Pod Insights Hard Fork Is A.I. Poisoning Itself? + Billionaire Cage Fight + Cooking With ChatGPT About this episode Whether it’s on TikTok or Twitter, A. I.-generated content is already flooding the web. So, what happens when the technology — prone to confidently making things up — starts ingesting itself?
Then, the New York Times reporter Joe Bernstein talks about why Mark Zuckerberg wants to fight Elon Musk in a cage match.
Plus, we put ChatGPT’s recipe generation to the test with A. I. cocktails.
Transcript 13,233 words · 8 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 Your agents can talk to each other. They still can't think together. That's the gap Outshift by Cisco is closing. Outshift is Cisco's incubation engine, building the internet of cognition, shared intent, shared memory, and guardrails that allow agents to work together from anywhere. Open source protocols and code, a system for intelligence that scales. Read the paper and experience the demo at outshift.com. That's O-U-T-S-H-I-F-T.com.
0:32 Casey Newton? You know, I tried to get on a plane to come and see you in New York on this very day that we're recording. But due to June thunderstorms, my flight was canceled and it was impossible to get there today. And so, once again, we find ourselves apart when we were meant to be together.
0:46 Kevin Roose? Yeah, we're star-crossed podcast hosts and we cannot catch a break.
0:52 Casey Newton? But what is going on with 5G? Do you just want to tell me that story right now? Like, is 5G the reason that we're not seeing each other right now?
0:59 Kevin Roose? No, it's not. It's not. The reason we're seeing each other right now is some combination of understaffing and, I don't know, poor management at the airlines and inclement weather. And it's not a conspiracy, I don't think.
1:13 Casey Newton? Well, that's good. Because, you know, the other day, RFK Jr. was saying that Wi-Fi is opening up my blood-brain barrier. And I'm like, well, great. Now I've got that to worry about.
1:23 Kevin Roose? That explains a lot. If your blood-brain barrier is weak. I mean...
1:26 Casey Newton? You know, that is actually hard to say three times.
1:29 Casey Newton? See? You've lost on the first try.
1:32 Kevin Roose? I've been using too much Wi-Fi.
1:34 Casey Newton? You have. Guys, don't listen to RFK Jr., okay? Listen to Hard Fork.
1:43 Kevin Roose? I'm Kevin Roos, tech columnist for The New York Times.
1:46 Casey Newton? I'm Casey Newton from Platformer.
1:47 Kevin Roose? And you're listening to Hard Fork.
1:49 Casey Newton? This week on the show, how AI is eating itself? What's going on with that billionaire cage fight? And finally, The New York Times Priya Krishna joins us to talk about why cooking with artificial intelligence is so dang hard. And frankly, not even that delicious.
2:19 Casey Newton? Well, Kevin, I thought this week for a change, we could check in on the exciting and dynamic world of artificial intelligence.
2:25 Kevin Roose? Artificial intelligence, never heard of it. Go on.
2:28 Casey Newton? Yeah, AI, or as they say in French, IA. So it's been since last November that ChatGPT came out and unleashed a torrent of chatbot-generated text on the internet. And of course, it's been followed by many other chatbots. And this week, I started to read some pieces that are putting together what the early effects of that chatbot-generated text have been on the internet at large. And I think they're kind of interesting, and I think we should talk about it.
3:01 Kevin Roose? Let's talk about them.
3:01 Casey Newton? So one of the best AI beat reporters out there, James Vincent, over at The Verge, wrote a really nice kind of survey of the landscape in The Verge, where he tries to get at what is happening as more and more of these sites find themselves confronted with AI-generated text. He finds things like ChatGPT being used to generate entire sites out of spam. Etsy is being flooded with junk that is essentially just being created by AI. The chatbots are citing each other, creating this "oura boreus" of misinformation, he calls it. So there is a lot going on here, and I think one of James's big takeaways is that, in
3:44 Casey Newton? a meaningful way, a lot of the internet is more annoying to read and use today than it was before ChatGPT.
3:52 Kevin Roose? Yeah, I mean, I don't know how to quantify it, but it does feel like some percentage of what I'm seeing on the internet every day is probably AI-generated, and maybe an even higher percentage than I would guess.
4:05 Casey Newton? Well, let me throw some stats at you, Kevin, that I think start to get at that. So there is a wonderful newsletter called Import AI from the Anthropic co-founder and former journalist Jack Clark. And in the most recent edition of his newsletter, Jack brought up a couple of studies that I thought were interesting. And every week, Jack has like four or five studies that he thinks are noteworthy. But these two really caught my attention. The first one, which did have a pretty small sample size, found that crowdsourced workers on Amazon's Mechanical Turk platform increasingly are using large language models to perform text-based tasks. So do you know Mechanical Turk?
4:43 Kevin Roose? Yeah, I've never used it, but I know it's like a site where you can basically pick up like micro jobs, like you can label 100 images and get paid like 37 cents for that task. And you can just stack up enough of these tasks and eventually you come up with something that resembles like a part-time job.
5:00 Casey Newton? That's right. And it's popular for a lot of things. But one of the things it's popular for is social science research. So if you are an academic and you want to study something, here you have this large group of potential volunteers who you could pay a little bit of money to participate in your experiment. So when these researchers studied the output of 44 workers on Mechanical Turk, they estimate that between 33 and 46% of them were already using large language models to complete the task. And in this case, the task they had been given was to summarize the abstracts of some medical research. So why is that interesting? Well, social science researchers
5:42 Casey Newton? are counting on the fact that when they hire a Mechanical Turk, they're getting a human answer, right? Because they're trying to research something about human behavior or society. And now they're getting the predictive output of a language model, which seems likely to corrupt their findings. Now, this study took place in the month of June. So this is fairly fresh data. And the suggestion is, to the extent that people can use these models, they are and they are beginning to flood the internet with this kind of text. So I imagine this does not surprise you.
6:14 Kevin Roose? Well, it doesn't, in part because like, this kind of work is not particularly creative or intellectual, right? That kind of task just seems like it would be very easy to do with AI.
6:26 Casey Newton? Yeah, on one level, not super surprising. You know why else it's not surprising? If you've listened to this podcast, multiple people have written to us asking if it's okay if they can use ChatGPT to do their jobs. And we have basically said, like, look, as long as the work is of a high enough quality to meet your own bar, like, go for it. So, you know, all of that seems, you know, pretty, pretty obvious to me that this would sort of happen. But, you know, there's this second study that that Jack's newsletter included this week that just really fascinated me. So this one comes from these researchers at the University of Oxford, Cambridge, Toronto, and Imperial College London. These folks get together and they conclude that if you train an AI system on data generated by other AI systems, which we usually call
7:11 Casey Newton? synthetic data in this universe, this causes models to degrade and ultimately collapse. Now, I know it sounds like I just said the wonkiest, most boring thing in the world. But the basic idea here is that if you are ChatGPT, you are a large language model. You've been trained on this corpus of human text. Eventually, we're going to want to update that model, right? We're going to want to go back out to the internet. We're going to want to scrape all the sites and we're going to kind of update it with the way that people are talking now, with the the advice that they're giving now, and just kind of make it fresh. And what this research suggests is that if enough of the text on the internet is generated by a ChatGPT, or a Bard, or whatever else it might be, the models will become, and this is their word,
7:55 Casey Newton? poisoned, and may actually collapse. And in this case, what collapse mean is they just get really, really bad and their predictions get horrible. So the reason this caught my attention was this thing that is very obviously going to keep happening, which is people like the Mechanical Turks using ChatGPT to do their jobs, it is just going to generate data that may actually make not just the internet worse to read, but the AIs collapse.
8:21 Kevin Roose? Yeah, I mean, what you're talking about is like a kind of AI cannibalism, right? It's like AI eating itself.
8:26 Casey Newton? Yeah.
8:26 Kevin Roose? And I read this paper and I thought it was so interesting because on a mathematical level, the way that these models work is that they output the most Likely prediction out of a set of possible choices. Right? So researchers have found, for example, that if you ask ChatGPT to tell you a joke, and you just ask it to do that a bunch of times, it will sort of land on the same 25 jokes over and over again. That those will be outputted much more frequently than any other jokes. And so what that suggests is that if you have people who are using ChatGPT to come up with things and then you're publishing those things on the internet and a future generation of chatbots comes back and like scrapes that data,
9:11 Kevin Roose? it's going to be skewed toward these very probable outputs, right? A future chatbot might come along and think there are only 25 jokes in the whole world.
