Every's Dan Shipper on the "dirty secret" of writing with AI

He runs a magazine that reviews the future and a lab that is building it. So what is Dan Shipper seeing before the rest of us? That's this week on Platformer. This episode is brought to you by Jira by Atlassian, where teams and coding agents get the context they need to do the right work. Try it free at jira.dev. That's J-I-R-A dot D-E-V. Welcome to Platformer. I'm Casey Newton. This week, our guest is Dan Shipper, co-founder and CEO of Every, a company that is genuinely hard to classify, but is also maybe a sneak peek of the future. Every is a publication about AI that's read across industry that also builds and sells its own software, including an AI email assistant, a file organizer, a dictation app, and more. The writing feeds the products, the products feed the writing, and Dan is sitting in the middle insisting that it is all one unified thing. We're going to get into that very interesting business and also how Dan's unique vantage point helps him see around corners a little before the rest of us. First, though, here to give us some data on the state of AI and jobs, as always, is Platformer fellow and Gen Z AI correspondent Ella Marquianes. Ella, how are you this week? I'm doing pretty well. How are you, Casey? I'm doing well. I was sad to hear that you've recently misplaced your AirPods. It's true. It sounds like you're, like, coming to me from the Marianas Trench right now. And emotionally, I am. But let me ask you this. Where was the last place you remember having your AirPods? At my old apartment that I recently moved out of. Okay. Well, I think we know maybe one place where you could look. But before we send you to reclaim your misplaced AirPods, Ella, what have you learned this week on the subject of AI and jobs? Yeah, so this week I was really excited to see a one-year update from some economists at Stanford that I'm a super fan of, because last year they published this thing called Canaries in the Coal Mine. And this was while we were, like, kind of having all of these debates of, like, is there any real effect with, like, AI and recent grads, like, seeming like they have difficulty getting jobs? And previously, like, nobody really had a big enough sample to say anything about this, and they got, like, a sample of literally millions of people's payroll data, which is kind of rare. And that was able to show initially that young people in AI-exposed jobs have, like, a disproportionate— basically their headcount is just, like, lower than either the group of young people whose jobs aren't AI-exposed, and then with more experienced workers, you don't see as much of a decline from AI. Interesting. So maybe one of the first places that AI-related disruption to jobs is showing up is among junior workers. This is an idea that has been, you know, theorized over the past year or so. But give us the data. What changed over the last year? Yeah. So when these researchers initially reported their data, roughly the group of AI-exposed young workers were 15% less employed than you would expect for workers that were not AI-exposed. Now, a year later, they are 19% less employed than you would expect. Okay. So it hasn't gone up hugely, but it is still, that's a meaningful increase. Yeah. So basically, and, you know, they have like a long graph of how this has developed, and it's basically just steadily this population of workers is less employed. Now, some people might say, Hey, how do we know this isn't like something to do with interest rates being high or like a hangover from COVID overhiring? Yeah. So I think the interesting thing about like this year's update is like now that we've seen the trend like for a year, we can actually like at least have a little more probability that some of this is actually due to AI and not due to other effects. One is like this time the researchers have like run an analysis of how sensitive this population of jobs is to interest rates, and it's actually like less sensitive than average. So you like wouldn't expect it to be because of interest rates. Another thing is the like AI-exposed young worker population actually basically recovered proportionally from COVID around 2022, but since 2022 has actually been doing worse. So it's just like overall a steady trend line, like at about the time you would expect AI to start mattering from, you know, like an enormous sample. Got it. And then so what does it mean now that companies may be less willing to hire entry-level workers? Yeah. Well, first of all, it just sucks. But other than that, so there's this interesting thing where these researchers and some other people are, like, trying to figure out, like, what about young workers makes them, like, less good complements in the age of AI? Like, for example, you might expect the opposite because, like, you know, young people know stuff about technology. Maybe they're more willing to use AI. And yet that's not what's happening. One thing that the Stanford authors have a preliminary theory about is basically tacit versus codified knowledge, where, like, you know, AIs, they are, like, for example, shockingly good at, like, math. Like, they're, like, disproving novel conjectures. They're, like, actually doing stuff on the level of a trained mathematician. And that's partially because we have, like, so much written-down information about how math works, how to do math. And so they're finding, like, when they're looking at the data, that things that involve more tacit knowledge, so stuff like, you know, learning how to work in a team, like soft skills, things that you can kind of, like, only learn on the job and not just, like, academically or, like, from a textbook, tend to do better in the job market. Interesting. But it sounds like the bar is rising for employers to hire entry-level workers. Yeah. And so there's also this survey that came out from ZipRecruiter recently. It was just a thousand employers, so I don't think it's, like, as kind of, like, unquestionable. But, you know, like, 31% of those employers said that they're increasing the experience requirements for entry-level employees. Which, of course, the whole problem of being an entry-level employee is that you don't have experience. Yeah. Yeah, I really love, yeah, I mean, like, you know, when I was like a recent graduate, I was like applying for jobs. Even like a little pre-AI being such a huge deal, it was like you'd look at a job listing and it'd be like, new hires, entry level, please four years of experience. It's like, like a tautologically impossible thing they're asking for. Okay, so what are we supposed to do about this, Ella? Yeah, so obviously the, like, acquiring tacit knowledge by having secretly had a full-time job the entire time you were also full-time in college is not very tractable. There's some other stuff you can do. For example, 74% of recruiters on ZipRecruiter said that AI skills would be a strong advantage for candidates. Okay. And then there are also, like, now the thing is, like, I do kind of wonder about this because I'm like, what are AI skills? And, like, what actually, like, allows you to have an advantage using AI in your work? Where, like, my experience is, like, unfortunately