DHH’s new way of writing code

• URL: https://www.youtube.com/watch?v=JiWgKRgdgpI
• Channel: The Pragmatic Engineer
• Fetched: 2026-08-28T13:31:08+00:00
• Language: English (auto-generated)

Transcript

[00:00] I feel that you very much value software
[00:03] engineering as a craft
[00:04] >> hugely. I mean I think aesthetics is
[00:07] truth. When something is beautiful, it’s
[00:09] likely to be correct. I think this is
[00:11] true in mathematics. This is true in
[00:12] physics. This is true in a lot of
[00:14] different domains.
[00:15] >> I wonder if there’s a part of AI about
[00:17] the impact of doing work that we would
[00:18] have not done before.
[00:20] >> The number of projects we have tackled
[00:21] internally that we would never even have
[00:23] contemplated starting on or Legion.
[00:25] Jeremy, one of our most agent
[00:27] accelerated people, went like, “We’re
[00:29] going to do P1. We’re going to optimize
[00:30] P1.” Literally the fastest 1% of
[00:33] requests, we’re going to make them even
[00:35] faster. There’s a bit of tension right
[00:36] now is that most of the people I find
[00:38] who are all in, they’re working harder
[00:39] than they ever have. And I’ve seen that
[00:41] with myself now, too. When you can be
[00:44] this effective and impactful on an hour
[00:46] of supervision of these agents, it’s
[00:48] really intoxicating. And I need to go,
[00:51] do you know what? This is not like a
[00:52] limited sale. on Lex Freriedman’s
[00:54] podcast. You were still rightfully so
[00:56] very skeptical of AI.
[00:57] >> This is a nuance point and maybe it’s
[00:59] self- serving, but I don’t actually
[01:00] think my opinions have changed. What
[01:02] have changed is
[01:05] [music]
[01:07] how has the creator of Ruby on Rails
[01:08] changed how he builds software now with
[01:10] AI [music] agents? David Hayam Hire
[01:12] Hansen often referred to as DHH created
[01:14] Ruby on Rails Omachi and is a co-founder
[01:16] of 37 signals. He bashed capabilities of
[01:19] AI coding tools on Lex Friezen’s podcast
[01:21] 6 months ago. Then over the course of a
[01:23] [music] few weeks over the winter break,
[01:25] he did a 180 turn and went AI first on
[01:27] everything. In today’s conversation, we
[01:29] cover how David and his team at 37
[01:31] Signals build software today and how AI
[01:34] tools are making them more ambitious
[01:35] than ever [music] before. Why Ruby on
[01:37] Rails Analytics could become even more
[01:38] popular than they are today as they are
[01:40] both well suited working with AI agents.
[01:42] why taste and beautiful software are
[01:44] becoming more important and why [music]
[01:45] both standout designers and engineers
[01:47] who care about the craft could become
[01:49] more in demand and many more. If you’re
[01:51] interested in what one of the most
[01:52] experienced builders in the tech
[01:54] industry thinks [music] about the
[01:55] practical utility of AI tools and how
[01:57] these tools could impact software
[01:58] engineers who care about the craft then
[02:00] this episode is for you. This episode is
[02:02] presented by static the unified platform
[02:04] for flags analytics experiments and
[02:05] more. Check out the show notes to learn
[02:07] more about them and our other season
[02:09] sponsors sonar and works. David, it’s
[02:12] awesome to have you here.
[02:13] >> Thanks for having me. Thanks for coming.
[02:15] I should actually say you’re in
[02:16] Copenhagen. That’s my city of choice at
[02:19] the moment. It’s a beautiful city. It’s
[02:21] got so much going for it. And so, what
[02:23] have you been up to?
[02:24] >> I’m always building stuff. I have been
[02:26] building stuff for a good damn three
[02:29] decades now on the internet. I got
[02:31] started back in 94, I think it was, when
[02:34] I first got exposed to it and basically
[02:36] just never stopped. And in the past six
[02:38] months, I’ve been building a variety of
[02:41] things. One of them is a new Linux
[02:44] distribution called Umachi.
[02:46] I switched to Linux about a little over
[02:49] two years ago, I think, now. First spent
[02:51] some time on Ubuntu, having fun with
[02:53] that, and then realizing I actually
[02:55] wanted to make my own system from
[02:56] scratch, building it on top of Arch and
[02:59] Hyperland. So, put a lot of time into
[03:01] Amachi. It got started as a summer
[03:03] project in between racing at the 24
[03:05] hours of Lama. there’s a lot of downtime
[03:07] in that week. So, I just started hacking
[03:10] on it and it really took off very
[03:12] quickly thereafter. It’s been a truly
[03:15] inspiring ride to see that even in a
[03:18] market as crowded as Linux
[03:20] distributions, there’s about 7,000
[03:22] different distributions out there, some
[03:24] of them with long pedigrees and many of
[03:27] them even based on sort of kind of
[03:29] similar vibes to some extent. There’s
[03:32] room for something new and it’s a great
[03:34] reminder that all the ideas in the world
[03:37] may be taken and it doesn’t matter
[03:38] because your spin on it isn’t. And I put
[03:41] my spin on Linux. Billachi built the
[03:44] perfect computer system for me and saw
[03:48] exactly the same thing I’ve ever seen.
[03:50] And whenever I build something that
[03:52] really just hits the spot for me
[03:54] personally, there are thousands of
[03:57] others just like me or close enough to
[03:59] what I like that they find the same
[04:01] pleasure and joy in it. Whether it was
[04:03] Ruby on Rails, Kamal, getting out of the
[04:05] cloud, any of these things, it’s the
[04:06] same syndrome. Yeah. With with Rails,
[04:09] you were literally scratching your own
[04:10] itch. You were just building your own
[04:12] components and then open sourcing them.
[04:14] Is that how it started? Basically, I
[04:16] picked up Ruby in the early 2000s and
[04:20] really put it to the test in 2003 when
[04:23] we started building Base Camp and I did
[04:26] not have a mandate of what to use to
[04:28] build it. Prior to that, I’d been
[04:30] working for a lot of client projects
[04:32] that would say, well, we’re building
[04:34] this in PHP because we have someone who
[04:37] knows that. So, this is what you have to
[04:39] use. And then we were building our own
[04:40] system. We’re building Basec Camp. And I
[04:42] was free to choose. So I chose Ruby and
[04:46] at the time Ruby didn’t have any tooling
[04:49] or ve not very much when it came to web
[04:52] applications. So I had to build it all
[04:53] myself and that turned into Ruby on
[04:55] Rails which is still going strong. I’m
[04:57] still very heavily involved with that. I
[04:59] think in some ways Ruby Rails is having
[05:01] a little bit of a renaissance now that
[05:04] it is one of the most token efficient
[05:06] ways of building web apps. It’s ideally
[05:09] suited for the agent workflows we’re
[05:11] dealing with now. We’ll see how long
[05:13] that lasts. Maybe all the agents are
[05:15] going to be writing machine code or
[05:16] assembler in about five minutes. So
[05:18] maybe that comes to an end. But for the
[05:20] moment, token efficiency still matters.
[05:22] And it still matters whether the agents
[05:24] produce code that humans are able to
[05:27] read and verify. That may also come to
[05:29] an end at some point. But as it is right
[05:32] now, it’s uh been a fun ride to just see
[05:35] these kinds of projects where I’m
[05:37] scratching my own itch resonate with a
[05:39] much larger community of people who then
[05:42] show up and want to help. I mean for
[05:44] Umachi which has only been around for
[05:46] what is that just over 6 months now we
[05:48] have what 400 contributors who’ve made
[05:51] code changes to the distribution and on
[05:54] top of that we have tens of thousands of
[05:56] people who’ve installed it and uses as
[05:57] their daily driver. So, I always love
[06:00] that discovery of something new, novel,
[06:04] and inspiring like Ruby or it sounds
[06:07] weird to talk about discovery of a
[06:09] operating system that’s been around
[06:11] since what 91, but for a lot of people,
[06:13] Linux now is that discovery because they
[06:16] have not been using it on their personal
[06:18] computer. So, they’re seeing it for the
[06:19] first time. And for me to help a new
[06:22] cohort of Linux users and hopefully even
[06:26] enthusiasts come to be because I’m
[06:29] flattening the curve a little bit. I’m
[06:30] making it easier to get started. I’m
[06:33] making the default installation just
[06:35] look amazing so that they don’t feel
[06:37] like they have to invest 100 hours into
[06:40] tweaking the system to get going is
[06:43] really fun. But what’s also fun of
[06:45] course is that both of these things,
[06:47] both Ruby and Rails and Amachi were not
[06:50] just hobby projects. I love hobby
[06:52] product and I will always do those, but
[06:54] I also like to apply them to business.
[06:55] So at 37 Signals, we built an entire
[06:58] business for 20 plus years on top of
[07:01] Ruby and Rails. We’re now running Linux
[07:04] on the majority of developer machines
[07:06] because we now have our own distro.
[07:09] >> So it’s obi
[07:11] people can choose, right? can they
[07:13] >> well sort of kind of we started with a
[07:15] with an open choice and then at some
[07:17] point it just doesn’t make sense anymore
[07:19] in the same way it would not make sense
[07:21] for someone to be at 37 single and say I
[07:23] want to write this thing in Django we’re
[07:25] going to use Python and this other
[07:26] framework even if you have Ruby and
[07:28] Rails and you’re doing that so we
[07:30] pivoted from an early invitation to play
[07:33] around that was what when I first
[07:34] switched to Linux just said like hey if
[07:36] you want to check it out check it out
[07:37] then when things got a little more
[07:39] serious with Amachi I just said let’s go
[07:42] all in for everyone who’s on the
[07:44] technical side of things, not the iOS
[07:46] developers of course, but anyone who’s
[07:48] working with the web, who’s working with
[07:49] Ruby, who’s doing DevOps, they should be
[07:53] on Linux because first of all, that’s
[07:56] closer to what we deploy. We’ve always
[07:58] deployed on Linux. We’ve been a Linux
[07:59] shop on the server side since day one.
[08:02] For developers and system operators, I
[08:06] actually think it is a material
[08:07] advantage to be closer to your
[08:08] production environment and just be more
[08:10] familiar with the tools. Then on top of
[08:12] that, of course, we are building this
[08:14] distribution and we should have as many
[08:16] hands help out as possible. And given
[08:18] the fact that I’m the CTO of this
[08:20] company, I get to set the technical
[08:22] direction and this is the direction
[08:23] we’re going. Can you just like do a like
[08:26] just a very short recap of of you know
[08:28] like how you right and right now where
[08:30] are you like where where is the business
[08:31] as a whole and you know you keep you
[08:33] keep building you keep launching new and
[08:35] exciting and just cool stuff. I think
[08:36] Fizzy was the latest one.
[08:38] >> Yes. So 37 signals was founded in 1999.
[08:41] It started as a web design firm and then
[08:45] I joined up in 2001, two years after and
[08:49] for a couple years, collaborated with
[08:51] Jason on these consulting projects and
[08:53] then it was in 2003. We started work on
[08:56] Base Camp, released it in 2004.
[08:58] Actually, either the day after or the
[09:00] day before Facebook went live, which is
[09:03] kind of a funny coincidence that we were
[09:06] of that same time and cohort. And within
[09:09] about a year, we realized this thing was
[09:11] taking off and we went full-time and
[09:14] switched from being a consultancy to
[09:15] being a software company.
[09:16] >> Awesome.
[09:17] >> And that’s now 22 years ago, a little
[09:20] more than that. And in that time, we’ve
[09:24] released a ton of products. Basec camp
[09:26] was the first. Remains the biggest and
[09:30] most important, which is also kind of
[09:32] funny because you sometimes perhaps have
[09:35] this delusion that as you learn more and
[09:38] as you get more experience, you’ll get
[09:40] smarter and you’ll have better ideas.
[09:41] And like, no, there’s tons of people for
[09:44] whom their first idea was the best idea.
[09:47] And I have no shame in saying that Base
[09:49] Camp was the best idea objectively in
[09:52] terms of a business that we’ve ever had.
[09:54] And I’m incredibly proud that we’ve been
[09:56] able to keep that going and growing and
[09:59] flourishing for over 20 years. Very few
[10:02] software companies, let alone software
[10:04] products, can boast of that longevity
[10:07] and legacy. But we’ve tried a ton of
[10:09] things over those years and had some
[10:11] other great successes. We launched
[10:12] hey.com our email service back in 2020
[10:16] which was a crazy mission when you think
[10:18] about it.
[10:19] >> Here is a sector completely dominated by
[10:22] a single player Google with Gmail that’s
[10:24] a good product.
[10:26] >> It hasn’t really changed in 17 years but
[10:28] it was really solid and lots of people
[10:30] are perfectly content with it. They
[10:33] think they hold this duality in their
[10:35] head where at once they both hate email
[10:38] but somehow don’t connect it to the fact
[10:40] that they’re using Gmail which I find
[10:42] curious but either way we launched this
[10:45] that is not only a competitor to this
[10:47] very entrenched product that has
[10:50] probably a greater grasp on market share
[10:54] in any major category than any other
[10:56] product I can come to mind of in the US
[10:59] I think Gmail is something like 85% of
[11:01] all email traffic, which sounds insane.
[11:03] Maybe it’s 80%. It’s incredibly high.
[11:06] It’s basically Gmail and then all the
[11:10] rest is in this tiny little part of the
[11:12] graph. So, we thought that after using
[11:16] Gmail, I used it since I don’t know when
[11:19] I signed up, a few weeks into it, I got
[11:20] one of those invite codes. That was a
[11:22] really clever launch and I used it ever
[11:23] since. So, that’s literally 17 years or
[11:26] something of that of Gmail usage. And
[11:27] over that time, I built up a lot of
[11:29] opinions about things that didn’t work
[11:30] quite like I would prefer it to work.
[11:32] And we put all those opinions into a new
[11:34] software product. Spend about almost 2
[11:36] years developing it. Millions of dollars
[11:38] in accumulative R&D funds. And launched
[11:41] it in the summer of 2020, which by the
[11:44] way, time to launch a product.
[11:47] 2020 wasn’t great for a whole host of
[11:50] different reasons. We were kind of
[11:51] trying to slot in a can there just be a
[11:54] week where the whole world is not just
[11:56] insane.
[11:57] >> Yeah.
[11:57] >> We finally picked a week. We went live
[11:59] and then we had the battle of our lives
[12:02] with Apple.
[12:02] >> With Apple. I remember that.
[12:04] >> And ultimately
[12:06] >> they didn’t want to approve your your
[12:07] app.
[12:07] >> They didn’t want to approve our app
[12:08] unless we paid the toll fee, the 30%.
[12:12] >> And they were basically willing to say
[12:14] you can’t be in the app store, which for
[12:16] an email product like that is a death
[12:18] sentence.
[12:18] >> Yes. you have to be on not just mobile
[12:21] phones but specifically the iPhone. This
[12:24] is true today. The majority of hey
[12:27] paying customers are iPhone users
[12:29] because that’s the largest most affluent
[12:32] market in the US and the US is the most
[12:34] affluent and market software market in
[12:36] the world. So for that business to work
[12:38] we needed to be on the iPhone. After a
[12:41] twow weekek epic struggle back and
[12:44] forth, thankfully time to perfection
[12:47] with WWDC where Apple preferably didn’t
[12:51] want to look like the Goliath squashing
[12:54] a
[12:55] >> developer, tiny developer, we ended up
[12:58] being allowed in and Apple sort of
[12:59] rewrote the rules after the fact to make
[13:02] it fit. Um, it was a small victory, not
[13:05] the ultimate victory, but at least it
[13:07] allowed us to to be there. And hey,
[13:10] ended up being an enormous success. In
[13:12] part, ironically, because Apple gave us
[13:14] wall-to-wall coverage for two weeks.
[13:16] When I look back upon that, I think I
[13:19] wouldn’t have gambled like that because
[13:20] the outcome would have been zero, right?
[13:23] Like, uh, Apple refuses our app.
[13:25] >> We sign up 200 people and the app is
[13:29] dead. What instead happened was they
[13:31] gave us a multi-million dollar launch
[13:34] campaign and coverage in all major media
[13:36] and we signed up tens of thousands of
[13:38] people in those first weeks. That was uh
[13:41] an insane event but uh also very
[13:44] satisfying. And the other satisfying
[13:46] thing was I just love Hey. I use it
[13:49] every day. I basically use base camp in
[13:52] terms of web applications. That’s where
[13:54] we do all our collaborative work. And
[13:55] then my number two app and many days
[13:57] it’s my number one app is hey because I
[13:59] just do all my stuff in email. I am
[14:01] constantly communicating with people.
[14:03] I’m writing. I’m doing a lot of stuff in
[14:05] email as many people do. And having that
[14:08] be a pleasurable experience and a nice
[14:11] environment and my inbox being a little
[14:14] more sacred than what happens with Gmail
[14:16] where total strangers around the world
[14:19] can just make your pocket buzz if you
[14:20] have notifications turned on which they
[14:22] are by default. Just seems insane to me.
