Live with Dwarkesh Patel | INBOUND 2025

Kevin. Hi. How's it going? Good.

Should we jump in?

Yeah, let's jump in. All right.

Thanks for doing this.

This is a rare treat for me not to be the one interviewing someone. I get to just chill out here.

I like this. This is my comfort zone.

Yeah.

Okay. What do you think the role of journalists will be now that these AIs are getting better at summarizing, analyzing information, putting two and two together? What will reading the front page in five years look like?

It's a good question. It's one I think about probably more than I should. It tends to create kind of an existential despair. But I think actually journalism is going to be okay. And I think what will happen over time with these AI models is that the job of journalism will sort of skew in the direction of collecting information, new facts, digging through documents, opening up a reluctant source. And maybe you'll do some sort of notebook dump to an AI model, and it'll kind of assemble that into a story or series of stories. Maybe there's a version at 400 words and a version at 800 words and a version at 1200 words. Back in the day, before my time in journalism, there was this thing called the rewrite desk where at the New York Times and other publications, foreign correspondents, people stationed all over the world, would literally call in stories to a group of people in New York or Washington or wherever that would just take the dictation and transcribe it and format it into a story. So you'd call in from the Jerusalem bureau and you'd say, hey, a bomb just went off. Or this thing just happened. Or I see a. I see this thing. And the rewrite desk would sit there and rewrite. So I think one possibility is that AI just kind of takes over that rewrite desk function and frees up the journalists to actually do the thing that is still, I think, highly valuable. Which is like going out into the world, talking to people, learning things, synthesizing them, and bringing people that news in a way that they can access.

The learning things makes a lot of sense, just coaxing information out of people, building up context. The synthesizing thing feels like it's the exact comparative advantage of the AIs, right. So I wonder if the flow will be things happen in the world, human interaction with the world to figure out what's happening. And the AI is the thing that converts human output into the final. The final output.

Yeah, I think that's plausible. And just because people have so many different ways that they like to absorb information So I can imagine a version of a New York Times where, you know, you go in every day and your, your homepage is totally personalized to you. It has only the topics you care about, but then within that, it also has stories in the format that you like to read them. So, you know, you and I could be sitting next to each other, each with our app open, and you're reading an 800 word story because you really like 800 word stories. And I'm watching a short form AI generated video because I'm illiterate and I like short form AI generated videos. And both of those are sort of coming out of the same reporting process. So I can imagine that future too. But I'm curious, what do you think?

Yeah, I don't know if it's sort of interesting that other platforms where there is this recommender system, you would think that there'd be a lot more siloing, but it seems like actually if you look at Twitter or YouTube or something, people are kind of watching the same stuff. Like there's a shared embedding space of like, what humans find interesting.

Oh, I totally disagree. Mr.

Beast goes viral everywhere, right? And then, and then like this idea that everybody's like, got their own niche thing, I don't think has actually happened.

I totally disagree. It's an interesting hypothesis and I kind of wish it were true because I like mass culture and I like watching things with other people. Like, I like watching White Lotus when everyone else is watching White Lotus. But I'll never forget, years ago I was doing a story on PewDiePie, the like, huge YouTube celebrity guy. And I was in the UK interviewing him and we were sitting at a coffee shop and it was like there were two different sets of eyes on him. Like, everyone over 20 had no idea who he was. Everyone under 20 was like, lining up for selfies with him and it was just this, like, bizarre. Not even micro celebrity. He was probably one of the most mass personalities on YouTube at that time. But it was still like, it wasn't like sitting with Tom Hanks where like, everyone would know who he was. It was like a demographic subset of people was obsessed with him and everyone else to everyone else who is a total stranger.

Yeah, yeah. I mean, I also wonder what implications, like the fact that there's so many more micro celebrities who are famous to at least some people, I think has interesting implications. Like if we live in a world where only a thousand people get to be famous, especially in a world where other things matter less.

Right.

Like if AI is doing A whole bunch of things like attention matters a lot more. The fact that it's possible for people, you know, millions of people to be famous, to serve certain subsets, I think is a positive development.

I think that's true, although I'm not sure I aspire to be famous. You know, I know people who I think would be considered famous and I'm not sure that they are happier for that in their lives. I don't know. Famous. I don't think that's true.

But he says while on stage.

