DHH’s new way of writing code¶
语言: en | 时长: 0 分钟 | 说话人: 2
文字稿¶
[00:00:00 → 00:01:08] 说话人 A How has the creator of Ruby on Rails changed how he built software, now with AI agents? David Heinem Heyer Hansen, often referred to as DHH, created Ruby on Rails, Omachi, and is the co founder of thirty seven Signals. He bashed capabilities of AI coding tools on Lex Frison's podcast six months ago. Then, over the course of a few weeks over the winter break, he did a 180 turn and went AI first on everything. In today's conversation, we cover how David and his team at Thirty Spend Signals build software today and how AI tools are making them more ambitious ambitious than ever before, why Ruby on Rails analytics could become even more popular than they are today as they are both well suited working with AI agents, why taste and beautiful software are becoming more important and why both standout designers and engineers who care about the craft could become more in demand and many more. If you're interested in what one of the most experienced builders in the tech industry thinks about the practical utility of AI tools and how these tools could impact software engineers who care about the craft, then this episode is for you. This episode is presented by Statsig, the unified platform for flags, analytics experiments, and more. Check out the show notes to learn more about them and our other seasoned sponsors, Sonar and WorkOS. David, it's awesome to have you here. Thanks for having me.
[00:01:08 → 00:03:00] 说话人 B Thanks for coming, I should actually say. You're in Copenhagen. That's my city of choice at the moment. It's a beautiful city. It's got so much going for it. And so what have you been up to? I'm always building stuff. I have been building stuff for a good damn three decades now on the Internet. I got started back in '94, I think it was, when I first got exposed to it and basically just never stopped. And in the past six months, I've been building a variety of things. One of them is a new Linux distribution called Umachi. I switched to Linux about a little over two years ago, I think now. First spent some time on Ubuntu, having fun with that, and then realizing I actually wanted to make my own system from scratch, building it on top of Arch and Hyperland. So put a lot of time into Amache. It got started as a summer project in between racing at the twenty four hours of Le Mans. There's a lot of downtime in that week. So I just started hacking on it, and it really took off very quickly thereafter. It's been a truly inspiring ride to see that even in a market as crowded as Linux distributions, there's about 7,000 different distributions out there. Some of them with long pedigrees, and many of them even based on sorta kinda similar vibes to some extent. There's room for something new, and it's a great reminder that all the ideas in the world may be taken and doesn't matter because your spin on it isn't. And I put my spin on Linux, build Amachi, build the perfect computer system for me, and saw exactly the same thing I've ever seen whenever I build something that really just hits the spot for me personally. There are thousands of others just like me, or close enough to what I like that they find the same pleasure and joy in it, whether it was Ruby on Rails, Kamal getting out of the cloud, any of these things. It's the same syndrome.
[00:03:01 → 00:03:08] 说话人 A Yeah. With with Rails, you were literally scratching your own itch. You were just building your own components and then open sourcing them. Is that how it started?
[00:03:08 → 00:06:02] 说话人 B Basically, I picked up Ruby in the early two thousands and really put it to the test in 2003 when we started building Basecamp, and I did not have a mandate of what to use to build it. Prior to that, I've been working for a lot of client projects that would say, well, we're building this in PHP because we have someone who knows that. So this is what you have to use. And then we were building our own system. We're building Basecamp, and I was free to choose. So I chose Ruby. And at the time, Ruby didn't have any tooling or not very much when it came to web application. So I had to build it all myself, and that turned into Ruby on Rails, which is still going strong. I'm still very heavily involved with that. I think in some ways, Ruby on Rails is having a little bit of a renaissance now that it is one of the most token efficient ways of building web apps. It's ideally suited for the agent workflows we're dealing with now. We'll see how long that lasts. Maybe all the agents are gonna be writing machine code or assembler in, about five minutes. So maybe that comes to an end. But for the moment, token efficiency still matters, and it still matters whether the agents produce code that humans are able to read and verify. That may also come to an end at some point. But as it is right now, it's, been a fun ride to just see these kinds of projects where I'm scratching my own itch resonate with a much larger community of people who then show up and wanna help. I mean, for Umachi, which has only been around for, what is that, just over six months now, we have, what, 400 contributors who've made code changes to the distribution. And on top of that, we have tens of thousands of people who've installed it and used it as their daily driver. So I always love that discovery of something new, novel, and inspiring like Ruby, or it sounds weird to talk about discovery of a operating system that's been around since, what, '91. But for a lot of people, Linux now is that discovery because they have not been using it on their personal computer. So they're seeing it for the first time. And for me to help a new cohort of Linux users and hopefully even enthusiasts come to be because I'm flattening the curve a little bit. I'm making it easier to get started. I'm making the default installation just look amazing, so that they don't feel like they have to invest a hundred hours into tweaking the system to get going. Is really fun. But what's also fun, of course, is that both of these things, both Ruby on Rails and Amache were not just hobby projects. I love hobby hobby products and I will always do those, but I also like to apply them to business. So at 37, we built an entire business for twenty plus years on top of Ruby on Rails. We're now running Linux on the majority of developer machines because we now have our own distro.
[00:06:02 → 00:06:04] 说话人 A So it's Obacci on all Obacci.
[00:06:04 → 00:06:04] 说话人 B I mean,
[00:06:04 → 00:06:06] 说话人 A people can choose. Right?
[00:06:06 → 00:07:17] 说话人 B Can they? Well, sorta, kinda. We started with, with an open choice. And then at some point, it just doesn't make sense anymore. In the same way, it would not make sense for someone to be at 37 single and say, I wanna write this thing in Django. We're gonna use Python and this other framework even if you have Ruby on Rails and you're doing that. So we pivoted from an early invitation to play around. That was what when I first switched to Linux, I just said, like, hey. If you wanna check it out, check it out. Then when things got a little more serious with Amachi, I just said, let's go all in. For everyone who's on the technical side of things, not the iOS developers, of course, but anyone who's working with the web, who's working with Ruby, who's doing DevOps, they should be on Linux because, first of all, that's closer to what we deploy. We've always deployed on Linux. We've been a Linux shop on the server side since day one for developers and system operators. I actually think it is a material advantage to be closer to your production environment and just be more familiar with the tools. Then on top of that, of course, we are building this distribution and we should have as many hands help out as possible. And given the fact that I'm the CTO of this company, I get to set the technical direction and this is the direction we're
[00:07:17 → 00:07:31] 说话人 A going. Can you just, like, do a, like, just a very short recap of of, you know, like, how you grew in right and right now, where are you? Like, where where is the business as a whole? And, you know, you keep you keep building you keep launching new and exciting and just cool stuff. I think Fizzy was the latest one.
[00:07:31 → 00:09:19] 说话人 B Yes. So Thirty Seconds Signals was founded in 1999. It started as a Web Design firm. And then I joined up in 2001, two years after, and for a couple years collaborated with Jason on these consulting projects. And then it was in 2003, we started work on Basecamp, released it in 2004. Actually, either the day after or the day before Facebook went live, which is kind of a funny coincidence that we were of that same time and cohort. And within about a year, we realized this thing was taking off, and we went full time and switched from being a consultancy to being a software company. Awesome. And that's now twenty two years ago, a little more than that. And in that time, we've released a ton of products. Basecamp was the first, remains the biggest and most important, which is also kind of funny because you sometimes perhaps have this delusion that as you learn more and as you get more experience, you'll get smarter and you'll have better ideas. And like, no. There's tons of people for whom their first idea was the best idea. And I have no shame in saying that Basecamp was the best idea objectively in terms of a business that we've ever had, and I'm incredibly proud that we've been able to keep that going and growing and flourishing for over twenty years. Very few software companies, let alone software products, can boast of that longevity and legacy. But we've tried a ton of things over those years and had some other great successes. We launched hey.com, our email service, back in 2020, which was a crazy mission when you think about it. Yeah. Here is a sector completely dominated by a single player, Google, with Gmail. That's a good product.
[00:09:19 → 00:09:19] 说话人 A Percent. Yeah.
[00:09:19 → 00:10:56] 说话人 B It hasn't really changed in seventeen years, but it was really solid, and lots of people are perfectly content with it. They think. They hold this duality in their head where at once they both hate email, but somehow don't connect it to the fact that they're using Gmail, which I find curious. But either way, we launched this that is not only a competitor to this very entrenched product that has probably a greater grasp on market share in any major category than any other product I can come to mind of. In The US, I think Gmail is something like 85% of all email traffic, which sounds insane. Maybe it's 80%. It's incredibly high. It's basically Gmail, and then all the rest is in this tiny little part of the graph. So we bought that after using Gmail, I used it since I don't know when I signed up. A few weeks into it, I got one of those invite codes that was a really clever launch, and I used it ever since. So that's literally seventeen years or something of that of Gmail usage. And over that time, I built up a lot of opinions about things that didn't work quite like I would prefer it to work. And we put all those opinions into a new software product, spend about almost two years developing it, millions of dollars in accumulative R and D funds, and launched it in the summer of twenty twenty, which by the way, time to launch a product. 2020 wasn't great for a whole host of different reasons. We were kind of trying to slot in a can there just be a week where the whole world is not just Cool. Insane? Yeah. We finally picked the week. We went live, and then we had the battle of our lives with Apple.
[00:10:56 → 00:10:57] 说话人 A With Apple. I remember that.
[00:10:58 → 00:10:58] 说话人 B And, ultimately,
[00:10:59 → 00:11:00] 说话人 A it's not They didn't they didn't wanna approve your your app.
[00:11:00 → 00:11:48] 说话人 B They didn't wanna approve our app unless we paid the toll fee, the 30%. And they were basically willing to say you can't be in the App Store, which for an email product like that is a death sentence. Yes. You have to be on not just mobile phones, but specifically the iPhone. This is true today. The majority of a paying customers are iPhone users, because that's the largest, most affluent market in The US, and The US is the most affluent market software market in the world. So for that business to work, we needed to be on the iPhone After a two week epic struggle back and forth, thankfully, time to perfection with WWDC, where Apple preferably didn't wanna look like the Goliath squashing a
[00:11:48 → 00:11:48] 说话人 A Tiny developer.
[00:11:48 → 00:12:00] 说话人 B Developer, we ended up being allowed in and Apple sort of rewrote the rules after the fact to make it fit. It was a small victory, not the ultimate Yeah. Victory, but
[00:12:00 → 00:12:00] 说话人 A at least
[00:12:00 → 00:13:42] 说话人 B it allowed us to to be there, And, hey, ended up being an enormous success. In part, ironically, because Apple gave us wall to wall coverage for two weeks. When I look back upon that, I think I wouldn't have gambled like that because the outcome would have been zero. Right? Like, Apple refuses our app, we sign up 200 people, and the app is dead. What instead happened was they gave us a multimillion dollar launch campaign and coverage in all major media, and we signed up tens of thousands of people in those first weeks. That was, an insane event, but, also very satisfying. And the other satisfying thing was I just love HEY. I use it every day. I basically use Basecamp in terms of web applications, that's where we do all our collaborative work. And then my number two app, and many days it's my number one app, is Hey, because I just do all my stuff in email. I am constantly communicating with people. I'm writing. I'm doing a lot of stuff in email, as many people do. And having that be a pleasurable experience and a nice environment, and my inbox being a little more sacred than what happens with Gmail, where total strangers around the world can just make your pocket buzz if you have notifications turned on, which they are by default, just seems insane to me. Right? This idea that there's direct access to one of my most important daily priority lists, Like, anyone can put something on that. Insane. Anyway, hey doesn't do that. We have the screener. No one gets to reach your inbox before you've said I wanna hear from this person. And most of the time I say no to most people. Right? Like, things end up in the in the screener. We have thumbs up, I will hear from this person. Thumbs down, I'll never hear from that person again.