the best power users of AI I know are people who have been in their jobs at least 10 years because they, like, know what their job entails. And so they, like, know all of the steps that need to be done to automate it. I mean, the other thing that I see the most in AI power users, rather than, like, having read every Substack post about, like, the new way you prompt Claude, which, like, changes every year, is just people who are obsessed with AI and so are, like, willing to do stuff with it all the time. Like, I think the biggest issue employers who want their employees to use AI actually have is just that people aren't willing to adopt it or, like, excited about trying new stuff with it. That makes sense. And I do think that, you know, younger workers should think about those things. At the same time, like, if this trend continues, the government is going to have to get involved here, right? Like, I don't think this is going to be something that, you know, people are going to be able to solve at the individual level. Like, there's going to need to be some sort of policy response. Yeah. Okay, Casey, how have you acquired tacit knowledge? I assume you have a lot of it at this point. Yeah, I mean, I think my answer would be, like, working in newsrooms. I mean, like, this was the great thing about coming up in newspapers was that you would just sit in a group of other writers for eight or 10 hours a day, and I just— I just did that for many years, and you would, you know, in those days, we didn't even have phone rooms. You would conduct all of your interviews just like on the phone that was at your desk. And so I got to hear so many different interviewing styles over the years. And so, I mean, it's things like that that sort of gave me a sense of what it meant to be a reporter and let me sort of figure out my own version of that. And it was only possible because, you know, some very nice people took a chance on me when I was in my late teens and early twenties. Yeah. Hmm. I wonder, yeah, I wonder how we're supposed to acquire tacit knowledge while unemployed from our computers. Maybe, like— Have you considered asking ChatGPT? Sorry, bad joke. But it was right there. Well, we're not going to solve this one today, but we are going to keep our eye on it. And in the meantime, we're going to talk to somebody who is working on his own suite of productivity tools and, in his own way, challenging some of the big frontier labs. So when we come back, my conversation with Dan Shipper. Jira by Atlassian is where agent speed meets team intelligence. Assign any work item to your favorite coding agent and see exactly what it's doing without leaving your flow. When something stalls, you're not digging through logs. You can see what's running, unblock what's stuck, and stay in control at scale. Don't let your agents start cold. The teamwork graph feeds them context from across your entire stack, delivering 44% more accurate results with 48% less token usage. So the work they pick up is the right work done right the first time. Try it free at jira.dev. That's J-I-R-A dot D-E-V. Welcome back to Platformer. My guest today is Dan Shipper. He's the co-founder and CEO of Every, which may be one of the weirdest businesses in media, and I do mean that as a compliment. Every publishes a daily newsletter about AI, including the reviews of new models that the industry actually reads. They call them vibe checks. Plus Dan's own column, which is called Chain of Thought, and a podcast, AI and I, where he's interviewed everyone from Microsoft CTO Kevin Scott to Kevin Kelly to Granola's Chris Pedregal, who is now a friend of this show as well. And then inside the same company, which has about 30 people, Everyone runs a product studio. They make Cora, an email assistant. They make Sparkle, which cleans your desktop. Spiral, which is a writing assistant, and Monologue, a dictation app. It's all bundled with the journalism into one $20 a month subscription. Every describes its strategy as a flywheel with four steps: live in the future, write what you see, build what's missing, teach what works. They say AI writes essentially all of their code; the engineers manage agents instead. But the essays are still mostly written by humans, for now. What I want to understand today is whether this weird hybrid, half product lab, half media company, is a preview of what more companies will become in the AI era. So with that, here's my conversation with Dan Shipper. Dan Shipper, welcome to Platformer. Thank you. Delighted to be here. You run an unusual company, so I want to start by hearing about an ordinary day in the life for you. You have this company where you have AI writing the code and humans writing the essays. So I just want to hear about, like, your actual morning yesterday. What did you do and what did the AI do? That's actually a really good formulation. I gotta write that down: AIs write the code, humans write the essays. I love that. I'm actually looking at my calendar, so let's see. So actually, so Mondays are a little bit more unusual for me because I have more meetings on Mondays, but— And because we're recording on Tuesday, but I would say the average day, and I'll go through Monday, but like the average day, I try not to do meetings before noon. I spend, from when I wake up to around noon, usually writing. It's like my creative time. And then all day after that is like different meetings and talking to people and all that kind of stuff, and that's when I do, like, operating the company. And then for sure I am, you know, I've found that if I am coordinating agents, like actively coordinating agents before noon, it is very often like something that I get sucked into, and it's hard to write if I'm doing that at the same time. But what I will sometimes do is I will send off an agent on like a big chunky, meaty thing, and then I will go off and do my writing. So one of the things I've been working on that is like the big chunky, meaty thing is we have a conference we just launched, which I told you a little bit about, called Thesis. We got a lot of people apply to the conference, and I've been using Fable to help me visualize who are the different people at the conference, and like that's like a big new interesting, you know, realm of study for me. And so that's something I would send Fable off on. It does a bunch of stuff, and then I come back. I also just had it vibe code for me my own personal to-do app that's like on my iPhone. It's agent native, so like I can use it inside Codex and that kind of stuff. Right. Yeah. I have also vibe coded a to-do app. Very fun weekend activity for a certain kind of nerd. So you gave us an example of something you have an agent do that is maybe kind of like a one-off project. I'm curious if you have, like, surrendered part of your actual job to an agent at this point, or whether it's sort of more these sort of like, you know, project by project, go off and do this. Um, I think it depends on what you mean by surrendering part of your job. Like, I think the way that I've been talking about how we use agents, and this goes will use and use now, and this goes back to like 2022, 2023, 2024, is thinking about using AI, using agents. Using