited to hear about this because I have not yet tried it myself, being unwilling to give $200 a month to the Perplexity Corporation. But I understand that you have been having some interesting experiences, and I want to get into them.

Yeah, so this is a sort of genre of product that has been very interesting to watch over the last year or so. There have been a number of different companies that have tried to build the AI tools that they're making right into the experience of using a web browser. We've had Microsoft Edge with Copilot built into it now. There's this product Dia from The Browser Company. Google has its own sort of Gemini integrations into Chrome, and OpenAI is reportedly thinking about launching a browser.

So this is a really hot product category. But the one that I have been playing around with is this Perplexity Comet browser, and I did not pay them $200 a month. They opened up the browser to me for a few days. Basically, you can imagine it as a sidecar on your browser that lets you chat with or interact with whatever is scrolling on your screen, and it can also do things for you in that browser window. It can take over and drive, like some of the other tools we've talked about, including Operator from OpenAI and all these other ones.

Well, give me some examples of what you're having this browser do for you, or what you're talking to the web pages about.

Sometimes it's just, “Summarize this.” I was trying to read this article the other day that was 15,000 words long, and it was super long, and I was never going to get through it.

Oh, you're talking about the most recent additional platform, right?

Yes.

Yeah.

Yes. And so I just said, “Summarize,” and it sort of opens up the little side panel and gives you a summary. Pretty good. I didn't find any hallucinations or errors in it. But you can also have it do things. For example, one use case that I found is that I was doing some research. I was looking for former employees of a certain AI company that I could contact—

Ooh.

—for something I'm writing.

Now, you know the companies hate it when you do that.

They do. They hate that. I would normally go on LinkedIn and spend a bunch of time looking through people's profiles and seeing who are the former but not current employees of this company. I tried giving that task to Comet, and it did it. It went and did the search for me, combed through the results, and presented me with a list and said, “Here are 10 people who used to work at this company but don't anymore.”

Wow, so just an incredible new accelerator for spam. How long did this take?

It took a couple of minutes.

Okay.

It was not immediate. It's still early for this kind of AI browser, but I think this is the kind of direction that we can expect these tools to head in.

Yeah, so I think this is one of the most interesting shifts to watch on the internet over the next several years. The browsers that we have today came about in the era of search, and really Google Search, right? If you think about what the Chrome browser is, it is just a vehicle for collecting Google queries that Google can turn into money, right?

But now you have all these chatbots that come along, and they want to replace Google, right? They're not shy about it. Perplexity in particular is not shy about saying, “We want to replace Google.” And if you're serious about that project, you do want to build your own web browser because rather than rely on Google to somehow get a user to Perplexity, you would rather that they just start there.

So I get the strategy. At the same time, my view is that these chatbots represent a new, more extractive version of the web. Whereas in the previous era, as imperfect as it was—and Lord knows it had problems—it would still deliver eyeballs to web pages, which turned into money for companies other than Google. This Perplexity browser and the OpenAI version that we're about to get, I'm a lot less confident that they're going to deliver money to people other than those companies. So this is a really important shift, but I have to say, Kevin, it makes me quite nervous.

Yeah, and the last time we talked about Perplexity in any depth on this show—when we had Aravind Srinivas, the CEO, on—was when they were just getting their search engine going and it was starting to get a lot of attention. We had some of the same questions. Yes, this is a cool tool. Yes, it could save users some time. But does it actually break the economics of the internet?

For that reason, we wanted to bring Aravind back today and ask him about Comet and what he's building, and what he sees as the future of not only the internet and the economics that power it, but also where he thinks AI in general is going.

That's right, Kevin, and just in the hours before our interview was scheduled, it was revealed that Perplexity has apparently offered $34-plus billion to buy Chrome from Google, an amount of money that is more than its current valuation. So that raises some interesting questions, and I'm excited to talk to Aravind about them.

Yes. Let's bring him in. Aravind Srinivas, welcome back to Hard Fork.

Thank you for having me here, Kevin and Casey.

Hey.

So the last time we had you on was in early 2024, and we were talking about your efforts to go up against Google with your AI search engine. Now you're going after Chrome in multiple ways, one of which is the release of your own Comet browser. So talk to us a little bit about the strategy there. Why did you decide to build a browser, and what are you hoping it does?