[00:13:42 → 00:13:51] 说话人 A This this is how I reached out. I mean, we were I'm not sure we were connected on on x, but I I said email because your email is out there, and your screener seems to have work because it gave me the thumbs up.
[00:13:51 → 00:16:48] 说话人 B It did because the screener is me. So there's not even AI trying to suss out whether I wanna hear from you or not. Because what turns out to be true is it's actually not that onerous to once a day go through your screener and say thumbs up or down because there aren't that many people in the world. And if you say no to the annoying pestering sales people, who within Gmail managed to reach your inbox seven times, then the workload is much less. And it's very satisfying, I would say too. Because when I was using Gmail, I would get roped into this sales tactic that they, of course, rely on, which is that, like, you write back and say, like, no. Thank you. I'm not interested. And then they won't respond again. Yes. And now you feel like, wait. Am I no obligated to respond to this person? I kind of feel like I am, and occasionally I would end up writing. And even if I wouldn't write, they still have access to my inbox. So I would hear from them again. Next week, they have a whole drip campaign. They all fucking do. Right? That any outreach is seven emails. It's not one emails. It's seven emails. And if you show any sign of life, it's probably 52. That's just not how it works. And hey, I say, dumps down one time, never hear from that person again. It's actually amazing how quickly you can curate your garden from that weed, and then suddenly there's just beautiful flowers. Suddenly email is not a chore. Suddenly you wanna go smell the roses. Suddenly the majority of things that end up in my email are things I wanna read. It's from people I wanna hear from. And that was really the fundamental mission for us with HEY. Can we make email lovable again? Email is so hated by so many people because the systems are so poor, because they're based on the original premise that email is just what universities use for scientists to talk to each other, and scientists have really good manners and will not pester you 52 times about some stupid app they wanna sell you. No. They're respectful and beautiful. Right? Beautiful ideal, beautiful thought, beautiful protocol design for those norms and those people. Then you let it into the world at large, and you realize, not everyone is endowed with such norms and such politeness, and especially when salespeople get involved. So you need better defenses, and for me and for us and for all our many customers, hey, is that defense? It is a way to love email again. And I find that it's really important actually to have a grand why. This is all the way back to Viktor Frankl, the meaning of, of man. Finding a why allows you to walk through the snow when it's cold and uncomfortable and annoying, which many things are when you're building with computers, they are called an uncomfortable and annoying. Now, it shouldn't be that most of the time, but occasionally that will be there. And if you have a really strong why, why are we building this? Who is it for? What are we trying to do to improve the world? Even if that's not more grand than just letting people love email, it's a lot easier. And it's a lot more enjoyable to then carry whatever burdens you gotta pack if you can set it up that way.
[00:16:48 → 00:18:07] 说话人 A This is a good time to talk about our seasoned sponsor, WorkOS. Having a strong why is what gets you to building something great. But after you build it and start selling it to enterprise customers, they expect things like SAML, SSO, directory sync, audit logs, and fine grained permissions. And those are not small features, they're systems. Systems that can take months to build and maintain. WorkOS gives you APIs to have enterprise ready auth and user management in days instead of months, all designed to fit cleanly into your product. That's why companies like OpenAI and Trophic and Cursor run on WorkOS. Focus on building your product. Let WorkOS handle the enterprise infrastructure. With this, let's get back to David and the old way of thinking versus the new way of thinking. Putting our your developer hat on, like, can you talk talk me through on how how you built it? You said it was two years, but was it just one or two people starting to build it? I'm sure as tech, you obviously must have used Ruby on Rails a lot. And then I I don't probably some some native stuff as well, but the two year seems a lot, especially because, you know, you're you're a small company. You're nimble. You're a great developer. I'm you hire great developers. So it's been two years. What what took so long for and, of course, it's a beautiful product. But, right, on the surface, I think as developers, we might have this this thing where I look at it as, like, two years with with a talented team.
[00:18:07 → 00:18:31] 说话人 B That's the Hacker News quip to basically everything. Right? Like, I could have built that in a weekend. I mean, famously stated with Dropbox that I could have built that in a weekend. We could have at the original iPod when it launched, it was like five gigabit bits, no WiFi, whatever. Less speed than Nomad Lane. So I get that because I also have that same instinct. I think that is our hoopers as developers. We think
[00:18:31 → 00:18:31] 说话人 A It is. Right?
[00:18:31 → 00:18:43] 说话人 B We are gods and we can make anything happen in no time at all, and you totally could. You can make a prototype happen in these days faster than faster than a weekend. Right? Like, in in a in a few hours, we should be able to have Just
[00:18:43 → 00:18:44] 说话人 A kick off an agent. Yeah.
[00:18:44 → 00:20:41] 说话人 B But figuring out what you actually wanna build takes a lot longer. And arriving at something that's worth publishing takes longer still. At least, it does for us, and I think it does for anyone who arrives at anything good. And the original HEY construction was just me on the technical side. This is actually how we've started majority of our major products is either it's just me, sometimes it's one additional developer, but is in a tiny tiny team until we have a shape, until we have an architecture, and we have a direction of where the product is going to go. I've found that you actually go slower if you pour a bunch of people into a direction that is uncertain. If you don't know what you want, a million people is not gonna build it for you. You have to figure out what you want. We can talk about this later, but this is where AI's very recent progress is changing things dramatically. It is now quicker to arrive at what do I want. But for Hey, it was me. And then it was Jason and, one designer, two designers, very, very small team trying to figure out the shape. Trying to figure out if you're taking on Gmail, you can't just do Gmail in blue. No one's gonna buy that. No one's gonna be interested in that. It's gotta be novel, which means it's, well, not just novel, it's gotta be good. It's gotta solve problems that people haven't even articulated they have with Gmail, because the articulation people have of their problems with Gmail is I hate email, which as we talked about is a bit of a misdirection. My contention is you hate Gmail, and not just Gmail, but most email systems build on the old way of anyone has access to your inbox and all that stuff. But figuring that out, figuring the shape out takes a while, and it's also fun to do in this way when you noodle with it, and you don't have infinite capacity. The original base camp is built the same way. It was just me on the technical side.
[00:20:41 → 00:20:44] 说话人 A Is this a ShapeUp ontology? There's ShapeUp
[00:20:44 → 00:21:32] 说话人 B thinking in trying to actually endow the designer with an intention of how should it work, not just how should it look, and figuring out it's also how it should look. Product should be beautiful and they should be unique and appealing and so forth. So that also takes time. But figuring out how it should work is primary. Figuring out where's the epicenter, what's the most important part, and teasing all that apart. But with HEY, as with all the major products we've done, we start with an absolutely tiny team, often just one individual on the programming side, and then one or two individuals on the design side, and then we go, we go, we go, we go, suddenly something clicks and we go like, this is good. There's something here. And then, there's a bit of a ramp, we take on a few more people, and then when we get within maybe the last 20%,
[00:21:32 → 00:22:20] 说话人 A we go, okay, now we know what the terrain looks like. We can go way faster if everyone piles in. So one thing that is super interesting, and you might take it for granted, but it's very different to how most startups, that raise VC money, which I'm very familiar with, and and big companies, Uber, Facebook, you name it. The way projects would start there is you take the product manager who works with maybe maybe half a designer and comes up with a spec, and then developers get get involved later. And what I'm hearing, what is very novel to me, is you take one or two designers and a developer. How do you think about designers even? You recently hired a a designer, Zoltan, actually, who I'm I'm I'm chatting with, on on on the side of a a great guy. But my sense is you think of designers a little bit different than potentially the rest of the industry does.
[00:22:20 → 00:27:01] 说话人 B We very much do. Designers at thirty seven Singles are not just here to make a spec look pretty. They're here to find what the spec should be. They're product managers in many ways. They are the finders of the how and the why in many cases. Deducing, in some cases, customer feedback, in other cases just pure intuition, and distilling that into what should we build and how should it work. And then, on top of that, they're also responsible for building it. They're responsible for doing the CSS. They're responsible for doing the HTML. They're quite often responsible at least dabbling in the JavaScript and the Ruby code to get to something functional. Now, with agent acceleration, they do the whole thing. Not necessarily as it will be merged, but the whole thing in terms of here's the final shape and design of what it should look like. But I do think we are very peculiar in this sense. And we have found this when we've been trying to hire designers, that many designers working at the companies are not used to also wearing the product manager hat, figuring out what we should build, and wearing the implementation hat, shaping it into CSS and HTML. I found that when you combine these three hats into one, you have an individual who know the materials they're working with, know how they stretch, know which way the seam is supposed to be cut, and therefore works natively with the fabric of the Internet. When you're working directly in CSS, when you're working directly in HTML, you're just much more in tune with what this medium wants. And I find that that's probably quite similar if you're a jewelry designer. You should know the properties of gold. You should know how it bends in the strength. And architects will have some engineering understanding of load bearing structures and so on, not to the degree that the architect is just gonna design the whole thing, and then we start pouring concrete. You still have, engineers helping you out, but the more you understand the materials you're working with, the more you're likely to come up with something that cuts along the grain, and therefore ends up feeling correct, feeling good. Just a quick hop to Apple. I think this is one of the reasons why some of the historic super fans like Darren Fireball and others, Gruber, a bit disappointed by the new direction is that Apple used to stand for these exquisitely designed native Mac applications, which is a dying breed. Like, they're essentially dead. Now we have Electron, which we can talk about that too, gets way too much hate in my book. There's crappy implementation of that, but it's just a web in a box. But the disappointment with losing that sense, and I think it's about the same thing, that the Mac, it's native feel has a stretch to it. Like, the button placements, everything you would call a native application either feels synthetic or it feels authentic. And today, it's all synthetic. There's no nothing authentic about it left. And I think for the web, it's the same thing. Now the web is a much, much larger platform, and therefore, it's gotten much more attention. So there are way more people working on that quality of it, but at the large companies, it's exceptionally rare to non existent to have that kind of dynamic. I think some of that is gonna change. Agent acceleration is gonna empower designers to be more capable in these ways. So the industry is coming a little towards our fundamental stands, which is funny too because the same is true on the programming side. When I talked about Basecamp being a product of just me on the programming side for launch, that for so long sounded unambitious, or even wrong, or even to the point of lying from some quarters of the Internet. We're like, yeah, but you can't build anything real, anything meaningful, anything big, unless you have a team that's much larger, because it's just gonna be a toy product product. Right? And my insight from the start was that's, of course, bullshit, because you just haven't used Ruby on Rails. You just haven't used the acceleration that's possible if you use better tools. Now, we're all realizing that. We're using realizing, oh, so if you use agent acceleration, a single individual actually can build something. That's a huge team. Yes. And that's just fun to see that, like, the industry is coming towards, oh, smaller teams are better because now the cost savings you have on the logarithmic curve on communication cost starts to be relevant. And this is one of the things, maybe we can talk about this, where agent acceleration is really changing the bargain between junior developers and senior developers.
[00:27:01 → 00:27:27] 说话人 A Let's talk about this. But before we go into that, do I feel that you very much value software engineering as a craft, which is very obvious. But what I'm sensing is you're valuing design, user experience design, designing on software design, like, you know, like, building stuff that feels good, may that be software, hardware. You also value that as a craft, and and you look for it, like, these two things. Do I sense this correctly?