AI, using agents as like hiring an employee and like managing that employee. And so it is often actually very difficult for first-time managers to give up their job. There's a choice that you have to make when you have a new hire, which is, do I micromanage and get the result I want, or do I delegate and get leverage? And that's actually really hard. That's really hard for early managers. Same thing is true in AI. But I think there's a bit more agita about the AI version because it feels like you're not handing it off to another human. You're handing it off to a thing, and if a thing can do what you do, that's very scary. And I think what I have felt is if you take the plunge and you sort of play with the thing and you're like, oh, like, what if I can get it to do the thing that I'm doing all the time? There's always a new thing above the task that you were previously doing that is like hard and complicated, and that itself is its own job. You know, managers actually tend to be very busy people despite having people doing the work for them. It changes the work, and I think that there's a lot of legitimate things to say about when and why you might want to change the work. If you have a particular relationship, for example, to writing, which I do, there are certain times where I don't want AI involved. Like, I want to just handwrite it because the way that my brain connects to my fingers is a specific thing that I don't want to lose. But also, I mean, I wrote— Like, I don't know, probably 500 words this morning before all the craziness of my day started, and I used AI for a lot of that. It didn't do it for me, but I couldn't have written those lines without the support of this partner that I have. Right. So, Every grew out of Super Organizers, your interview series about how productive people organize their lives. You've spent years asking brilliant people about their systems. Now you're building software that, in part, can replace some of those systems. So what has all your interviewing taught you that has ended up in your products? That's a really good question. One of the things I think is really interesting is, yeah, so I had this thing called Super Organizers, a newsletter called Super Organizers. I was always interested in tools for thought, so it's like Roam or Notion or, like, how do you organize notes? You know, there's all these, like, memes of nerds where it's like on the one side of the midwit curve is like, I just use Apple Notes, and on the other side of the midwit curve is, I just use Apple Notes, and in the middle it's like, I have this very complex system that I'm, like, you know, building. And that was, like, me, and that was, like, what I built with Super Organizers. And now I, honestly, I do use Apple Notes, but, so I made it all the way to the other side of the midwit curve. Congratulations. My wife, thank you. But what is really interesting about AI is it's like all of the things that I most desperately wanted and wanted to build just happened with AI. It's like the biggest gift to nerds like me, and there's a lot of us. I think I started off—what every productivity nerd goes through is you start off thinking that it's about the tools, and then you end up realizing that it's about emotions. And people who are really into productivity have— Often a feeling about themselves that they're really talented, but stuff slips through the cracks. And if they could just, like, organize it well enough, they would be able to make sure that, you know, they didn't drop the ball on things. Or another thing is, like, sort of wanting to maximally make the right decision in any given circumstance. So if you just wrote everything down and you had all the right information at the time, you wouldn't make any mistakes. Like, those are the kinds of—I don't think anybody is realistically thinking that, like, explicitly, but those are the kind of, like, inner things that are going on that what can often make it really apparent is you get a god tool, like AI, that just, like, solves all those problems for you, and you're like, I'm still making lots of mistakes and dropping the ball, and I still have all these, like, fears. So there's something is happening with me. Yeah. Yeah. Well, you've also been open about being diagnosed with OCD in your late 20s and how being treated for that helped to give you your clarity back. I'm curious about the relationship between that and building productivity tools, and, like, how did maybe getting to, I don't know, some kind of, like, underlying—or yeah, how did that shape the tools you're building? You've definitely done your research. I appreciate it. I love it. Yeah. How did it shape the tools I'm building? Well, I definitely don't think—to some degree, a company goes as its founder goes, especially early on. And I definitely don't think had I not been lucky enough to go through enough therapists and treatment options to, like, find the right treatment, which it's actually really hard with OCD. I don't know how much experience you have with OCD, but it's typically not— Okay. It's typically not diagnosed for, like, seven years because it looks like anxiety. And the treatment for it is the opposite of the right treatment for anxiety, often. And people say they treat it, but they don't. Like, you really have to go to a specialist, and, yeah, like, I can just go down the list. It's, like, it's a pretty horrible condition to have. But for me, in case you have anything that might be OCD, exposure and response prevention and Zoloft, incredible combination. Highly recommend. Highly recommend. So, A, like, I would not be able to run this company without that. Companies go as their founders go. I think a big way that the company became what it is—we're about 30 people now—but a big way that the company became what it is is, like, two or three years in, it, like, almost fell apart. My co-founder left, and I had to really think to myself, what do I want to do with this company, and, like, how do I—what do I want to make? And I really said to myself, well, I really want to write. I really, like, want to be a writer. But also, I really like businesses and I like making products. Is that possible to do? Because I think there's a lot of—those worlds are usually separate, and there's a lot of Silicon Valley discourse that's like, you can't really write and run a good company, or at least there used to be. Now I think it's a little bit more accepted. But the way that I grew up, let's say, it was very not accepted. And so at the time, I just asked this little thing called ChatGPT, and it was like, yeah, there are people that do this, you know, like Sam Harris or Bill Simmons, or there's lots of examples. They're just not necessarily in the main Silicon Valley discourse. And what that required me to do—and I wrote a post about this—but what it required me to do is, like, be able to admit to myself, this is what I want, and then run the business in that way, even if it meant it was, like, kind of uncomfortable and felt unconventional. And I made that initial decision. Without the aid of Zoloft, but with the aid of a lot of therapy, I think that I was able to fully inhabit that and really actually stick to this is what I, how