[14:24] Right. this idea that there’s direct
[14:26] access to one of my most important daily
[14:29] priority lists like anyone can put
[14:31] something on that insane. Anyway, hey
[14:34] doesn’t do that. We have the screener
[14:35] and no one gets to reach your inbox
[14:37] before you’ve said I want to hear from
[14:39] this person. And most of the time I say
[14:41] no to most people, right? Like things
[14:42] end up in the in the screener and we
[14:43] have thumbs up. I will hear from this
[14:45] person, thumbs down, I’ll never hear
[14:47] from that person again.
[14:48] >> This this is how I reached out. I mean,
[14:50] we were I’m not sure we were connected
[14:52] on on X, but I I sent an email cuz your
[14:54] email is out there and your screener
[14:56] seems to have worked cuz it gave me the
[14:57] thumbs up.
[14:58] >> It did because the screener is me. So,
[15:00] there’s not even AI trying to sus out
[15:03] whether I want to hear from you or not.
[15:05] Because what turns out to be true is
[15:07] it’s actually not that ownorous to once
[15:08] a day go through your screener and say
[15:11] thumbs up or down because there aren’t
[15:13] that many people in the world. And if
[15:15] you say no to the annoying pestering
[15:18] salespeople who within Gmail managed to
[15:21] read your inbox seven times, then the
[15:24] workload is much less. And it’s very
[15:26] satisfying, I will say too, because when
[15:28] I was using Gmail, I would get roped
[15:30] into this sales tactic that they of
[15:32] course rely on, which is that like you
[15:34] write back and say like, “No, thank you.
[15:35] I’m not interested.” And then they would
[15:37] respond again. And now you feel like,
[15:38] “Wait, am I now obligated to respond to
[15:40] this person? I kind of feel like I am.”
[15:42] And occasionally I would end up writing
[15:44] and even if I wouldn’t write, they still
[15:46] have access to my inbox. So I would hear
[15:48] from them again next week. They have a
[15:49] whole drip campaign. They all [ __ ]
[15:50] do, right? That any outreach is seven
[15:53] emails. It’s not one emails. It’s seven
[15:55] emails. And if you show any sign of
[15:56] life, it’s probably 52. That’s just not
[15:59] how it works. And hey, I say thumbs down
[16:01] one time, never hear from that person
[16:02] again. It’s actually amazing how quickly
[16:04] you can curate your garden from that
[16:06] weed. And then suddenly there’s just
[16:08] beautiful flowers. Suddenly email is not
[16:10] a chore. So you want to go smell the
[16:12] roses. Suddenly the majority of things
[16:13] that end up in my email or things I want
[16:15] to read is from people I want to hear
[16:17] from. And that was really the
[16:19] fundamental mission for us with hey can
[16:21] we make email lovable again? Email is so
[16:24] hated by so many people because the
[16:27] systems are so poor because they’re
[16:29] based on the original premise that email
[16:32] is just what universities use for
[16:34] scientists to talk to each other and
[16:36] scientists have really good manners and
[16:37] will not pester you 52 times about some
[16:41] stupid app they want to sell you. No,
[16:43] they’re respectful and beautiful, right?
[16:46] beautiful ideal, beautiful thought,
[16:49] beautiful protocol designed for those
[16:52] norms and those people then you let it
[16:54] into the world at large and you realize
[16:56] ah not everyone is endowed with such
[16:58] norms and such politeness and especially
[17:00] when sales people get involved. you need
[17:02] better defenses and for me and for us
[17:04] and for all our many customers hey is
[17:06] that defense it is a way to love email
[17:08] again and I find that it’s really
[17:10] important actually to have a grand why
[17:13] this is all the way back to Victor
[17:15] Frankle the meaning of uh of man finding
[17:19] a why allows you to walk through the
[17:23] snow when it’s cold and uncomfortable
[17:26] and annoying which many things are when
[17:29] you’re building with computers they are
[17:31] cold and uncomfortable and annoying.
[17:33] Now, it shouldn’t be that most of the
[17:34] time, but occasionally that will be
[17:35] there. And if you have a really strong
[17:37] why, why are we building this? Who is it
[17:39] for? What are we trying to do to improve
[17:42] the world? Even if that’s not more grand
[17:45] than just letting people love email,
[17:47] it’s a lot easier and it’s a lot more
[17:49] enjoyable to then carry whatever burdens
[17:52] you got to pack if you can set it up
[17:54] that way.
[17:55] >> This is a good time to talk about our
[17:57] season sponsor, Work OS. Having a strong
[17:59] why is what gets you to building
[18:01] something great. But after you build it
[18:03] and start selling it to enterprise
[18:05] customers, they expect things like SAML,
[18:07] SSO, directory sync, audit logs, and
[18:10] fine grain permissions. And those are
[18:12] not small features. They’re systems.
[18:14] Systems that can take months to build
[18:16] and maintain. Works gives you APIs,
[18:19] enterprise ready off, and user
[18:20] management in days instead of months.
[18:22] All designed to fit cleanly into your
[18:24] product. That’s why companies like
[18:25] OpenAI and Traffic and Cursor run on
[18:28] work OS. Focus on building your product.
[18:30] Let Work OS handle the enterprise
[18:32] infrastructure. With this, let’s get
[18:34] back to David and the old way of
[18:35] thinking versus the new way of thinking.
[18:38] Putting our your developer hat on like
[18:40] can you talk talk me through on how how
[18:43] you built it? You said it was two years,
[18:45] but was it just one or two people
[18:46] starting to build it? I’m sure as tech
[18:48] you obviously must have used Ruby on
[18:49] Rails a lot. Uh, and then I I don’t
[18:52] probably some some native stuff as well,
[18:54] but the two years seems a lot especially
[18:55] because you know you’re you’re a small
[18:57] company. You’re a nimble. You’re a great
[18:59] developer. I’m you hire great
[19:00] developers. Suddenly it’s been two
[19:02] years. What what took so long for and of
[19:05] course it’s beautiful product. But right
[19:06] on the surface I think as developers we
[19:08] might have this this thing where I look
[19:10] at it as like two years with with a
[19:13] talented team.
[19:13] >> That’s the hacker news quip to basically
[19:16] everything, right? Like I could have
[19:18] built that in a weekend. I mean famously
[19:20] stated with Dropbox that I could have
[19:23] built that in a weekend. We could have
[19:24] at the original iPod when it launched.
[19:26] It was like 5 GB bits uh no Wi-Fi uh
[19:30] whatever less speed than Nomat lane. So
[19:32] I get that because I also have that same
[19:34] instinct. I think that is our hoopers as
[19:36] developers. We think we are gods and we
[19:40] can make anything happen in no time at
[19:41] all. And you totally could. You can make
[19:43] a prototype happen in these days faster
[19:45] than faster than a weekend, right? like
[19:47] in in in a few hours we should be able
[19:49] to have
[19:50] >> kick off an agent. Yeah.
[19:51] >> But figuring out what you actually want
[19:53] to build takes a lot longer and arriving
[19:55] at something that’s worth publishing
[19:57] takes longer still. At least it does for
[20:00] us and I think it does for anyone who
[20:02] arrives at anything good and the
[20:04] original hey construction was just me on
[20:07] the technical side. This is actually how
[20:09] we’ve started majority of our major
[20:12] products is either it’s just me
[20:14] sometimes it’s one additional developer
[20:16] but is in a tiny tiny team until we have
[20:19] a shape until we have a an architecture
[20:23] and we have a direction of where the
[20:25] product is going to go. I’ve found that
[20:27] you actually go slower if you pour a
[20:29] bunch of people into a direction that is
[20:32] uncertain. If you don’t know what you
[20:34] want a million people is not going to
[20:36] build it for you. You have to figure out
[20:38] what you want. We can talk about this
[20:39] later, but this is where AI’s very
[20:42] recent progress is changing things
[20:44] dramatically. It is now quicker to
[20:45] arrive at what do I want? But for hey,
[20:49] it was me and then it was Jason and uh
[20:52] one designer, two designers, very very
[20:54] small team trying to figure out the
[20:57] shape, trying to figure out if you’re
[20:59] taking on Gmail, you can’t just do Gmail
[21:02] in blue. No one’s going to buy that. No
[21:05] one’s going to be interested in that.
[21:06] It’s got to be novel, which means it’s
[21:08] well, not just novel, it’s got to be
[21:09] good. It’s got to solve problems that
[21:12] people haven’t even articulated they
[21:14] have with Gmail because the articulation
[21:17] people have of their problems with Gmail
[21:19] is I hate email, which as we talked
[21:22] about is a bit of a misdirection. My
[21:24] contention is you hate Gmail. And not
[21:26] just Gmail, but most email systems built
[21:29] on the old way of anyone has access to
[21:31] your inbox and all that stuff. But
[21:33] figuring that out, figuring the shape
[21:35] out takes a while and it’s also fun to
[21:37] do in this way where you noodle with it
[21:41] and you don’t have infinite capacity.
[21:44] The original base camp is built the same
[21:45] way. It was just me on the technical
[21:48] side. Is this a shape uplogy? There’s
[21:50] shape up thinking in trying to actually
[21:55] endow the designer with an intention of
[21:57] how should it work not just how should
[21:59] it look and figuring out it’s also how
[22:01] it should look product should be
[22:03] beautiful and they should be unique and
[22:05] appealing and so forth. So that also
[22:06] takes time. But figuring out how it
[22:08] should work is primary. Figuring out
[22:10] where’s the epicenter, what’s the most
[22:11] important part and teasing all that
[22:13] apart. But with Hey, as with all the
[22:16] major products we’ve done, we start with
[22:18] an absolutely tiny team, often just one
[22:20] individual on the programming side and
[22:22] then one or two individuals on the
[22:24] design side. And then we go, we go, we
[22:26] go, we go. Suddenly something clicks and
[22:29] we go like, this is good. There’s
[22:31] something here. And then there’s a bit
[22:33] of a ramp. we take on a few more people
[22:35] and then when we get within maybe the
[22:37] last 20% we go okay now we know what the
[22:41] terrain looks like we can go way faster
[22:43] if everyone piles in. So one thing that
[22:46] is super interesting and you might take
[22:48] it for granted but it’s very different
[22:50] to how most startups uh that raise VC
[22:53] money which I’m very familiar with uh
[22:55] and and big companies Uber Facebook you
[22:58] name it the way projects would start
[22:59] there is you take the product manager
[23:02] >> who works with maybe maybe half a
[23:04] designer
[23:05] >> and comes up with a spec and then
[23:07] developers get involved later and what
[23:09] I’m hearing what is very novel to me is
[23:11] you take one or two designers and a
[23:13] developer how you think about designers
[23:14] Even you recently hired a designer,
[23:17] Zultan actually, who I’m I’m chatting
[23:19] with on the side. A great guy.
[23:22] >> But my sense is you think of designers a
[23:25] little bit different than potentially
[23:26] the rest of the industry does.
[23:27] >> We very much do. Designers at 37 Signals
[23:30] are not just here to make a spec look
[23:33] pretty. They’re here to find what the
[23:35] spec should be. They’re product managers
[23:37] in many ways. They are the finders of
[23:41] the how and the why in many cases
[23:44] deducing in some cases customer feedback
[23:47] in other cases just pure intuition and
[23:49] distilling that into what should we
[23:51] build and how should it work and then on
[23:53] top of that they’re also responsible for
[23:55] building it they’re responsible for
[23:57] doing the CSS they’re responsible for
[23:58] doing the HTML they’re quite often
[24:01] responsible at least dabbling in the
[24:02] JavaScript and the Ruby code to get to
[24:05] something functional now with agent
[24:08] acceleration. They do the whole thing,
[24:10] not necessarily as it will be merged,
[24:13] but the whole thing in terms of here’s
[24:15] the final shape and design of what it
[24:18] should look like. But I do think we are
[24:21] very peculiar in this sense. And we have
[24:22] found this when we’ve been trying to
[24:23] hire designers that many designers
[24:26] working other companies are not used to
[24:28] also wearing the product manager hat,
[24:30] figuring out what we should build and
[24:32] wearing the implementation hat, shaping
[24:35] it into CSS and HTML. I found that when
[24:38] you combine these three hats into one,
[24:40] you have an individual who know the
[24:43] materials they’re working with, know how
[24:45] they stretch, know which way the seam is
[24:48] supposed to be cut, and therefore works
[24:50] natively with the fabric of the
[24:53] internet. When you’re working directly
[24:55] in CSS, when you’re working directly in
[24:56] HTML, you’re just much more in tune with
[24:59] what this medium wants. And I find that
[25:02] that’s probably quite similar if you’re
[25:04] a jewelry designer. You should know the
[25:06] properties of gold. You should know how
[25:08] it bends and the strength. An architect
[25:10] should have some engineering
[25:12] understanding of loadbearing structures
[25:15] and so on. Not to the degree that the
[25:18] architect is just going to design the
[25:19] whole thing and then we start pouring
[25:21] concrete. you still have uh engineers
[25:23] helping you out, but the more you
[25:25] understand the materials you’re working
[25:27] with, the more you’re likely to come up
[25:29] with something that cuts along the grain
[25:32] and therefore ends up feeling correct,
[25:35] feeling good. Just a quick hop to Apple.
[25:37] I think this is one of the reasons why
[25:39] some of the historic super fans like
[25:41] Daring Fireball and others uh Gruber
[25:44] have been disappointed by the new
[25:46] direction is that Apple used to stand
[25:48] for these exquisitly designed native Mac
[25:53] applications which is an dying breed
[25:56] like they’re essentially dead. Now we
[25:58] have Electron which we can talk about
[26:00] that too gets way too much hate in my
[26:02] book. There’s crappy implementation of
[26:05] that, but it’s just a web in a box. But
[26:07] the disappointment with losing that
[26:10] sense, and I think it’s about the same
[26:11] thing that the Mac, its native
[26:15] feel has a stretch to it. Like the
[26:18] button placements, everything you would
[26:20] call a native application either feels
[26:23] synthetic or it feels authentic. And
[26:26] today, it’s all synthetic. There’s no
[26:28] nothing authentic about it left. And I
[26:31] think for the web it’s the same thing.
[26:33] Now the web is a much much larger
[26:35] platform and therefore it’s gotten much
[26:36] more attention. So there are way more
[26:39] people working on that quality of it.
[26:42] But at the large companies it’s
[26:43] exceptionally rare to non-existent to
[26:46] have that kind of dynamic. I think some
[26:48] of that is going to change. Agent
[26:50] acceleration is going to empower
[26:52] designers to be more capable in these
[26:54] ways. So the industry is coming a little
[26:56] towards our fundamental stance which is
[26:59] funny too because the same is true on
[27:01] the programming side. When I talked
[27:03] about base camp being a product of just
[27:05] me on the programming side for launch
[27:08] that for so long sounded unambitious or
[27:12] even wrong or even to the point of lying
[27:15] from some quarters of the internet like
[27:17] yeah but you can’t build anything real
[27:19] anything meaningful anything big unless
[27:21] you have a team that’s much larger
[27:24] because it’s just going to be a toy
[27:26] product right and my insight from the
[27:28] start was that’s of course [ __ ]
[27:30] because you just haven’t used Ruby on
[27:31] Rails you just haven’t used the
[27:33] acceleration that’s possible if you use
[27:35] better tools. Now we’re all realizing
[27:38] that we’re using realizing oh so if you
[27:40] use agent acceleration a single
[27:42] individual actually can build something
[27:45] highly [clears throat] valuable team.
[27:47] >> Yes.
[27:47] >> And that’s just fun to see that like the
[27:50] industry is coming towards oh smaller
[27:52] teams are better because now the cost
[27:54] savings you have on the logarithmic
[27:56] curve on communication cost starts to be
[27:59] relevant. And this is one of the things
[28:00] maybe we can talk about this where agent
[28:02] acceleration is really changing the
[28:05] bargain between junior developers and
[28:07] senior developers. Let’s talk about
[28:08] this. But before we go into that, do I
[28:11] feel that you very much value software
[28:15] engineering as a craft, which is very
[28:16] obvious, but what I’m sensing is you’re
[28:19] valuing design, user experience, design,
[28:23] designing on software design, like you
[28:25] know, like building stuff that feels
[28:26] good. May that be software, hardware,
[28:29] you also value that as a craft and and
[28:31] you look for it like the these two
[28:32] things. Do I sense this correctly?