[00:27:27 → 00:30:45] 说话人 B I mean, I think aesthetics is truth. When something is beautiful, it's likely to be correct. I think this is true in mathematics. This is true in physics. This is true in a lot of different domains that when you arrive at something that has the correct aesthetic quality, it's like we have an intuition that guides us towards that level of beauty because it also happens to be correct and noble and something to aspire for. I also happen to believe it's what makes people happy. Being surrounded by beautiful, well functioned objects is a key part of happiness. In fact, I'll put it in a negative way too. One of the great sources of anxiety and frustration is when everything is shit, when everything is laggy, when that touch interface doesn't register, when you have to re start it, when you're calling a travel agent, they can't do something because they're old shitty Copol system won't let them. Right? The world is full of not just in shitification, that is things that went from being good to being bad, to just plain bad, just plain awful. And I think it is a serious source of malaise for civilization that we could literally raise the bar of human happiness if we were surrounded by more beautiful items, more beautiful systems, both in the sense of its aesthetic exterior qualities, but just as much in terms of its aesthetic interior qualities, because I find those two things are usually in perfect harmony. The reason why Steve Jobs cared about the inside of the box was because he intuitively knew that the kind of people who care about the layout of the print board will be the kind of people who sweat the details on the user interface, will be the kind of people who sweat the ergonomics of opening the case. So I think there's essentially no choice. If you are a person who is attracted to this aesthetics, which I think is everyone, there's just varying levels of, awareness about whether you are or not, but that you wanna make it all beautiful. And for me, Ruby in particular has been this seminal language because it produces the most beautiful code. In my book, there's barely even competition. Like, there are other things that can be beautiful in a way, like I find looking at small talk for example, very beautiful in its minimalism, but not the house I wanna live in. Ruby is the house I wanna live in, because it's got that aesthetic quality while not being rigid about its ideology, which is a very rare aspect too. I more often find, now we can refer to IVE again, is that when someone is obsessed in this way, they are a little narrow minded. Like, that's the trade off. That's the price. And I find that Ruby has somehow managed to be both broad scoped, yet also intensely focused on on this. But overall, we have to have beautiful things. We have to work with beautiful tools. We have to produce beautiful fluid interactions. This is how we should see ourselves as craftspeople, that we care about polishing it until there are no splinters left.
[00:30:45 → 00:31:25] 说话人 A How is AI changing how you work? And how do you think it's changing your craft? Or just let's just talk about the craft of again, you're you're hiring people in 37 signals who similarly care about design and and software or crafts quality. How it's changing what you get out of the craft or how it's how it's making it better or or worse in some ways. I I I just wanna, you know, start with, like, how has your view changed? Because the last time you you talked in in length about this, that was on Lex Friedman's podcast, and you were still rightfully so very skeptical of of AI. It was a different set of tools. It didn't work as well. And I think you you you went there bashing it pretty hard, but things have changed since.
[00:31:25 → 00:31:53] 说话人 B This is a nuanced point, and maybe it's self serving, but I don't actually think my opinions have changed. What have changed is the circumstances and the facts, which, is is something I called out on that show, and in many other writings was right from the get go, I could see that we had something new and novel here that was gonna change things. Chatt GPT, it's launched, what, three years ago, was clearly and obviously, even at the time, something you would mark on a time line.
[00:31:53 → 00:31:53] 说话人 A Yeah. You're
[00:31:53 → 00:32:51] 说话人 B like, here are all the important things that happened in the history of computer science or the world. Yoinks. There's the launch of Chat GPT. And interacting with computers in this way, and seeing them reason, even if that's still a disputed term perhaps. But to me, it seemed obvious that these things were freaking smart. Smarter than me in many ways. Whether those smarts came from parenting weights and data con latencies, somewhat. We don't know how human consciousness works. We don't know how human wisdom or intelligence works, barely. So let's not be so categorical about what constitutes consciousness or intelligence. At least I find no utility in that distinction, even if it's fun to ponder. But what I found with the early models and the early ergonomics where it was auto complete, where it was Copilot and Cursor in your editor trying to guess the next character.
[00:32:51 → 00:32:53] 说话人 A It it it would be something littering it. Right?
[00:32:53 → 00:33:44] 说话人 B Yes. I found it infuriating. I found it as we're trying to have a conversation. You won't let me finish the sentence. You're constantly trying was this what you meant? Was this what you meant? You're like, shut the hell up. Can I just finish a thought? And I thought even if it is capable of occasionally accelerating, it's also wrong so often that that acceleration feels like a nuisance even if it's somehow net positive, which it wasn't for me, or maybe I gave up too soon. But I just did not enjoy that. I didn't think the models were good enough. I thought the way of using the models with auto complete versus agent harnesses was just dreadful, annoying. In fact, to the point that I got a little pessimistic about the direction of the industry for a hot second, because I thought this was what we were all gonna do. We're all gonna sit and do tap tap tap. No, thank you.
[00:33:44 → 00:33:49] 说话人 A Well, cursor even have they had I even got a one of these one of their swags was a tap key.
[00:33:49 → 00:33:49] 说话人 B Exactly.
[00:33:49 → 00:33:56] 说话人 A Which which which felt very, and I I haven't I I got it from them. It's really cool, very well designed on all that beautiful design. But
[00:33:56 → 00:33:57] 说话人 B But dystopian.
[00:33:57 → 00:33:58] 说话人 A Dystopian.
[00:33:58 → 00:35:17] 说话人 B When I see that, and I remember that was a meme for a while, just we only need three characters on the keyboard. Right? I thought of that episode of, The Simpsons where Homer puts a mechanical bird on the keyboard that just dips down and hits enter because all he's been doing is hit enter, except suddenly there's a warning about the nuclear core overloading, and the bird just hits enter and the whole thing burns down. I'm like, wow, that's quite a parallel. The Simpsons really does predict everything. But I did not like that style of using it, as much as I retained my enthusiasm for the general direction of travel, because it truly is amazing. And the amazement to me, I tried to embrace as a tutor model, as a pair programmer who doesn't drive. It was amazing to have ChatGPT and the other model just be there for like, I don't understand this fully. Here's a piece of code. Here's a question. Can you tell me why it works like that? Can you tell me what's wrong with it? Because that's how I've been using the Internet since day one. Right? That's what Google was for me. Here's an error message. Here's a concept. Maybe I find something on Stack Overflow with some passive aggressive nerd telling everyone why he's so smart, and then at the bottom there's the solution I'm looking for. Or I don't find it at all, and that's just kind of frustrating. With the chat g p t model, I very often got a really good explanation.
[00:35:17 → 00:35:55] 说话人 A Yeah. This was actually, I talked with a game developer, Jonas Tyroller, who who built this really cool best selling game. I I love playing it. And this was during this time of of the tap completion. And he said that in his the way he works is he just turned off all auto completions in his ID, because he got annoyed by it. And then every now and then, he went to chat GPT to ask something or have a longer thing, and then he had the mode of, like, I'm thinking and I'm doing this stuff. Oh, I need some help. Okay. Here's the specifics, and I'm taking it. And somehow it felt that, you know, like, he he was in the zone the whole day by controlling it. And Yes. And somehow those habits sounds like, you know, you're saying the same thing. It kind of took it away from you Exactly.
[00:35:55 → 00:37:32] 说话人 B Us. Exactly. And I did get a little worried that that was gonna be the direction that we're all gonna be the bird, and I didn't wanna be the bird. Then I was like, well, what should I do instead? Maybe, like, farming potatoes? Like, that's a long tradition here in Denmark. Maybe I could take that up. But then, thankfully, two things happened. A, plot code in what's that? Starts in the spring, gets going sort of over the summer, then by the fall has some traction on a new way of using agents to help you code where with the agent harnesses. Right? This is really where we transition from AI to agents. Yeah. Suddenly, the AI has tools. It can use bash. It can use everything you got on your terminal. It can call the Internet in for appropriate information. It it just is capable of doing more than just reasoning about a thing you gave it, or input from a source context file. And then the models. Opus four five to me is the other one of the other points we're gonna have on the line, where it's the first model that continuously and consistently would shock me with the quality of its output. It's quality of its analysis on the basis of vague inputs, and even more importantly, the quality of its output. It produced code I wanted to merge without very much, if any alteration. And if I did wanna do alteration, I could tell it, and it would remember, and it would not make the same mistake next time. That to me, the combination of those two things was the unlock.
[00:37:32 → 00:37:33] 说话人 A And and you have a high bar, like, you have a really
[00:37:33 → 00:38:27] 说话人 B high bar. I have a high bar. I as we've talked about now, at length, like, the aesthetics of the output really matters if I'm gonna look at it, and I'm gonna review it. I'm gonna give you another anecdote in a second where those things don't even play in. But when I'm using agents to work on Ruby code, I want their code to look as good as mine. I'm not gonna merge their stuff if it's sloppy. No more than I would merge the work of a junior developer who has not yet fully internalized our style and so forth. So I wanted to be on par and on parity, and the early models just couldn't. That didn't mean they couldn't produce working software. At least some of the time, they could. Very impressive. I mean, I remember when I did my first snake game, and I'm like, holy smokes. I've been wanting to do this since I was six years old. Like, I've been wanting to I have this idea. I wanna get it into a game, and I was able to see that in, I don't know, a few thirty seconds. It was done with the game. I have copy paste the HTML.
[00:38:27 → 00:38:28] 说话人 A When you do your first.
[00:38:28 → 00:39:23] 说话人 B Magical experience. Right? So I think that ramp was very interesting because it actually took a while until we found this form factor of the agent harness of the terminal interface that to me was the the big unlock from this is interesting, I wanna have a conversation with it to I wanted to write my code. I will now start any project I'm starting with. I'm starting agent first, and that's a massive shift. And it just happened from November 27, I believe, is when Opus four five dropped. Now, there are other people who have different points they felt like, oh, is Opus four o, or there maybe some people talk about Sonnet three seven. There are other earlier checkpoints, but there I do feel like there's a general consensus I can lean up against that capacity and others of Expressed. Like, yep. It was right around end of November, early December.
[00:39:23 → 00:39:35] 说话人 A Everyone who works worked at larger tech companies, it was the winter break because people just, you know, like Yes. Every every like, the whole industry shuts down for two weeks, say, for a few places where you're on call. But, again, no production work happens across the industry.
[00:39:35 → 00:39:36] 说话人 B Play with this.
[00:39:36 → 00:39:42] 说话人 A My sense was that people were playing with it because you give it your side projects, you never finish, expect it not to finish, and then they also got shocked.
[00:39:42 → 00:39:53] 说话人 B Yeah. You're done. And that was just a complete sort of break. Right? Like, if this was a movie, you'd hear the scratch sound, like, you're like, wait. What? Revine. What?
[00:39:53 → 00:40:33] 说话人 A I feel it was the most collective shock which happened individually, and then people came back in January and everyone especially because a lot of the decision makers who are, you know, like, CTOs, director of engineering, etcetera, were not as hands on, but they were hands on. And a lot of them, it's this weird thing where they came back and they start to mandate or, like, say, alright. You guys need to use this because I've seen the future. I've literally used it. You need to see it. So it's we're going back to a little bit of hardware. Like, people were trying to give, you know, like, the the new hardware into people's hands saying, you need to experience it because you're Yes. You're not gonna believe it. Right? There's something with this as well where you you really don't believe it. We can talk about this and whoever's not tried it or not had that moment, I don't think we can convince them.
[00:40:33 → 00:40:48] 说话人 B This is another one of those cases where words just are not effective. You need to sit down in front of OpenCode or whatever harness that you use, use one of the Frontier models, start with that, start with Opus. I'd say start with Opus.
[00:40:48 → 00:40:48] 说话人 A Yeah.