I want to spend my time and what I'm good at. And even if it's like a little bit weird to say, like, I'm going to spend half my time writing, I was able to, like, really bring that, my full shape of myself to the company and the job and stick to it. And I think that has been, like, a massive, massive part of our growth and success. There's lots of—there's many other things and many other people involved, but that's one big thing. I mean, I really appreciate the point because I'm somebody who is still in the middle of the midwit curve designing ever more complicated systems for myself, mostly just because I enjoy it at this point. That we do focus mostly on software tools when we talk about productivity, and I think it's important to say, as you just did, and I will say for myself that therapy is an incredible productivity tool. And if you're not—if you haven't gotten involved in your life yet, it might be a good time to start. I'm getting a lot more done thanks to therapy. I agree. And, you know, that's one of the things that therapists will be like, well, if you go to therapy to just get more done, there's an interesting therapeutic question there. True. But I agree with you. I'm right there with you. I will say, though, like one other thing, just on this sort of Zoloft point that I think is really interesting that I think dovetails with AI is when I was getting on Zoloft, I love to read. I'm a huge just reader, book nerd. And I read this book that's not very popular now, but was super popular in the '90s called Listening to Prozac. Fantastic book about—it's a psychiatrist. He's starting to use... SSRIs, antidepressants, of which Zoloft is one, on patients. And he's starting to realize not only is it beginning to lower their depression and OCD symptoms, but it is also, like, changing their personality. And it's like a deep New Yorker-y, over many, many pages, like, investigation of what that is and what it means. And what the book ultimately is about is how does technology change how you see yourself? And, for example, how we split up the world into categories. So a really interesting example is he talks about this idea of what he calls pharmacological dissection. So to tell the difference between something like bipolar and schizophrenia is actually quite difficult, especially there are lots of borderline cases. But one of the ways that they tell is lithium is helpful for schizophrenia, but not for bipolar. And so the fact that we use this tool and it affects schizophrenic people but not bipolar people in a positive way helps us draw a category to be like, this is schizophrenia. And I think there's a lot, a lot, a lot of resonances with AI and how it changes how we see ourselves and who we are and what we do and all that kind of stuff. And that, I think, is a big thread for me. Like, having seen a lot of that and being interested in it because of all my mental health issues and the fact that I take Zoloft and read this book has been a big way for me to see, oh, this is why I love this stuff and what it does and all that kind of stuff. That's super interesting. Well, so you talked about a little bit of the origin story of how you came to have a business that is both editorial and a product lab. You have a master plan on your website that talks about this flywheel. You say that, you know, every helps you live in the future. You guys write what you see, you build what's missing, and you teach what works. I want to hear about that flywheel in action. Like, is there something that started as an essay and ended up as a product, or started as a product problem and turned into the story? Like, how does that work? I mean, basically everything. I think of ourselves as—and I actually just made this a little bit more concrete for our whole team, because we're 30 people now, so a lot of people have not been around. And one of the things I have to grow on as a CEO is, if you've only been here for six months or three months, you're probably not going to know all the things I know, but I probably assume that you do. I think people coming in actually have more context on Every than the average company because they start as readers, but still. Anyway, we have this—it's, yeah, it's a flywheel. It's a pipeline. There's this pipeline from Dan or someone else inside the company—we have a frontier team—so Dan or someone else on the frontier team comes up with some weird idea, or it can be someone else in the company too. I think of the whole company as being a little bit like a lab. We come up with a little idea. It's usually something that is—it's like half-baked. It's very weird. It's not something that the average person would use. We spend some time—usually, like one person gets excited about it—and then we see if it spreads inside of the org, and I think of Every itself as being a little bit of a... Good bellwether for, because we have different kinds of people in the org. They're all AI-pilled, but they're different levels and types of AI-pilled, so it gives us a good sense of who might like this. If it spreads, we send it to the audience. If that works, we, like, turn it into a product, all that kind of stuff. And there's different form factors for the different stages. So it starts off as just like, it's a tweet. It's like some little idea that I'm, like, talking about excitedly. We often turn it into an article. We often turn that into a guide. We turn that into—and then it becomes a product, like that kind of thing, you know? So our job as a company is to find those things and then, like, help them along as they go from frontier to regular knowledge workers. Right, and you pointed out an interesting way how it doesn't just become one thing. It can become a set of things that you guys can make various products or sell in different ways. You mentioned vibe checking. Your vibe checks are the early previews of new models that you guys do, which I really have begun to read religiously. I used to work for a gadget review site, shout out to The Verge, where we would get access to products early. And I found that there is sometimes tension between what the companies want from those early reviews and what you give them as critics. So you published what I would say was a fairly critical review of Sonnet 5, which you called, A model pitched for everyone impresses no one. How did that affect your relationship with Anthropic, and how much did you think about that before you hit publish? I've got to say, like, obviously there's a—I know a lot of people, we know a lot of people at OpenAI and Anthropic, and obviously you never want to be just, like, totally mean to people who you're friends with, you know? So there's a little bit of that. But I actually think— They, even before we publish anything, they're like, What do you think? Because they want to make the model better. And they know that if we don't like it, it means something. And they'd rather know beforehand, honestly, than, like, find out from a ton of other people who use it and just give them a ton of bad feedback. So I actually find that they welcome and want our actual real feedback because they want to make the models better. They obviously probably would also prefer that we didn't, like, publish a big