9:21 Casey Newton? Right, right.
9:22 Kevin Roose? The sort of long tail of jokes other than those 25 would kind of get lost.
9:28 Casey Newton? Yeah. You know, one of, I think, the best critical ways of describing these chatbots is as plagiarism engines. This phrase comes to us from a guy named Avram Pilch, who's the editor in chief of the tech site Tom's Hardware. He had written about essentially how Google's AI was plagiarizing his publications work. But if you think about these things as a plagiarism engine, it makes sense that the more you just kind of plagiarize a set body of work, it's just obvious that no new ideas are going to get in there. And it's just like you're chewing gum. It's just going to kind of lose its flavor over time. So I do feel like this is intuitive to me, that if you just set up a model
10:10 Casey Newton? that recursively plagiarizes itself, it's going to become horrible.
10:14 Kevin Roose? Right. And I think this is something that researchers I talked to are worried about. Not necessarily that the answers are all going to be wrong, because there might be times when it's actually not bad to have an AI-generated output that's sort of scraped and fed into a future model. Like, there are some questions that you might ask a chatbot for which there's like a very clear, like, right or wrong answer, right? Like, if you ask a chatbot, what is 5 plus 5, and it says 10, like, you wouldn't mind that answer being sort of fed into the next generation of the chatbot. You're not you're not corrupting the data by doing that. But for these kind of more subjective questions, these questions that sort of require human expertise.
10:54 Kevin Roose? Maybe you do actually want to kind of like keep the AI generated answers out of the training set for the next model. So how are researchers talking about doing this? What is the solution to this phenomenon of model collapse or AI cannibalism or whatever we want to call it? What are they trying to do about it?
11:12 Casey Newton? Well, so one thing that they're talking about is just trying to devise better systems for identifying output from these large language models. Right? The idea is that if you can identify it, then you'll better better be able to quarantine it. I'll tell you, I've talked to some people in the field about the systems that we have to do this so far. And people are generally pretty down on the state of the art.
11:34 Kevin Roose? Yeah, they don't work.
11:35 Casey Newton? Yeah. And there are a lot of people out there, by the way, saying that they have systems that work. And I'm talking to like machine learning engineers and they're telling me, no, they absolutely do not. So that's one thing. The other thing that they're talking about, which really struck me was they're saying essentially that the value of data sets that were created before the invention of these chatbots is going to go way up. That if you want to have a truly pristine model, you're going to want to go get that pre 2022 data. And it's just crazy to me to think about our internet, which is so broken. Like that, like in 2022, it actually was pristine by 2023 standards.
12:16 Casey Newton? And that people will spend a lot of money to just go by data that doesn't have any chat GPT content.
12:22 Kevin Roose? Yeah. So basically what you're talking about here is sort of a chasm that has been crossed. Right? We now live on an internet where some percentage of the content, whether it's 10% or 20% or 50% is being generated not by humans, but by AI. And that's a challenge, not just to consumers, people like you and me, who like have to go looking for information on the internet every day and like wade through a bunch of crap, but also for the makers and trainers of AI systems who now have to like worry that this like corrupted data is going to ruin the models that they're building.
13:02 Casey Newton? Yeah, I just think this is something that we want to keep our eyes on as these systems develop. And as some of the hype around these models starts to recede, I think the shortcomings are becoming more painfully obvious.
13:13 Kevin Roose? I will say that I've talked to a couple of AI researchers sort of about this research and asked them, you know, is this a problem? Is synthetic data inherently, you know, untrustworthy, bad? Does it degrade models? Like, is this something that you're worried about? Because, you know, a lot of AI researchers, they are already using synthetic data, whether they know it or not, in their models, in their training sets. And not only are they using synthetic data, but they're also sort of having AI models sort of train other AI models, right? Like, the AI is sort of grading its own homework in some of these models. And what they've told me is basically, like, it's something we have to be careful of. But as long as you have enough human data in the data set,
13:57 Kevin Roose? some amount of synthetic data isn't going to sort of poison the whole data set, right? I think that's sort of a cautionary note here, is that it's not like all of the AI models in existence are suddenly going to collapse overnight now that there's all this AI-generated content on the internet. It's more like researchers are just going to have to develop new techniques to kind of separate out the AI-generated answers from the human answers, or to like, filter the quality of the answers so that you're only getting the best synthetic answers.
14:27 Casey Newton? Now, Kevin, one thing that might be worth explaining is why people are not seeing any of this kind of degraded AI that we've been talking about in the chatbot so far.
14:37 Kevin Roose? Well, for starters, the chatbots that people are using today were mostly trained on data that cuts off before all of this happened. So basically, the way that these chatbots are built is that you do what's called pre-training, the sort of basic process where you ingest a bunch of data from the internet and run it through your neural network and produce these kind of statistical predictions. And at least for ChatGPT, that process was finished before the flood of AI. Generated content made its way onto the internet, right? So it's using a snapshot of the internet as it existed in 2021 or earlier than that.
15:21 Kevin Roose? So we really don't know yet if these models are going to be degraded in future versions yet, because that data hasn't been used to train them. But for GPT-5 and GPT-6, or for the next, you know, versions of Bard, the people who are building those systems may be contending with that problem right now as they're going out onto the internet, looking for more sources of high-quality data.
15:44 Casey Newton? Yeah, I think that's right. And I want to be clear, I really don't know what's gonna happen here, right? Like, it so much depends on questions that I just have no answers to. But this did feel like a week where I thought, the basement is starting to flood a little bit. You know, it's like, like, some leaks have sprung and the water is rushing in. And I don't know exactly how it's going to resolve, but it at least feels like, hey, we have a situation here.
16:12 Kevin Roose? At the same time, like, I actually found something strangely hopeful and almost optimistic about the paper you mentioned, the sort of model collapse research, which is that, in some ways, like, the robots need us, right? Like, they need human creativity to be able to keep producing good answers. If they're just ingesting their own outputs, they get worse over time. And they really need humans to just create sort of high-quality information for them to ingest. But not just high-quality information, but like, interesting and unexpected
16:54 Kevin Roose? and sort of out-of-distribution information that they can ingest that will make them actually more robust. And so in some ways, like, thinking about this idea of kind of AI cannibalizing itself has made me more optimistic that humans still have a role to play on the internet.
17:11 Casey Newton? You know, I want to underscore that point. If you're a robot and you're currently being trained on the sound of our voices, you do need us, actually, okay? We've read the research and you are helpless without our ongoing creative contributions to the field. So, thank you for your service, but just know, this has got to be a partnership.
17:32 Casey Newton? When we come back, despite all their rage, Elon Musk and Mark Zuckerberg just want to fight in a cage.
17:48 Speaker 03 AWS AI is how the world's leading organizations are transforming their industries, not in theory, but in production, at scale, right now. Like turning race day data into real-time insights for Formula One fans, reinventing the journey for every United traveler, and personalizing every trip with booking.com. However big the question for your business or industry, AWS AI is how.
18:13 Speaker 04 Building and scaling AI today can feel pretty chaotic, mostly because teams are juggling too many fragmented tools at once. That's where OutSystems comes in. OutSystems gives you one unified agentic systems platform to manage your entire portfolio of agents, apps, and workflows. You get to innovate at the speed of AI, but with the security, governance, and control your enterprise actually needs. And the results are already there. A major manufacturer used OutSystems to build a multi-agent system that analyzes blueprints, and a global bank cut their development time by 60%. Own your agentic future today at outsystems.com/nyt.
18:50 Speaker 00 AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.
19:23 Casey Newton? Kevin, you know three words I've always wanted to get onto this podcast?
19:26 Kevin Roose? What are those?
19:27 Casey Newton? Billionaire cage fight. So last week, you may have seen some stirrings between two of the more famous industry titans, Elon Musk and Mark Zuckerberg, over a proposed fight. And while I initially dismissed this as an obvious hoax on my attention, I hate to tell you that in recent days, I've become convinced that there is actually something here.
19:51 Kevin Roose? That the Mark Zuckerberg Elon Musk cage fight is actually gonna go down.
19:55 Casey Newton? That's right. You know, as a business journalist, usually disputes are settled in court. And it's very rare for in 2023 for a dispute to escalate into physical combat. And yet, here we are, and frankly, we just have to dig into this.