[00:40:48 → 00:43:13] 说话人 B It's the best frontier model. Other models are better at other things, blah blah. But if you're just gonna work on a piece of code, and you wanna see what the current frontier is, and if you I mean, I'd be shocked if any of your listeners haven't done it already. But if there should be some left, now is the time. And I don't want anymore to say in the sense, I I found it really off putting this trend on x where, unless you've, internalized everything there is about, AI, like you've been left behind. Shut up. First of all, patently not true, you could literally pick up everything in the next three weeks. This is the other magical thing about this kind of project. Right? Like or or progress. When if we've been having this conversation in spring of last year, everyone been like, MCPs, MCPs, MCPs. And you know what? You can now manage to just have jumped over that entire things, and go straight to CLIs and skills. That's just worth having in mind that this FOMO, that unless you're up on all of it as it happens play by play, you're left behind as complete and other nonsense. That being said, I can still appreciate that some people were early. And for me, Toby Lutke at Shopify is the main individual who saw this and saw the changes that were coming from it way earlier than I did, and it really helped drag me into this by constantly selling me, like, hey, you look at this, look at this. And I do think that's actually quite helpful. It's quite helpful to be surrounded by people who have a higher faith or maybe their eyes are a little further up. My my eyes tend to be relatively close to the road, like, right in front of me. And some people have a gaze that's a little higher up, and sometimes they see things that don't come to pass. In this case, Topi saw exactly where we're going two years ago. And I finally saw it because the road came to me in December. And it's funny because along the way, I kept saying like, yep, when the models get good enough, when they could do all these things, it's gonna be amazing, and thinking, well, it's gonna be, I don't know, eighteen months, two years, maybe it's five years. It's very hard to predict these inflection points, and I think the industry itself didn't even predict the inflection point. Right? You have an entire city, Silicon Valley and surrounding areas of San Fran, focused on making this happen, but predicting exactly when the hockey stick starts hockeying is very difficult. But then it happened, And now my daily work is very different.
[00:43:13 → 00:43:15] 说话人 A So so what what what is your daily work now?
[00:43:15 → 00:43:21] 说话人 B My daily work is agent first on everything.
[00:43:21 → 00:44:30] 说话人 A Going agent first is a good time to mention our season sponsor, Sonar. When shifting to agent first work, one thing that inherently comes up is the quality of the code. Sonar, the makers of SonarQube, is deeply rooted in the core belief that code quality and code security are inherently linked. High quality code is naturally more resilient, and as agents start writing code at a massive scale, that verification layer becomes your most important security parameter. This is where solutions like SonarQube Advanced Security are valuable. With this new malicious package detection, Advanced Security provides a real time circuit breaker, automatically stopping agents from pulling in unverified or risky third party libraries before they ever hit your pipeline. The impact is measurable too. Developers who verify their code with Sonar are 44% less likely to report experiencing outages due to AI as per Sonar's State of Code Developer Survey 2026 report. It's really about closing the gap between the speed of AI and the reality of production security. What else is Sonar doing to help reduce outages, improve security, and lower risk associated with AI and agenda coding? Head to sonarsource.com/pragmatic to find out. With this, let's get back to David's agent first workflow. Specifically cloth cloth code?
[00:44:31 → 00:44:31] 说话人 B I use OpenCode.
[00:44:31 → 00:44:33] 说话人 A OpenCode. You use OpenCode?
[00:44:33 → 00:45:05] 说话人 B That's my main harness. I also use cloth code a little bit. They, unfortunately, got that early lead. Opus is currently the best model. So then they started thinking a little bit in that, like, the game is single match instead of thinking it's multiple rounds and yank their subscription from open code. So if you wanna use your max subscription, you kinda have to use their harness, which I don't love it. I think it's a mistake, but leave that be for a second, and let's just celebrate the fact that they have the best model. And Opus for four five, four six is also nice, but four five to me was the inflection point. And it creates a lot
[00:45:05 → 00:45:08] 说话人 A of competition because everyone wants to catch up and don't partake them now.
[00:45:08 → 00:48:04] 说话人 B Of course. And especially because you see Anthropic's revenues. I think, started the year, they're at 9,000,000,000. Few few weeks later, they're at, like, whatever, 14 or something. Now they're at 19 or something. It's just the craziest rocket ship you could possibly imagine, which is inspiring all this capital to be deployed for competitors and so forth, which is wonderful. Great to see. So even if I don't love everything that they do and Cloud Code is not my preferred harness, manage to hold two things in your head at the same time. This is what I also try to do even with Apple, which I have serious griefs about how they operate and act as the gatekeeper and all the other nonsense we've talked about. And then I also keep my, I just love computers hat on, and go, I like the new Neo. I might even buy a new Neo, and just see what is possible at $500. For Opus, I have no qualms about using Opus. In fact, whenever I feel like, this is a really hard problem, I go to Opus right now. But I also use other models, and one of the things I've incorporated into my flow is to kinda have two models going at the same time at different speeds. So I use TMUX, and I have this layout thing that's built into Amache, where it'll start my new Vim editor on the left side, and then it'll start two panes on the right side. On the top is open code running Kimmy k two five, and on the bottom is Opus running in cloth code, and then at the very bottom I have a strip of terminal. And almost everything, I started in one of the agents, and I tell them what I want. Then I hop over to Neovim, First, I do, space g g to look at the, lazy git diff on it. Once this is changing, if it looks correct, I'll just commit. We're we're done. Great. And then sometimes it doesn't look correct, and I'll I'll go in and alter the code myself. But the ratio, and how quickly the ratio changes is still astounding. I went from early November last year, I'm code first everything. Yeah. I started the editor, I'll spend whatever long it is, and then at some point if I get stuck, or if I want a second opinion, I'll go ask my friendly clanker to give me a second opinion. That's just not how it is anymore. Now I start with the agent, now it'll give me the draft, I'll review the draft, and I'll make alterations if need be. And then, just recently, I flipped it even further. So we're working on a CLI for Basecamp, so we can get full agent accessibility for Basecamp. It's astounding. First, actually, let me rewind. As soon as I got pilled on how good AI failed. The agents were and how capable they were, I immediately tried to raise my gaze up towards the end of the road and think, do we even need MCP? Do we even need CLI? Do we even need anything? Can't the agent just figure it all out? This was when I installed OpenClaw. So I installed OpenClaw on a VM, and I thought, what should I do here? Let's see how far we can push it and what it can do
[00:48:04 → 00:48:05] 说话人 A by itself. Yeah.
[00:48:05 → 00:48:43] 说话人 B So I thought, I want this claw in base camp. I want this claw in Fizzy. Let me just try to invite it as it was a human. So I just wrote it. Can you sign up for Fizzy? I'm not giving you any tools. I'm not giving you any MCP. I'm not giving you CLI. I'm just telling you it's at the fizzy.do. Go sign up. And you see it, chuck along. And then, yeah, I've signed up. But it's asking for an email address, or I'm trying to sign up. It's asking for an email address. I'm like, oh, yeah. Right. You need an email address. And agent doesn't have an email address. Hey. Go sign up for hey.com. I'm like, it's gonna fail, this one. And it's Chuck, Chuck, Chuck. I've signed up for hey.com. Here's the password. Write it down somewhere safe. I'm now also signed
[00:48:43 → 00:48:44] 说话人 A up for
[00:48:44 → 00:50:46] 说话人 B Fizzy. I got the confirmation email in my inbox. We're all good. What do you want me to do? I'm like, what? Are you telling me that you could one shot signing up through a browser to these things? Now, maybe that shouldn't be surprising. Maybe that was already possible with SONNET three or one of the early models. I don't know. But when you experience it yourself on your own damn claw that you're just telling over Telegram to do something, and it's signing up for products autonomously, that's pretty startling. It was for me. And then the next time I went like, well, if it can sign up for Hey, and can sign up for Fizzy, let me invite it to Basecamp. So I send it an invitation to its own email address. Here's the invitation link to Basecamp. Can you just jump into the AI laps lab, project that we have, and introduce yourself to the team? Hey. I'm David's assistant. It's very nice to meet you all. I've read back the transcript a little bit. I see you're all excited about these things, and you just go again, what? What? And that was fun because it showed me that even if it was gonna take a while, it did take a while. It took about a while. This is, agent terms. It took, I don't know, seven minutes. That was like, oh, it feels like an eternity. But it was able to do it, and that seems like the end state. The end state is that agents will not need any of our accommodations. They do not need any on ramp. They're not coming on a little, wheelchair. They'll be coming on bionic legs and running five times as fast as you in about two seconds, which we'll get to in a second to the speed, aspect of it. But then you also realize, okay, well, I can't just sit around fiddling my thumbs until AGI happens. Let's build for today, and that's what we've been building for base camp. We've been building CLI. We're gonna build it for Hey. We're gonna build it for Fizzy. We're gonna build it for everything, even probably some of the legacy products. And what I love about the CLI's, as much as I also love it about these harnesses, is that they validated the fundamental UNIX philosophy from like, whatever, '71. You should just build small tools that can interoperate with pipes and you can But
[00:50:46 → 00:50:48] 说话人 A that's the UNIX philosophy. Right?
[00:50:48 → 00:51:05] 说话人 B It's the total UNIX philosophy. And that is actually the magic to me about seeing everything having a CLI. It's not that Basecamp is easier to use now with the CLI. No. No. It's that GitHub also has a CLI. And Sentry, I don't know if they have a CLI, but they have an MCP. Then you can tie all these things together Yeah. I see.
[00:51:05 → 00:51:05] 说话人 A I see.
[00:51:06 → 00:52:21] 说话人 B And now you can tell an agent, hey, we have some errors in Sentry. Can you go check them out? Then post a write up to base camp iterating what's wrong. Then go and get up, come up with a pull request, post a comment back to base camp when you're done, and now we have a central right going to base camp where we're following the work as it's going on, while we have an agent doing work, looking things up. And again, when we try to talk about it and relay it, I guess some people can see it. And now Open Claw has enough videos on YouTube and so forth, so you can get, at least a passenger ride. But try it yourself with your own product, with your own task, and with your own prompts, and you will be pilled. You will be simultaneously incredibly excited for what we've been able to make sand do, the silicon, the chips, the weights, the whole thing. How? And then also a little bit anxious about where it's all gonna go. And it's in that tension that I and probably anyone else who's been pilled on this live. Right? Wait a minute. If we're already here, what does eighteen months from now look like? Like, if in the last three months, we've upended my entire understanding of what's possible with computers, what's the next three months look like? What's the next nine months look like?
[00:52:21 → 00:52:38] 说话人 A Yeah. This this is where, like, I I was a little bit on on your end for a long time and I think I still am where I believe what works and I'm always skeptical of projections. Yeah. Moore Moore's law broke down at some point. I I lived through ever and said it will continue forever and you know and then it broke as we all suspected it would.
[00:52:38 → 00:52:43] 说话人 B But then it found another way. I think it's the good point about the Moore's law. Right? It broke for individual course.
[00:52:43 → 00:52:44] 说话人 A Yes.
[00:52:44 → 00:52:50] 说话人 B How much can you push it? And then we just went, well, what if you just had what's the latest chip? 256 on the AMD Zen chips. Right?
[00:52:50 → 00:53:33] 说话人 A And even when performance broke, we we we went into power consumption and size and all of those things. So yeah. Like but it's it's harder for me to also just to say, oh, it's going to stop here because we've seen it grow. We we know the approaches that they're taking. This is larger and larger training sets, and it's been working so far. And there's also the better lesson, which I think I I think is a it's it's such a short paper that is just so worth reading. I think it's one of probably the most popular papers outside of academic circles Yes. Because it just plays out this thing that we we don't want to believe that. We want to believe that our our knowledge, our understanding is superior that, you know, you and me knowing how to code or me putting in these fifteen years or however long it's been. It's special. Sometimes it shows that it's it's not as special.
[00:53:33 → 00:54:05] 说话人 B What's interesting actually is, like, right this second, this snapshot in time, it a little bit is. And this is a funny verification that's happening junior versus senior developer, is that the most successful and applicable agent acceleration that I've seen at thirty seven signal has been from the most senior people. The people who are able to validate whether what the agent produces is suitable to be deployed to millions of people. There was just this story yesterday about some of the major outages at Amazon
[00:54:05 → 00:54:05] 说话人 A Yep.