thing being like, This model sucks. But that's sort of just like, they know that we're not trying to be mean. We just have to say what we think, and if we think that, it's probably—it's pretty likely a lot of other people are going to feel that way. And my goal is always, it's not to shit on them. It is to help make better AI happen. And I think we do that in partnership with them. Sometimes it can get, like, a little bit heated every once in a while. Like, they're like, I don't see how you feel— They're like, I don't see how you feel that way or whatever. But that's, I would say that's mostly the exception to the rule. Do you find that some labs are more sensitive to criticism than others? I don't think that it's a lab thing. I think it's actually, who are the people? Like, who are the comms people? Who are the, you know, creator program people? Like, all that kind of stuff is going to—and what is their disposition towards conflict? What is their desire for knowing the truth versus, like, just—it is sort of a comms person's job sometimes to just, like, yell at you. You know, like, it is sort of their job, even if they know that you're right. And, like, everyone knows that they're right, that we're right or whoever's right, but, like, they just have to do it. And you're just like, I know you're doing your job, you know? So it sort of depends on the person. I think it's less organizational. I've found that both the big labs, which we work with very closely, are in general, like, extremely open to feedback and are not super sensitive. Yeah, interesting. Well, another interesting aspect of your dynamic, I would say, is the way that the labs are sort of both enabling you to do these very cool vibe checks. You're using their models to build products, and also it seems like inevitably they're just encroaching on your terrain and everyone else's. Every ran a piece in June called Built on Moving Ground, which sort of got at some of the vertigo of building products on these models that you don't control. The capabilities are changing continuously, and there's just always this risk that the labs will release something that you spent the last year on as a feature. So how do you guys think about that risk, and where do you hope your customers will find the durable value in a bundle like the one that you make? That's a really good question, and I will say I don't have an answer to it. And there is just this dynamic of— They're gonna make their models better, and their models getting better does just actually take a lot of the stuff that you build and make it less relevant. And so they also have application layers, so there is these different parts of the same company that are, like, supporting you and then also are sort of competing with you, and, like, it's just a little bit of a messy dynamic. I think that I have a couple feelings about this. One is the thing that we can do that no model company can do is tell you which models are good. No one trusts a model company to tell you where they objectively sit. How could they? So we, I think, have a good position to be an arbiter between them, and that's not something that model progress will get rid of. I also think that we, as a company, have gotten really good at staking out, like, we actually have a vision for what really great human work looks like after AGI or after automation. And that is something that obviously I think the labs would like, but it takes a very specific company and culture and just vibe to do well. And I think that they actually look to us to help them figure it out, because just because you make the model doesn't mean you know exactly how to use it or how to use it well. I liken them a little bit sometimes to oven makers. Like, you can make the oven, but it doesn't mean you know how to make a soufflé. And our job is to take an oven and be like, what is the coolest thing that we could make that would be good, you know? And they're like, cool, great, we'll make the oven better for that. But then sometimes they're like, you know, we'll make a—well, maybe we'll make a soufflé too. And so that's the other part of it, right? I think we can adjudicate between them and be the sort of standard bearer for what happens with this stuff. But I think another important part of this is that we do just live in this zone where things are moving really fast, and that means that we can't just, like, rest in any one particular place, which was always true. Like, businesses have never really been able to rest, but it's, I think, more true in AI, and it means that we have to use the access that we have and our, I think our even more important than that, our special sort of taste for what does this model enable to allow us to just throw things out every three to six months and go find something new. And by the way, that's something that happens in the model companies too. It's just a general feeling in AI that you have to both really want to make something awesome and high quality, and you have to be willing to throw it out every three to six months as the capabilities change and what is possible changes. And that's a hard thing to do, but I think it's possible. It also just strikes me that a benefit of the fact that you guys are a bundle is that you can just sort of, like, swap in new stuff all the time, right? Like, you haven't made one single giant bet. Like, you have many, many different baskets with many different eggs. Yes, and I think we are also, the company itself is an ecosystem. It's thick. There's a lot of different reasons why you might be into Every. Some of them are the tools, and some of them are the vibes, some of them are the writing, some of them are the YouTube. And so that whole bundle of things creates something that people just want to be a part of. And so any one tool, yeah, maybe it's, like, less useful now because a new model came out, but that's okay, because you're right. Like, we have all these different things that you can do and all this different value we bring you. Yeah. So I want to talk more about how you guys use AI at work and how it changes things for you as you're sort of on the leading edge of testing the latest models. In, like, general terms, where does AI make every measurably more productive, and where has it still not changed very much? Great question. I mean, we just actually would never have been able to do almost all of the things that we do without it. So we—and now we're bigger, but, you know, for a while we were maybe like 12 or 15 people, and we were running like six software products and a daily newsletter. That's, like, that's insane. Even running a daily newsletter with—I mean, you know, it's hard. Yeah, yeah. You need a lot of people to do that well. You do. And not just do it, but actually have a newsletter that grows and people like and read all the time. It's really hard. And then to add software products on top of that, and without really very much funding, like we haven't raised very much money. And it only really became possible because we started to be able to get enough from a single engineer that you can have one person run an entire software product end to end, and that was certainly not possible before AI at any, like, real level of