[28:33] >> Hugely. I mean, I think aesthetics
[28:37] is truth. When something is beautiful,
[28:39] it’s likely to be correct. I think this
[28:41] is true in mathematics. This is true in
[28:43] physics. This is true in a lot of
[28:45] different domains that when you arrive
[28:47] at something that has the correct
[28:49] aesthetic quality. It’s like we have an
[28:53] intuition that guides us towards that
[28:55] level of beauty because it also happens
[28:58] to be correct and noble and something to
[29:01] aspire for. I also happen to believe
[29:03] it’s what makes people happy. Being
[29:05] surrounded by beautiful, well functioned
[29:09] objects is a key part of happiness. In
[29:12] fact, I’ll put it in a negative way,
[29:13] too. One of the great sources of anxiety
[29:16] and frustration is when everything is
[29:19] [ __ ] When everything is laggy, when
[29:22] that touch interface doesn’t register,
[29:25] when you have to restart it, when you’re
[29:27] calling a travel agent, they can’t do
[29:29] something because their old shitty
[29:30] cobalt system won’t let them. Right? The
[29:33] world is full of not just in
[29:36] shitification. That is things that went
[29:38] from being good to being bad to just
[29:40] plain bad, just plain awful. And I think
[29:44] it is a serious source of malaise for
[29:48] civilization. that we could literally
[29:51] raise the bar of human happiness if we
[29:54] were surrounded by more beautiful items,
[29:57] more beautiful systems. Both in the
[30:00] sense of its aesthetic exterior
[30:02] qualities, but just as much in terms of
[30:04] its aesthetic interior qualities,
[30:06] because I find those two things are
[30:08] usually in perfect harmony. The reason
[30:11] why Steve Jobs cared about the inside of
[30:13] the box was because he intuitively knew
[30:16] that the kind of people who care about
[30:18] the layout of the print board will be
[30:21] the kind of people who sweat the details
[30:23] on the user interface will be the kind
[30:25] of people who sweat the ergonomics of
[30:27] opening the case. So I think there’s
[30:31] essentially no choice if you are a
[30:34] person who is attracted to these
[30:38] aesthetics which I think is everyone.
[30:39] there’s just varying levels of u
[30:41] awareness about whether you are or not
[30:43] but that you want to make it all
[30:45] beautiful and for me Ruby in particular
[30:48] has been this seinal language because it
[30:50] produces the most beautiful code in my
[30:52] book there’s barely even competition
[30:54] like there are other things that can be
[30:56] beautiful in a way like I find looking
[30:59] at small talk for example very beautiful
[31:02] in its minimalism but not the house I
[31:05] want to live in Ruby is the house I want
[31:07] to live in because it’s got that
[31:08] aesthetic equality while not being rigid
[31:12] about its ideology which is a very rare
[31:14] aspect too. I more often find now we can
[31:18] refer to IV again is that when someone
[31:20] is obsessed in this way they are a
[31:22] little narrow-minded like that’s the
[31:24] trade-off that’s the price and I find
[31:26] that Ruby has somehow managed to be both
[31:29] broadscoped yet also intensely focused
[31:32] on on this but overall we have to have
[31:36] beautiful things we have to work with
[31:38] beautiful tools we have to produce
[31:41] beautiful fluid interactions this is how
[31:45] we should see ourselves as crafts people
[31:48] that we care about polishing it until
[31:50] there are no splinters left. How is AI
[31:53] changing how you work and how do you
[31:56] think it’s changing your craft or just
[31:58] let’s just talk about the craft of again
[32:01] you’re you’re hiring people in 37
[32:03] signals who similarly care about design
[32:05] and and software craft’s quality how
[32:07] it’s changing what you get out of the
[32:09] craft or how it’s how it’s making it
[32:11] better or or worse in some ways I I I
[32:13] just want to you know start with like
[32:15] how has your view changed because the
[32:17] last time you you talked in in length
[32:19] about this that was on Lex Freriedman’s
[32:21] cast and you were still rightfully so
[32:23] very skeptical of of AI. It was a
[32:25] different set of tools. It didn’t work
[32:26] as well and I think you you went there
[32:29] bashing it pretty hard but things have
[32:31] changed since.
[32:31] >> This is a nuance point and maybe it’s
[32:33] self- serving but I don’t actually think
[32:35] my opinions have changed. What have
[32:36] changed is the circumstances and the
[32:38] facts which uh is is something I called
[32:41] out on that show and in many other
[32:43] writings was right from the get-go I
[32:46] could see that we had something new and
[32:48] novel here that was going to change
[32:50] things. Chat GBT its launch what three
[32:54] years ago was clearly and obviously even
[32:57] at the time something you would mark on
[32:59] a timeline. You’re like here are all the
[33:01] important things that happened in the
[33:02] history of computer science or the
[33:04] world. Yoinks, there is the launch of
[33:06] Chat GBT and interacting with computers
[33:09] in this way and seeing them reason, even
[33:13] if that’s still a disputed term perhaps,
[33:15] but to me it seemed obvious that these
[33:17] things were freaking smart, smarter than
[33:19] me in many ways, whether those smarts
[33:22] came from parenting
[33:25] weights and data.
[33:27] So what? We don’t know how human
[33:29] consciousness works. We don’t know how
[33:31] human wisdom or intelligence works.
[33:33] barely. So, let’s not be so categorical
[33:36] about what constitutes consciousness or
[33:39] intelligence. At least, I find no
[33:41] utility in that distinction, even if
[33:42] it’s fun to ponder. But what I found
[33:45] with the early models and the early
[33:49] ergonomics where it was autocomplete,
[33:51] where it was co-pilot and cursor in your
[33:55] editor trying to guess the next
[33:57] character,
[33:58] >> it it would be sometimes littering it.
[33:59] Right.
[34:00] >> Yes. I found it infuriating. I found it
[34:02] as we’re trying to have a conversation.
[34:04] You won’t let me finish a sentence.
[34:06] You’re constantly trying. Was this what
[34:07] you meant? Was this what you meant?
[34:09] You’re like, shut the hell up. Can I
[34:11] just finish a thought? And I thought,
[34:13] even if it is capable of occasionally
[34:16] accelerating, it’s also wrong so often
[34:20] that that acceleration feels like a
[34:22] nuisance, even if it’s somehow net
[34:24] positive, which it wasn’t for me. Or
[34:26] maybe I gave up too soon. But I just did
[34:28] not enjoy that. I didn’t think the
[34:30] models were good enough. I thought the
[34:32] way of using the models with
[34:33] autocomplete versus agent harnesses was
[34:36] just dreadful, annoying. In fact, to the
[34:40] point that I got a little pessimistic
[34:42] about the direction of the industry for
[34:44] a hot second because I thought this was
[34:46] what we were all going to do. We’re all
[34:47] going to sit and do tap tap tap.
[34:50] No, thank you.
[34:51] >> Well, cursor even have they had I even
[34:53] got one of these one of their swags was
[34:55] a tap key.
[34:56] >> Exactly. which which felt very and I I
[34:59] haven’t I I got it from them. It’s
[35:00] really cool, very well designed and all
[35:01] that beautiful design, but
[35:03] >> but dystopian
[35:04] >> dystopian
[35:05] >> when I see that and I remember that was
[35:07] a meme for a while just we only need
[35:08] three characters on the keyboard, right?
[35:10] I thought of that episode of uh The
[35:12] Simpsons where Homer puts a mechanical
[35:16] bird on the keyboard that just dips down
[35:19] and hits enter because all he’s been
[35:22] doing is hit enter. Except suddenly
[35:25] there’s a warning about the nuclear core
[35:28] overloading and the bird just hits enter
[35:30] and the whole thing burns down. I’m
[35:31] like, “Wow, that’s quite a parallel.”
[35:33] The Simpsons really does predict
[35:35] everything. But I did not like that
[35:37] style of using it. As much as I retained
[35:40] my enthusiasm for the general direction
[35:42] of travel because it truly is amazing
[35:44] and the amazement to me I tried to
[35:47] embrace as a tutor model as a pair
[35:50] programmer who doesn’t drive it was
[35:52] amazing to have chat GBT and the other
[35:55] model just be there for like I don’t
[35:56] understand this fully here’s a piece of
[35:58] code here’s a question can you tell me
[36:00] why it works like that can you tell me
[36:02] what’s wrong with it because that’s how
[36:04] I’ve been using the internet since day
[36:06] one right that’s what Google was for
[36:07] here’s an error message. Here’s a
[36:09] concept. Maybe I find something on Stack
[36:11] Overflow with some passive aggressive
[36:13] nerd telling everyone why he’s so smart
[36:14] and then at the bottom there’s the
[36:16] solution I’m looking for. Or I don’t
[36:18] find it at all and that’s just kind of
[36:19] frustrating. With the chat GBT model, I
[36:22] very often got a really good
[36:23] explanation. Yeah, this was actually I
[36:25] talked with a game developer Jonas
[36:27] Tyroller who who built this really cool
[36:29] bestselling game. I loved playing it and
[36:31] this was during this time of of the tab
[36:33] completion and he said that in his the
[36:35] way he works is he just turned off all
[36:37] auto completions in his ID uh because he
[36:39] got annoyed by it and then every now and
[36:41] then he went to chat GPT to ask
[36:43] something or have a longer thing and
[36:45] then he had the mode of like I’m
[36:46] thinking and I’m doing this stuff oh I
[36:49] need some help okay here’s the specifics
[36:51] and I’m taking and somehow it felt that
[36:53] you know like he just he was in the zone
[36:55] the whole day by controlling it and and
[36:57] somehow those habits sounds like You
[36:59] know, you’re saying the same thing. It
[37:00] kind of took it away from you. Us.
[37:02] >> Exly. Exactly. And I did get a little
[37:04] worried that that was going to be the
[37:05] direction that we were all going to be
[37:07] the bird and I didn’t want to be the
[37:09] bird. Then I was like, well, what should
[37:10] I do instead? Maybe like farming
[37:11] potatoes. Like that’s a long tradition
[37:13] here in Denmark. Maybe I could take
[37:14] [laughter] that up.
[37:16] >> But then thankfully two things happened.
[37:19] A clot code in what is that? starts in
[37:22] the spring, gets going sort of over the
[37:25] summer, then by the fall has some
[37:27] traction on a new way of using agents to
[37:30] help you code where with the agent
[37:31] harnesses, right? This is really where
[37:33] we transition from AI to agents.
[37:36] Suddenly the AI has tools. It can use
[37:39] bash. It can use everything you got on
[37:42] your terminal. It can call the internet
[37:44] in for appropriate information. it it
[37:46] just is capable of doing more than just
[37:48] reasoning about a thing you gave it uh
[37:51] or input from a source context file. And
[37:54] then the models opus 45 to me is the
[37:58] other one of the other points we’re
[37:59] going to have on the line where it’s the
[38:02] first model that continuously and
[38:05] consistently would shock me with the
[38:08] quality of its output. it quality of its
[38:12] analysis on the basis of vague inputs
[38:16] and even more importantly the quality of
[38:20] its output. It produced code I wanted to
[38:22] merge without
[38:25] very much if any alteration and if I did
[38:29] want to do alteration I could tell it
[38:31] and it would remember and it would not
[38:33] make the same mistake next time. that to
[38:35] me the combination of those two things
[38:38] was the unlock
[38:39] >> and and you have a high bar like you
[38:40] have a really high bar
[38:40] >> incredibly high bar I as we’ve talked
[38:42] about now at length like the aesthetics
[38:44] of the output really matters if I’m
[38:47] going to look at it and I’m going to
[38:48] review it I’m going to give you another
[38:49] anecdote in a second where those things
[38:51] don’t even play in but when I’m using
[38:54] agents to work on Ruby code I want their
[38:57] code to look as good as mine I’m not
[38:58] going to merge their stuff if it’s
[39:00] sloppy no more than I would merge the
[39:03] work of a junior developer who has not
[39:06] yet fully internalized our style and so
[39:08] forth. So I wanted to be on par and on
[39:10] parody and the early models just
[39:12] couldn’t. That didn’t mean they couldn’t
[39:13] produce working software. At least some
[39:15] of the time they could. I’m very
[39:17] impressive. I mean I remember when I did
[39:19] my first snake game and I’m like holy
[39:21] smokes. I’ve been wanting to do this
[39:23] since I was 6 years old. Like I’ve been
[39:25] wanting to I have this idea. I want to
[39:27] get it into a game and I was able to see
[39:29] that in I don’t know a few 30 seconds.
[39:31] It was done with the game. copy paste
[39:33] the HTML magical experience, right?
[39:37] >> So, I think that ramp was very
[39:41] interesting because it actually took a
[39:43] while until we found this form factor of
[39:46] the agent harness of the terminal
[39:49] interface.
[39:50] That to me was the the big unlock from
[39:53] this is interesting. I want to have a
[39:55] conversation with it to I wanted to
[39:57] write my code. I will now start any
[40:01] project I’m starting with. I’m starting
[40:02] agent first and that’s a massive shift
[40:06] and it just happened from November 27th
[40:09] I believe is when Opus 45 dropped. Now
[40:11] there are other people who have
[40:12] different points they felt like oh is
[40:14] Opus 40 or
[40:16] >> maybe some people talk about Sonic 37.
[40:19] There are other earlier checkpoints but
[40:21] there I do feel like there’s a general
[40:22] consensus I can lean up against that
[40:25] capy and others have expressed like yep
[40:27] it was right around end of November
[40:29] early December. Everyone who works
[40:30] worked at larger tech companies, it was
[40:32] the winter break because people just you
[40:35] know like like the whole industry shuts
[40:37] down for 2 weeks say for a few places
[40:39] where you’re on call but again no
[40:40] production work happens across the
[40:41] >> industry to play with this.
[40:43] >> My sense was that people were playing
[40:44] with it because you give it your side
[40:45] project you never finish expecting not
[40:47] to finish and then they also got
[40:49] shocked.
[40:49] >> Yeah.
[40:50] >> You’re done and that was just a complete
[40:54] sort of break, right? Right? Like if
[40:55] this was a movie, you’d hear the scratch
[40:56] sound like you’re like, “Wait, what?
[40:59] Revine, what happened?”
[41:00] >> I feel it was the most collective shock
[41:02] which happened individually and then
[41:04] people came back in January and everyone
[41:08] especially because a lot of the decision
[41:09] makers who are you know like CTO
[41:12] engineers etc were not as hands-on but
[41:14] they were hands-on and a lot of them
[41:16] it’s this weird thing where they came
[41:17] back and they start to mandate or like
[41:19] say all right you guys need to use this
[41:20] because I’ve seen the future. I’ve
[41:22] literally used it you need to see it.
[41:23] So, it’s we’re going back to a little
[41:24] bit of hardware like people were trying
[41:26] to give, you know, like the the new
[41:27] hardware into people’s hands saying you
[41:29] need to experience it cuz you’re you’re
[41:31] not going to believe it, right? There’s
[41:32] something with this as well where you
[41:33] you really don’t believe it. We can talk
[41:35] about this and whoever’s not tried it or
[41:37] not had that aha moment. I don’t think
[41:38] we can convince them.
[41:40] >> This is another one of those cases where
[41:42] words just are not effective. You need
[41:45] to sit down in front of Open Code or
[41:48] whatever harness that you use, use one
[41:50] of the frontier models, start with that.
[41:52] Start with Opus. I’d say start with
[41:54] Opus. It’s the best frontier model.
[41:56] Other models are better at other things,
[41:58] blah blah. But if you’re just going to
[41:59] work on a piece of code and you want to
[42:01] see what the current frontier is and if
[42:03] you I mean I’d be shocked if any of your
[42:05] listeners haven’t done it already, but
[42:06] if there should be some left, now is the
[42:08] time. And I don’t even want to say in
[42:10] the sense I I found it really offputting
[42:12] this trend on X where unless you’ve
[42:15] internalized everything there is about
[42:16] AI, like you’ve been left behind. Shut
[42:18] up. First of all, patently not true. you
[42:21] could literally pick up everything in
[42:23] the next three weeks. This is the other
[42:25] magical thing about this kind of
[42:26] project, right? Like or or progress when
[42:29] if we had been having this conversation
[42:30] in spring of last year, everyone been
[42:32] like MCPs, MCPS, MCPS. And do you know
[42:36] what? You can now manage to just have
[42:38] jumped over that entire things and go
[42:40] straight to CLI and skills. That’s just
[42:43] worth having in mind that this FOMO that
[42:46] unless you’re up on all of it as it
[42:48] happens play by play, you’re left behind
[42:51] is complete and utter nonsense. That
[42:53] being said, I can still appreciate that
[42:55] some people were early. And for me, Toby
[42:59] Lutki at Shopify is the main individual
[43:02] who saw this and saw the changes that
[43:06] were coming from it way earlier than I
[43:08] did and have really helped drag me into
[43:11] this by constantly sending me like,
[43:13] “Hey, you look at this, look at this.”