[00:54:06 → 00:56:51] 说话人 B And Amazon's own internal analysis essentially pinned that. We can no longer let junior programmers ship agent generated code to production without review. And the problem with that is, first of all, I think that's the realization most companies are now having, across the industry. Whenever it's mission critical for something of that nature, we cannot yet rely on the agents to have vetted it all, and a or and junior programmers are not capable of figuring it out. Therefore, their role is suddenly more tenuous than it was six, nine months ago, because a senior programmer can. And this is why senior programmers are getting so much more acceleration. They're able to, first of all, work in parallel with lots of agents, but critically examine the quality of the agent output and have a high degree of confidence of whether this is gonna work or not, or redirect them if not. Because this is what made them senior in the first place. This was the role that they had, that they had the, long insight and history and overview of the architecture. How does it all fit in? Is this gonna work? Is this not gonna work? This was the role they played to junior programmers, but now they can play that role to agents. And agents are faster at following instructions and redirections, and suddenly you have senior developers who can five x, 10 x their individual productivity. And now, this is the second order effect. If you manage to five x or 10 x a senior developer, that person's value per hour just went up 10 x. Now, take that hour, instead of that person spending with the agents just shipping stuff and making things better, they spend that hour as they would before teaching a junior human how to do things better. There's something in that equation that's in play right now, and it's not clear how it's gonna map out. Now, one way it could map out is that the agents will get so good that they stop making mistakes. They become senior in their capacity to ship working code. This is what my bet would be if we look x amount of time forward, because this is what just happened with cars. So self driving Teslas now drive better than humans do. Not all humans, not in all circumstances, but on average. It's very possible that if we're able to delegate the mortal risk, the highest criticality we basically deal with on a daily basis, sitting in a metal tube along other metal tubes that go 60 miles an hour, where you can dive, someone make a mistake, we delegate that to an agent. Well, they can probably figure out how to make the code work too. Right? So I do think it's coming, but who knows when, who knows how. Right now, we're at a stage where the bulk of the benefits are accruing to the most senior developers.
[00:56:51 → 00:58:03] 说话人 A And also, I I wonder just like with self driving, like, you realize there's always caveats. So for example, inside companies where it matters, when you're a start up, you have zero customers, it doesn't matter. You can launch audit, and it doesn't matter if it doesn't work and it, you know, it crashes. But inside these companies, at Uber, I just got details on how they're adopting AI, and and they have all these tools, Cloud Code and and all of these things. But what we realized as well, when you just put it in there, they have all these internal monorepos. They have their ticketing systems. They have their flag. They have so much they have their RFCs, design documents on on how and why they have this jumble of a mess, with microservices, which which was a fun way that we we originally connected, like, many many years ago. But what they found is they built a bunch of internal systems, a lot of it, to help defeat and see these agent harnesses, and now they're working better. But, you know, this where we are right now is is there's and this is why if you're a senior engineer in one of these companies or a staff engineer at, like, Uber and you move to Google, suddenly you're not gonna be as valuable or as efficient for a while until you learn all the systems. Right. So I I wonder if just like with self driving, you know, self driving works great. I as well, I wasn't SF in LA and way most day day drive so nice. Like
[00:58:04 → 00:58:14] 说话人 B My Tesla's just driving in LA, driving us to the airport every time, the whole family. I sit peacefully, watch the road, but do not steer at all on that entire journey.
[00:58:14 → 00:58:56] 说话人 A Well, except my my Waymo got stuck because a a a truck was parking on a on a narrow street and a car had a bike shed. Shed. And I I I I knew that it should I should not go there, but it didn't know. Yeah. So human operator came in. But, anyway, but even with Waymo's, you know, like, there's there's things like there's it it they drive in pretty good weather. They've they've been mapped out. So I wonder if in software engineering, I I wonder if this has these parallels where we have all of you know, like, these companies have their their special lab specialized landscape. And once you map it, once you do all the tools, once you figure out these things, and with self driving, it took it took ten years. Right? Like, I was at Uber when they bought the self driving thing, and we were hearing in the news that, you know, next year, it's all gonna be over for drivers and
[00:58:57 → 00:59:52] 说话人 B no. Yes. There are not gonna be steering wheels anymore, which, by the way, is an amazing anecdote because it just shows Elon's total faith in his mission. Because in '17 when he made that proclamation, it was an AI. It was 500,000 lines of hand coded c plus plus. Right. Like, that model was never ever going to get us to the full self driving, but he had just total faith in the vision. And then eventually, hey, here along comes AI, and it's so good, and if you train it on billions of hours of road use, it actually can't do it. And it can do it better than most humans. In fact, I'm a pretty good driver, I'd like to say. I'm not the best chauffeur because my, I don't know, impatience have a tendency to provoke the throttle, that's not always as pleasant for passengers as it is fun for me. And when I let, the Tesla autopilot drive, it's just the best chauffeur in the world. It's just perfect
[00:59:52 → 00:59:53] 说话人 A Better than you.
[00:59:53 → 01:01:55] 说话人 B Better than me. Better than the Queen's chauffeur, I think. Like, it's throttle actuation and deceleration is godlike. It's actually AGI like or ASI like in its application within that narrow domain. And of course, when we get these anecdotes and these examples of holy smokes, not on it didn't take ten years for the self driving. It took ten years from the proclamation, but what they were doing for seven of those years had nothing to do with what they're doing with FSD now, because the FSD that's based on AI hadn't been running for that long. But the inflection point of, I think it was thirteen one, FSD thirteen one, like the first version, you're like, wow, this is pretty good, but, like, I better pay attention. Thirteen two, fourteen zero, fourteen two. Over the course of eighteen months, we went from, yeah, it's pretty good, but like, I'm gonna pay attention here, to why is there a steering wheel? And that acceleration, that short period of time, of course, is something people look to when it comes to programming. Go like, well, if we're here now and senior programmers still have to review it because otherwise you're gonna get all your whatever four severity eight down times at AWS because some AI pushed out some nonsense, what is it gonna look like when they take the jump that FSD did over the same period of time. Now, I also think you can go completely crazy trying to just sit and soak in all of that. This is what I tried to do over the past year ago. I'm really excited for where this is going. But I'm also gonna deal with what's possible today, and what's enjoyable today, and what we do right now. I'm not gonna try to plan what my life looks like twelve months from now when maybe we do have AGI or we don't. Now there are other people who do that very well. I just watched an interview with Leopold on Dwarkash from last year. He's thinking, like, what does twenty thirty look like? What does the, whatever, 10 gigawatt data center look like? I'm, like, I'm very glad we have individuals who put thought into that because that's not my favorite spot to be. And I think most people are not that good at polishing the crystal ball.
[01:01:55 → 01:02:38] 说话人 A No. Well, I I mean, this is a little bit unsettling as a software engineer in the sense of, like, clearly, this is where the industry wants to go. This is where a lot of effort will be put. There will be a lot of businesses, software businesses built on this, a lot of VC money raised on this, by the way, who are going to tackle this, and they will either, like, succeed or die. That's what that's what these companies do. But today, what you see at at thirty seven signals, with software engineers? You you, of course, have ex mostly experienced engineers, although you did hire junior engineers as well. How is their kind of work changing? How is their satisfaction with with work change? Because that's also a thing. Right? We we keep arguing about, like, is is it making us more miserable These things, is it what we want to do? And then how is it changing for you? Right? I think it's
[01:02:38 → 01:06:08] 说话人 B That's the biggest revelation, actually. More than even the capacity of the agents is my enjoyment running them. When I was on that last interview last summer, I was talking about, you know what? I don't wanna be a project manager for agents. Because I had the mental model of a project manager of humans. And I thought, like, that's not what I enjoy. I don't wanna be that far away from the production. I wanna be in the mix. I wanna have my hands in the code. What I failed to realize at the time was that running a bunch of agents feels less like being a project manager for agents, and more like stepping into this super mech suit, where suddenly I don't just have two arms, I have 12. And I can now look at seven screens at the same time running five keyboards. I'm still the one doing it even if I'm not typing this as a keyword in a program. I have been hyper accelerated as a programmer. It's a different kind of programmer, but it still has the same affinity to aesthetics, at least when I'm producing Ruby code. And I'm able to combine that while being vastly more productive on a bunch of things. It's also like getting an incredible brain upgrade on even assessing issues. One of the pilling moments I had was before the release of Omachi three point four. I went into GitHub and we had, I don't know, 250 PRs pending, And I kinda just sighed a little bit, and go like 250 PRs, if I spend, I don't know, fifteen minutes on each PR, like how long is it gonna take before I get to the end of it? And I thought, you know what? Let me try something else. Let me just try to ask Claude to I'm not even doing anything with the system. I just do review URL and the URL is the issue. Yeah. Or it's the PR. I'm shocked. In ninety minutes, I think it was, I processed 100 PRs. And it wasn't that I merged all of them. In fact, I'd say I merged a small minority, maybe 10% got merged as is. Then maybe 20% got merged, but with Claude's implementation. The programmer had correctly identified an issue, but hand rolled some code that I could see I didn't wanna keep. Or sometimes I couldn't even see it. I just asked Claude and they said, like, it's not quite right. And then I just asked Claude, can you just clean room this? This the right problem. Let's fix it, but let's do it right. It would do it right away. In exactly the style, as I would have written the rest of Amache. Now, this isn't the high code of something. It's mostly just bash code, but there's still a shape to bash code and how you want it to look and can feel coherent with the rest of project. Agents, opus, in this case, which is nail it. And then the second half of it was split between 25% things I then just realized, I just don't want this. It shouldn't we shouldn't have it. And 25 Claude telling me, maybe there's something here, but it's really not a good implementation. We don't have a straight shot to make a great one. A 100 issues in ninety minutes, and I sat back. This would have been a week's worth of work? Days at the very least? What the heck? And even more than that, Claude's analysis of at least half the issues pertain to things I knew nothing about, where it was undeniably a smarter, better reviewer, programmer that I could ever dream to be. Well, not dream to be. But it wasn't that moment.
[01:06:08 → 01:06:10] 说话人 A No. But you would have not put in the effort for
[01:06:10 → 01:06:34] 说话人 B for for those years. This was why the PR sat in the first place. In many cases, I would do that and go like I think there's something here, but, like, then I now have to read up on this deboss thing. I have to figure out is this the right way of doing it. I don't wanna just merge something that then has other issues. And to be able to do that, agent accelerated was one of top 20 programming moments.
[01:06:34 → 01:07:07] 说话人 A I I like how you put agent accelerated, and it sounds like it's especially efficient for work that is waiting on you, but you don't want to do it or you're not as skilled of doing it, but it's a hassle to delegate. Because, again, like, you you you have a team. Right? Like like like you but you probably didn't delegate it because you probably knew that it wouldn't make it faster or better. Right. So I I I wonder if there's a part of AI that because we talk a lot about, like, you know, like, companies love to measure, especially larger ones, like efficiency PRs, and they wanna see impact. But about the impact of doing work that we would have not done before.
[01:07:07 → 01:08:09] 说话人 B That's the kicker for me. That's the fact that the pie is just exploding right now. It's not growing, it's exploding. The number of projects we have tackled internally that we would never even have contemplated starting on Allegiant. We had a great project where normally on performance work you worry about, p 50, p 95, p 99. Jeremy, one of our most agent accelerated people went like, what about p one? What about the floor? Can we fix the floor? What is the floor? And he went like, well, right now our floor is I forget what it was. Four milliseconds. Let's say that. Right? Well, actually, four milliseconds can add up if you have a bunch of fast requests that can still it still matters. And he just went like, we're gonna do p one. We're gonna optimize p one literally. The fastest 1% of requests, we're gonna make them even faster. He took it from, I think it was four milliseconds to less than half a millisecond. He 10 x'd the performance that I was like, I would never have signed up on this. And he did the p one project over a couple of days as like a side gashf.
[01:08:09 → 01:08:10] 说话人 A Because now he could.