scale. And now that that is possible, it turns out, like, once you have one person, you start to hire more people, and so we have products that have more than one person on them. But you can get Signal and really serve an actual customer base with a real product with one person, and that has meant we can expand and do all these things in a way that we totally could never have done before. I think a lot of—in a lot of ways, you can think of every—there's a lot of overlaps between The New York Times and every—but The New York Times was only able to do the games bundle and cooking and athletic and all that kind of stuff after 150 years and a lot of scale. And we can start to do that much, much earlier and more quickly with less money. I like that answer because when I ask it of other CEOs, I typically get something really vague about, well, you know, the engineers are shipping more code, or, like, they're less focused on the drudgery and they're doing more of the work that's meaningful to them. But you're just like, we have individuals that are, like, running entire products. Like, that sounds like a much more concrete answer. On the flip side, though, I'm curious if there is something that you keep throwing models at that either they're just terrible at or that you feel like it's just a stubbornly human job that needs to be done. Yes, all the time. I mean, I have a lot. There's a lot to say. Let me start simple. One thing that we have been throwing models at for a long time that only just started to work is we have an editor-in-chief, Kate Lee, who's fantastic, who you may know, who I've been trying to automate out of a job for years in an extremely benevolent way. Basically, she just does a ton of copy editing for us. She has, like, the best copy editing taste of anyone at the company. And as the company has grown, she's no longer just copy editing the articles that go out. She's doing, like, launch emails and landing pages and, like, all that kind of stuff. And she's in charge of making sure they all adhere to a standard. But her time is limited. She's an editor-in-chief. Like, she has many other responsibilities other than doing that. And I've been, like, since GPT-3, been like, I think we can make this better. And the answer has been, no, you can't, for a really long time. And it just started to work. And part of that is the models are good enough at instruction following that you can make a good enough prompt that it actually knows what to do in any given situation and it's not totally dumb. Another thing is they're good enough at browser use that, or computer use, that they can actually go into Google Doc and make suggested changes, which is wild and crazy when you see it. Another thing is they're good enough now that I collected a dataset of, you know, 30,000 of her historical edits and then used that to make a prompt. And then I could backtest the prompt on all of the previous documents to just, like, hill climb and make it better and better and better. And then now we have a skill. We have an agent internally, the every agent that we use, that anytime someone has a piece they're working on or a landing page or whatever, they just @ the every agent and they're like, do a copy edit on it, and it does it. And it's not perfect, but it's much better than having heard you everything, and it gets better automatically over time. So as it makes edits and then she goes in and says, you know, makes more edits, it automatically learns, like, here are the things I missed, and gets better over time. So that's one thing that just became—we call it compounding—so it just became compoundable. But, you know, one of the things we're now working on is, is copy editing like a special, more mechanical, more rules-based thing? And what's the layer above that where it's maybe there are some rules or principles, but it's not quite as rule-based? So, you know, and I'll say, like, even copy editing, which is rules-based, is super, super complicated and not fully automatable even now. Totally. Yeah. And, like, taste is a component to copy editing. Like, there are definitely rules about where to place commas and em dashes, but sometimes I place them in different places because that's how I like it, right? And that's just, like, part of writing to me. It totally is. And so what I would say about this is I see what we do less as like, oh, we're going to automate all copy editors, and actually more as— Kate has a specific set of skills as an expert inside of Every that she can only apply by spending her time right now. And what we do with compounding is we allow her to get some of that taste and some of that viewpoint and set of skills into a little tool that allows her to spread that throughout more of the org where she doesn't have to spend her time to do more work. And that is a really valuable thing when you start seeing it that way. You're like, of course I want that. But it's only if you sort of see it in that way instead of, oh, I'm automating candidate of a job, which I know I said at the beginning in a sort of tongue-in-cheek way, that it begins to feel like, oh, this is actually exactly what I would probably want. Like, of course I want a tool I can work with to teach it my taste so that in situations where I would be needed to spend my time, I don't have to spend as much time so I can do higher-level, more interesting things. I will say I'm trying to compound myself right now. Can we take, when I'm in a meeting or when someone asks me for feedback on a launch document or whatever, can we get it to like 80% of what I would say? The answer is it's really hard. It's actually really hard. I think we'll make some progress here, but stuff like that is, in any place that's not really well defined, it gets harder and harder. Right. So speaking of automating, my impression is that you guys aren't automating anyone out of a job. In fact, I think that you went from about 15 people in the middle of last year to around 30 this spring. So you've doubled while automating. I think you've called this the AI paradox, where sort of the more things that you automate, the more humans that you need to do more things. And I think you've sort of given us part of the answer, which is like you can now have one person do a single product, and you're a company that wants to make a lot of products, so it makes sense that you would sort of bring on more people. But did you expect to double in size, and do you think you'll double again? My big thing or my big observation over the last year is, as a company, we try to automate everything we possibly can. Why did we double in size in terms of human employees? Not that, you know, I don't necessarily—my ideal world is not one where I only hire agents or whatever and we have no humans. I'm not, like, weird like that. But also, we don't have a ton of funding. You would expect a company like ours to, you know, try to be efficient, let's say, with what we do and not hire people unless we have to, and we've had to hire people. And part of that is, yeah, we're, like, growing, so we can and we should, and so there's more work to do. But I think there are also deeper structural reasons why automation weirdly creates more work for humans, especially