0:04 Please welcome to the stage David Shu, 0:07 Retool CEO, and Patrick Coulison, Stripe 0:11 co-founder and CEO. 0:14 [Music] 0:28 Hi, Patrick. 0:29 >> Hello. 0:30 >> How's it going? 0:31 >> Thanks for Thanks for having me. 0:33 >> Thanks for being here. 0:34 >> Um, this is this is Regel's first major 0:36 event. 0:36 >> It is our first ever. Yes. 0:38 >> It looks so profession. 0:40 >> Thanks. Thanks. It's okay. Yeah. 0:42 >> So much more professional than Stripe's 0:43 first ever event. 0:44 >> Tell us about that. 0:46 >> Does any of you know Zeitgeist? The bar? 0:48 >> It was at Zeitgeist. 0:50 >> Okay, great. 0:51 >> SF Jazz is much more refined, dignified. 0:54 >> Yeah. Great. Well, uh, Patrick needs no 0:57 introduction. Patrick is the co-founder 0:59 of Stripe and the co-founder of the Arc 1:01 Institute. Welcome, Patrick. 1:03 >> It's great to be here. 1:04 >> I wanted to start off with, um, this 1:08 programming language you created back in 1:09 the day. I think when you were 16, you 1:11 won an award. I think it was Ireland's 1:12 Young Scientist Award, and it was for 1:15 creating this programming language 1:16 called Chroma. 1:17 >> That's right. 1:17 >> What was Chroma? 1:19 >> Gosh, I don't think I've ever been asked 1:20 that this. Uh so um well back in the day 1:24 uh so this was this was I guess this is 1:26 exactly 20 years ago in fact um but um 1:30 uh 1:31 well I'd gotten interested in in AI I 1:34 guess I was you know a decade or two too 1:37 early um and in the course of getting 1:39 interested in AI I discovered lisp uh 1:42 this um this obscure role any list 1:44 programmers here one two okay yes there 1:48 are no list programmers here fine um so 1:50 u I I think I heard one. Woo. Okay. Uh, 1:52 so I got really interested in Lisp. Uh, 1:54 and I, um, I wanted to you, you know, 1:56 every Lisp programmer swiftly comes to 1:58 realize that Lisp is the, um, the best 2:00 programming language. Um, and that 2:03 everybody else labor in an 2:04 unenlightened, uh, sort of medieval 2:06 state, not using Lisp. Um, and I wanted 2:09 to use Lisp for web application 2:10 development because, you know, I want to 2:11 build stuff my friends could use. And, 2:13 uh, the the, you know, AI of the day was 2:15 not really going anywhere. Um and uh and 2:17 so you know in the choice between PHP 2:19 and Lisp lisp seemed much more 2:20 attractive. Uh but you know there 2:22 weren't really any good web development 2:24 libraries for lisp and it was kind of 2:25 annoying and hard to build web 2:26 applications in it. So chroma was 2:28 designed to be uh a dialect of lisp um 2:31 adapted for web applications. 2:33 >> Okay. And what did you build with it? 2:35 >> Um well like any good programing 2:37 language designer I became much more 2:38 interested in building the programming 2:39 language than you know the dirty details 2:41 of actual applications. Uh so mostly 2:44 mostly worked on the um on the uh on the 2:47 on the language itself and you know that 2:49 that turns out to have all sorts of you 2:51 know interesting details uh like um you 2:53 know well scheme for example another 2:56 list dialect supports these things 2:57 called continuations which is kind of a 2:59 first class raification of the call 3:01 stack. Uh but you can do sort of neat 3:03 things with 3:05 >> continuations that you know um well are 3:07 useful in web applications but also in 3:09 other places. actually. So, we started a 3:11 a couple years later I we started a 3:13 company before Stripe uh called 3:14 Octamatic. You guys have never heard of 3:16 it. Um and it didn't do very well. Um 3:18 but um but the whole journey was very um 3:23 was very educational for us because we 3:26 for for our first company Octomatic we 3:27 wrote everything in small talk um 3:30 another dead obscure programming 3:31 language um uh but you know I wasn't 3:33 afraid of that from the lip experience 3:35 and everything at Octomatic was so 3:37 nicely architected um like it was really 3:39 good like for example a thing that I 3:41 still miss is if a um if any customer 3:44 user hit uh like an exception you know 3:46 some problem with the web application 3:47 some 500 3:48 um we would capture a one of these 3:50 continuations. We would store it, you 3:53 know, send a notification to us that, 3:54 you know, somebody encountered a problem 3:56 and you could load up that like the 3:58 stack trace essentially uh in the small 4:00 talk environment and resume the web 4:02 request, but you're not just like 4:04 looking at a stack trace. You're 4:05 actually like you've resumed the 4:07 request. You have the complete state. 