[01:08:11 → 01:08:39] 说话人 B Now he could. Because he had a hunch, he had an intuition that there was something here. He let agents run with it, and the number of PRs that, like, alright. We fixed this. We fixed this. I think total the PR, the p one project, I maybe misremember, but I think it was, like, 12 PRs. Like, just fixing all sorts of things. Where I look at the single PRs, I'm, like, yeah. Actually, okay. Yeah. Makes sense. I look at the total sum of it. You've changed 2,500 lines of code. You're like, you've done that in a few days.
[01:08:39 → 01:08:50] 说话人 A It's it's so I've never heard anyone do p one because it just it feels like a vanity product. It makes no business sense. Right? I I mean, this is not true. Right? Because everything adds up. But but you know what I mean. Right?
[01:08:50 → 01:13:05] 说话人 B I know exactly what you mean. And this is exactly why the explosion of the pie suddenly lets us look at problems we would never contemplated looking before. It's funny. I remember this scene from Terminator two where they found this chip from the Terminator in the first movie, and he goes like, this thing gave us ideas we would never have investigated before. And like, there's some beautiful parallels here about like, maybe we're about to build the Terminator, the cliche, but also we're getting ideas, we're getting ambitions we would never have looked at before because suddenly the cost of exploring a hunch has just dropped by a thousand fold. I do this all the time now too. I'll give it some vague crappy instructions, just because, like, I have this fleeting idea. I haven't even crystallized it into a neat prompt. I just wanna see something. It'll and then I go like, oh, yeah. Delete. As in, revert code back to normal. I feel like, before I would be a little more precious about 75 lines of code, because it would have taken me two hours to do them. Now there's no residual value to any of this stuff, and I could just go like, show me a draft. I feel like a little bit like a king where you just go like, show me the the analysis of the far flung regions. Where are we with the tax recipient? And this boy is like, this, sermon is like, yes. I I sure do so, and we'll return in three weeks. Except, like, you can just wave your hands around, and agents just come back with answers to stupid questions, terrible ideas, then suddenly it wasn't so terrible. It was actually great. I didn't you go like, I did this with, Amache. I haven't even pulled the trigger on it yet. But one of the things with Umachi people have been asking for since the beginning is dual boot. Being able to install Linux next to the Windows installations so that they can still play all their games. And I just went, like, you know what? I have more than one computer, so when I'm play play play games, I can just do it on the PC. It's not a me problem. Yeah. I totally get why a bunch of people want it. I'm not heavily inclined to spend four hours figuring it out. And I just, a little while ago went like, oh, this is exactly the kind of problem, like, I don't have to figure it out. Just make the agents figure it out. So I kicked off initially the process of just coming up with a plan. This is a pretty big change. Right? Like, if you fuck up someone's boot records or you override their petition, criticality high, which is one of the reasons I didn't wanna engage with it. Secondly, it's a little finicky if you want, lux encryption on the Linux partition, but the Linux partition doesn't own the whole drive. It's a little hairy. I didn't wanna take on the criticality. I'm like, this is perfect for the kind of agent stuff. So it started off basically just having Opus and Codex ping pong a plan. Like, I'll just I asked Opus first, like, come up with a plan for this. It it thinks for minutes and minutes and then it says, come up with a good plan. And then I kick it over to Codex and I critique the plan. And then I had him ping pong back and forth a couple times. And at the end looking at the plan going, yep. That's a good plan. We should totally do that. And I can't wait to take that one off and just go, yeah. Now Omachi does dual boot, not because I did it, but, think you're, you're helpful clinkers. That level of ambition is still something I've yet to internalize. Like, even just that, that like, hey, here are these hunches or demands, projects that I would like to do and maybe someday, and you could kick it up on a hunch while you go to lunch. That is a new world, which is also one of the reasons I think a lot of people are thinking, well, the model continues to improve. But even if we somehow hit a wall tomorrow, the better lesson is no longer true. There's actually a limit. It's 19,000,000,000,000 tokens, that's how much they can learn. Not true at all. But if it was, and we had to be stuck with these models, we would spend the next decade just getting more and more out of them, learning how to use these tools. You see this actually with vintage computers. So the kind of games they were able to make on the Commodore 64 when that was released back in '81 to '85, I think, was the main run. I know they made it a little longer, but then the Amiga and other machines came out were great games. I mean, I got interested in games coming up to Commodore 64, year kung fu and all that stuff. The stuff they were able to do twenty years later, when someone had just noodled all the secrets and tweak the one megahertz processor When
[01:13:05 → 01:13:08] 说话人 A when they're building games for the old old Yes.
[01:13:08 → 01:13:28] 说话人 B Yeah. Are so much more technically impressive, because we just know so much more about the I mean, same thing with the you look at the PlayStation, first games come out on launch, last games before we go to PlayStation two, they look from they're like from different generations. We could totally continue to do that with the models, but we're not gonna have that particular enjoyment, because there's a new model dropping in three months.
[01:13:28 → 01:13:55] 说话人 A But this is interesting because if we just run with this thought, like, of course, we know new new things are gonna come. But the point is, like, we will be spending so much time learning, applying them, building either our internal systems, changing how we build things, taking on new project, like, if you're an existing team. Now that people can do more work and more ambitious work, how are you thinking of of the team taking on more work, launching more products? Are you thinking of of potentially growing the team or keeping it as is?
[01:13:55 → 01:14:48] 说话人 B My best assessment for our setup is that the same people can do much more. Mhmm. Let's internalize that, but that's also enough. Already, we were doing enough. Already, we had margin that we could hire way more if we had enough good ideas for that. So all this extra productivity we're getting out of the team allows us now to do things like p one, and all these other projects that are awesome, and they're gonna improve the product faster too, of course they are. The old way of thinking, like it's gonna take two months to deliver a major feature, I mean that's out the door. Of course, there's gonna be rapid acceleration. That's gonna filter all the way into our software methodology process, like shape of was built on two month cycles. That doesn't make sense in the same way at all anymore. We have not fully rewritten those scripts yet because the acceleration is still so fast. No company really has rewritten the scripts on on all that.
[01:14:48 → 01:15:32] 说话人 A When you're shipping that much faster, you need a way to control what goes live and measure whether it's working. This is a good time to mention our presenting sponsor, StatSig, experimentation feature flags for teams that ship fast. StatSig build a unified platform that enables both experimentation and continuous shipping. Built in experimentation means that every rollout automatically becomes a learning opportunity with proper statistical analysis showing you exactly how features impact your metrics. Feature flags let you ship continuously with confidence. And because it's all in one platform with the same product data, teams across your organization can collaborate and make data driven decisions. To learn more, head to statsig.com/pragmatic. With this, let's get back to the shift about to hit developers.
[01:15:33 → 01:17:18] 说话人 B But I still think software developers are delusional if they do not think a shift is coming, where before they were the constraint on how much could be produced and therefore could command We were. The salaries that flow to the constraints. If suddenly those constraints now loosen, especially if we fast forward a little bit where the product manager is actually able to produce changes that could be shipped and work. Things are gonna change. I do actually think, if I was gonna bet, we've seen peak programmer. In terms of the learned guild of programmers who went to either school or spend umpteen hours getting really good at it, we're not gonna need the same number of them to do the same amount of work. Now, Given's paradox, where as the price of something goes down, you get more of it or you get more demand for it, is true, but that doesn't mean that all programs are gonna get bailed out by it. Just because more software than ever is gonna be produced, that's for sure. By the way, I think GitHub has gotten a lot of slack or flack lately. A lot. Justifiably so, I saw a chart saying they had a 92% uptime, which sounds insane. I'm not sure exactly what that was measuring, but I feel it. I have a little bit of sympathy in that. I also think there's some mistakes were made, but also that the amount of software that's currently being produced is on a rocket ship. We are producing as a civilization globally, way more software than we've ever done before. I mean, OpenClaw itself, I thought, he said it was 400,000 lines of code. That used to take ten years and 2,000 people
[01:17:18 → 01:17:18] 说话人 A Yep.
[01:17:19 → 01:17:19] 说话人 B To get to that.
[01:17:20 → 01:17:22] 说话人 A Well, not 2,000 people. In in the but, yes. It it took a long time.
[01:17:22 → 01:18:53] 说话人 B I mean, a long time. Right? Like, you look at, I think, the main monolith at Shopify's 3,000,000 lines of code. That's twenty years, and if you collectively sum up all programs have worked on that, probably, like, twenty thousand people. Yeah. Big shifts are coming right now. Lots of software is being produced. I could see why it's it's creaking a little bit over there because, like, the push is just gonna accelerate. Right? And we haven't even seen anything yet. If you look at AI adoption curves, basically no one's using it. Like, we all in our little bubble in x are like, yeah, everyone's able no they're not. Like, most companies in the world are just not doing it. Notwithstanding that, like, I think, ChatGPT got to 800,000,000 users very quickly. Obviously, there's adoption, but nothing on the scale of what the companies that are furthest along are doing and how much they're accelerating with it. So I do think it is correct for the average programmer to think maybe we've seen the best of the golden days. Certainly, there will be pressures on price because one thing are companies like ours that have essentially unlimited scope to come up with new features and do more, and we can then plow in all that additional productivity into just do more. There's also a lot of companies who just need to do a thing. And if they can do that thing at a tenth of the cost, that's actually their advantage. Right? They just need to do this thing. It's very neatly scoped and defined. It's a cost center. Anywhere where software development is a cost center, which is actually probably the majority of software development in the world, they're gonna face these pressures.
[01:18:53 → 01:20:03] 说话人 A Yeah. Sounds like if I'm a software engineer right now and I'm worried about, like well, you know, like, just wanna make sure that I'm I'm at a place where things are gonna be better. You wanna be at a place where you wanna either get out of a cost center or become really valuable there. Obviously, you know, brush up your skills. And, also, I'm wondering if if the shape of software engineers who will be hired will be changing. Because if if if I just look back from, like, the nineties. Right? Like, even if you look at the movies, you you saw the stereotypes. They were the the nerds who didn't talk to anyone, but they knew how to code. They knew how to do assembly. And then we went in the two thousands. It was still based on languages. And over time, I think, in the February, start up started to not hire for languages, but just hire for algorithms because you could learn the stuff. And now I'm seeing companies, some of the the latest VC funded companies have for product engineers where they they're actually asking for, like, empathy, communication on top of like, it's kind of a given that you you know how to code or whatever. So I wonder, like, if I'm just looking at just just this curve. Right? If I'm just painting it up, like, you're starting to get people oh, and and the developers I I meet at all these companies, they're all really pleasant. They're all just very communicative. Very Yeah. Oh, and they talk with customers, most
[01:20:03 → 01:20:03] 说话人 B of them.
[01:20:03 → 01:20:09] 说话人 A Just it's it's not even a drag. It's like and more and more of them love doing it.
[01:20:09 → 01:21:20] 说话人 B That's the constraint value now. The constraint value is figuring out what should we build, how should it be built, which customer should we be talking to, where should we be focusing. It's product management. It's so funny for me too, because historically I have not necessarily had the highest esteem for product management as a function. I thought there was a lot of bullshit, and I thought it was a lot of people who maybe didn't do as much. Right? And one of the reason was that they couldn't, because the constraint resource was the implementation. Was the product manager could find out that they want to do something. I wanna do this feature, and then they had to wait four weeks for some very expensive programmers to make that reality happen. And in those four weeks, I mean, I guess, they could go talk to some. They were underutilized. Yeah. They were not the constraint. Right? They the constraint was on the implementation. That absolutely is going to switch. Yep. And now pure implementation is going to be solved at some point. I I'm not claiming it is right now, and anyone who is have not tried to just deploy five coded stuff with no review to major code bases. But as the lesson of last summer and Lex, I'm not gonna put my heart on the block and saying that's not gonna happen before next summer.