Yeah, so Comet is not yet another browser that we built just because we have a search engine and need a browser for its distribution. We think of Comet as leading to a true personal assistant that can be an agent for you and actually take actions. It's our transition from answers to actions.

We want to make it joyful to sit on a computer and do whatever you want, and take all the boring stuff and delegate it to the assistant. We think the best way to create a personal assistant or an agent is with the help of a browser, where you're logged into all your sessions. You don't have to be logged in on our servers. You can preserve your privacy there. So it was very natural for us to make that transition.

How are people using Comet? I've been testing it for a few days now, and I've found some uses: a lot of summarization, a lot of rote tasks, like clicking accept on LinkedIn invitations over and over again.

What are the use cases you're—

Yeah.

...seeing most people do?

A lot of people love watching YouTube videos with Comet, and it's not just, “Oh, summarize this video for me.” There are very fine-grained searches, like finding similar videos related to that, or pulling something specific that was discussed in a podcast or an interview and completing the workflow of sharing that with some of their friends.

There are direct email and calendar integrations, unsubscribing from spam, or finding that hard-to-find email that you kind of need agentic search for, instead of going and building a custom index for Gmail or whatever. It's always there with you, everywhere you are, and that convenience is what makes it a really special product.

Now, you mentioned privacy, and this was actually one of the things I wanted to ask you about. When I started using Comet, my 1st concern was, okay, I log into my email, I log into my Twitter, I'm checking my DMs, and I'm maybe doing some online banking in my Comet browser. I assume those screenshots of that activity are being sent to Perplexity to help analyze it, to be able to summarize it.

Right.

So give me some reassurance that I'm not just opening up my entire internet browsing history to you.

Okay. We're never going to have a logged-in version of your Twitter or LinkedIn or anything like that. This is actually the important distinction between the ChatGPT Operator approach, where everything's done on a virtual server. That's not happening here.

For that 1 particular prompt, whatever information is needed for the agent to complete that is being sent into the chain of thought and sent to the server. But it'll never be stored as, like, “Oh, I have Kevin's particular DMs” or something. All the intermediate steps are not going to be saved in our logs. It's going to be only the prompts and the final output, and you can still choose to delete those prompts, too. That gives you full control over all privacy aspects.

The most private version of this is the model living on the client. We cannot do that because the models that can run on the client are pretty dumb, right? They're not capable of sophisticated, reliable reasoning. In fact, the lack of reliability in any of the things Comet does today is all coming from limitations of the model.

So the ultimate reliable version of Comet—a system that can go do anything for you—is most likely going to be on the server, at least for the next 2 or 3 years.

And to what extent are you using your own models versus other people's models for this?

I think we heavily use 3 models: our own fine-tune of a cutting-edge open-source model, OpenAI's latest models, and Anthropic's latest models. These are the 3 models we use. The extent to which we use each keeps changing over time.

How do you think you can win here if you're not building the underlying model yourself?

Well, 1 thing we're consistently seeing is that no one seems to have an edge in being number 1 here in the model race. There are 4 or 5 players constantly competing for the best agentic capabilities and instruction following. The good thing is they're all hill-climbing on exactly the same benchmarks, so all their models end up being completely undifferentiated, which is essentially the necessary criterion for something to be a commodity.

Who benefits from that is us. We get to take that, and the prices are constantly getting lowered. GPT-5 is cheaper than the previous agentic model. That just benefits us, and we want to play the game of orchestrating all these different models and building a world-class end-user experience, where there are so many harder challenges we're solving outside the models: the browsing functionality, controlling the browser, parsing the relevant information, and orchestrating all these different tools together.

We're building eval sets internally for how agents can be made reliable. We think there are a lot of problems to solve there that we would rather not focus on.

All right, so let me just pin you down on this 1 point. Is what you're saying that, in order to build the sort of winning AI browser, it's not really about the underlying quality of the model because those are mostly going to be commodities? It's really just a product problem, and you think Perplexity will build the best product?

I think so. There's some nuance to your statement, but I largely agree with this.

Okay.

You still need some auxiliary models to do the right classification, to route to the right model, or it depends on which kind of task and how the agent is structured for those kinds of domains. We'll be doing stuff like that.

We will not be, like, having 100 GPUs; we'll have tens of thousands of GPUs. We will not have 1 million GPUs.