But I think that like, you should do a poll of people who have experienced both like micro fame and macro fame. And actually I think the macro fame might be easier on people because when you're Tom Hanks, when you're Scarlett Johansson, when you're the Rock, you like have a whole apparatus around you to help you with that fame. Whereas if you just are like the guy who goes viral on TikTok and then all of a sudden you're famous, I think it can be very disorienting.

I disagree with this take. I mean, as somebody, I think both of us are in the category of like, there's a certain group of people who are like aware of our stuff, et cetera.

Right.

And I think that's like, that's very, very healthy because almost nobody cares at all. So that doesn't really impact your day to day life. And that way, I mean, if you're the Rock or something, I think it's like easier to go crazy because every interaction you have with people is colored by this weird, the fact that they know what you do, whereas we just like almost no interaction involves that except these very particular ones.

I don't know. I've heard you get stopped on the street in San Francisco, so I think you may be a little, a little more than a micro celebrity. But yeah, I take your point. I think there are definitely things about attention and the attention economy that have created all these micro celebrities. And I think that's probably good for those individuals and their career prospects. But yeah, I don't know. I don't know that I would. That I wake up in the morning yearning for either of those things.

Okay, so if I look at the front pages of the New York Times right now, I doubt I'd see something about AI or on any given random day, if what you and I think is plausible with AGI actually comes to pass, there will be some moment where what Covid became after March 2020 will happen for AI. Like it becomes a sort of main thing. People in the world are thinking about, do you have some sense of what would get AI in that position? Because that may be the end state, but what is the thing that causes that? March 2020 moment.

I thought about this a lot. You know, I'm, I'm, you know this, but I'm like, fairly AGI pilled as journalists go. I believe that we are going to get smarter than human AI systems, you know, within the next handful of years. And I keep waiting for that to become as big a deal to other people as it feels like it is to me and probably to you. And I think there's just a lot of sort of factors that have made that not happen yet. I think one of them is just like, you know, I was a reporter during the sort of rise of social media, and what happened with social media wasn't that like, all of a sudden it became front page news? It was actually that it started to creep into all the other sections sort of bit by bit. So, like, you would have a story in the style section about some fashion influencer who had blown up on Instagram, or you would have a story in the real estate section about a private Facebook group where they were trading rental information or something like that. And it just, it sort of like came from the sides rather than down the middle.

Right?

And I think that could be what happens with AI where, like, it won't be like one day something happens and everyone's like, okay, AGI, it's important. It's the biggest story in the world. I think it'll be more like, how is it being used in medicine, how is it being used in restaurants, how is it being used in schools? And just the way that newsrooms, at least my newsroom is organized. There's not like, I mean, there are AI reporters, but there are many more people who just like, they cover AI twice a year when something happens on their beat related to AI. And so I think that's sort of the way it'll show up.

Do you think it's better that there is not going to be some crisis moment where people are forced to pay attention to it and rather it'll just be this thing which becomes a fixture of our lives. I guess we can ask the question, is it good that social media had this, like, slow percolation rather than in 2012, everybody got together and like, what do we think about social media?

Well, social media did have that moment. It just came in 2016, after Donald Trump was elected, and everyone said, oh, shit, like, social media got the president elected. And I think Some parts of that narrative were directionally true and some parts were overblown. But there was that moment with social media, but it came like after the actual arc of influence had peaked, which is somewhat interesting. I think with AI there will be sort of minor catastrophes that happen that will be sort of traceable back to AI. I hope there are not major catastrophes. I think one of my worries about the AI future is like, I think a lot about like the post 911 sort of crisis state and like all of the things like the Patriot act that were passed in the wake of 911 that I don't think reflected like a thoughtful and deliberate process of rulemaking. And my worry, the thing that I'm constantly telling AI executives and people who are like in government thinking about AI is like, do not wait until the crisis to figure out the rules of the road because you will end up with something that is reactive, that is probably worse and like worse for the industry than just like accepting some regulation now. So I think depends who you ask. But my like, I guess best case scenario would be there is no like one major moment where everyone kind of wakes up and starts to panic because I think that just creates a lot of bad policy and bad incentives right.

Now of the pool of AI coverage, what do you, you know, if some fraction of the front page in New York Times is going to get AI compared to what things actually end up about AI in the front page? What, what do you think should be more covered? What, what is undercovered even relative to the fact that AI as a whole is under covered?