4:08 You can inspect every variable. You can 4:10 edit the code dynamically and you know 4:12 test whether your your fix actually 4:14 fixes the problem or not. So it was 4:15 super powerful. Um, nobody ever 4:18 benefited from it because Automatic had 4:19 no users. Uh, um, and so it was the, uh, 4:22 the, you know, world's most elegantly 4:24 designed, uh, userless, uh, company. Um, 4:28 and, uh, which, you know, turned out to 4:29 be a bad strategy for a company. Uh, 4:32 and, uh, and that was very informative 4:33 for Stripe because for Stripe, we 4:36 decided to kind of take the opposite 4:38 approach. I mean, um, we like every time 4:41 for when we were starting Stripe, we we 4:43 said every time there's, you know, a a 4:46 super elegant way to do things and then 4:48 a practical, pragmatic way to do things, 4:51 we're just going to cut the corner. Um, 4:53 uh, at least until we validate that 4:54 there's actual, you know, um, uh, uh, 4:58 you know, user value here, user demand. 5:00 Uh, and so Stripe is written in Ruby, 5:02 which is obviously a much worse 5:03 programming language uh, than than 5:04 either of those for small talk. uh we 5:06 used MongoDB for the first versions of 5:07 Stripe which you know was was a very 5:10 easy to use and kind of quick to 5:11 prototype with database uh but uh you 5:13 know at the time did not have great 5:15 resiliency you know primitives and so 5:16 we've had to do a whole lot of work in 5:18 the subsequent 15 years to make uh 5:20 excuse me manga to be extremely you know 5:22 durable and and you know to make 5:23 replication very uh effective. So um 5:25 anyway 5:27 falling down this rabbit hole ended up 5:29 shaping how we built Stripe. Hm. So, it 5:32 turns out that we're actually big fans 5:34 of functional programming here at Ritual 5:36 2. And actually, the first version of 5:37 Retool 5:38 >> was written in Closure Script actually, 5:40 but then we rewrote it. 5:41 >> Okay. You you survived that grievous 5:44 early mistake. 5:45 >> Yeah, indeed. Yeah. Um 5:46 >> Well, I don't know. May maybe there's 5:48 lots of ways to to, you know, skin cats 5:50 and so forth. Um can we still say skin a 5:53 cat or that offensive to the cats? But 5:55 um uh but um uh New Bank, I understand. 5:58 Um they use a lot of I mean you know 6:00 super successful fintech company in in 6:02 LATAM and I I understand they use a lot 6:04 of closure uh so um so maybe maybe we 6:07 were we learned the wrong lesson. 6:09 >> Yeah, Anthony was too worried that we 6:10 couldn't hire anyone if we wrote it in 6:11 closure scripts. That's why we then 6:13 rewrote it in JavaScript. We we we we 6:14 found the opposite uh that I mean I mean 6:16 yes it's true that for these super 6:17 obscure programing languages you cannot 6:19 hire anyone who knows them but good 6:21 engineers are intrigued by them and so 6:24 it ended up being a selling point and 6:25 you know we dangled the prospect that 6:26 you could develop in small talk and you 6:27 people liked that. 6:29 >> Yeah I've heard actually that uh Jane 6:30 Street engineering was very high because 6:32 of so yeah great 6:35 >> uh awesome well how did you get started 6:37 with programming in the first place? 