for human experts in a lot of domains, and we feel this a lot. So the way that I talk about it, and I wrote this big essay on it called After Automation, the way that I talk about it is very much from our own—like, I really try to start with what do we observe, and then what does that mean for the rest of the economy as much as we can. And there are many, you know, caveats and flaws with that, but I do think that one of the things we've been really successful at is if you have a company that operates in this way, in this very ground-up, we just have our hands on the tools and we just figure out what we can do with them without any preconceived notions of how you should run a company or anything like that, it does tend to tell us a lot about what will happen in the broader economy as people begin to adopt them. And what we found is, yeah, we end up hiring more people, having more work to do. You know, you would expect that we would have no engineers or very few engineers or whatever, and we have a lot. We've hired a big engineering team. And I think the—I was trying to unpack the reasons for that, and I think the reason is something like— The way that AI works is it is trained on the residue of human expertise. It's like it's trained on problems that have already... Like, it's trained on problems that have already been solved. And one of the beauties of AI is that now you have this corpus of, or this thing that knows about how to solve every problem that's ever been solved, and now you're trying to apply it to your problem. The interesting thing is that your problem is slightly different than any other problem that's ever been solved. It's slightly different. And what that does is it creates this situation where tons of people are just, like, mashing on their keyboard, being like, Solve my problem, and it solves it, but, like, only sort of. It's, like, close, but not quite there. And that creates a ton of slop, and that's not really valuable, right? You have this glut of things that look impressive on first blush, but then are eventually, like, you realize they're kind of worthless and the market reprices. So then, well, what do you do? Now you need an expert to come in and solve the problem for, like, this particular situation, like really think it through and use AI to do that. But you need someone to, like, actually go and make it really good. And who does that work? Oh, it's an expert who knows how to do it really well. And you can see that the way that experts now are involved in the economy, or at least internally at Every, is, one, they are, because everybody can do something that's sort of like what they do now, like everyone is a programmer to some extent, experts are involved, expert programmers, for example, are involved in building systems to help take the people who want to program and contribute and make that actually productive. Right. And did that sort of unfold in the way that you expected, or did it take you by surprise? I don't know that I had, like, any strong—I had a strong intuition that it would raise the ambition level of what individuals would be able to do, especially generalists. And I just, I got there sort of by looking at my own use of, for example, GPT-3, and then thinking about what's an example of people who already have a power like this that you could buy. And I was like, oh, it's just hiring people. Like, any writer that has a team, like, can do a lot more than I can as a single writer. And so I definitely saw that. I don't know that I had a particular vision or thing for, other than I think we can stay smaller for longer. But it was definitely surprising to be like, there's way more engineering work, even though we've automated most—like, if I take a step back and I'm like, you told me in 2020 that you could just send a fable off and vibe code an entire to-do app in a day that normally would take many, many years and lots of engineers to do at this level of quality, what would happen to engineering? I'd be like, I don't know, that sounds nuts. And the reality is we still have engineering. It's just, it has moved up a level. So I do think that was surprising. Let me ask about writing and AI. You guys are sort of leaning very hard into having AI do as much as possible, but you're retaining some level of authorship, and I want to get it exactly when it comes to the AI. So how do you think about, like, how much of the writing, like let's say of your vibe checks, should be a person typing words on a keyboard, and how much of it they can outsource? And has that changed? I imagine that's maybe changed over the past year. I will tell you, but first I want to ask, like, have you ever used an editor? What kind of editor? Like a human? Developmental editor, copy editor, any kind of editor. Anyone that's going into your Google Doc and making suggested changes. Yes. Yes. Clicked accept on any of those changes. So that is a similar, I think that's a similar dynamic to, especially for public writing that has your name on it, to appropriate use of AI. Yeah. You want someone that understands what you think, and you know, you have to know what you think, which sometimes you can get to with an AI, and is helping you create the best version of that, but it's yours. And whether or not you type the words does not really matter to me, but they have to be yours. Well, how are they yours if you didn't type them? Well, how are they yours if you just pressed accept on a change that your editor made? I mean, it's a fair question, and I think that there's probably, like, many—well, there are different versions of it, right? Because, like, if I'm working with an editor and, you know, they say, you know, you've used this word, but isn't this phrase more exact? I would say yes. And if somebody came to me later and said, Well, are those your words? I would say, Well, like, my editor helped me with that, but, like, no one would get mad because it's, you know, fairly common for writers to work with editors. People do have really strong feelings about AI, and while I don't think most people would get mad about AI suggesting a different phrase, I could imagine if they suggested, like, an entire chunk of your vibe check and sort of, you know, wrote the first version of it, and, like, you guys published it as is, maybe people would have feelings about that. So how much, like, theorizing, if you had to do around, like, where the lines are, and, like, are they drawn in pen or in pencil? I think a lot. A lot of theorizing, a lot of trying different things. I think they're definitely in pencil because things are changing. I also really want to separate out what are— People's reactions from what do I think is the long-term norm here? Yeah, yeah. Because I think that there's a big difference there, and there's a big difference between, for example, our audience and a mainstream audience, which would be, I think, much more sensitive to this. But I am also thinking about what do I think is the right long-term norm from people who are just used to this technology and it just feels like a part of everyday life, as opposed to this new big threatening thing. Because I think there's a long history of this. A simple example is when I started originally