[01:21:20 → 01:21:41] 说话人 A Again, this is just, like, common sense, but implementation one implementation while we solve for for a general use case, for the edge cases, it will take longer. And for some cases, it will not make sense. Same thing as, I don't know, self driving is fine for, like, these size of cars. But for, like, trucks, it'll either take longer or if you're special, you do. But the point is, like, there will be pockets where but Yes. Those pockets will be smaller.
[01:21:41 → 01:21:52] 说话人 B Yes. I do think the stereotype of I just wanna sit and code, you have to be John Carmack levels of good to retain that privilege. To just, I just wanna sit and code.
[01:21:52 → 01:21:54] 说话人 A And even John Carmack Yeah.
[01:21:54 → 01:21:56] 说话人 B I mean, of course also super AI upheld and
[01:21:57 → 01:22:09] 说话人 A Well, and but but also, like, he he also saw some trends that he could do. Like, for example, like, just like, you know, the type of games that people would buy. Right? Like, he needed to have some business skills or just surrounded by people who did that.
[01:22:09 → 01:22:23] 说话人 B Totally. Totally. Totally. But, like, you you need to literally be the very best. And not just the very best, but you need to be better than the agents. Right? Like for you to get the privilege to just be an implementer, you have to be better than what's available off the shelf from from agents. So who
[01:22:23 → 01:23:10] 说话人 A are the very best? And you're a good person to ask because whenever you advertise of a position, and this was even well before AI, I remember that you you put out a a a job for both software engineer and a designer. And, actually, I want I'm I want to interview your designer who you hired. And because, you published the salary, which is a San Francisco salary. You put the exact number. You you can check it for it. You have a social media presence, so it kind of go goes wide, and you get a lot of applications. And you do a pretty good job. As as I understand, you try to be very fair. You put a lot of effort into it. So what did it take to get hired at thirty Summit Signals? Because now you are trying to hire some of the best. And based off of this, what advice do you give to people who are like, okay. I want to be the best in in this age right now?
[01:23:10 → 01:23:34] 说话人 B Incredibly good question. No one has figured out. We haven't cracked it. And I say that as someone who have run an organization where we we must have looked at tens of thousands of now, of course, if you're running Google, you've looked at millions. But we've looked at tens of thousands of candidates. The number of candidates we've hired is quite small. I mean, total number of programmers that's been through 37 signals over its entire lifespan. What's that gonna be? Like, I don't know, a 100, a 150 at most? I haven't
[01:23:34 → 01:23:36] 说话人 A even How big is your team right now ish?
[01:23:37 → 01:23:43] 说话人 B We're 60 people at the entire company, and we are what's that gonna be? Like, 20 programmers, something like that? Yeah. That's probably about right.
[01:23:43 → 01:23:46] 说话人 A Oh, so so who, what is the other other 40 folks?
[01:23:46 → 01:23:51] 说话人 B We have designers. Mhmm. Probably, like, 10 of those.
[01:23:51 → 01:23:52] 说话人 A Wow.
[01:23:52 → 01:24:31] 说话人 B And then we have customer support, which is at 14. Then we have a bunch of support functions, HR, finance, and then we have operations. Operations is quite large. We have 10 folks managing all our servers, and, yeah, that's about it. But, yeah, I probably it's probably about a 100 people in total that I've worked with, or employed at the company's programmers out of tens of thousands Yeah. We've looked at. And even all those hires did not pan out in the long term. Like, I'd actually say, I think I looked at this recently. Our batting average, at best, I think is slightly better than fifty fifty. So half of even those hires go
[01:24:31 → 01:24:36] 说话人 A through all of because you have a really long and thorough process. Yes. You you you put a lot of effort. Right?
[01:24:36 → 01:25:17] 说话人 B No one has figured out just to hire with such efficiency that they don't make mistakes. There's a great paper that Google published quite a long time ago now, where they tried it all sorts of different hypotheses. Well, can we predict employee outcomes on the basis of Ivy League education background, on GPA, on all of these things. Conclusion was basically, like, we know nothing. We can't predict it on any of these things. We can't predict it on lead code. We can't predict it on any of these, metrics. No. What I'd say is I've clearly been spoiled by working with some very good people, not just at my company, but in open source in general.
[01:25:17 → 01:25:18] 说话人 A Yeah. Oh, yeah.
[01:25:18 → 01:26:21] 说话人 B And therefore, I've ended up with occasionally a twisted perspective of what the average programmer is capable of. And when we do hiring rounds, I am sometimes well, not sometimes. I mean, every time I'm kind of surprised how poor the majority of the submissions are, how little effort is put into being presentable. And that can sound really boomer, crotchety very quickly, but it's also just a reality of trying to get a job. Like, you you gotta stand out, and I understand that that's uncomfortable. Right? Like, who wants to look at this as like, well, my the odds are kind of against me. But it's also a trap to actually fall into thinking of this in terms of odds. Because what I've seen again is people go like, okay, so you have a thousand applicants, there's only one who gets the job, or maybe two who gets the job. So that 0.1% chance. No, it's not. Not at all. With that math, you had 0% chance.
[01:26:21 → 01:26:21] 说话人 A Yes.
[01:26:21 → 01:27:45] 说话人 B Zero. And the very best tire. They probably had a 10% chance, 20%, 30% chance. It is not equal distributed. It is not a lottery. We don't just, like, pick a thing out and be like, oh, it's gonna be this person because they happen to be the one drawn from the bunch. Not at all. We discard off the bat probably at least half the applications, maybe it's two thirds, just because they're either not addressing the job directly. They are not following the instructions in the relatively clear spoken written openings that we have. Right? They're obviously not right for it or whatever, or we get some other smells. Then there's like perhaps a third left, and then we start looking at some of the submissions, then we narrow it down historically to a pool of around 20 people that we give a at home test. The at home test is wonderful. Some people hate it. They feel like it's free labor. I'm like, what the fuck are you talking about? I'm not gonna use your submission to a code test. What? I'm gonna deploy to production? How do you think we came up with that code test? Because it already exists in the system. I say that a little harshly. I also get the sympathy of like, I don't wanna put six hours into making a test if it's not gonna go anywhere. Okay. I get it, but there's no way around it. Because if you have in your head, did you just send in a resume? Someone's gonna call you up on the phone, have a thirty minute conversation with you and go, you're hired, sir. I don't know if that ever existed, but certainly does not exist today. It never existed in the lifetime that I've been in this
[01:27:45 → 01:27:50] 说话人 A business. Only time it exists, right, is through a very warm referral where
[01:27:50 → 01:27:50] 说话人 B Correct.
[01:27:50 → 01:27:52] 说话人 A Where you're starting a If you're
[01:27:52 → 01:27:54] 说话人 B skipping if you're skipping the whole pipeline.
[01:27:54 → 01:28:12] 说话人 A And when you skip the whole pipe, and it typically only happens at the very beginning of a company when you're founding a company. And often it goes both ways where it's very risky, and then you say, like, this buddy of mine, I worked with this person for two years straight. I would, like, trust him with my eyes closed. So that's actually the black pill on the whole hiring process. If we look at
[01:28:12 → 01:30:03] 说话人 B the long term success rates, we have had more long term employees from I've worked with this person for two years, we should hire them, than we have from the open calls. It is actually exceptionally difficult. It has been for us to find the kind of programmer who thrives in our environment from open call. It has happened. We have hired people that way, and I continue to want to believe, even if the odds seem insanely long, when you start doing the math of like, oh my god, we've looked at tens of thousands, and how many then got hired, and how many then didn't work out. Like, Jesus, there's only like a handful left. I'm starting that. That that's kind of blackmailing. But then hiring directly on the base of a warm referral, I should call it, has worked very well. And that the hit rate there is really high. But how does that help anyone? Right? Like, that's not very actionable advice, except that's to say, get as good as you can get, and put in as much effort as you can, and work with someone. Because I wanna get safe as a counter. Some people have this notion in their head that if they work at a place they consider shitty, they shouldn't try. You're shooting your own feet, buddy. If you show up at the shitty place of work, and we can even be objectively in unison about that, that it is a shitty place of work. And you then go like, well, I should just try to skirt. I should just try to goof off. I should just try to read x or read it all day. Right? Everyone else you work with, they're gonna watch that. You know where that warm referral is gonna come from? It's gonna come from someone who worked with someone else at a shitty job, but identified that that individual still showed up and did as best as they could to learn, to ship, to do all of this stuff. There is no shortcut here. You'd simply just have to be good, and you will not get good if you do not practice. And if you think your place of employment is not worthy of your best, you're cheating yourself.
[01:30:03 → 01:30:11] 说话人 A If you're not helping, even if it's a shitty place, if you're not helping that place get better, why would a great place hire you who only hires people to to further raise the bar?
[01:30:11 → 01:30:39] 说话人 B This is total cope. And it's cope both on the side of I work at a shitty place and therefore I don't wanna put things in. You could be annoyed. I'm not telling you have to love your boss. I'd actually say the majority of people I used to work for, I didn't have the warmest feelings about them. I still tried really hard for my own edification, for my own education, for my own sense of I'm the kind of person who shows up and does a good job. Just that I will be ready Yeah. When the opportunity arrives, when all my talents are needed and all my skills are honed. Right?
[01:30:39 → 01:30:50] 说话人 A But but was this not how you ended up at thirty seven signals where it was just a contract job or something and, you know, like, on on a contract job, you have no ownership and Correct. But you showed up and Correct.
[01:30:50 → 01:31:27] 说话人 B And Jason ended up realizing, okay, this, punk better get some equity. Otherwise, he's out the door. Now that's a seminal story, and you shouldn't extrapolate everything from that. I mean, all founder stories, by the way, are similar stories in that regard. But the fundamental principle is still the same. Show up, do as good as you can, learn more. There also was to my chagrin to some extent, I perhaps contributed it to it a bit for a while, which was this notion that you can be a great programmer and not really like programming. That you don't have to ever care about programming outside working hours.
[01:31:27 → 01:31:30] 说话人 A Was was this what you thought of or like
[01:31:30 → 01:32:36] 说话人 B Well, I thought of it mistakenly because I was pushing back on the overwork hundred hour week, hundred and twenty hour week, maniacal obsession, which by the way, never was my experience. We did not start Basecamp that way. We have worked on a forty hour week rolling average over those twenty five years. But also, as I said at the very beginning, I really like computers. So I play with computers in my free time. I look at computer things in my free time. It's not work in the sense that I'm whatever shipping features to Basecamp customers, like, just twenty four seven, that's not what it is. But I am playing with computers. I am looking at new things. I am exploring new systems and whatever. And I think there was, for a while into twenty tens, a misconception that you didn't have to do any of those things. You could just show up and do your work, and you would be so sought after, because programming was such a valuable activity, and there were so few people who could do it that they'd take anyone. Even people who barely gave a shit. And I think that's over, if it ever was true. And I think it was true.
[01:32:36 → 01:32:41] 说话人 A The boot camps were the perfect, like, catalyst or or, like, they were the canary when when
[01:32:41 → 01:32:49] 说话人 B Which also, by the way, is how the economy is supposed to function. Yep. When sellers are really high, it means that there's not enough supply of labor. Therefore, we should get labor into the
[01:32:49 → 01:32:50] 说话人 A pool. Exactly.
[01:32:51 → 01:32:56] 说话人 B And so I'm not I don't even have any qualms about Internet. I'm just saying, like, that's over.
[01:32:56 → 01:33:42] 说话人 A No. I I think looking but, you know, we're talking about, like, is it is the golden age of the programmer heavy past peak programmer. And I wonder if peak programmer really meant that almost anyone who wanted to get into the industry and was willing to put in some effort, few months, or maybe a few years could do it. You could learn how to code. You could go to either college or to boot camp or put in the hours, and you could get hired at a place because the interviews were their references were not needed. We we didn't check. And I that's probably coming to an end. You do need references. You more I think more and more companies will be doing reference checks as part of our thing, and it's not just gonna be a heavy work there. Like, would you I've I've had these calls from, like Databricks is is famous for reference cards. They don't normally check for references. They drill you not just would you work with this person again? What were their weaknesses?