[43:15] And I do think that’s actually quite
[43:16] helpful. It’s quite helpful to be
[43:18] surrounded by people who have a higher
[43:22] faith or maybe their eyes are a little
[43:23] further up. Like my eyes tend to be
[43:26] relatively close to the road like right
[43:28] in front of me and some people have a
[43:29] gaze that a little higher up and
[43:30] sometimes they see things that don’t
[43:32] come to pass. In this case, Toby saw
[43:35] exactly where we were going two years
[43:37] ago. And I finally saw it because the
[43:41] road came to me in December. And it’s
[43:45] funny because along the way I kept
[43:47] saying like, “Yep, when the models get
[43:49] good enough, when they can do all this
[43:50] thing, it’s going to be amazing.” and
[43:52] thinking wow it’s going to be I don’t
[43:54] know 18 months two years maybe it’s five
[43:56] years it’s very hard to predict these
[43:58] infliction points and I think the
[43:59] industry itself didn’t even predict the
[44:01] infliction point right you have an
[44:03] entire city Silicon Valley and
[44:05] surrounding areas San Fran focused on
[44:09] making this happen but predicting
[44:11] exactly when the hockey stick starts
[44:13] hockeying is very difficult but then it
[44:16] happened and now my daily work is very
[44:19] different
[44:20] >> so so what is your daily work. Now,
[44:22] >> my daily work is
[44:25] agent first on everything.
[44:28] >> Going agent first is a good time to
[44:30] mention our season sponsor, Sonar. When
[44:32] shifting to agent first work, one thing
[44:34] that inherently comes up is the quality
[44:36] of the code. Sonar, the makers of Sonar
[44:38] Cube, is deeply rooted in the core
[44:40] belief that code quality and code
[44:42] security are inherently linked. High
[44:44] quality code is naturally more resilient
[44:46] and as agents start writing code at a
[44:48] massive scale that verification layer
[44:50] becomes your most important security
[44:52] parameter. This is where solutions like
[44:54] Sonar Cube advanced security are
[44:55] valuable. With this new malicious
[44:57] package detection, advanced security
[44:59] provides a real-time circuit breaker
[45:01] automatically stopping agents from
[45:02] pulling in unverified or risky
[45:04] thirdparty libraries before they ever
[45:06] hit your pipeline. The impact is
[45:08] measurable, too. Developers who verify
[45:10] their code with Sonar are 44% less
[45:12] likely to report experiencing outages
[45:14] due to AI as per Sonar state of code
[45:17] developer survey 2026 report. It’s
[45:19] really about closing the gap between the
[45:20] speed of AI and the reality of
[45:22] production security. What else is Sonar
[45:24] doing to help reduce outages, improve
[45:26] security, and lower risk associated with
[45:28] AI and agenda coding? Head to
[45:29] sonarsource.com/pragmatic
[45:31] to find out. With this, let’s get back
[45:33] to David’s agent first workflow.
[45:35] >> So, specifically cloth cloth code. Uh I
[45:38] use open code open code. You use open
[45:40] code.
[45:40] >> That’s my main harness. I also use cloud
[45:41] code a little bit. They unfortunately
[45:43] got that early lead. Opus is currently
[45:45] the best model. So then they started
[45:47] thinking a little bit in that like the
[45:49] game is single match instead of thinking
[45:52] it’s multiple rounds and yanked their
[45:55] subscription from open code. So if you
[45:56] want to use your Mac subscription, you
[45:58] kind of have to use their harness, which
[46:00] I don’t love it. I think it’s a mistake.
[46:02] But leave that be for a second and let’s
[46:04] just celebrate the fact that they have
[46:05] the best model and Opus for 45 46 is
[46:09] also nice but 45 to me was the
[46:11] inflection point
[46:11] >> and it creates a lot of competition
[46:12] because everyone wants to catch up and
[46:14] overtake them now
[46:14] >> of course and especially because you see
[46:16] Anthropic’s revenues I think start of
[46:18] the year they’re at 9 billion few weeks
[46:21] later they’re at like whatever 14 now
[46:23] they’re at 19 or something it’s just the
[46:25] craziest rocket ship you could possibly
[46:26] imagine which is inspiring all this
[46:29] capital to be deployed for competitors
[46:32] and so forth which is wonderful great to
[46:34] see so even if I don’t love everything
[46:36] that they do and cloud code is not my
[46:38] preferred harness manage to hold two
[46:41] things in your head at the same time
[46:42] this is what I also try to do even with
[46:44] Apple which I have serious griefs about
[46:47] how they operate and act as the
[46:48] gatekeeper and all the other nonsense
[46:49] we’ve talked about and then I also keep
[46:51] my I just love computers hat on and go I
[46:54] like the new Neo I might even buy a new
[46:56] Neo and just see what is possible at
[46:59] $500 for Opus I have no qualms s about
[47:01] using Opus. In fact, whenever I feel
[47:03] like uh this is a really hard problem, I
[47:06] go to Opus right now, but I also use
[47:08] other models. And one of the things I’ve
[47:10] incorporated into my flow is to kind of
[47:12] have two models going at the same time
[47:15] at different speeds. So, I use T-Max and
[47:19] I have this layout thing that’s built
[47:21] into Amachi where it’ll start my new Vim
[47:24] editor on the left side and then it’ll
[47:26] start two panes on the right side. On
[47:28] the top is open code running Kimmy K25
[47:32] and on the bottom is Opus running in
[47:35] cloud code and then at the very bottom I
[47:37] have a strip of terminal and almost
[47:39] everything I started in one of the
[47:41] agents and I tell them what I want. Then
[47:43] I hop over to Neoim. First I do uh space
[47:47] gg to look at the um lacy git diff on
[47:52] it. [snorts] Once this is changing if it
[47:54] looks correct I’ll just commit. We’re
[47:56] we’re done. Great. and then sometimes it
[47:58] doesn’t. It’ll correct and I’ll I’ll go
[47:59] in and alter the code myself. But the
[48:01] ratio and how quickly the ratio changes
[48:04] is still astounding. I went from early
[48:07] November last year, I’m code first
[48:10] everything. I started the editor,
[48:13] I’ll spend whatever long it is and then
[48:14] at some point if I get stuck or if I
[48:16] want a second opinion, I’ll go ask my
[48:19] friendly clanker to give me a second
[48:21] opinion. That’s just not how it is
[48:22] anymore. Now I start with the agent. Now
[48:24] he’ll give me the draft. I’ll review the
[48:26] draft and I’ll make alterations if need
[48:29] be. And then just recently I flipped it
[48:32] even further. So we’re working on a CLI
[48:35] for Base Camp so we can get full agent
[48:37] accessibility for Base Camp. It’s
[48:40] astounding. First, actually, let me
[48:42] rewind. As soon as I got pilled on how
[48:46] good the agents were and how capable
[48:50] they were, I immediately tried to raise
[48:53] my gaze up towards the end of the road
[48:55] and think, do we even need MCP? Do we
[48:57] even need CLI? Do we even need anything?
[48:59] Can’t the agent just figure it all out?
[49:01] This was when I installed OpenClaw. So,
[49:03] I installed OpenClaw on a VM and I
[49:07] thought, what should I do here? Let’s
[49:09] see how far we can push it and what it
[49:10] can do by itself. So I thought I want
[49:13] this claw in base camp. I want this claw
[49:15] in Fizzy. Let me just try to invite it
[49:17] as it was a human. So I just wrote it.
[49:20] Can you sign up for Fizzy? I’m not
[49:22] giving you any tools. I’m not giving you
[49:23] any MCP. I’m not giving you CLI. I’m
[49:25] just telling you it’s at fizzy.do. Go
[49:28] sign up. And you see it. Chuck along.
[49:30] And then yeah, I’ve signed up, but it’s
[49:32] asking for an email address or I’m
[49:33] trying to sign up. It’s asking for an
[49:34] email address. I’m like, oh yeah, right.
[49:36] You need an email address. An agent
[49:37] doesn’t have an email address. Hey, go
[49:39] sign up for hey.com. I’m like, it’s
[49:41] going to fail this one. And it’s Chuck,
[49:43] Chuck, Chuck. Uh, I’ve signed up for
[49:45] Hey.com. Here’s the password. Write it
[49:48] down somewhere safe. I’m now also signed
[49:50] up for Fizzy. I got the confirmation
[49:53] email in my inbox. We’re all good. What
[49:56] do you want me to do? I’m like, what?
[49:58] Are you telling me that you could
[50:00] one-shot signing up through a browser to
[50:03] these things? Now, maybe that shouldn’t
[50:05] be surprising. Maybe that was already
[50:07] possible with Sonnet 3 or one of the
[50:09] early models. I don’t know. But when you
[50:11] experience it yourself on your own damn
[50:12] claw that you’re just telling over
[50:14] Telegram to do something and it’s
[50:17] signing up for products autonomously,
[50:21] that’s pretty startling. It was for me.
[50:23] And then the next step I went like,
[50:24] well, if it can sign up for Hey, and can
[50:26] sign up for Fizzy, let me invite it to
[50:28] Base Camp. So, I send it an invitation
[50:30] to its own email address. Here’s the
[50:32] invitation link to Base Camp. Can you
[50:34] just jump into the AI labs lab uh
[50:36] project that we have and introduce
[50:37] yourself to the team? go, “Hey, I’m
[50:40] David’s assistant. It’s very nice to
[50:43] meet you all. I’ve read back the
[50:45] transcript a little bit. I see you’re
[50:46] all excited about these things.” And you
[50:49] just go again, “What? What?” And that
[50:53] was fun because it showed me that even
[50:56] if it was going to take a while, it did
[50:58] take a while. It took a while. This is
[51:01] um agent terms. It took I don’t know,
[51:02] seven minutes. That was like, “Oh, it
[51:04] feels like eternity.” But it was able to
[51:07] do it. And that seems like the end
[51:08] state. The end state is that agents will
[51:10] not need any of our accommodations. They
[51:12] do not need any on-ramp. They’re not
[51:13] coming on a little uh wheelchair.
[51:16] They’ll be coming on bionic legs and
[51:17] running five times as fast as you in
[51:19] about 2 seconds, which we’ll get to in a
[51:22] second to the speed aspect of it. But
[51:24] then you also realize, okay, well, I
[51:26] can’t just sit around fiddling my thumbs
[51:28] until AGI happens. Let’s build for
[51:31] today. And that’s what we’ve been
[51:33] building for Basecam. We’ve been
[51:34] building CLI. We’re going to build it
[51:35] for Hey, we’re going to build it for
[51:36] Fizzy. We’re going to build for
[51:37] everything, even probably some of the
[51:38] legacy products. And what I love about
[51:40] the CLI, as much as I also love it about
[51:43] these harnesses, is that they validated
[51:45] the fundamental Unix philosophy from
[51:47] like whatever 71. You should just build
[51:50] small tools that can interoperate with
[51:51] pipes and you can
[51:53] >> that’s philosophy, right?
[51:54] >> It’s the total Unix philosophy. And that
[51:57] is actually the magic to me about seeing
[52:00] everything having a CLI. It’s not that
[52:03] Base Camp is easier to use now with a
[52:04] CLI. No, no, is that GitHub also has a
[52:06] CLI and Sentry, I don’t know if they
[52:09] have an CLI, but they have an MCP that
[52:10] you can tie all these things together
[52:12] and now you can tell an agent, hey, we
[52:14] have some errors in Sentry. Can you go
[52:16] check them out? Then post a write up to
[52:18] Basecam iterating what’s wrong. Then go
[52:21] in GitHub, come up with a pull request,
[52:24] post a comment back to Base Camp when
[52:26] you’re done. And now we have a central
[52:27] right base cam where we’re following the
[52:29] work as it’s going on while we have an
[52:31] agent doing work looking things up. And
[52:34] again, when we try to talk about it and
[52:37] relay it, I guess some people can see
[52:39] it. And now OpenClaw has enough videos
[52:41] on YouTube and so forth so you can get
[52:43] at least a passenger ride. But try it
[52:45] yourself with your own product, with
[52:47] your own tasks and with your own prompts
[52:50] and you will be pilled. You will be
[52:54] simultaneously
[52:56] incredibly excited for what we’ve been
[52:59] able to make sand do. The silicon, the
[53:02] chips, the weights, the whole thing.
[53:07] And then also a little bit anxious about
[53:09] where it’s all going to go. And it’s in
[53:10] that tension that I and probably anyone
[53:13] else who’s been pilled on this live,
[53:15] right? Wait a minute. If we’re already
[53:17] here, what does n 18 months from now
[53:19] look like? Like if at the last 3 months
[53:22] we’ve upended my entire understanding of
[53:24] what’s possible with computers, what’s
[53:26] the next 3 months look like? What the
[53:27] next nine months look like?
[53:28] >> Yeah. This this this is where like I I
[53:30] was a little bit on on your end for a
[53:32] long time and I think I still am where I
[53:33] believe what works and I’m always
[53:35] skeptical of projections. Mo Moore’s law
[53:38] broke down at some point. I I live
[53:40] through everyone said it will continue
[53:41] forever and you know and then it broke
[53:43] as we all suspected it would
[53:45] >> but then it found another way. I think
[53:46] it’s the good point about the Moors law,
[53:48] right? It broke for individual cores.
[53:50] Yes. How much can you push that? And
[53:51] then we just went, well, what if you
[53:53] just had what’s the latest chip? 256 on
[53:55] the AMD 10 chips, right?
[53:57] >> And even when performance broke, we we
[53:59] we went into power consumption and size
[54:01] and all of those things. So, yeah, like
[54:03] but it’s it’s harder for me to also just
[54:05] to say, oh, it’s going to stop here
[54:07] because we’ve seen it grow. We we know
[54:10] the approaches that they’re taking this
[54:12] larger and larger training sets and it’s
[54:14] been working so far. And there’s also
[54:15] the bitter lesson which I think I I
[54:17] think is a it’s it’s such a short paper
[54:20] that it’s just so worth reading. I think
[54:21] it’s one of probably the most popular
[54:23] papers outside of academic circles.
[54:25] >> Yes.
[54:25] >> Because it just lays out this thing that
[54:27] we we don’t want to believe that. We
[54:29] want to believe that our knowledge our
[54:31] understanding is superior that you know
[54:32] you and me knowing how to code or me
[54:34] putting in these 15 years or however
[54:36] long it’s been it’s special. Sometimes
[54:38] it shows that it’s it’s not as special.
[54:40] What’s interesting actually is like
[54:42] right this second this snapshot in time
[54:44] it a little bit is and this is a funny
[54:46] bification that’s happening junior
[54:48] versus senior developer is that the most
[54:53] successful and applicable agent
[54:55] acceleration that I’ve seen at 37 signal
[54:57] has been from the most senior people the
[54:59] people who are able to validate whether
[55:02] what the agent produces is suitable to
[55:04] be deployed to millions of people. There
[55:07] was just this story yesterday about some
[55:09] of the major outages at Amazon.
[55:12] >> Yeah.
[55:12] >> And Amazon’s own internal analysis
[55:15] essentially pinned that we can no longer
[55:17] let junior programmers ship agent
[55:19] generated code to production without
[55:21] review. And the problem with that is
[55:24] first of all I think that’s the
[55:26] realization most companies are now
[55:27] having across the industry. Whenever
[55:30] it’s mission critical for something of
[55:31] that nature, we cannot yet rely on the
[55:35] agents to abet it at all and a and
[55:39] junior programmers are not capable of
[55:40] figuring it out. Therefore, their role
[55:44] is suddenly more tenuous than it was 6 n
[55:49] months ago because a senior programmer
[55:51] can and this is why senior programmers
[55:53] are getting so much more acceleration.
[55:55] They’re able to first of all work in
[55:57] parallel with lots of agents but
[55:59] critically examine the quality of the
[56:01] agent output and have a high degree of
[56:03] confidence of whether this is going to
[56:04] work or not and redirect them if not
[56:06] because this is what made them senior in
[56:07] the first place. This was the role that
[56:09] they had that they had the uh long
[56:12] insight and history and overview of the
[56:14] architecture. How does it all fit in? Is
[56:15] this going to work? Is this not going to
[56:17] work? This was the role they played to
[56:18] junior programmers. But now they can
[56:19] play that role to agents and agents are
[56:24] faster at following instructions and
[56:28] redirections. And suddenly you have
[56:32] senior developers who can 5x 10x their
[56:37] individual productivity. And now this is
[56:40] the second order effect. If you manage
[56:42] to 5x or 10x a senior developer, that
[56:46] person’s value per hour just went up
[56:48] 10x. Now take that hour instead of that
[56:51] person spending it with the agents just
[56:53] shipping stuff and making things better.
[56:55] They spend that hour as they would
[56:57] before teaching a junior human how to do
[57:00] things better. There’s something in that
[57:01] equation that’s in play right now and
[57:03] it’s not clear how it’s going to map
[57:06] out. Now one way it could map out is
[57:08] that the agents will get so good that
[57:10] they stop making mistakes. They become
[57:14] senior in their capacity to ship working
[57:16] code. This is what my bet would be if we
[57:18] look x amount of time forward because
[57:21] this is what just happened with cars. So
[57:23] self-driving Teslas now drive better
[57:25] than humans do. Not all humans, not in
[57:27] all circumstances, but on average. It’s
[57:29] very possible that if we’re able to
[57:31] delegate the mortal risk, the highest
[57:34] criticality we basically deal with on a
[57:36] daily basis sitting in a metal tube
[57:38] along other metal tubes that go 60 m
[57:41] hour where you can die if someone makes
[57:42] a mistake, we delegate that to an agent.