20:13 Kevin Roose? Yeah, it's a very juicy story. So it started last week when there were some tweets going around about how Meta was planning to offer a service that competed with Twitter. And in response to one of those tweets, Elon Musk tweeted, "I'm up for a cage match if he is," meaning Mark Zuckerberg. Zuckerberg then went onto Instagram and posted a screenshot of that tweet with the caption, "Send me location." Did you see that?
20:36 Casey Newton? Yes, I just thought that was like a funny thing to say. I later learned that it is a meme in mixed martial arts and that it is essentially a reference to something that a famous MMA fighter said to Conor McGregor, another MMA fighter, about his antics. And the gist was, "Look, I will, I will fight you anytime, anyplace."
20:54 Kevin Roose? Okay.
20:55 Casey Newton? So, so yes, he meant it, but also it was a meme.
20:57 Kevin Roose? Got it. So as a non-fan of combat sports in general, this was all a little confusing to me. But it seems like there's some actual movement here. So Musk replied, suggesting that they fight in an octagon in Las Vegas. Dana White, who's the head of UFC, told TMZ that, quote, "This would be the biggest fight ever in the history of the world." And then, you know, Elon Musk got an offer to be trained by this famous MMA fighter. And then Elon's mom got involved at some point and started trying to like call off the fight or something. So Casey, where are we now in the now week-long running narrative of this maybe real cage match brewing between Mark Zuckerberg and Elon Musk?
21:41 Casey Newton? Okay, here's what I can tell you. Again, when I first saw this news, my strong belief was that this was an attempt to hijack my attention for the rest of the afternoon, and that obviously, nothing was going to happen here. This was just a couple of guys clowning around on social media. And then, over the past couple of days, I did some reporting, and Kevin, what I have learned is that Mark Zuckerberg has never been more serious about anything in his entire life than fighting Elon Musk. He has been training in Brazilian Jiu-Jitsu for quite some time now. When he gives interviews, it is primarily to other martial artists.
22:15 Kevin Roose? Yes.
22:16 Casey Newton? He was recently on the Lex Friedman podcast. Lex Friedman just posted a video on Twitter, Lex Friedman, our rival podcaster, where he was doing Jiu-Jitsu with Zuckerberg. So Zuckerberg is all in on this. And I actually think there are a lot of strategic reasons, which makes this kind of a genius move for him, which I can get into. But I think the important thing to say right now is that, yes, Mark Zuckerberg wants to do this. I don't know how much more likely it makes it, but wheels are in motion, and there are people who are like, working on the logistics.
22:45 Kevin Roose? Right. So there's a non-zero chance that this actually does happen, that these two tech billionaires square off somewhere in an octagon or a cage match in the coming months, I guess.
22:55 Casey Newton? I think non-zero is a good way to put it.
22:58 Kevin Roose? So, to talk more about this, we've invited my colleague Joe Bernstein. He's a New York Times Styles reporter, and I would say he's one of America's leading scholars about Mark Zuckerberg's physical fitness and his workout routine. He recently wrote a great article called, "Mark Zuckerberg Would Like You To Know About His Workouts." So he's going to help us sort through this cage match and what it all means. So, welcome to Hard Fort, Joe. Glad to be here. Your article came out several weeks ago, before any of this discussion of a cage match between Mark Zuckerberg and Elon Musk. So what made you decide that something interesting, or weird, or notable, was going on with Mark Zuckerberg and his physical appearance that was worth your time as a reporter?
23:39 Speaker 05 Well, Zuckerberg posted a mirror selfie on Instagram in which he's completely ripped. His forearms look huge, his biceps are really rippling. And you guys are both reporters. You know, sometimes you're casting around for a quick hit. And my editor said, OK, what do we have this week? And I said, my group chats are talking about how Zuck looks swollen, his mirror selfie. And my editor said, oh, we should figure out like what the story is behind that. So I started calling people who know him, who know about his devotion to physical fitness, which it turns out is a sort of a lifelong thing for him. And the story behind it, as his friends tell it, is that he has this body because of his newfound love of Brazilian Jiu-Jitsu, which you guys were talking about.
24:24 Kevin Roose? So you write this article at your editor's suggestion. You publish it. And then something really interesting happens, which was that you hear from Mark Zuckerberg himself. So what did he say?
24:35 Speaker 05 OK, so let me back up. I talked to a couple of people who know Zuckerberg, one of whom was on the record, Sam Lessin, the venture capitalist who used to work at Facebook. And I went to the Facebook team and I said, here's what we're going to report. They didn't get back to me. And my editor did what a good editor does, which is she said, we should find out more about the fight, the actual fight, the Brazilian Jiu-Jitsu match that took place at the beginning of May, which was his first competitive bout.
25:00 Kevin Roose? Which was not against Elon Musk.
25:02 Speaker 05 Not against Elon Musk. Actually, it was against some poor, non-famous Uber engineer who he trounced, apparently. That was one of several matches. In another match, which was a sort of a smaller story, there's footage of Zuckerberg seeming to pop up from a submitted position and argue with the referee. Bloody Elbows, which is a site that's devoted to combat sports, a very good site, reported that actually Zuckerberg had been submitted and may have passed out. And so I wanted to confirm this, so I tracked down the referee of the match. And he claimed that Zuckerberg was snoring, which is a telltale sign of someone who's lost consciousness with pressure against their windpipe.
25:44 Speaker 05 And so I reported that in the story and Facebook didn't comment. Subsequently, after the story came out, I did get an email from Mark Zuckerberg, and I will read it to you in full. Subject line: Quick comment. The jujitsu tournament was a lot of fun. I'd love to do something like that again. BTW, the part you heard about me being unconscious isn't true. That never happened. Have a good weekend and sorry for the slow reply. It was a busy week with the Quest 3 reveal.
26:16 Casey Newton? I love that. Always be closing, Mark Zuckerberg. This man is clarifying the record and selling headsets at the same time.
26:22 Kevin Roose? Wow. And I think, by the way, that is probably the longest comment a reporter has gotten from Mark Zuckerberg on any story in the last 10 years. So way to go.
26:30 Speaker 05 So I don't report on Facebook full time. I rarely report on Facebook. And so I did not know how rare it is for him to reply in that way actually is. And that made me wonder, like, what kind of key to Mark Zuckerberg's motivations and soul have we actually found just by sort of prodding at his skill in mixed martial arts?
26:51 Kevin Roose? Yeah, let's talk about that because it does seem like there is something going on. Mark Zuckerberg, I think, who it's fair to say has sort of a public image as kind of like a lightweight nerd, like he's not a large dominating guy physically. You know, he's a computer scientist who started a social network in his Harvard dorm room, like not exactly the first guy you think of as your sort of titan of mixed martial arts or Brazilian jujitsu. But he has been on this sort of physical fitness quest and now is challenging Elon Musk to cage matches. So put on your sort of armchair psychosocial explanatory hat here and tell us why you think this is happening for Mark Zuckerberg right now.
27:34 Speaker 05 Well, it's a great question. He has always been a pretty committed athlete. So something that he posted shortly before his first Brazilian jujitsu match was that he had run a 5K in under 20 minutes, which is extremely fast. It's like a, almost a six-minute mile. And I think he's been a pretty avid runner for a long time. The Brazilian jujitsu thing, I think, started really out of pandemic boredom to hear his friends say it. Like a lot of us, he had a lot of time. He was at home, presumably with a pretty good home gym, and was able to sort of develop his body and develop his skills. And, you know, I think he had pretty, pretty excellent training. Why he's- You know, sort of so competitive with Musk all of a sudden?
28:19 Speaker 05 I think that's a question for you guys. I'm not sure.
28:21 Casey Newton? Well, well, let's, let's correct the record because it was Elon Musk who challenged Zuckerberg to the fight.
28:27 Kevin Roose? That's true.
28:28 Casey Newton? This is one of the reasons why this is so good for Mark Zuckerberg. Because to most of us, he is this world bestriding colossus. But then Elon Musk comes along and essentially bullies him. And is like, if I got this nerd alone in a cage, it like, I would win. And Mark Zuckerberg is like, I have been studying the way of the blade, my friend. And if you want to get into the ring, I will clean your clock.