writing about writing and AI, like a couple years ago, I did some research into the history of the typewriter, for example. And Mark Twain was, like, huge on typewriters. He was, like, the first American author who really loved the typewriter. And he also, I think, spent a bunch of money and got bankrupted himself trying to make typewriters a thing. So maybe not as good of a businessman as he was a writer. But at that time, that was a big deal for him to be into it because people got offended if he sent them a typewritten letter because it wasn't your handwriting, so it looked like an advertisement, it looked impersonal. And so this is a very common thing in the history of technology. Same thing for if I send my mom a text message, like, it doesn't feel as personal as a call. But a call didn't feel as personal to her mother as talking in person, right? So I'm trying to think about where is the norm going to go, and that's why I think about cases like, well, if you have an editor who replaces a sentence and rewrites it and you're like, that's better, is that still yours? And my feeling about it is generally yes, it is still yours. And so when we think about AI, And I do think this changes a lot. There's a whole range of different circumstances you have to consider, and we do different things in different circumstances. When we think about external writing, different from internal, external writing that has your name on it and has, like, is from your perspective. So that might be different, for example, than if we write, like, a long-form guide that is intended to be mostly informational rather than narrative-driven. We often include the AI as a co-writer, and I'm much more fine with having chunks of it be AI-written because I think a lot of it's going to be AI-read. It's informational. The point is put it in your agent and help it, like, help it help you at the time when you need it. And so the norm is much less sensitive to whether this is written by AI or not, as long as you stand behind it. For stuff that is from you or from, like, from a person and feels like it's yours, that's different. You know, if I see an X not Y construction, even if you wrote that, I'm like, get rid of that. Like, I want you to think about this, right? Like, what is good writing at its core? And George Saunders says this, which I love, is it is just, like, applying your taste, like, on every word over and over and over again until it is the most pure expression of what you thought over and over as it can be. And the only way you can do that is write a sentence, take a look at it, be like, do I like this or not? Change it, keep going, change it, keep going. And I think you can use AI to help you with that. I do that all the time, but it requires a lot of your time. Like, writing, this is a cliche to say, but it's about thinking. And I often don't know what I think until I write something. I'm writing this long-form... piece about OpenAI and how Codex happened, because I think it's like one of the most interesting stories in business history. And I had an angle for it that I went in with, but then I did a ton of interviews with people, and I still had the same angle. And now I'm like 4,000 words into the piece, and I'm like, I'm starting to see some stuff that I hadn't seen before, and see some stuff about myself and, like, how I think about this story that I hadn't seen before. And you don't get that unless you're really thinking about it and really going line by line and word by word and being like, when they said that in the room, what did that make me feel, and why? And what does that say about them? You know? And 100%, you can use AI to help you with that. Like, I'm sitting there this morning with voice mode, and I'm, like, doing a scene with me, and I'm thinking about a conversation I had with Tibo, who's the head of Codex and head of ChatGPT. And then I'm thinking about, he said this thing. What was he probably doing when he said that to me? And then I had voice mode go, like, go look up a bunch of Tibo interviews and find a similar kind of physical dynamic. And then I just, like, watched that, and I was like, how would Tolstoy write that? And what's, like, the inner versus outer here? And that is crazy that I can do that. I would never have been able to spend the time, ordinarily, to do that unless I was, like, living on my Russian, you know, manor, whatever, and I didn't have any work to do. But I can do that, but I am also thinking about it. It's like I'm in the driver's seat, and I think that's an important distinction. Yeah. Well, maybe to bring this into a landing here, and building off of that, you guys are doing this thesis conference several weeks from now where you will be sort of trying to think through what does great work mean in a world where work gets increasingly automated. And leading up to this, you have a show, AI and I, where you have interviewed lots of creatives about their creative process. And I wonder. While this is a very fraught topic in the creative community, for those who are curious about using AI as part of their creative process, what has separated the writers who get better with AI from the ones who sort of lose themselves in it? So here's the thing. I think there is a real dirty secret right now that almost every writer is using it. Just most of them are not saying. And I almost want to have a little Writers Anonymous support group to just have people come and confess that they use AI, because they're all—some more than others, and some are truly still, like, with pen and paper. But, like, George R. R. Martin still writes in DOS. That's just a common thing. Writers have very particular preferences for how they do their thing. Yeah, but he hasn't finished a novel in, like, 20 years, so I'm not sure we want to be holding him up as a productivity model. He's definitely not a productivity model. But I think that, I mean, the writers that do it well, it's the same thing for using AI well in general. It's this—it's like putty. It's like you can do anything with it, and your goal is to find something that you're excited about and then play around with it and take risks and just be like, What if I did this? How would it work? And could it help me? and whatever. And to, as much as you can, I think people feel weird about, Well, what if it doesn't work? And what if I waste my time? Like, just allow yourself, if you can, to get into it and to take the risk. And allow it to change what it means to write for you a little bit, and know that you can go back, but there's a whole new world of things that are possible that might be scary, but once you get into it, it's, like, really awesome, that it changes you, it changes the work that you do, and I think it's for the better. All right, well, we will leave it there. The company is Every. Dan, thanks for joining us. Thanks for having me. Platformer is produced by Lindsay Chu and edited by Fitz Harris at Story and Sound. You can watch this whole episode on YouTube at youtube.com slash Casey Newton. My email is casey@platformer.news, and we'll see you next week. Jira by Atlassian is where your team and your agents work from the same context. Try it free at jira.dev. That's J-I-R-A dot D-E-V.