[57:44] Well, they can probably figure out how
[57:45] to make the code work too, right? So, I
[57:47] do think it’s coming, but who knows
[57:49] when, who knows how. Right now, we’re at
[57:52] a stage where the bulk of the benefits
[57:55] are acrewing to the most senior
[57:57] developers. And also I wonder just like
[57:59] with self-driving like you realize
[58:01] there’s always KV. So, for example,
[58:03] inside companies where it matters. When
[58:05] you’re a startup, you have zero
[58:06] customers. It doesn’t matter. You can
[58:07] oneshot it and it doesn’t matter if it
[58:09] doesn’t work and it, you know, it
[58:10] crashes. But inside these companies, uh,
[58:12] at Uber, um, I just got details on how
[58:15] they’re adopting AI and and they have
[58:17] all these tools, cloud code and and all
[58:19] these things. But what we realize as
[58:20] well when you just put it in there, they
[58:22] have all these internal monor repos.
[58:24] They have their ticketing systems. They
[58:25] have their slack. They have so much.
[58:27] They have their RFC’s design documents
[58:29] on on how and why they have this jumble
[58:31] of a mess uh with microservices which
[58:33] which was fun way that we we originally
[58:36] connected like many many years ago. But
[58:38] what they found is they built a bunch of
[58:40] internal systems, a lot of it to help to
[58:43] feed NCD’s agent harnesses and now
[58:45] they’re working better. But you know
[58:47] this where we are right now is is
[58:49] there’s and this is why if you’re a
[58:51] senior engineer in one of these
[58:52] companies or a staff engineer at like
[58:54] Uber and you move to Google suddenly
[58:57] you’re not going to be as valuable as
[58:59] efficient for a while until you learn
[59:01] all the systems. So I I wonder if just
[59:03] like with self-driving, you know,
[59:04] self-driving works great as well. I was
[59:06] in SFN and LA and way most they they
[59:08] drive so nice. Like
[59:11] >> my Teslas was driving in LA driving us
[59:13] to the airport every time. The whole
[59:14] family I sit peacefully watch the road
[59:18] but do not steer at all on that entire
[59:20] journey. Well, except my my weimo got
[59:23] stuck because a a truck was parking on
[59:26] on a narrow street and a car had a bike
[59:28] shed and I I I I knew that it should I
[59:30] should not go there, but it didn’t know.
[59:32] So, human oper operator came in. But
[59:34] anyway, but even with Whimos, you know,
[59:35] like there’s there’s things like there’s
[59:37] they drive in pretty good weather.
[59:39] They’ve been mapped out. So I wonder if
[59:41] in software engineering I I wonder if
[59:43] this has these parallels where we have
[59:44] all of you know like these companies
[59:46] have their their specialized
[59:49] landscape and once you map it once you
[59:51] do all the tools once you figure out
[59:52] these things and with self-driving it
[59:54] took it took 10 years right like I was
[59:55] at Uber when they bought the
[59:57] self-driving thing and we were hearing
[59:59] in the news that you know next year it’s
[01:00:00] all going to be over for drivers and no
[01:00:04] >> yes there are not going to be steering
[01:00:05] wheels anymore which by the way is an
[01:00:07] amazing anecdote because it just shows
[01:00:09] Elon ‘s total faith in his mission
[01:00:13] because in 17 when he made that
[01:00:15] proclamation, it was an AI. It was
[01:00:17] 500,000 lines of handcoded C++,
[01:00:20] >> right?
[01:00:20] >> Like that model was never ever going to
[01:00:23] get us to the full self-driving. But he
[01:00:25] had just total faith in the vision. And
[01:00:27] then eventually, hey, here come along
[01:00:29] comes AI and it’s so good. And if you
[01:00:31] train it on billions of hours of road
[01:00:34] use, it actually can do it. And it can
[01:00:36] do it better than most humans. In fact,
[01:00:38] I’m a pretty good driver. I’d like to
[01:00:40] say I’m not the best chauffeur because
[01:00:44] my I don’t know impatience have a
[01:00:47] tendency to provoke the throttle. Uh
[01:00:49] that’s not always as pleasant for
[01:00:51] passengers as it is fun for me. And when
[01:00:53] I let uh the Tesla autopilot drive, it’s
[01:00:56] just the best chauffeur in the world.
[01:00:58] It’s just perfectly
[01:00:59] >> better than you.
[01:01:00] >> Better than me, better than the queen’s
[01:01:02] chauffeur, I think. like it’s throttle
[01:01:04] actuation and deceleration is godlike.
[01:01:08] It’s actually agi like or as like in its
[01:01:12] application within that narrow domain.
[01:01:14] And of course, when we get these
[01:01:16] anecdotes and these examples of holy
[01:01:20] smokes, not it didn’t take 10 years for
[01:01:23] the self. It took 10 years from the
[01:01:25] proclamation, but what they were doing
[01:01:26] for seven of those years had nothing to
[01:01:28] do with what they’re doing with FSD now
[01:01:30] because the FSD that’s based on AI
[01:01:33] hadn’t been running for that long. But
[01:01:35] the inflection point of I think it was
[01:01:37] 131, FSD 131, like the first version,
[01:01:40] you’re like, “Wow, this is pretty good,
[01:01:41] but like I better pay attention.” 132
[01:01:45] 140 142
[01:01:48] over the course of 18 months we went
[01:01:50] from yeah it’s pretty good but like I’m
[01:01:52] going to pay attention here to why is
[01:01:54] there steering wheel and that
[01:01:57] acceleration that short period of time
[01:01:59] of course is something people look to
[01:02:01] when it comes to programming go like
[01:02:02] well if we’re here now and senior
[01:02:05] programmers still have to review it
[01:02:06] because otherwise you’re going to get
[01:02:07] all your whatever four severity eight
[01:02:10] down times at AWS because some AI pushed
[01:02:12] out some nonsense. What is it going to
[01:02:14] look like when they take the jump that
[01:02:16] FSD did over the same period of time?
[01:02:17] Now, I also think you can go completely
[01:02:19] crazy trying to just sit and soak in all
[01:02:22] of that. This is what I tried to do over
[01:02:25] the past year. Go, I’m really excited
[01:02:27] for where this is going, but I’m also
[01:02:29] going to deal with what’s possible today
[01:02:30] and what’s enjoyable today and what we
[01:02:32] do right now. I’m not going to try to
[01:02:34] plan what my life looks like 12 months
[01:02:36] from now when maybe we do have AGI or we
[01:02:38] don’t. Now, there are other people who
[01:02:40] do that very well. I just watched an
[01:02:42] interview with Leopold on Drakesh from
[01:02:45] last year. He’s thinking like what does
[01:02:47] 2030 look like? What does the whatever
[01:02:49] 10 gawatt data center look like? I’m
[01:02:51] like I I’m very glad we have individuals
[01:02:53] who put thought into that because that’s
[01:02:55] not my favorite spot to be and I think
[01:02:59] most people are not that good at
[01:03:01] polishing the crystal ball.
[01:03:02] >> No. Well, I I mean this is a little bit
[01:03:04] unsettling as a software engineer in the
[01:03:06] sense of like clearly this is where the
[01:03:07] industry wants to go. This is where a
[01:03:09] lot of effort will be put. There will be
[01:03:10] a lot of businesses, software businesses
[01:03:12] built on this. A lot of VC money raised
[01:03:13] on this by the way who are going to
[01:03:15] tackle this and they will either like
[01:03:17] succeed or die. That’s what that’s what
[01:03:19] these companies do. But today, what do
[01:03:21] you see at at 37 signals uh with
[01:03:25] software engineers? You you of course
[01:03:27] have mostly experienced engineers,
[01:03:29] although you did hire junior engineers
[01:03:30] as well. How is their kind of work
[01:03:32] changing? How is their satisfaction with
[01:03:35] with work change? Because that’s also a
[01:03:37] thing, right? We keep arguing about like
[01:03:39] is is it making us more miserable at
[01:03:41] these things? Is it what we want to do?
[01:03:42] And how’s it changing for you? Right. I
[01:03:44] think it’s
[01:03:45] >> that’s the biggest revelation actually
[01:03:47] more than even the capacity of the
[01:03:49] agents is my enjoyment running them.
[01:03:51] When I was on that leg interview last
[01:03:53] summer, I was talking about you know
[01:03:54] what I don’t want to be a project
[01:03:55] manager for agents because I had the
[01:03:58] mental model of a project manager of
[01:04:00] humans and I thought like that’s not
[01:04:01] what I enjoy. I don’t want to be that
[01:04:03] far away from the production. I want to
[01:04:05] be in the mix. I want to have my hands
[01:04:07] in the code. What I failed to realize at
[01:04:10] the time was that running a bunch of
[01:04:13] agents feels less like being a project
[01:04:16] manager for agents and more like
[01:04:18] stepping into this super mech suit where
[01:04:21] suddenly I don’t just have two arms. I
[01:04:23] have 12 and I can now look at seven
[01:04:25] screens at the same time running five
[01:04:27] keyboards. I’m still the one doing it
[01:04:30] even if I’m not typing this as a keyword
[01:04:33] in a program. I have been hyper
[01:04:35] accelerated as a programmer. It’s a
[01:04:38] different kind of programmer, but it
[01:04:39] still has the same affinity to
[01:04:42] aesthetics, at least when I’m producing
[01:04:43] Ruby code. And I’m able to combine that
[01:04:46] while being vastly more productive on a
[01:04:49] bunch of things. It’s also like getting
[01:04:51] an incredible brain upgrade on even
[01:04:55] assessing issues. One of the pilling
[01:04:58] moments I had was before the release of
[01:05:01] Omachi 3.4.
[01:05:03] I went into GitHub and we had I don’t
[01:05:06] know 250 PRs pending and I kind of just
[01:05:11] sighed a little bit and like 250 PRs if
[01:05:13] I spend I don’t know 15 minutes on each
[01:05:16] PR like how long is it going to take
[01:05:18] before I get to the end of it and I
[01:05:20] thought you know what let me try
[01:05:21] something else let me just try to ask
[01:05:24] Claude to I’m not even doing anything
[01:05:26] with a system I just do review URL and
[01:05:29] the URL is the issue or is the PR are
[01:05:32] shocked In
[01:05:34] 90 minutes, I think it was, I processed
[01:05:38] 100 PRs. And it wasn’t that I merged all
[01:05:41] of them. In fact, I’d say I merged a
[01:05:43] small minority. Maybe 10% got merged as
[01:05:46] is. Then maybe 20% got merged. But with
[01:05:51] Claude’s implementation,
[01:05:54] >> the programmer had correctly identified
[01:05:56] an issue
[01:05:57] >> but hand rolled some code that I could
[01:06:00] see I didn’t want to keep or sometimes I
[01:06:02] couldn’t even see it. I just asked
[01:06:03] Claude and they say like ah it’s not
[01:06:05] quite right. And then I just asked
[01:06:06] Claude, can you just clean room this?
[01:06:08] >> This is the right problem. Let’s fix it
[01:06:10] but let’s do it right. It would do it
[01:06:11] right away in exactly the style as I
[01:06:15] would have written the rest of Amachi.
[01:06:16] Now this isn’t the high code of
[01:06:17] something. It’s mostly just bash code,
[01:06:19] but there’s still a shape to bash code
[01:06:21] and how you want it to look and can it
[01:06:23] feel coherent with the rest of project.
[01:06:24] Agents opus in this case would just nail
[01:06:27] it. And then the second half of it was
[01:06:30] split between 25% thinks I then just
[01:06:32] realized I just don’t want this. It
[01:06:34] shouldn’t we shouldn’t have it. And 25%
[01:06:36] claude telling me maybe there’s
[01:06:38] something here, but it’s really not a
[01:06:39] good implementation. We don’t have a
[01:06:41] straight shot to making a great one. 100
[01:06:44] issues in 90 minutes. And I sat back.
[01:06:47] This would have been a week’s worth of
[01:06:49] work, days at the very least. What the
[01:06:53] heck? And even more than that, Claude’s
[01:06:56] analysis of at least half the issues
[01:07:00] pertained to things I knew nothing about
[01:07:03] where it was undeniably
[01:07:06] a smarter, better reviewer, programmer
[01:07:11] that I could ever dream to be. Well, not
[01:07:13] dreamed to be, but wasn’t that
[01:07:15] >> moment? No, but you would have not put
[01:07:16] in the effort. This was why the PR sat
[01:07:19] in the first place. In many cases, I
[01:07:21] would look at it and go that
[01:07:23] >> I think there’s something here, but like
[01:07:25] then I now have to read up on this debug
[01:07:27] thing. I have to figure out is this the
[01:07:29] right way of doing it. I don’t want to
[01:07:30] just merge something that then has other
[01:07:32] issues. And to be able to do that agent
[01:07:35] accelerated was one of top 20
[01:07:39] programming moments. I I like how you
[01:07:42] put agent accelerated and it sounds like
[01:07:45] it’s especially efficient for work that
[01:07:47] is waiting on you but you don’t want to
[01:07:49] do it or you’re not as skilled of doing
[01:07:51] it but it’s a hassle to delegate because
[01:07:53] again like you have a team right like
[01:07:55] like like you but you probably didn’t
[01:07:57] delegate it because you probably knew
[01:07:58] that it wouldn’t make it faster or
[01:08:00] better. So I I I wonder if there’s a
[01:08:02] part of AI that because we talk a lot
[01:08:04] about like you know like companies love
[01:08:06] to measure especially larger ones like
[01:08:08] efficiency PRs and they want to see
[01:08:09] impact but about the impact of doing
[01:08:12] work that we would have not done before.
[01:08:14] >> That’s the kicker for me. That’s the
[01:08:16] fact that the pie is just exploding
[01:08:18] right now. It’s not growing. It’s
[01:08:20] exploding. The number of projects we
[01:08:21] have tackled internally that we would
[01:08:24] never even have contemplated starting on
[01:08:28] are legion. We had a great project where
[01:08:30] normally on performance work you worry
[01:08:32] about uh P50, P95, P99. Jeremy, one of
[01:08:36] our most agent accelerated people went
[01:08:39] like what about P1? What about the
[01:08:41] floor? Can we fix the floor? What is the
[01:08:44] floor? And he went like well right now
[01:08:46] our floor is I forget what it was 4
[01:08:48] milliseconds. Let’s say that, right?
[01:08:50] Well, actually 4 milliseconds can add up
[01:08:53] if you have a bunch of fast requests.
[01:08:55] They can still it still matters. and he
[01:08:57] just went like, “We’re gonna do P1.
[01:08:59] We’re gonna optimize P1 literally the
[01:09:01] fastest 1% of requests. We’re going to
[01:09:03] make them even faster.” He took it from,
[01:09:05] I think it was four milliseconds to less
[01:09:07] than half a milliseconds. He 10x the
[01:09:09] performance that I was like, I would
[01:09:11] never have signed up on this and he did
[01:09:13] the P1 project over a couple of days as
[01:09:14] like a side gef because now he could.
[01:09:17] >> Now he could because he had a hunch. He
[01:09:20] had an intuition that there was
[01:09:22] something here. He let agents run with
[01:09:25] it and the number of PRs that like all
[01:09:28] right we fixed this we fixed this I
[01:09:29] think total the PR the P1 project I
[01:09:32] maybe misremember but I think it was
[01:09:34] like 12 PRs like just fixing all sorts
[01:09:36] of things where I look at the single PRs
[01:09:38] I’m like yeah actually okay yeah makes
[01:09:40] sense I look at the total sum of it
[01:09:42] you’ve changed 2500 lines of code you’re
[01:09:44] like you’ve done that in a few days
[01:09:46] >> it’s so I’ve never heard anyone do P1
[01:09:48] because it just it feels like a vanity
[01:09:51] experience it makes no business sense I
[01:09:53] I This is not true, right? Cuz
[01:09:54] everything adds up. But but you know
[01:09:55] what I mean, right?