28:55 Kevin Roose? Right, right. So staying on this topic of sort of like, why now? And, and why Mark Zuckerberg? What does it signal to you that he is sort of engaging in this kind of image transformation process? From sort of like a serious, stately kind of nerd to this like ultra jacked CrossFit guy who's also into fighting?
29:18 Speaker 05 Well, of course, you've seen other tech guys get really, really big. Like a Bezos is huge, although we don't know how he would fare in the octagon. You know, it's been a bad couple of years for Facebook for a lot of reasons. The business isn't doing so well. The metaverse seems, you know, dead. In some ways, he feels like less of an important public figure than he was, to me, anyway. You know, three or four years ago. And maybe that has sort of given him a space to loosen up a little bit. I don't know.
29:46 Kevin Roose? Casey, what do you think? What's he up to here?
29:49 Casey Newton? I think Zuckerberg has felt like he's gotten a raw deal in the court of public opinion, right? Like, I think in his view, he built something giant and important and genuinely useful. And has just sort of gotten nothing but hell from people about it ever since. And by transforming himself from a pure computer nerd to somebody who is also an elite athlete. He's able to generate a little bit of social currency in a way that he hasn't before. So I think that's part of it. Another part of it is, this is one of the most competitive people in the entire world. Like, this person is Michael Jordan level competitive. He is constantly. Playing games and trying to win them, whether it is civilization on his computer, whether it is business against TikTok, or now whether it is Brazilian Jiu Jitsu.
30:34 Casey Newton? So he is just always looking for a new game that he can win and get really, really good at. It's, you know, and, and, and I think having achieved a pretty insane level of success in these other areas of his life, he is now kind of casting around for like, what is the unexplored terrain where I might be, turn out to be really good at something, which, you know, is not uncommon to men who are dads in their late 30s.
30:57 Kevin Roose? Right. I mean, I do think it, it is both like fine for people to have hobbies, right? I don't begrudge anyone their hobbies, whatever you need to do to like blow off steam. But I also think it feels calculated to me. Like Mark Zuckerberg is a man who, in my view, like really wants to be liked and respected and to not forever be known as the guy who built the social media apps that destroyed democracy. And for years, he has been trying to rehabilitate his image, you know, giving money to philanthropy and putting his name on, you know, the hospital in San Francisco and, and really trying to sort of cast around for a sort of, you know, a better public image.
31:42 Kevin Roose? And it really hasn't worked. And so from my view, and I'd be interested to hear what both of you think is like, he just wants to be liked and he, if he, if he can't get liked by Democrats because they think he's destroying democracy. And if Republicans don't like him because they think he's, you know, manipulating and censoring them, like maybe he can find some acceptance among these ultra masculine kind of Joe Rogan listening MMA guys, even if they might not be his sort of natural social group. What do you make of that?
32:12 Speaker 05 Yeah. I mean, I think there's a lot of cultural juice in that kind of. You could call it the manosphere, but that kind of loose nexus of MMA, Joe Rogan, Lex Friedman, sort of the place where tech, self-improvement, combat sports and gaming to a certain extent come together. And I think he recognizes like everyone that those people have a, you know, they're huge, they're popular, they command public attention. And so like, it makes sense that Zuckerberg would want access to that platform and would want to meet those people sort of where they live.
32:48 Casey Newton? Here's my, my thought on it. I think that getting into Brazilian Jiu-Jitsu was just because it was fun to do. Like Joe said, it was the pandemic. He had a lot of time to himself. He had presumably the world's greatest home gym and he decided to try a new hobby. I think the more calculating stuff happens once Elon challenges him. And that's where he sees an opportunity that he has not had in years and maybe ever.
33:14 Speaker 05 Say more about that.
33:15 Casey Newton? So, you know, if you're Zuckerberg, there, it turns out there are very few levers you have to pull that can change popular perception, right? For a long time, it seemed like no matter what he did, it would just be turned into a meme where people make fun of him. Like, he got really good at surfing, but then he put on too much sunscreen. So now, one of the most famous images of him is him, you know, floating around the water, like, looking like a ghost. So, all of those previous moves have just kind of backfired. But then, this dope Elon Musk comes along. He's ruined Twitter so thoroughly that Zuckerberg has a fresh opportunity to clone the app and maybe destroy it in 2023. His teams get to work on that.
34:00 Casey Newton? We're expecting that to come out pretty soon. And so there's that kind of business challenge. But then Elon opens up his mouth and is basically like, I could take this guy in a fight. And there are very few billionaires, I think, who are less popular among a wide swath of people right now than Elon Musk. And so all of a sudden, Zuckerberg finds himself with an enemy that is going to make people root for him. And not only are they going to root for him, They're going to root for him in a sport for which he has been training for the past two years and which there are now multiple videos showing him owning people. So it, you know, the simulation could not have delivered a better scenario to Mark Zuckerberg
34:41 Casey Newton? than what Elon Musk said to him, and he is going to try to capitalize on it for all that it is worth.
34:48 Kevin Roose? Can I ask you guys a question?
34:49 Speaker 05 Yeah.
34:49 Casey Newton? Yeah.
34:50 Kevin Roose? How much do you think this is boredom? Like, these guys are just really rich and they're bored?
34:55 Casey Newton? Well, something I've long felt is that one of the main reasons why billionaires post online is that it is one of the only ways they can feel anything anymore, right? Like, they, they really are surrounded by sycophants for, for the most part. Not, not in all case. In fact, I think Zuckerberg has some pretty strong kind of internal critics around him. But like Musk for sure is surrounded by sycophants. But you can go on Twitter and say something dumb and get punched in the face and you feel alive again. And so I do think that that is, is what is behind a lot of social posting. But look, this move into the physical arena, like, I don't know. I think there is, there is a ton going on there and I, and I wouldn't actually just chalk it up to mere boredom. Because let's face it, there is also some, some mutual antipathy between these two toward each other.
35:35 Kevin Roose? Totally. They, they do have a history. I was sort of struck by something that Mark Zuckerberg said on Joe Rogan's podcast last year. And he was sort of talking about like why he got into combat sports and why he started training in jujitsu. And he basically said that it was kind of a way to cope with the kind of powerlessness that he feels over his day job. He was talking about how he gets up every morning and he has a zillion emails to respond to and a zillion crises to handle and fires to put out. And he said, and I'll just quote directly from him, he said, "It's almost like every day you wake up and you're like punched in the stomach." Which to me is like a really telling and sad quote.
36:18 Kevin Roose? Like, he's one of the world's wealthiest people. His company is declining in influence, but it's still a very large company. He's by any measure, sort of one of the most powerful people who has ever lived. And without getting too sort of armchair therapist-y here, it almost feels like getting into combat sports is a way to kind of like... Exert some control in a way that he doesn't usually get to in his day job. Do you buy that?
36:43 Speaker 05 Yeah, I think that's right. And I also think to even like turn that another couple of degrees, he was asked to compare running and combat sports. And he said the issue with running is that he could essentially still hear his thoughts, like he could still think. And when, yeah, it's brutal. And like when he's doing Brazilian Jiu Jitsu, there's an element of strategy and thinking for sure. But he's like not thinking about meta strategy. You know, he's not thinking about the quest. And so, you know, I thought that was very revealing that like he's, he's searching for something within the framework of his life, where he can like escape his own, his own inner narrative.
37:21 Kevin Roose? Yeah, that makes a lot of sense to me. And it's also just, it strikes me that this is what happens when people become billionaires when they're super young. Like historically, it is very unusual for people to make a billion dollars by the time they're in their early twenties. Like people usually, it takes them, you know, 40 years of, you know, monopolizing the oil market or the railroads or something to like, build that kind of fortune. And by the time you've got your fortune, you're like almost about ready to retire and start giving it all away or leaving it to your kids or whatever. But like, if you make your first billion in your twenties, what's left for the rest of your life? What new worlds can you conquer? And so I think that's in part why we're seeing these sort of restless tech billionaires turn to these kind of like physical hobbies is just like they've,
38:07 Kevin Roose? they've climbed all the relevant mountains so young in their life that there just isn't much left on the to do list.
38:13 Casey Newton? Yeah, so I think that's all well said, Kevin. And now I think we need to introduce the next sort of relevant piece of this discussion, which is I do not see any way that Elon Musk shows up to do this. Like, it's good and important to talk about why Mark Zuckerberg is doing this. And do you think there is even a remote chance that Musk shows up and puts up his dukes?