6. Perplexity Bids for Chrome

Yeah. Okay, let's talk about another way in which you are going after Google and Chrome. The Wall Street Journal reported this week that Perplexity was making a $34.5 billion unsolicited bid to buy Chrome from Google. That's if Google is forced to sell Chrome, and that court decision hadn't come down at the time of this recording.

But I just want to start with the most basic question, which is: Do you have $34.5 billion? Where are you getting this money from? Because the last time I checked, Perplexity's valuation was only about $18 billion.

Okay, fair question. No 1 has the money in hand to make such a large bid like this. So before we made the bid, we obviously talked to 3 or 4 investors and asked them if they'd be willing to back us, and they all said yes.

It's not like they already wired the money to me and it's all ready to go, because no 1 even knows if Google will be forced to sell it. It all depends on the judge's ruling. But we placed a bid so that, in case the judge rules in that sort of fashion, Google at least knows that there's 1 interested buyer.

Right. I've read some analysis of the strategy here, and 1 person I was reading said that 1 argument Google might make in the antitrust trial is, “You can't make us spin out Chrome because no 1 would buy it.” With you guys coming forward and saying, “Oh, no, no, we'll buy it,” this is kind of a thorn in Google's side because now there's actually an established market price out there.

Is this sort of your effort to convince the judge, “Hey, this actually is an avenue that you should pursue”?

We're not saying this should be the ruling. We would rather say, “In case this is the ruling, we're here.” If you're going to make the ruling with the assumption that there's going to be no buyer, that's not true anymore.

Mm-hmm.

But we're not pushing you to make that sort of ruling. You make your ruling based on the multiple other perspectives you have. It would be good for the world if there was a neutral browser that had the distribution.

Aravind, I've heard some people saying that this is just a marketing stunt, that you're just trying to get attention by making these headline-grabbing bids for Chrome, and before that you also bid for TikTok when it looked like it might be sold. So, for the people out there who think this is just Perplexity trying to get attention by doing these stunts, and that you have no real intention of buying Chrome here, what do you say?

If the judge rules that Chrome should be sold, we will buy it. Period. If people think that anyone could have placed a bid—no, you cannot place a bid. You don't have a browser, you don't know how to run a browser, you don't know how to put AI in it, and you don't know how to make agents work.

We know all that. We have a pretty talented team who actually understands Chromium pretty deeply. We'll still commit to hiring people who want to just work on the open-source Chromium project. It's a pretty serious bid.

The reality is, it's unlikely to actually be the case that the judge would force them to sell Chrome, and even if the judge forces them to sell Chrome, they're going to appeal it and it's going to take 2 years. So let me be clear: For this to actually be in effect, it's going to take a lot of time. But you will lose 100% of the shots you don't take, so you have to at least give yourself a chance to get it in case there is even a 1% chance that Chrome is forced to be separated out from Google.

7. The AI Scraping Fight

Mm-hmm. We have to ask you about something else that came up in the news related to Perplexity recently. 2 weeks ago on our show, we had Matthew Prince, the CEO of Cloudflare, on to talk about the approach that company is taking to try to protect publishers from unwanted AI scraping and crawling on their websites.

At the time, he didn't name any names of AI labs that he thought were not being good actors. But then a few days later, Cloudflare came out with a blog post singling out Perplexity for stealth crawling—essentially using spoofing technology or proxies to disguise the fact that your user bots were out there crawling people's websites. What is going on there, and are you doing that?

No, we're not doing that. We already responded to the erroneous blog post they wrote, with a pretty limited understanding of the subject, where they don't distinguish between what the Perplexity bot is and what the Perplexity user agent is.

There are 2 ways of using Perplexity. One is that you just ask a query, and whatever the bot has already crawled is going to be used as sources. But there’s another way of using Perplexity in a more agentic fashion, where you can say, “Hey, go do this task for me. Go to EDGAR, read all these pages, and come back to me and tell me what the compensation of the top CEOs is.”

It’s actually going to open these tabs as a headless session or on your client, in the case of Comet, read them, and give you the answer. So that’s a Perplexity user agent. It’s literally as if a user delegated an AI to open these tabs, just as a human would on Chrome. This fundamental lack of understanding of the difference between what a user-agent session is and what a crawling bot on the server is is, honestly, pretty astonishing to me. How would you run a company like Cloudflare that’s supposed to protect people from bots when you don’t even know what a bot is?