[01:09:56] >> I know exactly what you mean. And this
[01:09:58] is exactly why the explosion of the pie
[01:10:00] suddenly lets us look at problems we
[01:10:01] would never have contemplated looking
[01:10:03] before. It’s funny. I remember this
[01:10:05] scene from Terminator 2 where they found
[01:10:08] this chip from the Terminator in the
[01:10:10] first movie and he goes like, “This
[01:10:12] thing gave us ideas we would never have
[01:10:15] investigated before.” And like there’s
[01:10:17] some beautiful parallels here about like
[01:10:19] maybe we’re about to build the
[01:10:21] Terminator, the cliche, but also we’re
[01:10:24] getting ideas, we’re getting ambitions
[01:10:26] we would never have looked at before
[01:10:28] because suddenly the cost of exploring a
[01:10:32] hunch has just dropped by a
[01:10:34] thousandfold. I do this all the time
[01:10:37] now, too. I’ll give it some vague crappy
[01:10:40] instructions just because like I have
[01:10:42] this fleety idea. I haven’t even
[01:10:44] crystallized it into a neat prompt. I
[01:10:46] just want to see something. It’ll And
[01:10:48] then I go like, oh yeah, delete as in
[01:10:52] revert code back to normal. I like
[01:10:54] before I would be a little more precious
[01:10:56] about 75 lines of code because it would
[01:10:59] have taken me two hours to do him. Now
[01:11:02] there’s no residual value to any of this
[01:11:04] stuff and I can just go like show me a
[01:11:06] draft. I feel like a little bit like a
[01:11:07] king where you just go like show me the
[01:11:10] the analysis of the farrung regions.
[01:11:12] Where are we with the tax recip? And
[01:11:14] this boy is like, “All right, this uh
[01:11:16] servant is like, “Yes, I I shall do so
[01:11:18] and return in 3 weeks.” Except like you
[01:11:20] can just wave your hands around. And
[01:11:22] agents just come back with answers to
[01:11:25] stupid questions, terrible ideas. Then
[01:11:28] suddenly it wasn’t so terrible. It was
[01:11:29] actually a great idea. And you go like,
[01:11:30] “Wha, I did this with I haven’t even
[01:11:33] pulled the trigger on it yet.” But one
[01:11:35] of the things with the Machi people have
[01:11:36] been asking for since the beginning is
[01:11:38] dual boot. being able to install Linux
[01:11:41] next to the Windows installation so that
[01:11:43] they can still play all their games. And
[01:11:45] I just went like, do you know what? I
[01:11:46] have more than one computer so when I
[01:11:48] play play games, I can just do it on the
[01:11:50] PC. It’s not a me problem.
[01:11:52] >> Yeah,
[01:11:52] >> I totally get why a bunch of people
[01:11:53] wanted. I’m not heavily inclined to
[01:11:55] spend four hours figuring it out. And I
[01:11:59] just uh a little while ago went like,
[01:12:01] oh, this is exactly the kind of problem
[01:12:03] like I don’t have to figure it out. Just
[01:12:05] made the agents figure it out. So I
[01:12:07] kicked off initially the process of just
[01:12:09] coming up with a plan. This is a pretty
[01:12:10] good big change, right? Like if you [ __ ]
[01:12:12] up someone’s boot records or you
[01:12:13] overwrite their petition criticality
[01:12:15] high, which was one of the reasons I
[01:12:17] didn’t want to engage with it. Secondly,
[01:12:19] it’s a little finicky if you want uh Lux
[01:12:21] encryption on the Linux partition, but
[01:12:23] the Linux partition doesn’t own the
[01:12:25] whole drive. It’s a little hairy. I
[01:12:27] didn’t want to take on the criticality.
[01:12:28] I’m like, this is perfect for the kind
[01:12:30] of agent stuff. So it started off
[01:12:31] basically just having Opus and Codeex
[01:12:33] pingpong a plan. Like I’ll just I asked
[01:12:36] Opus first like come up with a plan for
[01:12:37] this it thinks for minutes and minutes
[01:12:39] and come up with a good plan and then I
[01:12:41] kick it over to Codeex and like critique
[01:12:42] the plan and then I had him ping pong
[01:12:44] back and forth a couple times and at the
[01:12:46] end looking at the plan going like yep
[01:12:48] that’s a good plan we should totally do
[01:12:50] that and I can’t wait to kick that one
[01:12:52] off and just go yeah now does dual boot
[01:12:57] not because I did it but uh thank your
[01:13:00] uh your helpful clinkers. That level of
[01:13:02] ambition is still something I’ve yet to
[01:13:05] internalize. Like even just that that
[01:13:07] like, hey, here are these hunches or
[01:13:10] demands, projects that I would like to
[01:13:13] do and maybe someday and you could kick
[01:13:16] it up on a hunch while you go to lunch.
[01:13:18] That is a new world. Which is also one
[01:13:22] of the reasons I think a lot of people
[01:13:23] are thinking, well, the model continues
[01:13:25] to improve, but even if we somehow hit a
[01:13:27] wall tomorrow, the bitter lesson is no
[01:13:30] longer true. There’s actually a limit.
[01:13:31] It’s 19 trillion tokens. That’s how much
[01:13:33] they can learn. Not true at all. But if
[01:13:35] it was and we had to be stuck with these
[01:13:37] models, we would spend the next decade
[01:13:40] just getting more and more out of them
[01:13:42] learning how to use these tools. You see
[01:13:44] this actually with vintage computers. So
[01:13:47] the kind of games they were able to make
[01:13:49] on the Commodore 64 when that was
[01:13:50] released back in 81 to 85 I think was
[01:13:54] the main run. I know they made it a
[01:13:56] little longer, but then the AmIgga and
[01:13:57] other machines came out. Were great
[01:13:59] games. I mean, I got interested in games
[01:14:01] of the Commodore 64, Yung Fu, and all
[01:14:03] that stuff. The stuff they were able to
[01:14:05] do 20 years later when someone had just
[01:14:08] noodled all the secrets and tweaked the
[01:14:11] one MHz processor
[01:14:12] >> when they’re building games for the old
[01:14:14] old
[01:14:15] >> Yeah.
[01:14:15] >> are so much more technically impressive
[01:14:17] because we just know so much more about
[01:14:19] the I mean, same thing with the you look
[01:14:21] at the PlayStation first games come out
[01:14:23] on launch, last games before we go to
[01:14:25] PlayStation 2, they look from they’re
[01:14:27] like from different generations. We
[01:14:28] could totally continue to do that with
[01:14:30] the models, but we’re not going to have
[01:14:32] that particular enjoyment because
[01:14:33] there’s a new model dropping in 3
[01:14:35] months. But this is interesting because
[01:14:36] if we just run with this thought like of
[01:14:38] course we know new new things are going
[01:14:40] to come but the point is like we will be
[01:14:41] spending so much time learning applying
[01:14:44] them building either our internal
[01:14:47] systems changing how we build things
[01:14:48] taking on new project like if you’re an
[01:14:50] existing team now that people can do
[01:14:52] more work and more ambitious work. How
[01:14:54] are you thinking of of the team taking
[01:14:57] on more work launching more products?
[01:14:59] Are you thinking of of potentially
[01:15:00] growing the team or keeping it as is? My
[01:15:02] best assessment for our setup is that
[01:15:05] the same people can do much more.
[01:15:07] >> Let’s internalize that. But that’s also
[01:15:09] enough. Already we were doing enough.
[01:15:13] Already we had margin that we could hire
[01:15:14] way more if we had enough good ideas for
[01:15:17] that. So all this extra productivity
[01:15:21] we’re getting out of the team allows us
[01:15:23] now to do things like P1 and these other
[01:15:25] projects that are awesome and they’re
[01:15:27] going to improve the product faster too.
[01:15:29] Of course they are. The old way of
[01:15:31] thinking like it’s going to take 2
[01:15:33] months to deliver a major feature. I
[01:15:34] mean that’s out the door. Of course
[01:15:36] there’s going to be rapid acceleration
[01:15:38] that’s going to filter all the way into
[01:15:39] our software methodology process like
[01:15:41] shape of was built on two-month cycles.
[01:15:44] That doesn’t make sense in the same way
[01:15:45] at all anymore. We have not fully
[01:15:47] rewritten those scripts yet because the
[01:15:49] acceleration is still so fast. No
[01:15:52] company really has rewritten the scripts
[01:15:54] on on all that. When you’re shipping
[01:15:56] that much faster, you need a way to
[01:15:58] control what goes live and measure
[01:16:00] whether it’s working. This is a good
[01:16:02] time to mention our presenting sponsor
[01:16:03] stats. Experimentation feature flags for
[01:16:06] teams that ship fast. Static build a
[01:16:08] unified platform that enables both
[01:16:10] experimentation and continuous shipping.
[01:16:12] Built-in experimentation means that
[01:16:14] every roll out automatically becomes a
[01:16:16] learning opportunity with proper
[01:16:17] statistical analysis showing you exactly
[01:16:19] how features impact your metrics.
[01:16:21] Feature flags let you ship continuously
[01:16:24] with confidence. And because it’s all in
[01:16:26] one platform with the same product data,
[01:16:28] teams across your organization can
[01:16:30] collaborate and make datadriven
[01:16:31] decisions. To learn more, head to
[01:16:34] stats.com/pragmatic.
[01:16:36] With this, let’s get back to the shift
[01:16:38] about to hit developers. But I still
[01:16:40] think software developers are delusional
[01:16:43] if they do not think a shift is coming
[01:16:45] where before they were the constraint on
[01:16:49] how much could be produced and therefore
[01:16:51] could command
[01:16:53] >> the salaries that flow to the
[01:16:56] constraints. If suddenly those
[01:16:58] constraints now loosen, especially if we
[01:17:02] fast forward a little bit where the
[01:17:03] product manager is actually able to
[01:17:05] produce changes that can be shipped and
[01:17:08] work, things are going to change. I do
[01:17:11] actually think if I was going to bet
[01:17:13] we’ve seen peak programmer in terms of
[01:17:16] the learned guilt of programmers who
[01:17:20] went to either school or spend hours
[01:17:24] getting really good at it. we’re not
[01:17:26] going to need the same number of them to
[01:17:30] do the same amount of work. Now, Javon’s
[01:17:32] paradox where as the price of something
[01:17:34] goes down, you get more of it or you get
[01:17:35] more demand for it is true, but that
[01:17:38] doesn’t mean that all programmers are
[01:17:40] going to get bailed out by it just
[01:17:41] because more software than ever is going
[01:17:44] to be produced. That’s for sure. By the
[01:17:45] way, I think
[01:17:47] >> GitHub has gotten a lot of slack or flak
[01:17:50] lately.
[01:17:50] >> A lot
[01:17:51] >> justifiably so. I saw a chart saying
[01:17:53] they had a 92% uptime, which sounds
[01:17:56] insane. I’m not sure exactly what that
[01:17:58] was measuring, but I feel it. I have a
[01:18:00] little bit of sympathy in that. I also
[01:18:02] think there’s some mistakes were made,
[01:18:04] but also that the amount of software
[01:18:08] that’s currently being produced is on a
[01:18:09] rocket ship. We are producing as a
[01:18:12] civilization globally way more software
[01:18:16] than we’ve ever done before. I mean,
[01:18:18] open claw itself, I thought um he said
[01:18:21] it was 400,000 lines of code. That used
[01:18:23] to take 10 years and 2,000 people. Yeah.
[01:18:25] To get to that.
[01:18:26] >> Well, not 2,000 in and but yes, it it
[01:18:29] took a long time.
[01:18:29] >> I mean, a long time, right? Like you
[01:18:31] look at uh I think uh the main monolith
[01:18:33] at Shopify is 3 million lines of code.
[01:18:35] That’s 20 years. And if you collectively
[01:18:37] sum up all programmers who’ve worked on
[01:18:39] that, probably like 20,000 people. Yeah.
[01:18:41] Big shifts are coming right now. Um lots
[01:18:43] of software is being produced. I can see
[01:18:44] why it’s it’s creaking a little bit over
[01:18:46] there because like the pushes are just
[01:18:47] going to accelerate, right? And we
[01:18:49] haven’t even seen anything yet. If you
[01:18:51] look at AI adoption
[01:18:53] curves, basically no one’s using it.
[01:18:55] Like we all in our little bubble in X
[01:18:58] are like, oh, everyone’s no they’re not
[01:19:00] like most companies in the world are
[01:19:02] just not doing it. Notwithstanding that
[01:19:04] like I think uh chatt got to 800 million
[01:19:07] users very quickly. Obviously, there’s
[01:19:08] adoption, but nothing on the scale of
[01:19:10] what the companies that are furthest
[01:19:12] along are doing and how much they’re
[01:19:13] accelerating with it. So, I do think it
[01:19:15] is correct for the average programmer to
[01:19:19] think maybe we’ve seen the best of the
[01:19:23] golden days. Certainly there will be
[01:19:26] pressures on price because one thing are
[01:19:29] companies like ours that have
[01:19:31] essentially unlimited scope to come up
[01:19:33] with new features and do more and we can
[01:19:35] then plow in all that additional
[01:19:36] productivity into just do more. There’s
[01:19:38] also a lot of companies who just need to
[01:19:40] do a thing and if they can do that thing
[01:19:43] at a tenth of the cost that’s actually
[01:19:46] their advantage, right? They just need
[01:19:47] to do this thing. It’s very neatly
[01:19:49] scoped and defined. It’s a cost center.
[01:19:53] Anywhere where software development is a
[01:19:55] cost center, which is actually probably
[01:19:57] the majority of software development in
[01:19:58] the world,
[01:19:59] >> they’re going to face these pressures.
[01:20:00] >> Yeah. Sounds like if I’m a software
[01:20:02] engineer right now and I’m worried about
[01:20:04] like well, you know, like just want to
[01:20:06] make sure that I’m I’m at a place where
[01:20:09] things are going to be better. You want
[01:20:11] to be at a place where you want to
[01:20:12] either get out of a cost center or
[01:20:14] become really valuable there. Obviously,
[01:20:16] you know, brush up your skills. And also
[01:20:17] I’m wondering if if the shape of
[01:20:20] software engineers who will be hired
[01:20:21] will be changing cuz if if if I just
[01:20:23] look back from like the ‘9s right like
[01:20:25] even if you look at the movies you you
[01:20:27] saw the stereotypes they were the nerd
[01:20:29] who didn’t talk to anyone but they knew
[01:20:30] how to code they knew how to do assembly
[01:20:32] and then we went in the 2000s it was
[01:20:35] still based on languages and over time I
[01:20:37] think in the 2010s startups started to
[01:20:40] not hire for languages but just hire for
[01:20:42] algorithms because you could learn the
[01:20:43] stuff and now I’m seeing companies uh
[01:20:46] some of the the the latest VC funded
[01:20:48] companies have for product engineers
[01:20:49] where they they’re actually asking for
[01:20:51] like empathy communication on top of
[01:20:53] like it’s kind of a given that you you
[01:20:55] know how to code or whatever. So I
[01:20:57] wonder if I’m just looking at just just
[01:20:59] this curve, right? If I’m just painting
[01:21:01] it up like you’re starting to get people
[01:21:03] Oh, and and the developers I I meet at
[01:21:05] all these companies, they’re all really
[01:21:06] pleasant. They’re all just very
[01:21:07] communicative, very oh and they talk
[01:21:09] with customers, most of them just it’s
[01:21:11] it’s not even a drag. It’s like and more
[01:21:14] and more of them love doing it. That’s
[01:21:16] the constraint value. Now the constraint
[01:21:19] value is figuring out what should we
[01:21:21] build, how should it be built, which
[01:21:23] customers should we be talking to, where
[01:21:24] should we be focusing. It’s product
[01:21:26] management. It’s so funny for me too
[01:21:28] because historically I’ve not
[01:21:30] necessarily had the highest esteem for
[01:21:32] product management as a function. I
[01:21:34] thought there was a lot of [ __ ] and
[01:21:35] I thought it was a lot of people who
[01:21:37] maybe didn’t do as much right and one of
[01:21:40] the reason was that they couldn’t
[01:21:42] because the constraint resource was the
[01:21:43] implementation was the product manager
[01:21:45] could find out that they want to do
[01:21:46] something I want to do this feature and
[01:21:48] then they had to wait four weeks for
[01:21:51] some very expensive programmers to make
[01:21:53] that reality happen and in those four
[01:21:55] weeks I mean I guess they could go talk
[01:21:56] to some they were underutilized they
[01:21:59] were not the constraint right they the
[01:22:00] constraint was on the implementation
[01:22:02] that absolutely absolutely is going to
[01:22:04] switch
[01:22:05] >> and now pure implementation
[01:22:08] is going to be solved at some point. I
[01:22:11] I’m not claiming it is right now and
[01:22:13] anyone who is have not tried to just
[01:22:16] deploy bipoded stuff with no review to
[01:22:19] major code bases but as the lesson of
[01:22:22] last summer on Lex I’m not going to put
[01:22:24] my heart on the block and saying that’s
[01:22:26] not going to happen before next summer
[01:22:27] >> again this is just like common sense but
[01:22:29] implementation one implementation will
[01:22:30] be solved for for a general use case for
[01:22:33] the edge cases it will take longer and
[01:22:35] for some cases it will not make sense
[01:22:37] same thing as I know self-driving is
[01:22:39] fine for like these size of cars but for
[01:22:41] like trucks it’ll either take longer or
[01:22:43] if you’re specializ you do but the point
[01:22:44] is like there will be pockets where but
[01:22:46] those pockets will be smaller. Yes, I do
[01:22:49] think the stereotype of I just want to
[01:22:51] sit and code. You have to be John Karmic
[01:22:55] levels of good to retain that privilege
[01:22:57] to just I just want to sit and code.