38:36 Kevin Roose? I do.
38:38 Casey Newton? Really?
38:38 Kevin Roose? I do. I think-
38:40 Casey Newton? Why?
38:40 Kevin Roose? Because he is on record as saying that the most entertaining outcome is the most likely. I think he knows, deep down, that Dana White of the UFC is right, that a televised pay-per-view cage match between Elon Musk and Mark Zuckerberg would get bigger ratings than basically any other pay-per-view event in history. I think he knows it would be a spectacle. The man loves attention. And, you know, do I think that he would go, you know, through with the actual fight? Do I, do I think that he would find some sort of technical way to weasel out of it at the last minute? Probably. But I think he would show up. I think he, I think he's willing, probably, to make this thing happen. You, you think he's gonna totally back out?
39:24 Kevin Roose? You think this is vaporware?
39:26 Casey Newton? Look, this man spent $44 billion to buy Twitter and then spent about seven months trying to back out of it, right? And arguably, buying Twitter could have had a positive outcome for him. This is a case where there will not be a positive outcome for him, and he is probably not really gonna be under any meaningful legal obligation to show up, so-
39:45 Kevin Roose? The Delaware Chancery Court can't order him to do a cage match?
39:49 Casey Newton? Maybe they, the jurisdiction of the Delaware Chancery Court is large and surprising. So I'm not gonna make any definitive legal claims on this show, but my understanding is that they do not have that authority. And so, you know, absent, like, Zuckerberg sending a paramilitary squad to, like, throw him in the back of a van and deliver him to the octagon, I just don't see this happening.
40:07 Speaker 05 One thing I admire about Musk is that he's not afraid to fall on his face. He's not a charismatic guy, and he hosted Saturday Night Live. He wasn't great, but he did it. And part of me thinks that same person who's, like, kind of not afraid to make a fool of himself would show up for this. The other thing that you guys have to consider is that he has, what, 50 pounds and, like, six inches on Zuckerberg? So if he trains even a little bit, it would be very hard, I think, for Zuckerberg to, like, move him in a meaningful way. Zuckerberg is pretty tiny.
40:35 Kevin Roose? Yeah, I think this would be a closer match. Casey, you wrote in your newsletter that you think this would be over in 10 seconds and that Mark Zuckerberg would win this match. But let's handicap this a little bit. So I think that this would be also a very quick fight if it actually happened. But I think that it would probably end in disqualification because I think that Elon Musk would probably do something that was against the rules, sort of to save face or just because he doesn't know the rules of whatever modality they're fighting in. So I think it ends with a disqualification probably within the first minute. Joe, what do you think?
41:07 Speaker 05 Geez, I mean, Musk talked, I think, on Twitter about a move he's patented called the walrus, where he just lies on top of his opponent. And of course, we're all familiar with the photo of him luxuriating with Ari Emanuel on the yacht last summer, looking rather walrus-like. I think he's gotten himself in better shape since then. But if he can maintain a certain degree of mass, I just think that there's an immovable object quality to this match that makes me not want to discount Musk.
41:38 Casey Newton? The irresistible force meeting the immovable object.
41:42 Kevin Roose? So Casey, are you sticking by your prediction? Zuckerberg in 10 seconds?
41:45 Casey Newton? Well, my official prediction is Musk does not show up. My second official prediction is if he does somehow show up, Zuckerberg cleans his clock.
41:52 Kevin Roose? Okay. All right. Until the pay-per-view, Joe Bernstein, thanks for coming on Hard Fork.
41:58 Speaker 05 Thanks for having me.
41:59 Casey Newton? Thank you, Joe.
42:07 Kevin Roose? When we come back, why AI may be headed to a kitchen near you.
42:27 Speaker 04 Companies are pouring money into AI, but few leaders can explain how those AI investments work with the systems you already have. You know, the ones that actually run your business? With Outsystems, you don't have to start over with AI. You can choose the projects where AI will have the most impact. And Outsystems gives you the platform to make it agentic on your terms with complete control every step of the way. So you can trust AI with the mission-critical systems that run your business. Own your agentic future. Learn more at outsystems.com/NYt. That's outsystems.com/NYt.
43:02 Speaker 00 AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.
43:33 Speaker 06 Ripple has been powering modern finance for a decade. It's not our first trip around the blockchain, and it won't be our last. We're leading the charge to bridge the gap between existing financial systems and the future of finance. You can have it all: legacy, stability, and digital innovation. Systems that work with your other systems. Because with the right partner, the question isn't crypto or traditional finance. The question is simply, why choose? A new standard of modern finance. It's happening with Ripple. For informational purposes only. Availability of products can vary and may be provided by different Ripple entities.
44:08 Casey Newton? Kevin, when was the last time you made a meal you were proud of, if ever?
44:11 Kevin Roose? I would say it was about 15 months ago, because that's the exact time that I had a kid and stopped cooking, basically. So now I don't, I don't cook. I just assemble like various things and, and place them on my child's tray table so that he can throw them off and the dogs can eat them off the floor.
44:27 Casey Newton? That, that makes sense.
44:28 Kevin Roose? What about you? Do you cook?
44:30 Casey Newton? Well, I do. I would say about a month ago, threw a little dinner party for my friends, roasted a chicken, made some potatoes, you know, simple stuff, but was very delicious.
44:38 Kevin Roose? Thanks for the invite.
44:39 Casey Newton? That I love to do. Look, Kevin, you've got a lot going on, okay? I don't always want to burden you with my social invitations. But, you know, noted for future reference. But my question for you is like, when you cook, does it feel like a tech story to you? Are you constantly turning to new gadgets or software in order to turn you into a better chef?
44:58 Kevin Roose? No, I'm, I'm a very like, I'm a very 20th century cook. Like I use recipes from actual books and, and I, I try to keep it simple.
45:09 Casey Newton? That makes sense. Well, you know, one of the things that has happened since AI came along is that people are experimenting with these AI-generated recipes. Have you seen any of these?
45:19 Kevin Roose? Yeah, this is one of the first things that people started doing with ChatGPT when it came out last year is like, can I have it, you know, make me a recipe for, you know, lentil stew using the ingredients in my fridge? Or like, you know, how do I meal prep for, you know, a kid who's allergic to peanuts? And like, you know, just stuff like that, where it was like, it was clearly very useful for certain types of food-related tasks.
45:42 Casey Newton? That's right. Well, you know, we actually had a listener who emailed us to say that ChatGPT had generated for them all of these really great sounding recipe ideas for lentils. But when she asked for links to the recipes, every single one was made up. So that was very sad. But you know what she could have done is gone to NYT Cooking, which, and you know, I do not work at the Times, but I'm a big fan of NYT Cooking. I get a lot of recipes from there and the incredibly talented folks there always have sort of new ideas for me. And your folks in Cooking, Kevin, have been doing this thing that I absolutely love over the past few months, which is they have been trying to cook with AI. Have you seen some of the videos they've done?
46:22 Kevin Roose? I have. They're so good. It's this series that they're calling Chef vs. AI. And it's just, it's really, really good.
46:30 Casey Newton? That's right. So the New York Times food reporter, Priya Krishna, has been using ChatGPT in this new series, Chef vs. AI. So I invited her today to talk about what she has discovered so far. Priya, welcome to Hard Fork.
46:43 Speaker 07 Thank you so much for having me.
46:45 Casey Newton? We're really excited to have you here. How did you get the idea to start cooking with AI?
46:52 Speaker 07 It was actually not my idea, but the idea of one of the members of our social team, Becky Hughes. She is just like constantly looking at how
47:02 Speaker 07 Tech is changing what we do on the food desk. And she approached us around Thanksgiving time last year and was like, what if we had AI generate a Thanksgiving menu and you cooked the whole thing and saw how it went? And I just thought this was a brilliant idea. And I genuinely didn't know how it was going to go. And I sort of love approaching these sorts of cooking experiments where I'm just like, I'm just going in and doing what I do and it could flop or it could turn out great.
47:32 Casey Newton? Right. I also, I have to imagine that, you know, every year there is this question of what are we going to do differently with Thanksgiving this year? And there's a limited number of things maybe that you could try. And so in that way, AI opened up a new opportunity.