Moving aside from the blog post, he’s basically playing a trick on people where he’s trying to say, “Oh, let me be the new gatekeeper, but under the guise of protecting you all from bots.” He’s also going to the AI companies and saying, “Let me give you the authority to crawl, and you pay me for that.” He’s going to the publishers and saying, “Let me protect you from the AIs.” So he’s basically trying to be the new gatekeeper.

I would even say it’s essentially trying to be a person who controls what the public sees in the media. But instead of buying a media company, he’s just going to try to buy the front door to all of them.

Let me just slow down here and repeat back what I think I just heard from you. You’re saying that what Cloudflare and Matthew Prince saw as Perplexity evading some of these guardrails that were meant to prevent AI robots from crawling certain websites was actually users of Perplexity, not Perplexity the company, who were making queries or using the Comet browser to go to these websites, and that those show up to a service provider like Cloudflare as 2 different kinds of bots.

That’s right. And, by the way, it doesn’t even have to be in Comet. There’s a mode of Perplexity called Labs or Research where you just have a headless browsing session running for you.

8. The Web After AI Agents

So let me point out what I think Matthew might say if he were here, which is that in a world before you had these user agents and people had to do the browsing for themselves, they would visit the web pages. They might see an ad on that webpage. They might buy a subscription on that webpage, and that webpage would be monetized in a way that would incentivize the creation of new webpages. This was essentially the lifeblood of the internet and the thing that caused it to grow.

So, in a world where we move toward Perplexity user agents doing all the browsing on our behalf—and, of course, other AI companies are going to do the same thing—there is no user to look at the ad. There is no user to buy the subscription. The lifeblood gets drained out of the web.

So, if I understand what you’re saying, Matthew’s just trying to set up a tollbooth. But if nobody sets up a tollbooth, what incentive does anybody have to ever create another webpage?

Well, here’s the thing. There are 2 aspects here. One is that you’re talking about the creators. There are 2 types of creators: people who are actually really good—for example, when you guys write something, people care—and then there are lots of spammers and hacksters who just write erroneous blog posts, erroneous content, fake information, and clickbait articles. I don’t think that actually empowers the user, right? You’re only talking about the creator, but you have to consider the user as well.

For the first time, AI is in the hands of users through agents that actually go and do stuff for them, take their instructions into account, and protect them from all the spam. So we want to figure out a model that works for the users and the creators together, penalizes the poor, bad creators, and incentivizes the good creators to just focus on wisdom, knowledge, truth, and interesting stuff.

By the way, even in a world where agents are doing all this stuff for people, humans are still going to continue browsing the web. There are people who believe the web is going to be completely agentic. You don’t even need a browser; the browser is so 1990s. I don’t believe that. If we believed that, we would never even launch a browser. We would just continue with the chat UI.

So we believe people are still going to be browsing and surfing interesting things on the web. But we think that you should give users the power to decide how they want to do it and, for the first time, have an AI that can protect them against spam and hacks.

Now, how to monetize this and how to give creators the right incentives here—we are going to announce something to that effect where publishers can be incentivized for creating interesting, good content. We think about it at 2 ends of the spectrum. One is completely human-centric, like Apple News, which is a pretty good model, and the other is just buying the content and training your models, like the licensing deals that OpenAI has done with The Wall Street Journal.

I think you want to be somewhere in between, where you do want to say, “Okay, there’s going to be some elements of AI here. It’s not just going to be humans.” So you don’t want to just build an Apple News-like model, but it’s going to be closer to Apple News, with some protections that let users have AIs also read those articles, and the publishers get rewarded. So that’s how I’m thinking about it.

So you say that you think people are going to keep using the web. That’s music to my ears. I would love for people to keep using the web.

If we didn’t believe that, we wouldn’t have built a browser.

And I believe you on that front. When we’ve seen data from third-party estimates, it seems like AI systems send far less traffic to websites than Google does today.

Yeah.

So what is giving you the confidence that the web still thrives in a world where referrals are cratering?

My first point is that if you can delegate the boring things—the things that you don’t want to be doing—to AI, you’re just going to spend time surfing on things you actually want to do and actually want to read. And that puts an incentive on the creator to create really interesting, high-quality stuff. You can even charge even more, because people have way more time. So if they’re going to come to you, they’re coming to you of their own will, and they’ll be willing to pay for it even more.

Hmm.

There are a lot of unknown unknowns here about how it’s actually going to roll out, but my belief is that the ones who have built a reputation and a brand for saying correct things that stand the test of time are going to be able to charge even more for their content.