[01:22:59] >> And even John Karmak is also super AI
[01:23:03] appeal and lead.
[01:23:03] >> Well, but also like he he also saw some
[01:23:06] trends that he could do like for example
[01:23:09] like just like you know the type of
[01:23:11] games that people would buy, right? Like
[01:23:13] he needed to have some business skills
[01:23:14] or just surrounded by people who did
[01:23:15] that. Totally, totally, totally.
[01:23:17] >> But like you you need to literally be
[01:23:19] the very best. And not just the very
[01:23:21] best, but you need to be better than the
[01:23:22] agents, right? Like for you to get the
[01:23:24] privilege to just be an implementer, you
[01:23:26] have to be better than what’s available
[01:23:28] off the shelf from from agents. So, who
[01:23:30] are the very best? And you’re a good
[01:23:32] person to ask because whenever you
[01:23:33] advertise a position, and this was even
[01:23:35] well before AI, I remember that you you
[01:23:37] put out a a a job for both software
[01:23:40] engineer and a designer. And actually I
[01:23:42] want I’m I want to interview your
[01:23:44] designer who you hired and because uh
[01:23:47] you published the salary which is a San
[01:23:49] Francisco salary. You put the exact
[01:23:52] number you can check it for it. You have
[01:23:53] a social media presence so it’s kind of
[01:23:55] go go goes wide and you get a lot of
[01:23:57] applications and you do a pretty good
[01:23:58] job as as I understand you try to be
[01:24:00] very fair. You put a lot of effort into
[01:24:02] it. So, what did it take to get hired at
[01:24:05] 37 signals? Because now you are trying
[01:24:07] to hire some of the best and based off
[01:24:10] of this, what advice do you give to
[01:24:12] people who are like, okay, I want to be
[01:24:14] the best in in this age right now.
[01:24:17] >> Incredibly good question. No one has
[01:24:19] figured out we haven’t cracked it. And I
[01:24:20] say that as someone who have run an
[01:24:23] organization where we we must have
[01:24:25] looked at tens of thousands of now, of
[01:24:26] course, if you’re running Google, you’ve
[01:24:27] looked at millions, but we’ve looked at
[01:24:29] tens of thousands of candidates. The
[01:24:31] number of candidates we’ve hired is
[01:24:33] quite small. I mean total number of
[01:24:35] programmers that’s been through 37
[01:24:36] signals over its entire lifespan. What’s
[01:24:38] that going to be like I don’t know 100
[01:24:40] 150 at the most? I haven’t even
[01:24:41] >> How big is your team right now is?
[01:24:43] >> Uh we’re 60 people at the entire company
[01:24:46] and we are what is that going to be like
[01:24:47] 20 programmers something like that.
[01:24:49] >> Yeah, that’s probably about right.
[01:24:50] >> Oh, so so who what is the other other 40
[01:24:52] folks?
[01:24:53] >> Uh we have designers
[01:24:56] >> uh probably like 10 of those.
[01:24:58] >> Wow. Wow. And then we have customer
[01:25:00] support which is at 14. Then we have a
[01:25:03] bunch of support functions, HR, finance,
[01:25:07] and then we have operations. Operations
[01:25:09] is quite large. We have 10 folks
[01:25:11] managing all our servers. And yeah,
[01:25:13] that’s about it. But yeah, I probably
[01:25:15] it’s probably about 100 people in total
[01:25:17] that I’ve worked with uh or employed at
[01:25:20] the company’s programmers
[01:25:21] >> out of tens of thousands
[01:25:23] >> we’ve looked at. And even all those
[01:25:25] hires did not pan out in the long term.
[01:25:28] Like I’d actually say I think I looked
[01:25:29] at this recently. Our batting average at
[01:25:33] best I think is slightly better than
[01:25:35] 50/50.
[01:25:36] So half of even those hires
[01:25:38] >> you go through all of because you have a
[01:25:39] really long and thorough process. You
[01:25:41] you you put in a lot of effort, right?
[01:25:43] >> No one has figured out just to hire with
[01:25:47] such efficiency that they don’t make
[01:25:50] mistakes. There’s a great paper that
[01:25:53] Google published quite a long time ago
[01:25:55] now where they tried it all sorts of
[01:25:58] different hypothesis. Well, can we
[01:26:00] predict employee outcomes on the basis
[01:26:02] of Ivory League education background on
[01:26:05] GPA on all of these things? And the
[01:26:08] conclusion was basically like we know
[01:26:09] nothing.
[01:26:10] >> We can’t predict it on any of these
[01:26:12] things. We can’t predict it on lead
[01:26:13] code. We can’t predict it on any of
[01:26:15] these metrics.
[01:26:17] What I’d say is I’ve clearly been
[01:26:19] spoiled by working with some very good
[01:26:21] people, not just at my company, but in
[01:26:23] open source in general.
[01:26:24] >> Yeah. Oh, yeah.
[01:26:25] >> And therefore, I’ve ended up with
[01:26:26] occasionally a twisted perspective of
[01:26:29] what the average programmer is capable
[01:26:32] of. And when we do hiring rounds, I am
[01:26:34] sometimes, well, not sometimes, I mean,
[01:26:36] every time, I’m kind of surprised how
[01:26:38] poor the majority of the submissions
[01:26:41] are, how little effort is put into
[01:26:45] being presentable. And that can sound
[01:26:48] really boomery very quickly, but it’s
[01:26:51] also just the reality of trying to get a
[01:26:54] job. Like, you got to stand out. And I
[01:26:58] understand that that’s uncomfortable,
[01:27:00] right? like who wants to look at this as
[01:27:02] like well my the odds are kind of
[01:27:04] against me but it’s also a trap to
[01:27:07] actually fall into thinking of this in
[01:27:09] terms of odds because what I’ve seen the
[01:27:12] miscalculation happened time and again
[01:27:14] is people go like okay so you have a
[01:27:16] thousand applicants there’s only one who
[01:27:19] gets the job or maybe two who gets the
[01:27:21] job so that 0.1% chance no it’s not not
[01:27:25] at all with that math you had 0% chance
[01:27:27] >> yes
[01:27:28] >> zero and the very
[01:27:30] They probably had a 10% chance, 20%, 30%
[01:27:33] chance. It is not equal distributed. It
[01:27:36] is not a lottery. We don’t just like
[01:27:38] pick a thing out and be like, “Oh, it’s
[01:27:40] going to be this person because they
[01:27:41] happen to be the one drawn from the
[01:27:43] bunch.” Not at all. We discard off the
[01:27:46] bat probably at least half the
[01:27:48] applications. Maybe it’s twothirds just
[01:27:50] because they’re either not addressing
[01:27:52] the job directly, they are not following
[01:27:54] the instructions in the relatively clear
[01:27:56] spoken written openings that we have,
[01:27:59] right? they’re obviously not right for
[01:28:01] it or or whatever or we get some other
[01:28:03] smells. Then there’s like perhaps a
[01:28:04] third left and then we start looking at
[01:28:06] some of the submissions. Then we narrow
[01:28:08] it down historically to a pool of around
[01:28:10] 20 people that we give a at home test.
[01:28:13] The at home test is wonderful. Some
[01:28:15] people hate it. They feel like it’s free
[01:28:17] labor. I’m like, what the [ __ ] are you
[01:28:19] talking about? I’m not going to use your
[01:28:20] submission to a code test. What? I’m
[01:28:22] going to deploy it to production. How do
[01:28:24] you think we came up with that code test
[01:28:25] because it already exists in the system?
[01:28:27] I say that a little harshly. I also get
[01:28:29] the sympathy of like I don’t want to put
[01:28:31] six hours into making a test if it’s not
[01:28:33] going to go anywhere. Okay, I get it.
[01:28:35] But there’s no way around it because if
[01:28:38] you have it in your head that you just
[01:28:40] send in a resume, someone’s going to
[01:28:42] call you up on the phone, have a
[01:28:43] 30-minute conversation with you and go,
[01:28:45] you’ve hired sir. I don’t know if that
[01:28:47] ever existed, but certainly does not
[01:28:49] exist today. It never existed in the
[01:28:50] lifetime I’ve been in this. Well, the
[01:28:52] only time it exists, right, is
[01:28:54] >> through a very warm referral where
[01:28:57] correct
[01:28:57] >> where you’re starting a typing if you’re
[01:29:00] skipping the whole pipeline.
[01:29:01] >> And when you skip the whole pipeline, it
[01:29:02] typically only happens at the very
[01:29:03] beginning of a company when you’re
[01:29:04] founding a company and often it goes
[01:29:06] both ways where it’s very risky and then
[01:29:08] you say like this buddy of mine, I work
[01:29:10] with this person for two years straight.
[01:29:12] I would like trust them with my eyes
[01:29:14] closed. So that’s actually the black
[01:29:17] pill on the whole hiring process. If we
[01:29:19] look at the long-term success rates, we
[01:29:21] have had more long-term employees from
[01:29:25] I’ve worked with this person for 2
[01:29:27] years, we should hire them than we have
[01:29:28] from the open calls. It is actually
[01:29:30] exceptionally difficult
[01:29:33] has been for us to find the kind of
[01:29:35] programmer who thrives in our
[01:29:37] environment from open call. It has
[01:29:39] happened. We have hired people that way
[01:29:41] and I continue to want to believe even
[01:29:44] if the odds seem insanely long when you
[01:29:47] start doing the math of like oh my god
[01:29:48] we’ve looked at tens of thousands and
[01:29:50] how many then got hired and how many
[01:29:51] then didn’t work out like Jesus there’s
[01:29:53] only like a handful left from starting
[01:29:55] with that that that’s kind of
[01:29:56] blackmailing but then hiring directly on
[01:29:59] the base of warm referral as you call it
[01:30:01] um has worked very well and that the hit
[01:30:03] rate there is really high but how does
[01:30:05] that help anyone right like that’s not a
[01:30:07] very actionable advice except that’s to
[01:30:09] say Get as good as you can get and put
[01:30:12] in as much effort as you can and work
[01:30:14] with someone because I want to say that
[01:30:17] as a counter. Some people have this
[01:30:18] notion in their head that if they work
[01:30:20] at a place they consider shitty, they
[01:30:22] shouldn’t try.
[01:30:24] You’re shooting your own feet, buddy. If
[01:30:27] you show up at the shitty place of work,
[01:30:29] and we can even be objectively in unison
[01:30:32] about that, that it is a shitty place of
[01:30:34] work, and you then go like, “Well, I
[01:30:36] should just try to skirt. I should just
[01:30:37] try to goof off. I should just try to
[01:30:39] read X or Reddit all day, right?
[01:30:41] Everyone else you work with, they’re
[01:30:43] going to watch that. You know where that
[01:30:44] warm refo is going to come from? It’s
[01:30:46] going to come from someone who worked
[01:30:47] with someone else at a shitty job, but
[01:30:50] identified that that individual still
[01:30:52] showed up and did as best as they could
[01:30:54] to learn, to ship, to do all of this
[01:30:57] stuff. There is no shortcut here. You
[01:31:00] simply just have to be good. And you
[01:31:02] will not get good if you do not
[01:31:03] practice. And if you think your place of
[01:31:05] employment is not worthy of your best,
[01:31:08] you’re cheating yourself.
[01:31:10] >> If you’re not helping, even if it’s a
[01:31:11] shitty place, if you’re not helping that
[01:31:13] place get better, why would a great
[01:31:14] place hire you who only hires people to
[01:31:17] to further raise the bar?
[01:31:18] >> This is total cope. And it’s cope both
[01:31:20] on the side of I work at a shitty place
[01:31:22] if I don’t want to put things in. You
[01:31:23] could be annoyed. I’m not telling you
[01:31:25] you have to love your boss. I’d actually
[01:31:27] say the majority of people I used to
[01:31:28] work for, I didn’t have the warmest
[01:31:29] feelings about them. I still tried
[01:31:31] really hard for my own edification, for
[01:31:34] my own education, for my own sense of
[01:31:37] I’m the kind of person who shows up and
[01:31:39] does a good job just that I will be
[01:31:41] ready when the opportunity arrives when
[01:31:43] all my talents are needed and all my
[01:31:45] skills are honed. Right. Well, well, was
[01:31:47] this not how you ended up at 37 Signals
[01:31:49] where it was just a contract job or
[01:31:51] something and you know like on a
[01:31:53] contract job you have no ownership and
[01:31:55] correct but you showed up and
[01:31:56] >> correct and Jason ended up realizing
[01:32:00] >> okay this uh punk better get some equity
[01:32:02] otherwise he’s out the door now that’s a
[01:32:04] siminal story and you shouldn’t
[01:32:06] extrapolate everything from that I mean
[01:32:07] all founder stories by the way are
[01:32:09] siminal stories in that regard but the
[01:32:11] fundamental principle is still the same
[01:32:13] show up do as good as you can learn
[01:32:16] more. There also was to my
[01:32:20] chagrin to some extent. I perhaps
[01:32:22] contributed it to it a bit for a while,
[01:32:24] which was this notion that you can be a
[01:32:27] great programmer and not really like
[01:32:28] programming. That you don’t have to ever
[01:32:32] care about programming outside working
[01:32:34] hours. Was was this what you thought of
[01:32:36] or like
[01:32:37] >> Well, I thought of it mistakenly because
[01:32:40] I was pushing back on the overwork
[01:32:43] 100hour week, 120 hour week maniacal
[01:32:46] obsession, which by the way never was my
[01:32:48] experience. We did not start base camp
[01:32:50] that way. We have worked on a 40hour
[01:32:53] week rolling average over those 25
[01:32:57] years. But also, as I said at the very
[01:32:59] beginning, I really like computers. So,
[01:33:01] I play with computers in my free time. I
[01:33:03] look at computer things in my free time.
[01:33:04] It’s not work in the sense that I’m
[01:33:07] whatever shipping features to basec camp
[01:33:09] customers like just 247. That’s not what
[01:33:12] it is. But I am playing with computers.
[01:33:13] I am looking at new things. I am
[01:33:15] exploring new systems and whatever. And
[01:33:18] I think there was for a while in the
[01:33:20] 2010s a misconception that you didn’t
[01:33:23] have to do any of those things. you
[01:33:25] could just show up and do your work and
[01:33:28] you would be so soughta because
[01:33:31] programming was such a valuable activity
[01:33:33] and there were so few people who could
[01:33:34] do it that they’d take anyone even
[01:33:36] people who barely gave a [ __ ] and I
[01:33:38] think that’s over if it ever was true
[01:33:41] and I think it was true
[01:33:42] >> the boot camps were the perfect uh like
[01:33:45] catalyst or or like they were the canary
[01:33:47] when
[01:33:48] >> which also by the way is how the economy
[01:33:50] is supposed to function when salaries
[01:33:51] are really high it means that there’s
[01:33:53] not enough supply of labor Therefore, we
[01:33:55] should get labor into the pool.
[01:33:57] >> Exactly.
[01:33:57] >> And so, I’m not I don’t even have any
[01:34:00] qualms about internet. I’m just saying
[01:34:02] like that’s over.
[01:34:03] >> No, I I think looking, you know, we’re
[01:34:04] talking about like is it is the golden
[01:34:06] age of the programmer? Have you passed
[01:34:08] peep programmer? And I wonder if peep
[01:34:09] programmer really meant that almost
[01:34:11] anyone who wanted to get into the
[01:34:13] industry and was willing to put in some
[01:34:15] effort, few months or maybe a few years
[01:34:18] could do it. You could learn how to
[01:34:19] code. You could go to either college or
[01:34:21] to a boot camp or put in the hours and
[01:34:23] you could get hired at a place because
[01:34:25] the interviews were the references were
[01:34:28] not needed. We we didn’t check and I
[01:34:30] it’s probably coming to an end. You do
[01:34:31] need references. You more I think more
[01:34:33] and more companies will be doing
[01:34:35] reference checks as part of our thing
[01:34:36] and it’s not just going to be have you
[01:34:38] worked there like would you I I’ve had
[01:34:39] these calls from like data bricks is is
[01:34:41] famous for reference cards. They don’t
[01:34:43] only check for references. They drill
[01:34:45] you not just would you work with this
[01:34:47] person again, how what were their
[01:34:48] weaknesses,
[01:34:49] >> right?
[01:34:50] >> Where would you hire them, etc., etc.
[01:34:51] And
[01:34:52] >> no, I understand it. The weird thing is
[01:34:56] peak programmer sounds like this is
[01:34:59] something that affects all programmers.