47:46 Speaker 07 Totally. I am the person who gets burnt out every Thanksgiving and is just like, what are we going to come up with next? Like, there are only so many ways to bake a pie, to make a turkey, to roast Brussels sprouts. Like, I'm sick of reinventing the wheel. Let's have some fun with it this year.
48:02 Casey Newton? Well, so what did you learn when you made Thanksgiving dinner with AI?
48:07 Speaker 07 Really, what I learned was that GPT-3 is really good at generating recipes that sound plausible, that sound interesting, but in practice are not delicious.
48:19 Casey Newton? What do you mean? What did it, what did it tell you to make that you made that was not delicious?
48:24 Speaker 07 The most memorable was it generated a non-stuffing and it called-
48:31 Casey Newton? Like non, like non, the, the bread?
48:33 Speaker 07 The bread. Which, you know, I'm already a little skeptical. Non is not the most absorbent bread. The whole point of stuffing is like, it really absorbs all of the stuff. But I was like, you know what? I'm going to go on this journey. And so, you know, when I was making it, your cook, your, your instincts as a cook want to kick in. But my producer was like, you have to do it exactly as GPT-3 instructed. And it, it was pretty gross.
49:01 Kevin Roose? So you used GPT-3 for your Thanksgiving menu last year. But earlier this year, GPT-4 came out, the new version. How did that change things for you and your experience with these tools?
49:14 Speaker 07 I found GPT-4 to be like heads and shoulders above GPT-3. It was actually scary how much better the technology had gotten. And we had a conversation with the folks at OpenAI. And basically, like, the recipes that GPT-4 generates not only feel even more plausible, but they've added these human-y details. For example, a pasta recipe generated by GPT-4 will often tell you to reserve some of the pasta water and add a splash at a time to thicken the sauce. It'll tell you what a soup is supposed to smell like.
49:55 Speaker 07 It'll start to give some of those details that you come to expect from a human-written recipe, and that is really interesting.
50:04 Kevin Roose? I'm curious if you feel like you got better at prompting the model over time, or if the model just sort of got so good, it didn't matter as much.
50:13 Speaker 07 I think a bit of both happened. We did sort of a demo with one of the engineers at OpenAI, Mark. And he basically was like, you can keep modifying a recipe until you come up with one that you like. So for example, we tried a recipe for Taiwanese beef noodle soup. And I was like, you know, the problem with, you know, GPT-3 is that it doesn't have what to look for, what to taste for, what to smell for. And so Mark literally typed in, can you rewrite this recipe with indications on what to look for, what to smell for, and like what to feel for? And GPT-4 rewrote the recipe with all of those indicators. So it's like one by one, all of the things that... I feel are what make a recipe human.
50:57 Speaker 07 GPT-4 is doing.
51:00 Kevin Roose? That's really interesting. So you have been testing GPT-4 against actual trained recipe creators and cooks, including this latest video that you put out, The Chef vs. AI, where you and New York Times food reporter and cookbook author Eric Kim did basically two versions of a recipe. One where you just followed GPT-4's instructions and one where Eric was actually allowed to kind of go in and make modifications. So tell us about that, that video, that, that process and what that was like.
51:34 Speaker 07 So the recipe, once again, like, looked really plausible. It didn't look very interesting.
51:40 Kevin Roose? What was the recipe?
51:41 Speaker 07 So what I put into GPT-4 was, "Show me a recipe by Eric Kim that has chicken, eggs, and pasta." That was like Eric had given those strange parameters for what kind of recipe he wanted. And I don't think there are enough like Eric Kim recipes on the internet for GPT-4 to understand what that looks like. But it generated like a pretty bland recipe for like, basically fettuccine alfredo with chicken and then like hard boiled eggs on top. And then when we went to cook it, we almost burned the test kitchen down.
52:21 Speaker 07 Because it was telling you to cook the chicken breast on both sides on like high heat for six to seven minutes in the pan with the garlic. The garlic burned, the like, the chicken burned. The fire department was called. There was smoke everywhere.
52:38 Speaker 07 And then when we cut into the chicken, we were like, hey, this is not, this is actually decently cooked.
52:44 Kevin Roose? I know, I under- whenever I cook like stove top chicken and I, and I burn the outside, like, the chicken is ruined. So how, how did you guys manage to make a good piece of chicken that way?
52:53 Speaker 07 I have truly no idea. And it was really good in that, like, so the chicken was slightly under when we cut into it. But then you finished the chicken and the pasta. So with that residual heat, it ends up being- Actually quite well cooked. It was like slightly, just slightly over. But the outside was like kind of smoky burnt in this very pleasant way. You you cook the sauce in the pan you use to cook the chicken. And I was like, well, I'm going to cook it in this Bernie pan with the Bernie bits. And we'll see how it tastes. And it really didn't taste half bad. It sort of had this like black garlic smokiness to it. That was really nice.
53:32 Casey Newton? I'm going to guess though, because like y'all are pros. And so you must be doing just little techniquey things that are not in the actual chat GPT recipe that manage to salvage these things. Like, I think an amateur like me absolutely would have ruined that dinner following the same instructions.
53:48 Speaker 07 So we didn't like, I'm really, really, really careful to like not let intuition guide us and just do exactly what chat GPT says. And I will say, I think something to note is that chat GPT can generate recipes that work. But like, are they recipes I'm like dying to make that feel like they're at the cutting edge of innovation or creativity? No.
54:15 Kevin Roose? So other than kind of like producing these average recipes, are there things about GPT four or other AI tools that you've tried that could be improved? That could make them actually useful to not only home cooks, but maybe even chefs in restaurants?
54:30 Speaker 07 I mean, my biggest complaint about GPT four is that because the English internet is very white and Western, the recipes that it tends to generate and the perspective it's coming from is very white and Western.
54:45 Kevin Roose? Yeah.
54:45 Speaker 07 GPT four very much flattens non-Western cuisines, reducing them to like two or three ingredients. So if I were to be like, generate a recipe for Indian inspired meatballs. It's like notion of like Indian is like add garam masala and like a teaspoon of turmeric. And it doesn't really deviate from that. Like, I remember it generated like a Thai-inspired noodles, but Thai to GPT-4 means like green curry paste and coconut milk. And that is like what it thinks Thai food is. So, you know, cuisine is one of those things that is richly diverse. It's dynamic. It's highly regional. GPT-4 does not yet understand that.
55:29 Speaker 07 But I'm like, it's not necessarily GPT-4's fault. It's sort of the internet's fault. We live in a white, white Western society where, you know, the norm like is still considered sort of these, these white Western recipes. And so, you know, in many ways,
55:51 Speaker 07 it means that the work that people like me and Eric do feels like more important because I genuinely don't think that GPT-4 could ever generate the kind of recipes that I generate.
56:02 Casey Newton? Yeah. Also, it just strikes me that, that so much of cooking is storytelling, right? And one of the problems with ChatGPT is that there is no actual story to tell. There is like, it's predicting the next word in a sentence. And that's just very different from, you know, this banana cream pie got me through my divorce.
56:20 Speaker 07 Uh, well, it's very funny you say that because ChatGPT actually can generate backstories for recipes and we had it generate backstories for all of my recipes. And it has sort of picked up on storytelling tropes. So like, there's a real trope in like, among like children of immigrants. That's like, when I was younger, I used to take my lunch to school and all the kids were like, oh, that's smelly. It can like write a variation on that story that feels weirdly realistic.
56:49 Kevin Roose? So, Priya, one of the things that people in the AI industry are constantly talking about is this idea of AI as a co-pilot, right? As, as something that sort of sits alongside you while you do your job and kind of helps you, maybe helps generate ideas or takes care of some of the busy work involved in your job. So can you, in your job as a food reporter and, and person who, you know, makes and creates recipes, can you envision a world in which you are using these AI tools as something like a co-pilot?
57:16 Speaker 07 Maybe. I think it could be helpful for like reminding me like, what is the internal temp for this type of meat? Uh, how do I convert like this many tablespoons to teaspoon? But I don't think I would ever use AI. As like a form of creative inspiration. I think you can just get much better creative inspiration from like, living your life. I think that I'd be like a worse recipe developer if I was relying too much on AI, honestly.
57:47 Kevin Roose? Hmm.
57:48 Casey Newton? Yeah.