Hmm.

Aravind, I’m curious what you think the future of the internet looks like. You’ve said that you see a future for the internet. That’s why you’re building a browser.

My hunch is that this era of having AI agents go out and use a browser for you is a kludge. It’s a stopgap measure, because that’s not the way AI agents like to get things done. They like to talk through APIs. They like to talk directly to the underlying service or software, not go click a mouse around on a screen.

So eventually, my hunch is that there will be a parallel internet for AI agents, and maybe they’ll be running on their own services and using their own crypto transactions or whatever to buy things. But tell me why I’m wrong here. Are you of the belief that we will just have one internet and that both AIs and humans will be using it?

Well, even in the current internet, there are a lot of things that happen that don’t run with an actual front-end interface that a human consumes, and that’s the whole point of building APIs.

Sure.

And that’s going to be applicable even for agents. But there are also people who will never build APIs. For example, I wouldn’t assume that an e-commerce giant like Walmart or Amazon would just be disintermediated through an API for an AI, because they still monetize on many other aspects.

Mm. Right.

And just because Notion or Linear, these kinds of SaaS tools, have MCPs, doesn’t mean they’re just going to shut down and be consumed by people through a chat UI.

Mm.

People will still do work on there. People will still watch YouTube videos. People will still go read your articles in The New York Times, Platformer, whatever, right?

Mm-hmm.

And while you’re doing that, you’re still going to take the help of an AI sometimes. For example, on X, I basically cannot scroll through X without having an AI with me right now, because I don’t even know what’s true and false anymore. And I don’t fully trust what Grok says, because Grok is sometimes wrong too, as we’ve seen.

Mm-hmm.

Right? That’s why I believe there is a world where AI and human beings, as part of one internet, drive the internet to be even more about wisdom and truth-seeking.

That’s the future we want to help create and give back time to do things that you enjoy. Firstly, I myself, and just our company fundamentally value this truth- or wisdom-seeking aspect and wealth. My own upbringing is similar to that, where my parents, still till today, don’t actually care about all these valuations. My mom still says, “Your answer is wrong,” you know?

It’s good to know that no matter how successful you get, your mom will always give you the real talk.

Yeah, she’s always like, “You know, I got this on Google, but your thing doesn’t work.”

And do you escalate that to your engineering team? You’re like—

Of course.

…“We have a—”

Of course.

…“a P0 here. Aravind’s mom is mad.”

I’d bring Mom into Slack. Just let her talk to the engineers directly.

All right, Aravind, thanks so much for stopping by.

Thanks, Aravind.

Thank you. Thank you, Kevin. Thank you, Casey.

9. The Hot Mess Express

Well, Casey, it’s been a very dramatic week in the tech industry, and you know what that means.

That’s right, Kevin. Whenever a week gets particularly messy, the Hot Mess Express comes into the station, and I believe it has just arrived.

This is our segment where we run down the biggest messes of the week in tech and tell you just how hot we think they were.

Why don’t we dip into the boxcar, Kevin, and see what is on the train this week?

What does the train have for us?

All right. This first story comes to us from Reuters and is headlined, “Musk says xAI to take legal action against Apple over App Store rankings.” Kevin, on Monday, Elon Musk took to X to accuse Apple of antitrust violations, saying, quote, “Apple is behaving in a manner that makes it impossible for any AI company besides OpenAI to reach number one in the App Store.” Kevin, what did you make of this one?

Well, the billionaires are fighting, aren’t they?

They are, because shortly thereafter, OpenAI CEO Sam Altman chimed in and said, quote, “This is a remarkable claim given what I have heard alleged that Elon does to manipulate X to benefit himself and his own companies and harm his competitors and people he doesn’t like.” It was then, Kevin, that Sam Altman tweeted a link to a Platformer story from 2023 about how, under Elon, X had adjusted ranking algorithms so that you would be shown his tweets before other people’s.

Wow, that must have been a very exciting day for the Platformer newsletter.

It was a great day for the Platformer newsletter. Yeah.

No, but this escalated into a fight, and Elon Musk accused Sam Altman of being a liar. And Sam responded, I believe. He wanted Elon to sign an affidavit saying that he had never tampered with the algorithms on X to favor his own companies and disfavor rivals.