6:39 Um, how do I sort of um, well, I love 6:43 Lego. Um, and programming is just like 6:45 Lego with infinite bricks. Um, and it it 6:48 was this funny thing, and I'm sure, you 6:50 know, some people in the audience uh, 6:51 have had a similar experience. But, um, 6:54 I was convinced I would love programming 6:56 before I ever did it. I just loved the 6:58 idea of programming. Um, and one day 7:00 coming home from school, I bought a a 7:02 PHP book, in fact, and I read it. Um, 7:04 and then I started writing PHP and just 7:08 all downhill after that. 7:10 >> Did you build anything you were proud of 7:11 with PHP? 7:16 >> Does anyone build something they're 7:17 proud of with PHP? Uh, no. Um, um, gosh. 7:21 Um, I mean it it was infinite sort of um 7:25 throwaway things. Um, 7:28 what was the first thing I was actually 7:30 proud of? Um 7:33 I mean it was probably Octomatic the 7:35 first company but uh that um 7:38 >> no 7:42 nothing in nothing in PHP. Okay. 7:46 >> Sorry. This is this has ended up being a 7:48 very um very uh uh um we're we're going 7:51 to incite many programming holy wars 7:53 here if we're not careful. 7:55 >> We better shift gears then. So 7:56 >> yes, 7:57 >> we'll shift gears into progress. Uh so 7:59 progress is something you've written a 8:00 bit about. What is progress? 8:03 >> Um 8:04 um well I guess that's the question. Um 8:06 the the um 8:09 the well Tyler Cohen this economist and 8:12 I we wrote this article back in I guess 8:14 it was 2018. 8:16 um trying to make the case that you know 8:19 in some sense the most important thing 8:21 that has ever happened is that we 8:24 figured out a recipe to generate 8:26 economic growth and progress more 8:29 broadly and that you know there are some 8:31 things adjacent to economic growth that 8:32 are not directly included within it and 8:35 that if you figure out I don't know 8:36 liberal values or something like that 8:38 that might uh reinforce growth but it's 8:40 it's not the same thing as it um but you 8:43 know for for for most of human history 8:45 uh we you know if if humans arose I 8:48 don't know 100,000 years ago or 8:50 something like that you know we had not 8:51 cracked that code uh and the exponential 8:53 takeoff really only started um you know 8:56 in the 18th centuryish or thereabouts 9:00 um and if it's you know on the one hand 9:04 the most important thing that ever 9:06 happened uh again in some sense it feels 9:09 a bit funny that there's no there's no 9:14 label or um or hashtag for the group of 9:18 people who are interested in the 9:20 dynamics around it. like how can we you 9:22 know what in what which kinds of 9:24 progress are we missing or in which ways 9:27 um are we not generating the kind of 9:30 progress that we have you know at some 9:31 points in the past or something and you 9:33 might say well okay maybe that's 9:35 intellectually true but you know it 9:37 doesn't matter we're we're we're now on 9:38 this rocket ship you know we don't need 9:40 to study the rocket ship um but but it's 9:43 not obviously the case if you look at 9:44 the time series uh that that everything 9:46 is in rude health you know between 1950 9:50 and 2000 In the US, real GDP per capita 9:53 grew by about 2.3% every year. And 9:55 between 2000 and today, uh so the the 9:58 first quarter of the 21st century, um 10:01 the uh average real uh uh GDP per cap 10:04 GDP per capita growth rate in the US has 10:06 been about 1.3%. So a full percentage 10:09 point lower. And you know obviously as 10:11 one compounds this out over many years 10:13 that amounts to uh far you know far less 10:17 prosperity um and a uh a much less 10:20 bountiful future. Uh and then the US in 10:23 some senses where things are working 10:24 well if you look at sort of the similar 10:26 statistics in Europe where I grew up um 10:29 uh there the the you know growth rates 10:31 in the 21st century have all been below 10:33 1% or in for for most of the the major 10:36 markets in Europe. You know that's not 10:37 true of uh Eastern European countries. 