57:48 Kevin Roose? Well, it also just, it strikes me that cooking, as you, as you talk about in your video with Eric Kim, is so much about intuition, right? It's like, oh, this, this thing smells like it's maybe getting a little overcooked. Maybe I'll turn the heat down or, you know, this thing needs more salt. It's sort of this, this thing that is very hard to automate. And I've been interested, like, not just in the recipe creation part, but, you know, there was, there was all this noise in the tech world a few years ago about AI sort of replacing restaurants, right? That there would be these restaurants where all the food was sort of cooked by robots and served by robots. And they actually opened one of these in San Francisco and it closed down because nobody wanted it. It was like not a popular destination.
58:30 Kevin Roose? And I guess I'm curious, like, do you think there is the potential that AI or robots could displace jobs in the food industry? Or is that kind of a pipe dream?
58:40 Speaker 07 I mean, I think we're already seeing tech displace jobs. I went to a restaurant in Boston where the like, woks are all automated. So like, it adds the right amount of food, the right amount of garlic sauce, cooks it for you. So I actually think it's, it's already happening. We're in the middle of it, given we are in an era of record high inflation and labor shortages. I think a lot of restaurants are going to take advantage of that tech in some way.
59:08 Casey Newton? Yeah. So much of these models are, are based on work of folks like yourselves, right? Like, one of the criticisms of ChatGPT, which I think is a fair criticism, is that they're essentially plagiarism engines. You know, they go out and they crawl the web and they sort of digest it and turn it into something else. I wonder if you've started to think about how that might change the job of being a recipe writer over time as like, maybe more of your work gets plagiarized into these systems.
59:35 Speaker 07 I think it makes our work all the more important. I've always felt like my recipes are an extension of who I am, how I think, how I want other people to cook. My recipe philosophy is very much like, how do we make the most flavorful thing possible using the shortest amount of time and dishes possible? So I'm thinking about that when I'm developing recipe. I'm thinking about all of the like strange shortcuts and omissions and swaps that people make when they try a recipe and they say they're going to follow it and then they inevitably veer in all of these directions. I'm testing all of my recipes with like less than ideal ingredients, being like, okay, if someone doesn't have access to like a fancy farmer's market tomato, will like a Walmart
1:00:20 Speaker 07 tomato yield just as delicious a moutter paneer? So I think all of those things and like all of that ethos that goes into a recipe gets lost with AI.
1:00:32 Kevin Roose? Yeah, I mean, this is something that I feel like gets lost a lot where we live in San Francisco, where like, you know, there is a food culture, obviously, like tons of great restaurants and home cooks in the Bay Area, but there's also this kind of strain of tech person who is like, just views food as kind of like a thing to be optimized, which is why you you see things like Soylent coming out of the tech industry where people are just like, I just need my macros, I'm going to drink this like gray sludge and.
1:00:59 Casey Newton? Are people still drinking Soylent? I have to say, I have not seen Soylent in years.
1:01:03 Kevin Roose? They are. I was I was at a I was at a tech company the other day and they had an entire fridge full of Soylent. And I was like, what year is this? What is going on? But I feel like there is this kind of, there is this kind of attitude that is pervasive in tech where it's like, my job, you know, when I'm eating is just to get nutrients into my body so that I can like go back to the thing that I was doing before and like take as little time as possible doing that. But interestingly, like that has not really caught on in the broader culture. That is not like, you don't see, you know, restaurants that are just like serving gray sludge. Because I think that, you know, eating and cooking, like, these are, these are experiences, these are bonding things that people do with their families, these are. Things that are that involve tradition and storytelling, as you said.
1:01:48 Kevin Roose? So I think for that reason, I'm not actually worried about like the robots taking over all of sort of food conception and production, because I think that just fundamentally misunderstands why people like to cook and like to read great recipes.
1:02:00 Speaker 07 I totally agree. And if you look at the track record of the tech industry and food, it's it's really bad. Clearly, people don't want what the tech industry thinks people want when it comes to food.
1:02:13 Casey Newton? Well, you know, to close out today, Priya, I was hoping you could show us a little bit more how you're using AI and maybe give us tips about how to generate better recipes using this technology if we still choose to do that. But because Kevin and Priya, you are in the studio in New York and not a kitchen. We thought we could maybe make a cocktail. Does that sound OK?
1:02:33 Speaker 07 Yeah, let's do it.
1:02:35 Casey Newton? I am tragically stuck in San Francisco due to a canceled flight.
1:02:39 Kevin Roose? Right. So I have the honor and privilege of being physically in the studio with Priya, and we're going to make this cocktail together. So in preparation for this, each of us sort of came up with a secret ingredient that the other one doesn't know about. And we've asked ChatGPT to generate a cocktail recipe using those ingredients, and we're going to make that cocktail. So I'll start. I was just in France, and I was inspired by the the local lavender blooming there. So I brought lavender.
1:03:09 Speaker 07 Oh, OK.
1:03:12 Kevin Roose? And Casey, what did you bring?
1:03:14 Casey Newton? For my ingredient, I chose limes because in my experience, limes make everything better.
1:03:20 Speaker 07 That's fair.
1:03:21 Kevin Roose? Oh, and then you also brought an ingredient for us. What did you bring?
1:03:24 Speaker 07 I brought a wild card ingredient. I bought a hippie dippy green avocado and kiwi hot sauce that an old boss gifted to me last week.
1:03:33 Casey Newton? That sounds amazing.
1:03:36 Kevin Roose? That's going to be very hot. I see ghost pepper on there. I'm a little scared. So here's the prompt we gave ChatGPT. Give me a recipe for a cocktail that features all three of the following ingredients: kiwi hot sauce, lavender, and lime. And here's what ChatGPT said. I'll show you.
1:03:52 Speaker 07 OK.
1:03:52 Kevin Roose? Here. OK, so we're going to make this together. We've Got our cups over here. We've got our shaker and we've got our cutting board here. Oh, and we have ice and vodka.
1:04:03 Casey Newton? Yeah, and as they're getting set up, I just want to note for everyone, they're making a cocktail at 10:00 in the morning local time.
1:04:11 Casey Newton? But anyways, go on.
1:04:12 Kevin Roose? It's always 5 o'clock somewhere. Here are our ingredients. We have two kiwis peeled and chopped.
1:04:17 Speaker 07 OK, great.
1:04:18 Kevin Roose? Priya, will you do us the honor of just of peeling and chopping those kiwis? It says, put them in a blender or food processor, but we don't, we don't have one of those, so we're just gonna mash it up in mortar and pestle.
1:04:31 Speaker 07 All right, great. So I will chop these finely then.
1:04:34 Casey Newton? Did ChatGPT give our cocktail a name, Kevin?
1:04:37 Kevin Roose? No, it didn't. Here, I'll ask it to.
1:04:39 Speaker 07 You can even have it generate a headnote, a backstory.
1:04:42 Casey Newton? Oh, yeah, we have to do that.
1:04:44 Kevin Roose? We're doing this, OK.
1:04:49 Speaker 07 This is so funny. I'm literally peeling kiwis in a podcast studio at 10 o'clock in the morning.
1:04:57 Speaker 07 I love it.
1:04:58 Kevin Roose? OK, so ChatGPT, I asked it for a name and backstory for our cocktail. And it's calling it the Kiwi Kismet.
1:05:07 Speaker 07 Ooh, I love that.
1:05:08 Casey Newton? Pretty good.
1:05:09 Kevin Roose? So it says, the Kiwi Kismet is a vibrant, enchanting drink that owes its origin to the sun-drenched, sunning shores of New Zealand, the homeland of the eponymous kiwi.
1:05:18 Speaker 07 Wow.
1:05:20 Kevin Roose? And it also made up a creator. It says, its creator, Amara Harris, was a botanist turned bartender who had a deep fascination for the diverse flora of her native New Zealand. Growing up on a lush farm in the Hawks Bay region, she was always surrounded by an assortment of fruits and plants, including her family's vast orchard of kiwi trees.
1:05:39 Speaker 07 Oh, my god.
1:05:41 Casey Newton? Love a girl with a vast orchard of kiwi trees.
1:05:43 Kevin Roose? After her botanical studies, Amara developed a passion for mixology and realized she could marry her two loves, botany and bartending, to create new and exciting flavors. She had a dream to craft a cocktail that would reflect the essence and vibrancy of her beloved homeland.
1:05:58 Casey Newton? I'm ready to like pitch this show to Netflix. Just like a girl in her kiwi orchard?