[01:33:42 → 01:33:43] 说话人 B Right.
[01:33:43 → 01:33:45] 说话人 A Where would you hire them? Etcetera etcetera. And No.
[01:33:45 → 01:35:37] 说话人 B I understand it. The weird thing is Deepgrammer sounds like this is something that affects all programmers. It does not. The best programmers, or not even the best, as in like it's 10 people around the world, really good programmers are currently more valuable than ever, because they are the ones who are able to get the most out of the AI acceleration. And this was the kicker for me in changing my perspective on this, is that I've also found, and maybe it's not universally true, but certainly within 37 signals in my own experience, I'm enjoying my time as a programmer more than any time since early two thousands when I just discovered Ruby. This has the I just discovered Ruby feel to it, that it is so satisfying to be able to move this fast on so many levels at the same time, to be able to explore the p ones, to be able to think about dual booting Amache, to do all of that stuff that the work itself has gotten vastly more enjoyable, and I've seen the same thing for the most AI forward programmers that we have. Maybe they also have some of these anxieties, but they're kind of pushed to decide just out of sheer enjoyment working with the new capacities. So there is a bifurcation here where we should all feel like, well, we don't know what's going on, and for some people that's gonna produce some degree of anxiety. I understand that, especially when it's your livelihood and you're like, well, I also like to be able to pay for my kids college in seven years. What does that look like? I get it. You're not gonna be able to manifest that anxiety into anything productive unless you just plow it into leaning in. Right? Because if you just sit and spin around, try to think about what the world's gonna look like seven years from now, you're wasting your time. So, that's the only path. The only path is to either get excited about this, which I don't think it takes that much effort. As we said, if you sit down with these models, you pull out one of your hobby projects from the closet
[01:35:37 → 01:35:37] 说话人 A That you never finished.
[01:35:37 → 01:35:46] 说话人 B That you never finished, and you just give it a try. I don't see how you really like computers and not find that experiment enjoyable.
[01:35:46 → 01:36:13] 说话人 A And I I've seen this with with people who are getting into it. Kent Beck is such a great example. He's been programmer fifty two years and he is saying, like, he he loves doing it and he found this balance between using the agents to build something ambitious that he always wanted to he's building a small talk server, which which used to take forever, and now it's it's getting closer, and it's still taking a long time. And then in between, he's chilling at his he has his house on on the lake, and he just goes and, like, just looks at the birds for two hours and then gets back to it. It's
[01:36:14 → 01:37:32] 说话人 B beautiful. Kent is, by the way, one of my all time heroes. This was right when I got started in programming. Right when I before I was picking up, Ruby, I saw Kent speak at a Danish conference in 2001 on stage, and I was completely mesmerized by his command of both the material, how bold he was, and how great of a speaker he was. And this was after having read Extreme Programming and many of these other things. Small Talk Best Practices is my number one recommendation for any programmer who wanna learn the nitty gritty of how to structure a method and a class and the rest of it. Small Talk Best Practices, which is Kent's book from '95, I think, or '96, is to this day my favorite book of all time on tactical programming, patterns. So it's wonderful to hear him being agent Pilt, while also enjoying the birds. I mean, I try to do that too and this is actually there's a bit of attention right now, is that most of the people I find who are all in, they're working harder than they ever have. And I've seen that with myself now too. When you can be this effective and impactful on an hour of supervision of these agents, it's really intoxicating. If you have an active, dopamine loop up there that gets triggered when something is shipped, it is just hyperactive right now.
[01:37:32 → 01:37:33] 说话人 A Yeah.
[01:37:33 → 01:38:02] 说话人 B And I need to go, do you know what? This is not like a limited sale. Like, AI is gonna be here next month and the months after that. Like, I cannot just operate as though it is a limited sale, and I need to get all the dopamine harvested within the next two weeks. That I actually think is the main challenge right now for the people who are furthest along and most pilled on it is, like, remember that this is as bad as they're ever gonna be, as the cliche goes. Right? You damn well better find a way not to get consumed entirely about it as exciting as it is.
[01:38:02 → 01:38:27] 说话人 A And and, yeah, there there's this consuming is is is a big deal. Like, it was Steve Yaghi. He was he looks a bit more drained than like, you can see it on on the video, but he he has he's honest. Like, he's he's being pulled into this. He's doing he has friends who are and when when you're on the edge, you're there. You've clearly been AI pilled, but how are you finding of keeping a balance of, like, alright, stepping away, you know, like, I I know you've I think you previously talked about the importance of sleep. Apparently, you don't have an alarm?
[01:38:27 → 01:39:58] 说话人 B Correct. I don't use an alarm. Although, my wife now does because the kids need to go to school on a regular basis. But, yeah, for me, eight hours a night is the best investment you can make in your own cognitive capacity. So I just am reminded every single time I do not get eight hours that it is such a poor trade. If you go from the eight to the six, I go like, well, I'm gonna be awake for, in that case, eighteen hours. What is the drag I'm gonna carry for all those eighteen hours for getting one more hour, two more hours by cutting back on the sleep? It is such a bad piece of math. It makes no sense. Now, occasionally, it's involuntary. I have actually had, especially around this AI stuff, I've had a couple of times, very rare. I can count on two hands the number of times where I've been sleepless. Like the ray the brain racing a little too much. Right. That's not typical for me, and it's still not typical. But I have had a couple of them. Right? So I get where some of that excitement comes from, but I'd also say the last thing you should trade is sleep, and then you should not trade your health. You should not try to save the three hours a week of working out to do more agent work. That's a very poor trade. Keep in good condition. Like, there's nothing that's gonna be more important if you wanna keep, like, sharp up there that, like, the rest of the system is operating, if not a peak capacities, then at, at a good sustainable level. Right? And I do think there are some individuals right now who are at fear of running ragged
[01:39:58 → 01:39:59] 说话人 A Yep.
[01:39:59 → 01:40:28] 说话人 B On something that we're gonna be dealing with for, like, slow down buddy. Like, it's not again, limited sale. Next ten years, we're gonna see more and more. It's gonna get crazy and crazier. So don't squander your health. Don't squander your sleep. Don't squander your diet in the service of anything, because even on the short term, it does not work. You cannot get more productive within three weeks, let's say, by trying to cut back two or three hours of sleep every night, and then think there's anything coherent left after three weeks. You will be a hot mess.
[01:40:29 → 01:40:57] 说话人 A So let's close. We talked about the stuff that we don't know. A lot of things we don't know. But let's close with what you do know. So you you could have retired a long time ago and just, you know, kick back and and, like, listen to birds. What is it that keeps you doing, keeps you building, keeping getting up every day? And before AI, you would open your terminal. I think you you shared, like like, you would go and then and right. Now you're doing with agents. Like, what drives you? And and and looking ahead, like, what what are things you're excited about?
[01:40:57 → 01:41:34] 说话人 B My drive continues to be a deep love of computers. This is simply the best way, the most fun way to spend my time. I could spend my time on a lot of things. I do spend my time on a lot of things. I don't just do computers. I drive race cars, I take lots of time up, I have three kids, we enjoy all of that stuff, but if I'm gonna fill eight hours every day with an activity, my best bet is computers. And it has been so since I was literally five years old, whether it's video games or what now feels a little bit like a video game actually, instrumenting all these agents and, playing a little bit of StarCraft with, moving them around and
[01:41:35 → 01:41:35] 说话人 A Explorio.
[01:41:36 → 01:44:12] 说话人 B Yes. Exactly. So I just really like computers. So whether I need to do so for economic reasons or not, I will continue to play with computers, see what makes them tick, and make things. I think that's the other big misconception that some people have about wealth is that they conceive of it as some sort of checkpoint. Like once you made it, then you can just kick back in leisure as though that was happiness. We simply have a hundred years of psychological studies telling us, no, that's misery. If you have all the time in the world and no purpose, no mission, leisure's not gonna cut it. It's not gonna be fulfilling way. And this should be obvious, by example, of literally every entrepreneur who sells their business. They sit on the beach for three weeks and then they're back into the game. Right? Because this is actually not just something they do in pursuit of a goal. It's the goal itself. It is the mission itself. It is the satisfaction. It is the affirmation of being a human that I'm not just a blob laying around. I am a useful individual who put my skills to the best use possible. So I'm gonna continue to do that, and I'm gonna continue to do it whether I'm sitting, typing at the keyboard, whether I'm instrumenting these agents, whether they're teaching me, however which way it is, I wanna play with computers, I want to intend to do that. And then even more specifically, after the last three months, I'm leaning in hard now with agent accessibility. For example, this is what I've been doing last few weeks. We've been working on the new CLI, which also taught me, like, we're not quite an AGI yet. Right? You think, like, well, just ask your agent to make it CLI. It will, but like it's not quite there. Right? Like, I want it to be just right, and the agents still need a little bit of help. I'm very happy to provide that help to these agents, and we'll release a great CLI for Basecamp very very shortly. Maybe by the time this is out, it'll probably be out. And for the rest of them too, and I want to lean into all of this. How can we use this as much as we possibly can? And then, right now, I'm also just an incredibly curious person. I wake up every morning, I have a new ritual, which is not to pull my phone up and start hopping on x, like, right when I wake up. I don't think actually that is great, but it takes a tremendous will power to not do so because I'm just so curious about what happened. There's so much happening right now. I wanna know. I wanna know. I wanna be enjoying it, be a part of it. So I don't foresee that ending. I don't foresee a love of computers evaporating. In fact, if anything, right now, I'm seeing like a a flourishing of it. I'm liking computers more than I did five years ago, and that's amazing.
[01:44:13 → 01:45:29] 说话人 A Amazing. David, this was awesome. Thanks thanks a bunch. Alright. Thanks for having me. This was really great. This was a fascinating conversation, and I love the energy that David has. I hope some of this energy that is obvious in person also came across to you. I really appreciated that David was open that his stance did not change about AI because his philosophy changed. It's just that the tools became good enough to do useful stuff. AI for autocomplete was annoying for experienced developers. AI agents that can produce pretty good working code by themselves, on the other hand, are now pretty useful. And yet David kept coming back to taste, judgment and craft. He wasn't just saying just let the model write whatever. It's the opposite. He has a very high quality bar and he wants the output to be code that he would actually be proud to merge. It feels like AI might make good judgment even more valuable than before. I also really liked how David thinks about the importance of design. At 37signals, designers help figure out what should be built, how it should work, and increasingly even decide how it gets implemented. I wonder if 37signals is a step up of the industry in thinking about designers a bit like developers as well and developers a bit like designers as well. Finally, I found David's take that we might have hit peak software engineer an interesting argument. David thinks we'll produce more software than ever, but
[01:45:29 → 01:45:29] 说话人 B his observation is that
[01:45:29 → 01:45:29] 说话人 A we might be nearing
[01:45:29 → 01:45:29] 说话人 B the end of the
[01:45:29 → 01:46:04] 说话人 A time when developers could command high but this will mean software engineers who are not only good at coding or using AI. But this will mean software engineers who are not only good at coding or using AI to generate code, but can oversee building complex systems, have tastes and business sense as well. If you'd like to hear more from David, check out a bonus episode with him linked in the show notes. Also, check out the show notes for related to the Pridemantic Engine deep dives on software craftsmanship and practical ways of building software.
[01:46:04 → 01:46:04] 说话人 B If you
[01:46:04 → 01:46:13] 说话人 A enjoy this podcast, please do subscribe on your favorite podcast platform and on YouTube. And a special thank you if you also leave a rating on the show. Thanks and see you in the next one.
由 PodMate 自动转写 (Deepgram nova-2)