[01:35:01] It does not. The best programmers are
[01:35:03] not even the best as in like it’s 10
[01:35:05] people around the world. really good
[01:35:07] programmers are currently more valuable
[01:35:10] than ever because they’re the ones who
[01:35:12] are able to get the most out of the AI
[01:35:14] acceleration. And this was the kicker
[01:35:16] for me in changing my perspective on
[01:35:19] this is that I’ve also found and maybe
[01:35:21] it’s not universally true, but certainly
[01:35:23] within 37 signals in my own experience,
[01:35:25] I’m enjoying my time as a programmer
[01:35:27] more than any time since early 2000s
[01:35:30] when I just discovered Ruby. This has
[01:35:32] the I just discovered Ruby feel to it
[01:35:35] that it is so satisfying to be able to
[01:35:38] move this fast on so many levels at the
[01:35:42] same time to be able to explore the P1s
[01:35:44] to be able to think about dual booting
[01:35:46] omachi to do all of that stuff that the
[01:35:49] work itself has gotten vastly more
[01:35:51] enjoyable and I’ve seen the same thing
[01:35:52] for the most AI forward programmers that
[01:35:55] we have maybe also have some of these
[01:35:57] anxieties but they’re kind of pushed to
[01:35:59] the side just out of sheer enjoyment
[01:36:01] working with the new capacities. So
[01:36:03] there is a bifocation here where we
[01:36:06] should all feel like well we don’t know
[01:36:07] what’s going on and for some people
[01:36:09] that’s going to produce some degree of
[01:36:11] anxiety. I understand that especially
[01:36:12] when it’s your livelihood and you’re
[01:36:14] like well I’d also like to be able to
[01:36:15] pay for my kids college in seven years.
[01:36:17] What does that look like? I get it.
[01:36:19] You’re not going to be able to manifest
[01:36:22] that anxiety into anything productive
[01:36:23] unless you just plow it into leaning in.
[01:36:25] Right? Because if you just sit and spin
[01:36:27] around, try to think about what the
[01:36:28] world’s going to look like seven years
[01:36:29] from now, you’re wasting your time. So
[01:36:32] that’s the only path. The only path is
[01:36:35] to either get excited about this, which
[01:36:36] I don’t even think takes that much
[01:36:38] effort. As we said, if you sit down with
[01:36:39] these models, you pull out one of your
[01:36:42] hobby projects from the closet
[01:36:43] >> that you never finished,
[01:36:44] >> that you never finished, and you just
[01:36:46] give it a try, I don’t see how you
[01:36:49] really like computers and not find that
[01:36:51] experiment enjoyable. And I I’ve seen
[01:36:54] this with with people who are getting
[01:36:56] into it. Kent Beck is such a great
[01:36:58] example. He’s been programmer 52 years
[01:37:00] and he is saying like he he loves doing
[01:37:02] it and he found this balance between
[01:37:04] using the agents to build something
[01:37:06] ambitious that he always wanted to b
[01:37:07] he’s building his small talk server
[01:37:09] which which used to take forever and now
[01:37:11] it’s it’s getting closer and it’s still
[01:37:12] taking a long time and then in between
[01:37:14] he’s chilling at his he has his house on
[01:37:15] on the lake and he just goes and like
[01:37:17] just looks at the birds for two hours
[01:37:18] and then gets back to it. It’s
[01:37:20] beautiful. Kent is, by the way, one of
[01:37:22] my all-time heroes. This was right when
[01:37:24] I got started in programming right when
[01:37:27] I before I was picking up uh Ruby, I saw
[01:37:30] Kent speak at a Danish conference in
[01:37:33] 2001 on stage and I was completely
[01:37:35] mesmerized by his command of both the
[01:37:38] material, how bold he was and how great
[01:37:42] of a speaker he was. And this was after
[01:37:45] having read Extreme Programming and many
[01:37:48] of these other things. Small Talk Best
[01:37:50] Practices is my number one
[01:37:51] recommendation for any programmer who
[01:37:53] want to learn the nitty-gritty of how to
[01:37:54] structure a method and a class and the
[01:37:57] rest of it. Small Talk Best Practices,
[01:37:58] which is Kent’s book from 95, I think,
[01:38:01] or 96, is to this day my favorite book
[01:38:04] of all time on tactical programming
[01:38:07] uh patterns. So, it’s wonderful to hear
[01:38:10] him being agent pilt while also enjoying
[01:38:13] the birds. I mean, I try to do that,
[01:38:15] too. And this is actually there’s a bit
[01:38:16] of attention right now is that most of
[01:38:18] the people I find who are allin, they’re
[01:38:20] working harder than they ever have. And
[01:38:23] I’ve seen that with myself now too. When
[01:38:25] you can be this effective and impactful
[01:38:28] on an hour of supervision of these
[01:38:30] agents, it’s really intoxicating. If you
[01:38:32] have an active uh dopamine loop up there
[01:38:34] that gets triggered when something is
[01:38:36] shipped, it is just hyperactive right
[01:38:39] now. And I need to go, do you know what?
[01:38:41] This is not like a limited sale. like AI
[01:38:44] is going to be here next month and the
[01:38:46] months after that. Like I cannot just
[01:38:48] operate as though it is a limited sale
[01:38:49] and I need to get all the dopamine in
[01:38:51] harvested within the next two weeks.
[01:38:53] That I actually think is the main
[01:38:55] challenge right now for the people who
[01:38:56] are furthest along and most pled on it
[01:38:58] is like remember that this is as bad as
[01:39:01] they’re ever going to be as the cliche
[01:39:03] goes, right? You damn well better find a
[01:39:05] way not to get consumed entirely about
[01:39:07] it as exciting as it is. And and then
[01:39:09] yeah, there there’s this consuming is is
[01:39:11] is a big deal. Like with Steve Yaggi, he
[01:39:13] was he looks a bit more drained than
[01:39:15] like you can see it on the video, but he
[01:39:17] he has he’s honest like he’s he’s being
[01:39:19] pulled into this. He’s doing he has
[01:39:21] friends who are and when when you’re on
[01:39:22] the edge, you’re there. You’ve clearly
[01:39:24] been AI pill, but how are you finding of
[01:39:27] keeping a balance of like all right,
[01:39:28] stepping away, you know, like I I know
[01:39:30] you’ve I think you previously talked
[01:39:32] about the importance of sleep.
[01:39:33] Apparently you don’t have an alarm.
[01:39:34] >> Correct. I don’t use an alarm, although
[01:39:36] my wife now does because the kids need
[01:39:38] to go to school on a regular basis. But
[01:39:42] yeah, for me, eight hours a night is the
[01:39:45] best investment you can make in your own
[01:39:46] cognitive capacity. So, I just am
[01:39:49] reminded every single time I do not get
[01:39:51] eight hours that it is such a poor
[01:39:54] trade. If you go from the eight to the
[01:39:56] six, I go like, well, I’m going to be
[01:39:58] awake for in that case 18 hours. What is
[01:40:01] the drag I’m gonna carry for all those
[01:40:04] 18 hours for getting one more hour, two
[01:40:07] more hours by cutting back on the sleep?
[01:40:09] It is such a bad piece of math. It makes
[01:40:12] no sense. Now, occasionally it’s
[01:40:14] involuntary. I have actually had,
[01:40:15] especially around this AI stuff, I’ve
[01:40:17] had a couple of times, very rare, I can
[01:40:19] count on two hands the number of times
[01:40:21] where I’ve been sleepless, like the ra
[01:40:24] the brain racing a little too much.
[01:40:27] That’s not typical for me and it’s still
[01:40:29] not typical. But I have had a couple of
[01:40:30] them, right? So, I get where some of
[01:40:32] that excitement comes from. But I’d also
[01:40:34] say the last thing you should trade is
[01:40:36] sleep and then you should not trade your
[01:40:39] health. You should not try to save the 3
[01:40:42] hours a week of working out to do more
[01:40:45] agent work. That’s a very poor trade.
[01:40:47] Keep in good condition. like there’s
[01:40:49] nothing this can be more important if
[01:40:51] you want to keep like sharp up there
[01:40:53] that like the rest of the system is
[01:40:55] operating if not at peak capacities than
[01:40:58] at uh at a good sustainable level right
[01:41:01] and I do think there are some
[01:41:02] individuals right now who are at fear of
[01:41:04] running ragged
[01:41:06] >> on something that we’re going to be
[01:41:07] dealing with for like slow down buddy
[01:41:09] like it’s not again a limited sale the
[01:41:12] next 10 years we’re going to see more
[01:41:13] and more it’s going to get crazier and
[01:41:14] crazier so don’t squander your health
[01:41:16] don’t squander your sleep don’t squander
[01:41:18] your diet in the service of anything
[01:41:20] because even on the short term, it does
[01:41:22] not work. You cannot get more productive
[01:41:25] within 3 weeks, let’s say, by trying to
[01:41:28] cut back two or three hours of sleep
[01:41:29] every night and then think there’s
[01:41:31] anything coherent left after 3 weeks.
[01:41:34] You will be a hot mess. So, let’s close.
[01:41:36] We talked about the stuff that we don’t
[01:41:38] know. A lot of things we don’t know, but
[01:41:39] let’s close with what you do know. So
[01:41:42] you you could have retired a long time
[01:41:44] ago and just you know kick back and and
[01:41:46] like listen to birds. What is it that
[01:41:49] keeps you doing keeps you building
[01:41:51] keeping getting up every day and before
[01:41:53] AI you would open your terminal I think
[01:41:55] you you shared like like you would go
[01:41:56] and and write now you’re doing with
[01:41:58] agents like what drives you and and and
[01:42:01] looking ahead like what what are things
[01:42:02] you’re excited about?
[01:42:03] >> My drive continues to be a deep love of
[01:42:07] computers. This is simply the best way,
[01:42:10] the most fun way to spend my time. I
[01:42:12] could spend my time on a lot of things.
[01:42:14] I do spend my time on a lot of things. I
[01:42:15] don’t just do computers. I drive race
[01:42:18] cars. I take lots of time up. I have
[01:42:20] three kids. We enjoy all of that stuff.
[01:42:23] But if I’m going to fill eight hours
[01:42:25] every day with an activity, my best bet
[01:42:28] is computers. And it has been so since I
[01:42:30] was literally 5 years old. Whether it’s
[01:42:33] video games or what now feels a little
[01:42:35] bit like a video game actually
[01:42:37] instrumenting all these agents and uh
[01:42:39] playing a little bit of Starcraft with
[01:42:40] moving them around and
[01:42:42] >> Toro.
[01:42:42] >> Yes, exactly. So, I just really like
[01:42:45] computers. So, whether I need to do so
[01:42:47] for economic reasons or not, I will
[01:42:48] continue to play with computers, see
[01:42:51] what makes them tick and make things. I
[01:42:54] think that’s the other big misconception
[01:42:55] that some people have about wealth is
[01:42:58] that they conceive of it as some sort of
[01:43:01] checkpoint. Like once you’ve made it,
[01:43:03] then you can just kick back in leisure
[01:43:05] as though that was happiness. We simply
[01:43:07] have a hundred years of psychological
[01:43:09] studies telling us no, that’s misery. If
[01:43:12] you have all the time in the world and
[01:43:14] no purpose, no mission,
[01:43:18] leisure is not going to cut it. It’s not
[01:43:20] going to be fulfilling way. And this
[01:43:22] should be obvious by example of
[01:43:25] literally every entrepreneur who sells
[01:43:26] their business. They sit on the beach
[01:43:28] for three weeks and then they’re back
[01:43:29] into the game, right? Because this
[01:43:30] [clears throat] is actually not just
[01:43:31] something they do in pursuit of a goal.
[01:43:35] It’s the goal itself. It is the mission
[01:43:39] itself. It is the satisfaction. It is
[01:43:41] the affirmation of being a human that
[01:43:43] I’m not just a blob laying around. I am
[01:43:46] a useful individual who put my skills to
[01:43:51] the best use possible. So, I’m going to
[01:43:53] continue to do that. And I’m going to
[01:43:54] continue to do it whether I’m sitting
[01:43:56] typing at the keyboard, whether I’m
[01:43:57] instrumenting these agents, whether
[01:43:59] they’re teaching me, however which way
[01:44:01] it is, I want to play with computers,
[01:44:03] I’m going continue to do that. And then
[01:44:05] even more specifically after the last
[01:44:07] three months, I’m leaning in hard now
[01:44:09] with agent accessibility. For example,
[01:44:11] this is what I’ve been doing the last
[01:44:12] few weeks. We’ve been working on the new
[01:44:13] CLI, which also taught me like we’re not
[01:44:15] quite at AGI yet, right? You think like,
[01:44:17] well, just ask your agent to make a CLI.
[01:44:19] It will, but like it’s not quite there,
[01:44:22] right? like I want it to be just right
[01:44:24] and the agents still need a little bit
[01:44:26] of help. I’m very happy to provide that
[01:44:28] help to these agents and we’ll release a
[01:44:30] great CLI for Base Camp very very
[01:44:31] shortly. Maybe by the time this is out
[01:44:33] it’ll probably be out and for the rest
[01:44:34] of them too and I want to lean into all
[01:44:36] of this. How can we use this as much as
[01:44:37] we possibly can and then right now I’m
[01:44:42] also just an incredibly cur curious
[01:44:43] person. I wake up every morning I have a
[01:44:46] new ritual which is not to pull my phone
[01:44:49] up and start hopping on X. like right
[01:44:50] when I wake up. I don’t think actually
[01:44:52] that is great. But it takes a tremendous
[01:44:55] willpower to not do so because I’m just
[01:44:57] so curious about what happened. There’s
[01:44:58] so much happening right now. I want to
[01:45:00] know. I want to know. I want to be
[01:45:03] enjoying it. Be a part of it. So I don’t
[01:45:06] foresee that ending. I don’t foresee a
[01:45:08] love of computers evaporating. In fact,
[01:45:11] if anything, right now I’m seeing like a
[01:45:14] a flourishing of it. I’m liking
[01:45:16] computers more than I did five years
[01:45:18] ago. And that’s amazing. Amazing, David.
[01:45:21] This was awesome. Thanks. Thanks a
[01:45:22] bunch.
[01:45:23] >> All right. Thanks for having me. This is
[01:45:24] really great.
[01:45:25] >> This was a fascinating conversation and
[01:45:27] I love the energy that David has. I hope
[01:45:29] some of this energy that is obvious in
[01:45:31] person [music] also came across to you.
[01:45:32] I really appreciated that David was open
[01:45:34] that his stance did not change about AI
[01:45:36] because his [music] philosophy changed.
[01:45:38] It’s just that the tools became good
[01:45:39] enough to do useful stuff. AI for
[01:45:41] autocomplete was annoying for
[01:45:43] experienced developers. AI agent that
[01:45:45] can produce pretty good working code by
[01:45:46] themselves on the other hand are now
[01:45:48] [music] pretty useful. And yet David
[01:45:50] kept coming back to taste, judgment, and
[01:45:52] craft. He wasn’t just saying just let
[01:45:55] the model write whatever. It’s the
[01:45:56] opposite. He has a very high quality bar
[01:45:59] and he wants the output to be code that
[01:46:01] he would actually be proud to merge. It
[01:46:03] feels like AI might make good judgment
[01:46:05] even more valuable than before. I also
[01:46:07] really liked how David thinks about the
[01:46:09] importance of design. At 37 Signals,
[01:46:11] designers help figure out what should be
[01:46:13] built, how it should work, [music] and
[01:46:15] increasingly even decide how it gets
[01:46:17] implemented. I wonder if 37 signals is a
[01:46:20] step ahead of the industry in thinking
[01:46:22] about designers a bit like developers as
[01:46:24] well and developers [music] a bit like
[01:46:26] designers as well. Finally, I found
[01:46:28] David’s take that we might have hit peak
[01:46:31] software engineer an interesting
[01:46:33] argument. David thinks we’ll produce
[01:46:35] more software than ever. But [music] his
[01:46:37] observation is that we might be nearing
[01:46:39] the end of the time when developers
[01:46:40] could command high compensation simply
[01:46:42] [music] because they were the
[01:46:44] bottleneck. My two cents is that there
[01:46:46] will surely be high demand for
[01:46:48] professionals who can build profitable
[01:46:50] software. But this will mean software
[01:46:52] engineers who are not only good at
[01:46:53] coding or using AI to generate code, but
[01:46:56] can oversee building complex systems
[01:46:58] have taste and business sense as well.
[01:47:00] If you’d like to hear more from David,
[01:47:02] check out a bonus episode with him
[01:47:03] linked in the show notes. Also, check
[01:47:05] out the show notes for [music] related
[01:47:06] to pragmatic engineering deep dives on
[01:47:08] software craftsmanship and practical
[01:47:09] ways of building software. If you
[01:47:11] enjoyed this podcast, please do
[01:47:12] subscribe on your favorite podcast
[01:47:14] platform and on YouTube. And a special
[01:47:16] thank you if you also leave a rating on
[01:47:17] the show. Thanks and see you in the next