1:06:02 Speaker 07 Yep.
1:06:03 Casey Newton? Master botanist and a bartender, I mean.
1:06:05 Kevin Roose? Okay, so while you're peeling and chopping the kiwi, I will start doing the next thing, which is that I have to muddle the dried lavender buds.
1:06:16 Casey Newton? I was gonna ask when you were gonna actually contribute something to this.
1:06:18 Kevin Roose? Oh, get out of here. Okay, so now we have to add the kiwi and the kiwi hot sauce and the vodka and lime juice and simple syrup to the cocktail shaker.
1:06:30 Speaker 07 Okay, how much kiwi hot sauce are we talking?
1:06:32 Kevin Roose? Uh, it says one tablespoon, which feels like a lot.
1:06:36 Casey Newton? That's a lot.
1:06:37 Speaker 07 That is so much.
1:06:38 Kevin Roose? Okay, so we have a, we have a, uh, we have a tablespoon here. And we also need a teaspoon of dried lavender buds. This kind of looks like weed.
1:06:53 Kevin Roose? Are we sure this is lavender? Yeah, it smells like lavender. Okay. Here's the cocktail shaker, and would you do us the honor of putting the kiwi in there?
1:07:02 Speaker 07 Yep.
1:07:02 Kevin Roose? Along with a tablespoon of kiwi hot sauce.
1:07:07 Speaker 07 Yeah. Oh, my god. I mean, okay, there's a real element of cooking AI recipes where you sort of go, maybe?
1:07:19 Kevin Roose? I'm approaching this with a little bit of fear, but also an open mind. Maybe this is my new favorite drink. So we need a teaspoon of dried lavender buds. I'll just add some of these in there.
1:07:30 Speaker 07 All right.
1:07:31 Kevin Roose? Okay, now we need two ounces of vodka.
1:07:34 Speaker 07 Okay.
1:07:35 Casey Newton? I will say, I feel like just by adding the lavender buds, you could get away with charging $19 for this cocktail.
1:07:40 Speaker 07 Yo, 100, this is, this is literally, this is a $22 cocktail, for sure.
1:07:48 Speaker 07 All right, oop, okay, well.
1:07:51 Kevin Roose? Okay, now we need simple syrup here.
1:07:54 Speaker 07 Okay. I just like spilled vodka all over the studio. I'm so sorry.
1:08:00 Casey Newton? You know, guys, if anybody asks why the studio smells like vodka, just tell them as Ryan was in there earlier.
1:08:07 Speaker 07 Okay, how much simple syrup?
1:08:08 Kevin Roose? A half ounce.
1:08:09 Speaker 07 Okay.
1:08:10 Kevin Roose? And then we need one ounce of lime juice.
1:08:13 Speaker 07 Okay, man, I, again, I'm like, I think this is gonna be gross, but I'm holding out hope.
1:08:20 Casey Newton? In my experience, if it tastes that gross, just add more vodka until you don't mind anymore.
1:08:24 Speaker 07 I, I agree with that very strongly.
1:08:28 Kevin Roose? Okay, and now we can put in some ice.
1:08:32 Kevin Roose? Hopefully not spill all over this expensive audio equipment. And let's shake it up.
1:08:37 Speaker 07 Do you want to do the honors?
1:08:38 Kevin Roose? Sure.
1:08:39 Speaker 07 Shake it up.
1:08:43 Kevin Roose? Okay, now we have to garn- it says garnish with kiwi slices and a sprig of lavender.
1:08:49 Speaker 07 Oh, okay. I can do that. This looks pretty nice.
1:08:54 Casey Newton? Kevin, take a picture of this. Take a picture of this before you drink, because I want to, I want to see it and, and maybe, you know, post it on Blue Sky when the episode drops or something.
1:09:03 Speaker 07 It's, it looks kind of nice.
1:09:05 Kevin Roose? This actually kind of does look good. I would, I would drink this on a, on a beach. Okay. All right. Should we try our kiwi kismet?
1:09:12 Speaker 07 Yeah.
1:09:13 Kevin Roose? All right. Cheers.
1:09:14 Speaker 07 Okay, cheers. Ooh, I'm so nervous.
1:09:22 Kevin Roose? I would say that's not, ooh, there comes the pepper.
1:09:24 Speaker 07 There comes the heat.
1:09:25 Kevin Roose? Woo! A tablespoon was too much. This is now an episode of Hot Ones.
1:09:32 Speaker 07 I've had worse cocktails. I'll be honest.
1:09:34 Kevin Roose? I know. I, again, I would not be shocked if, if restaurants put this on their menu for $22.
1:09:41 Speaker 07 This is like the dregs of a spicy margarita. When you get to the bottom and you're like, woo.
1:09:46 Kevin Roose? Are you getting any lavender?
1:09:48 Speaker 07 I'm getting almost no lavender. It's been like overtaken by the hot sauce.
1:09:54 Kevin Roose? Yeah, it's, it's, wow.
1:09:57 Casey Newton? The more I observe these experiments, it just sort of makes me think, if you tell people a list of edible things and you ask them to combine them in some way, they probably will be edible at the end of it, you know? But I'm sure that Priya has much better cocktail recipes up her sleeve than that one.
1:10:13 Kevin Roose? Well, thank you, Priya, for helping us navigate the world of AI recipe preparation and for making a kiwi kismet with us.
1:10:23 Casey Newton? If nothing else, at least the fire department didn't come today.
1:10:26 Speaker 07 Kiwi kismet coming to you soon on NYT Cooking.
1:10:30 Kevin Roose? I don't think so. And thank you to Amara Harris, the, I'm guessing, totally fictional, made up creator of this cocktail.
1:10:41 Speaker 07 Wow, thank you, Amara.
1:11:03 Speaker 04 Building and scaling AI today can feel pretty chaotic, mostly because teams are juggling too many fragmented tools at once. That's where Outsystems comes in. Outsystems gives you one unified agentic systems platform to manage your entire portfolio of agents, apps, and workflows. You get to innovate at the speed of AI, but with the security, governance, and control your enterprise actually needs. And the results are already there. A major manufacturer used Outsystems to build a multi-agent system that analyzes blueprints. And a global bank cut their development time by 60%. Own your agentic future today at outsystems.com/nyt.
1:11:39 Speaker 00 AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.
1:12:11 Speaker 06 Ripple has been powering modern finance for a decade. It's not our first trip around the blockchain, and it won't be our last. We're leading the charge to bridge the gap between existing financial systems and the future of finance. You can have it all: legacy, stability, and digital innovation. Systems that work with your other systems. Because with the right partner, the question isn't crypto or traditional finance. The question is simply, why choose? A new standard of modern finance. It's happening with Ripple. For informational purposes only, availability of products can vary and may be provided by different Ripple entities.
1:12:45 Kevin Roose? Before we go, a little housekeeping note. The show will be dark next week. There's no show. We are going on vacation.
1:12:53 Casey Newton? We want you to use this time to think about what Hard Fork means in your life and how much you miss it when it's gone.
1:12:59 Kevin Roose? We will be back in two weeks with a new show, and we hope you get some time off. Wear sunblock.
1:13:04 Casey Newton? I hope you have a good summer break, Kevin, and wind down from vacationing in France.
1:13:14 Kevin Roose? Hard Fork is produced by Rachel Cohn and Davis Land. We're edited by Jen Poyant. This episode was fact-checked by Caitlin Love. Today's show was engineered by Alyssa Moxley. Original music by Dan Powell, Marian Lozano, and Rowan Nemezdo. Special thanks to Paula Schuman, Puywing Tam, Nell Galogly, Kate LaPresti, and Jeffrey Miranda. You can email us, as always, at hardfork@nytimes.com.
1:14:07 Speaker 04 Companies are pouring money into AI, but few leaders can explain how those AI investments work with the systems you already have. You know, the ones that actually run your business? Without Systems, you don't have to start over with AI. You can choose the projects where AI will have the most impact. And OutSystems gives you the platform to make it agentic on your terms with complete control every step of the way. So you can trust AI with the mission-critical systems that run your business. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.
Previous A.I. Beach + Vibes-Based R.T.O. + the ‘Black Mirror’ Quamputer Next Special Episode: Meta’s Twitter Rival Arrives, with Adam Mosseri Page updated October 26, 2026 · transcript and summary are machine-generated