I think the usage of AI is undercovered. I think there are so many stories about the companies building this stuff and I write a lot of them and I'm guilty of this as well. There's so much focus on the latest thing that Sam Altman said or the latest model release out of one of the big labs. I am always trying to find and tell stories of people out in the world using this stuff. One of my favorite podcast episodes we've ever done on my show is, was an interview with a 9th grade English teacher. This was just when sort of like a month or 2 After ChatGPT had come out and she just told us like what it was like in a 9th grade English class in the months after ChatGPT. And it was fascinating and it was actually she ended up being more eloquent and thoughtful about the role of AI in the classroom than like anyone I had talked to at the AI companies. So I think there's still a lot of like untold stories there. I think a lot of the social effects of AI are starting to be covered, but probably not enough like the AI companions, the delusional spirals that people can get into the sycophancy problem. I think those are all really interesting and important. And I think the job loss story is actually undercovered relative to its importance. It's hard, it's a hard story to tell because it's not like we have data saying like this many jobs were displaced by AI last month. Like, that's just not a form of data that anyone collects. But I think we are starting to see a lot of indicators that especially entry level college graduates in the workforce are starting to be displaced by AI. And so I would like to see a lot more coverage of that. And if I ran the New York Times, which I thankfully do not, I would assign a strike team to go out and pursue that story.

I would. You know, what I think might be going on here is at least historically you've believed in AGI. It's been such a wild thing to believe and so few people have been thinking about it that your head immediately goes to, oh, we'll have then super intelligence. And if you have super intelligence, we ought to think about what will God be like and how do we make sure God likes us, et cetera. And if you don't buy AI at all, then automation is not a thing to worry about. So this middle group of people who are like, I take AGI seriously and I'm thinking about not God, but just the impact on the economy, impact on the workforce. There just hasn't been anybody. You know, if you buy AGI, you just want to go for the, the biggest possible impact or you know, try, try to solve for that. That may change as the automation becomes more front and center and then there's enough people covering what we should do about God.

I have to ask you about this because this is my biggest beef with the AI safety community. Actually. This is the group of people, for those unfamiliar with our local tribal landscape, of warring AI factions in San Francisco. The AI safety people are the people who worry about the existential risks of AI. They talk about P Doom and things like that, and they are very worried about things like AI creating bioweapons or doing autonomous cyber attacks or deciding to wipe us out one day. And I take those risks somewhat seriously. But like when you ask those people about job loss or like the possibility that like there are going to be no human programmers in a couple of years, they sort of look at you like you're asking about something like, trivial and all. And like, I keep telling them, like, this is a safety issue. Like, it is not existential, but like a society with 10, 15, 20% unemployment is not a safe society, no matter where in the world you look. And so my plea to the AI safety people, with whom I agree on lots of stuff is like, consider some intermediate harms that we might reach on the way to, like the machine. God.

Yeah. I think what will actually happen is that there will be full employment up to a full automation.

Because.

And this is my answer to the question of, well, what. What is under covered is. I agree with you that people talk too much about the drama at the labs. People talk too much about, like, what models released this month. They're less focused on the underlying physical trends which are making this progress in AI possible in compute and hardware and energy. And if you extrapolate those forward a couple years out, you're looking at, you know, the biggest infrastructure build out that's ever been done in our civilization's history. And that will require a lot of physical labor. And, you know, if you don't think robotics is going to be ready by then, I think that might look like, well, you know, maybe we're out of a job as podcasters and journalists, but then we can make even more money just laying down concrete for data centers. So, I mean, from the economy perspective, it might be like full employment, people earning higher.

Wait, this is your optimistic vision, is that we're all laying down concrete at data centers?

From an automation perspective? Yes.

Oh, okay, got it.

For an immediate automation perspective, you remember.

That there was that old joke about, like, the factory of the future will have, like, a human and a dog working at it. And like, the human's there to feed the dog and the dog's there to bite the human if he tries to touch the robots. Like, that's sort of the scenario I envision. But, like, I think you're right. We do need a lot more reporting about the, like, the chips industry, the semiconductors, data centers, maybe give. Give people. Because I've heard you talk about these numbers, and at a certain point I think they start to like, just blend into like a soup of like, really big numbers in people's brains. But, like, what makes. What are the kinds of numbers you're thinking about when you say that this will be like the biggest infrastructure project in human history?