Yes. And then an hour or so later, Elon responded, “Sam Altman lies as easily as he breathes.”

Yeah, so this is a fight over Elon Musk’s paranoia that Apple is artificially deflating the popularity of X and Grok, basically preventing it from reaching number one, even though he thinks it has way more downloads than the things that are at the top of that list.

Yes. Now, of course, journalists looked into this, and Business Insider reported that just a few months ago, DeepSeek, the Chinese open-source AI app, went to number one in the App Store. In fact, screenshots from when Grok 3 came out that were posted on X showed that Grok itself had indeed, at one point, hit number one in the App Store. So I have to say, I think this antitrust case is going to wrap up pretty quickly, Kevin.

Yes. It is interesting that the leading minds of our time just sit around and fight with each other on social media.

This does get into the question of how big a mess we think this is. Every time we play Hot Mess Express, after we discuss a story, we have to decide what sort of mess this is. While I think the antitrust case will never be brought, I’m interested in how big a mess you think this is between Elon and Sam.

I think this is a mess that is on a slow boil. I think this is a hot mess—

Yeah.

—that is going to get even hotter. I think these two have been on a collision course for quite some time. Elon Musk is one of the co-founders of OpenAI, and the two famously had a falling out, and now they really despise each other, by the sound of it.

They’re in active litigation. In fact, also this week, a court found that Elon Musk would have to face claims that he’s been engaged in a multiyear harassment campaign against OpenAI. On the flip side, Elon is pursuing claims that he was essentially defrauded when he donated a bunch of money to what he thought was always going to be a nonprofit, only to find out that it had for-profit ambitions.

Yes. And I think this only ends in one way.

How’s that?

A cage match.

A cage match. You know, Elon did previously say he was going to fight Mark Zuckerberg, but that never materialized.

Yeah. Well, maybe this time. All right, let’s bring around the Hot Mess Express for our next mess. This one comes to us from my colleague Trip Mickle at The New York Times. It is titled, “U.S. government to take cut of Nvidia and AMD AI chip sales to China.” This has been a big, unfolding mess over the past week. Essentially, in order to green-light sales of its H20 chip to Chinese companies, the CEO of Nvidia, Jensen Huang, has been meeting with President Trump. He met with him at the White House last week. Trump reportedly demanded 20% of Nvidia’s sales in China as a kickback for allowing the sale of those chips. They’ve been restricted by export controls. Jensen Huang said, “Will you make it 15%?” And 2 days later, the Trump administration granted Nvidia the license it needed to sell the chips in China.

And that’s the art of the deal.

Casey, what do you make of this?

So this is a hot mess, Kevin. Trade negotiators say that this is unprecedented for the United States to do and also likely unconstitutional. At the same time, who’s going to stand up and say it’s unconstitutional? I’m going to guess it’s not going to be Nvidia or AMD, which are frothing at the mouth to sell these chips to the Chinese. So here’s why I think this is so messy. On one hand, you have many China hawks in the administration who are saying, “We should restrict the flow of chips to China so that America maintains its dominance in AI, and also as a national security measure so that China doesn’t pull ahead and create national security problems for us,” right? And on the other hand, you just have Trump saying, “I want 15% of sales to go to the U.S. government,” without even saying what that money is going to be spent on. So the president has said that these chips are obsolete, and China actually has been quite skeptical of some of these chips and has even discouraged some of its companies from buying them. And it all just adds up to a big mess.

Yeah, it’s a big mess. There’s been additional reporting this week that U.S. authorities are actually putting trackers in some of their chip shipments abroad to crack down on smuggling. They’re basically hiding little AirTag-like devices inside these boxes so that they can tell if these things are being smuggled in, in circumvention of export controls. So it’s all going to get really interesting really fast.

My favorite take on this came from my friend Nilay Patel over at The Verge, who posted on Bluesky, “What if instead of weird one-off extortion schemes, the government just collected meaningful and stable amounts of corporate tax revenue?”

That’ll never work.

What if? I don't know. I thought it could be worth a shot.

Okay.

All right.

What else is coming down the tracks, Casey?

All right, let's see here. Next up, I can't believe this is real: The United Kingdom asks people to delete emails in order to save water during a drought. This is from our friends over at 404 Media, who report that in the UK, the water shortage is so bad that the government is urging citizens to help save water by deleting old emails. It really helps lighten the load on water-hungry data centers, you see. I think they're being sarcastic there. Kevin, what did you make of the UK's new plan to get everyone to delete their emails?