There's a couple ways to look at it. The most simple and the one that's hardest hitting is the fraction of GDP that we'd be spending on AI right now it's less than 1%. If you extrapolate these numbers forward and assuming that you have AI systems which are earning lots of money, which justifies spending more on AI systems, et cetera, et cetera, it would make sense that 5, 10% of GDP is just being spent building more data centers, building more factories for AI. That's comparable. That's bigger than what the Internet boom was when there was a dot com build out in the 90s. And that's comparable to during the peak of the railway mania in the 19th century, how much we were spending on building railways.

And it's much bigger than something like the Apollo space program, right, Which I.

Think was 5% of GDP. And I think it's bigger than the Manhattan Project. Also the economy is way bigger. So just like as an absolute share, and I guess maybe the numbers are more absolute numbers, it would be trillions of dollars a year, 2, 3, 4, 5 trillion dollars a year. Like in the world where AI continues to get better.

You know why I think there isn't more coverage of like the physical infrastructure? This is really boring. Like, I mean this sort of facetiously, but I've tried to do stories about data centers and build outs and like, it is very hard to make people care about that. And I'm wondering like, maybe we need like outlets that are just finding like interesting human dramas in the data centers. Because one of my sort of hypotheses about journalism and media and storytelling is that like people really absorb information better through stories about people. And so maybe we just need to find some like capers and heist stories that are taking place in like Meta's data center.

I hate that. I'm just like whenever I'm reading a book and they're like they're talking about some science thing, I need to understand them. Like they give you a biography of the scientists. I'm like, come on, let's get to it. Who cares?

You might be a unique individual, but yeah, I think this bears out in all, all of my observations about which of my stories tend to do better. It's like always stories that are primarily about people.

You don't think solar panels, as far as the eye can see, would make.

For a great article? No, but I think if you can find the guy or the woman who's the sort of the biggest solar panel broker in the world and they're buying and selling and they're turning themselves into like the, you know, the biggest, like finding those dramas as a way of illustrating the larger trends is often very, very useful. And I'll, like. I'll compare it to something that actually was not my story, but was a colleague of my story during COVID Do you remember this guy who was like, hoarding hand sanitizer? Anyone remember this story? He was like, this guy in, I think I want to say, Kentucky, who, Who had just, like, cornered the market on hand sanitizer somehow. He had just, like, bought pallets and pallets of it from whatever overseas distributor. And that story got more attention, more traffic, more eyeballs than essentially anything else that I saw on Covid. Even the stuff where it was like, here's which states are like, locking down when, yeah, it was like this deeply hateable guy who had, like, was ripping off people, you know, by marking up hand sanitizer at a time when everyone was trying to buy it. Like, people just wanted to know about it and they wanted to, like, you know, have their feelings about it. And so I just, like, I want to believe that, like, you, Dwarkesh, are like the sort of platonic ideal of the news consumer who just wants, like, the readout of the data. But I think most people want, like, the juicy human dramas.

Any interesting lessons from COVID specifically about how media should cover crises better?

Yeah, I wasn't like a frontline Covid reporter, but I think that the media was late to taking the sort of story of COVID seriously in part because, like, we just don't think in terms. There were two problems with the COVID story. One was it took place on, like, an exponential curve, and, like, we're just not very good as, as humans, but also as journalists at thinking about the early stages, especially of an exponential curve and what that means going forward. And even if epidemiologists are telling you, like, oh, the, like, replication rate is this, which means in a month it's going to be this, like, we just don't think that way. And I also think there was like, there's this sort of, like, I don't know, I sometimes call it like, nothing ever happens ism that, like, is sort of a pervasive attitude among people in the media, but also politicians and I think consumers to some extent, where the sort of baseline assumption is that, like, the world tomorrow is going to look pretty similar to the world yesterday. And that, like, it would be weird and probably unpleasant if, like, too much change happened too quickly. And so, like, we are going to kind of marshal all the evidence we can that, like, things are not actually going to change. That, like, the people who are wearing the masks and avoiding handshakes are just being weirdos. And, like, nothing is actually going to lock down because, like, that would be weird and unprecedented and unexpected. And so, like, it would just be easier for us to assume that things are just going to revert to normal. But I don't know. What do you think? What's your. What's your diagnosis of, like, modern journalism and what we should have learned from things like Covid?

I think there was too much focus on stopgap solutions and the drama around lockdowns and politics and with the, you know, what governor is doing, what, what president is doing, X masking versus not masking. Where I think the real story is just like, get to the vaccines fast and which, you know, we did in a reasonable way. But I think there will be something similar for AI probably that maybe this contradicts what you were saying about, like, how do we solve automation and social unrest? And I think there will be. There will be a sense in which, like, there's a bigger project that needs to be achieved, which everything else sort of working hinges on. I imagine it will have something to do with making sure the next system is somewhat safe, reliable, et cetera. And yeah, I think, at least for Covid, it was just that, like, we weren't really paying attention to the very most important thing, which was a vaccine. Yeah. Yeah, I think that's right. Yeah.

What.

When you interview researchers and CEOs and executives of AI companies, what is it that you mentioned? The safety people are not paying attention enough attention to automation and social unrest? What are the people who are actually building systems?

Like?

You're just, like, you haven't thought about that yet. What is that?

I'm curious to compare notes with you on this because I know we interview a lot of similar folks. So I'm writing a book now about the race to create AGI. And so I'm talking with many, if not most of the people who have been instrumental in developing this technology. And it's so interesting to me how, like, these people seem to contain two seemingly contradictory ideas in their heads at once. One of which is the thing I'm working on is a project of world historic importance that will be as relevant to historians of this era as things like the Apollo space program and the Manhattan Project. And also, like, I. And my job and my role in this are not important at all. Like, I will. I interview people almost every week and I ask them, like, are you taking any notes? Like, do you keep a journal? Like, this meeting that you were in like a year ago where like something was decided that like altered the trajectory of AI, like do you remember who said what in that meeting? Or like what anyone was wearing or like what the food at the meeting was? And they're just like, no, like, I don't remember that stuff. My joke that I have is that like, I keep getting these excuses from researchers who say things like, you know, my brain just doesn't index things that way. Like they do not conceive of themselves as characters in a larger story, even though I think intellectually they know that what they're working on is very important. So I think that's a blind spot that many of them have.

Do you know if historically that's always been true? Like were the people working on the Manhattan Project or the Apollo Project, like, did they think similar things about their work?

Yes, because they kept notes and they all had lab. I mean they were physicists, right? And so physicists, I think more so than like AI scientists were like used to taking lab notes. They also, many of them worked for the government so they like had some official record keeping responsibility. But what's happening, AI companies now, it's all happening over slack. Slack in many of these cases is like set to self delete after a period of years unless the lawyers make you retain it. Everything that's not happening on slack is happening on signal. And so there's just like not a culture of record keeping. And so my new crusade that I'm on with these AI labs is to convince all of them to like hire teams of historians and archivists to just.

Like, their lawyers will love that.

Oh my God, they hate it so much. They're like, are you kidding me? We would like willingly document what happens here so that people in the future could sue us to get the records. So yeah, they're not like taking this idea, but I do think this would be important and selfishly it would also be very.

Didn't Nixon record the taste that led to Watergate out of a similar motivation?

No. I mean, I think this is why like government science projects are often so much better documented than private sector projects. Because they're just like when you're spending taxpayer dollars on something, you actually have to do reports to Congress and keep notes of what you're spending money on and stuff like that. So my worry is that like, and the reason I'm writing this book is because I just woke up one morning and was like, oh my God, it's all going to disappear. Like we're going to have no durable records of this incredibly important, incredibly strange period in the technology industry. And, like, someone should keep track of all that.

I was talking to a historian, and I asked them, oh, are you excited about the future where we'll have all these artifacts of. You know, people are, like, sending all these emails to each other, messages. Everything is digitized. And so the work of future historians in understanding our era will just be so much easier. They'll have a much better understanding of why things happen the way they did.

And they.

They thought that the opposite would happen. That historically, you're just writing letters back and forth. Those are preserved. Whereas now people are just, like, sending WhatsApp messages encrypted, that expire in a week. And so the amount of information that is going into making these decisions may be much more limited for future historians, which maybe is a good note to close on, since your work is to try to reverse that trend. So, Kevin, best of luck.

Thank you. Yeah, I'm excited and a little daunted, and I hope that what comes out of it will be a story that people and AIs can look back on for years and try to figure out what the heck was happening in San Francisco in 2025.

Writing for the AIs.

Yes.

Awesome.

Thank you. Thanks, Kevin. Thanks, everyone.