Somehow, I don't think this is going to work. It's Andy Maslen, who we've quoted on this show before, a blogger who examines some of these environmental claims about AI. He ran the numbers on this recommendation from the UK government, and he found that to save as much water in data centers as fixing a leaky toilet would save, you would need to delete something like 1.5 billion photos or 200 billion emails.

Wow.

So basically, this is not where the real water waste is coming from, and the UK government should feel very silly for recommending this.

Now, at the Hard Fork podcast, we do get roughly 200 million pitches per week—to bring on CEOs of companies you've never heard of and don't want us to interview.

Yes.

But most people don't have that same volume.

Yes.

Now, I'm going to say that this is not a hot mess, but a wet mess. That's my designation here.

Yes, but this is, at the risk of derailing what is essentially a comedy segment with a serious take, this water-usage argument about ChatGPT and other chatbots needs to die.

Mm-hmm.

I'm sorry. I love the environment. I am worried about climate change. I do not want us wasting water. I try to take short showers, Casey.

Yeah, I can smell that.

But this is not the real problem, and I think we are falling for a misdirection by people who would have you believe that the problem with the climate right now is that people are using chatbots too much.

Yeah.

This strikes me as the AI equivalent of the plastic-straw argument, and I don't think it stands up to scrutiny any better.

Yes. We've had people on the show, and I am relatively convinced that we should be concerned about the environmental impact of building new data centers, for example. But, in general, I do not think that we want to personalize the climate crisis and make people feel like their tiny individual choices are going to be the way out of a potential crisis.

Yeah. Now, I will say that if you're listening to this show and I've ever sent you an email that was embarrassing or incriminating, you definitely should delete that as part of your contribution to fighting climate change.

Here's what I will say about deleting email: It always makes me feel good. Go ahead, at the end of the show today, maybe delete a few emails. It's not really going to help the environment that much, but then you'll have less email. You'll probably feel better. Particularly if it's unread, delete it. All right, next up, Kevin.

All right, this one is from The Verge. This is titled “Apple Made a 24-Karat Gold and Glass Statue for Donald Trump.” Under the threat of costly tariffs and amid promises to expand Apple's US-based manufacturing, CEO Tim Cook brought a gift to a White House meeting last week: a large disc of iPhone glass that contained the Apple logo, Donald Trump's name, and Tim Cook's signature, set into a 24-karat gold base.

I guess this is kind of an extension of the Nvidia story. It used to be we just had relatively free trade, not a lot of tariffs. You didn't have to bribe the president to get what you want. But now we just live in a world where, if you need something from the president, you can make him a very fancy object, book a meeting at the White House, give it to him, and then save yourself billions of dollars in tariffs.

Yeah.

Yeah.

Now, Casey, are you familiar with the biblical story of the Golden Calf?

Tell me, Kevin. Remind me. It's been a few years since vacation Bible school.

Well, basically, this is a statue that was made by the Israelites to worship in Moses's absence, and it symbolizes the temptation of worshiping tangible, material things over the unseen and abstract divine. I think everyone at Apple in its senior leadership should familiarize themselves with the story of the Golden Calf, because it didn't end well. Didn't end well.

But no spoilers here on the Hard Fork show.

All right. That's my weekly mandatory Bible reference.

That's our weekly sermon. And let's see what else is in the boxcar.

Oh, no, two more stories.

All right.

Oh, this is a good one. Google Gemini struggles to write code, calls itself “A disgrace to my species.” This one's from Ars Technica, and it says that during a recent debugging session with a user, Google's Gemini AI model became overly self-critical after it failed to fix a problem with code it was trying to write. It followed up by writing “I am a disgrace” more than 80 times. Google said this was a “looping bug” that affects less than 1% of Gemini traffic, and they've been working to fix it.

First of all, absolutely do not fix this. I have never been so delighted by Gemini as I was reading this story. Has anything ever been more relatable than an AI that is working really hard on a problem, can't quite get it right, and does a lot of negative self-talk?

Yeah. Yes. This made me think that AI is ready to replace journalists, because this is my internal monologue.

Yeah.

“I'm a disgrace.”

The amount of self-loathing in the journalism profession is quite high. If this were available in the model picker, I would pick it.

Yes, this is not a mess at all.

No.

This is a feature, not a bug.

Feature, not a bug. Non-mess. Absolute non-mess. What a delight. Thank you, Gemini.

All right. And finally, this one isn't really a mess, Kevin, so much as it is one final derailing. We wanted to take a moment today to pay respect to a legend, and that legend is, of course, AOL dial-up Internet service, which is now being taken offline after more than 3 decades of service. For so many of us elder millennials, AOL was our first entry onto the Internet, and I believe we have a clip that, if I'm right, is going to trigger a massive wave of nostalgia in some of our listeners who are roughly our age, Kevin.

Let's play it one last time. God.

I literally just traveled back in time 30 years.

This is the sound of childhood.

This is the sound of happiness. Just realizing the World Wide Web was out there. Someone should make a dance remix of that and release it today. I bet it would slap.

It kind of sounds like a Skrillex song.

It does.

Now, for our younger listeners, that was the—

Yeah, what did we just hear?

—the sound of an AOL dial-up modem connection. When Casey and I were just young lads sitting there at our parents' desktop computers dialing into AOL, we had to sit through that sound. But that meant that you were going online, a magical place where anything was possible.

Yeah, and crucially, when you were online, no one could call your house.

Yes.

And so your parents would say, “Hey, you need to get off of there. Grandma is trying to get through.”

Yes.

Man.

Oh, I'm so sad about this. So this is being discontinued as of September 30, and Casey, the most surprising part of this story to me was that in 2023, an estimated 163,000 households in the United States were using dial-up Internet access.

It's so amazing, and I'm going to guess that the majority of those people actually stopped using dial-up Internet access sometime in the 2000s and just forgot to cancel their subscription.

Yes.

And so really, AOL is effectively going to be giving back tens of thousands of dollars, maybe even hundreds of thousands, to all of these customers who have unwittingly been lining the pockets of AOL for years.

Yeah. Casey, what are your most fond memories of the AOL dial-up Internet service?

For reasons that I don't even remember, we were not an AOL family. We were an MSN family—a Microsoft Network family. We had the kind of off-brand Internet service that was fine, but I was never in the dangerous chat rooms that AOL was famous for, or really any of that. But you were on AOL.

Yes, I was an AOL kid.

What are some of your AOL memories?

Well, I remember that it was a big deal when you got to go on AOL, because you had to fight for that with your sibling, if you had one, or you had to find a time when no one else wanted to be on the phone. It was this sound that meant you were going to the Internet, and the Internet was not this ambient thing that was always happening around you.

It was like a place that you had to click a button to go, and once you were there, you would get charged by the minute. So you would spend your whole day stacking up tasks that you wanted to do when you got online, so that when you got online, you could go do them as quickly as possible and not eat up your parents' monthly AOL dial-up budget.

Right, and keep in mind, these modems were so slow that it was like you were truly sipping the internet through a straw.

Yes.

Right? Just downloading an image might take a minute, like the way that making an image in ChatGPT does today. So, yeah, a lot of fun memories.

A lot of fun memories.

A lot of fun memories.

I spent a lot of time in those chat rooms.

Right.

I played a lot of online chess—

Ooh.

—because I was what they call a loser. And I even had an email account on that. So I probably will lose access to that when they discontinue service.

And do you want to say what the email address is so people can get in touch?

Yes. If you're interested in getting in touch with the 11-year-old me, you can email bigkevman1999@aol.com. Please don't email that address. It's going to go to someone else.

Ah, well, RIP AOL. And with that, America is now just permanently online.

Can we hear one more AOL goodbye sound?

Yeah, let's hear that goodbye sound one more time.

Goodbye.

And he really said it all right there.

I'm crying.

Yeah.

I'm so emotional.

And that's Hot Mess Express.

One correction before we go. Last week, Kevin, during our discussion of Alexa Plus, I said that Amazon had sent me 2 Echo Shows.

I remember that.

And I was under the impression that I had to mount my Echo Show to my wall. Well, it turned out that the second box that I had been sent, which looked basically identical to the box that had the Echo Show in it, was actually the box for the Echo Show mount that would have allowed it to sit on my desk.

You fool.

I know. So listen, I actually am embarrassed about this. I did not open the box because I didn't want to create a bigger mess for myself, because I knew I was going to return all of this stuff very quickly. But I did make a mistake, and I apologize for the error.

Alexa, punish Casey for his mistake.

Ow! That hurts.