10:39 uh and then if you look at you know 10:40 emerging markets, developing countries 10:42 uh etc. it's not obviously the case uh 10:44 that the kind of convergence we might 10:47 have imagined in the aftermath of World 10:48 War II is actually happening. And so 10:50 there's this kind of basic question of 10:52 you know will uh will you know will 10:55 Ghana be as rich as uh as the US in the 10:58 year 2011? Uh and you know if not why 11:01 not and you know what can we do to make 11:02 it so? Um so we're we're we're not 11:04 obviously just you know on a trajectory 11:06 for everything to work out. Um and I um 11:10 again the kind Tyler's uh and my uh 11:15 argument was was just that we should 11:17 have some kind of field uh for these 11:20 kinds of questions and some way for you 11:22 know people interested in these things 11:23 to to find each other uh and um and to 11:27 some extent you know to your question 11:28 was what is progress? Well, I think the 11:30 I think the simplest way to define it is 11:32 in terms of economic growth. Like I 11:34 think one of the most interesting uh you 11:35 know with Stripe, you know, our mission 11:37 is to grow the GDP of the internet. 11:38 Maybe I'm kind of a partisan for 11:40 economic lenses here. Uh but um but I 11:44 think one of the most interesting facts 11:45 about sociology or the intersection of 11:49 sociology and economics is almost every 11:51 good thing about a society that one 11:52 might care about. self-reported 11:54 happiness levels um or indices of 11:58 freedom and liberty and um again just 12:02 any any broad uh um uh female 12:04 empowerment you know what have you they 12:06 all tend to correlate at at extremely uh 12:10 uh high rates uh with uh with uh with 12:13 GDP and so like this is from memory but 12:16 I think the log of self-reported 12:17 happiness um uh against as I said 12:20 self-reported happiness against log GDP 12:22 excuse me and the core correlation of 12:23 those two GDP per capita. Um I think 12:25 it's like 75 or something. And you know 12:27 you you very rarely in sociology see 12:29 correlations that are that high. Um so I 12:32 think you can take an economic approach 12:33 to it but I think part of the question 12:36 is in fact coming up with better 12:38 understandings of well what is progress? 12:41 Uh and I think that's uh that's not yet 12:43 fully settled. 12:44 >> And how much progress has there been in 12:46 progress studies itself? 12:48 >> Uh fair question. Um it's been seven 12:50 years hoist for our own petard. Uh so um 12:55 well you know it's um it's it's early 12:59 days you know we're only 7 years in uh 13:02 but um 13:04 well if if you search progress studies 13:07 that that that was the suggested moniker 13:09 that that we put forth. There are now 13:11 there are conferences in this topic. 13:13 There are organizations there are groups 13:15 in many countries uh uh you know focused 13:17 on uh on these issues. uh there are 13:20 think tanks uh there is policy work 13:23 being conducted there's the institute 13:25 for progress uh in DC that I think is 13:27 doing some excellent work here they're 13:29 mainly focused on things like science 13:31 funding because when you think about it 13:32 um uh so so much of you know the the 13:35 prosperity we benefit from is is 13:37 downstream of discoveries in science so 13:39 I would say it's uh it's cautiously 13:42 encouraging but at the same time you 13:43 know I think there's I don't know I I I 13:46 I don't understand why everyone isn't 13:48 obsessed obsessed with this, right? It's 13:50 like, you know, it's well, it's the it's 13:52 the Twitter meme of, you know, the the 13:53 utopia at the flying cars, and it's 13:55 like, you know, well, are we on track 13:56 for the flying cars or not? And and, you 13:58 know, if if we're not, you know, what 14:00 can we do to fix it? Uh, and so uh I 14:02 think um you know, I um I I invite you 14:05 all to to join the field of progress. 14:08 >> So, we at Retool think that tooling is a 14:12 major lever for agree progress and 14:14 productivity. 14:15 >> Yes. And yet companies seem to 14:17 systematically underinvest in tooling. 14:20 >> Yeah. 14:20 >> Do you agree with that? And why? If so, 14:22 why? 14:23 >> Do they underinvest in tooling? Um, 14:29 I think they hm I think they probably I 14:34 think they probably underinvest in 14:38 um 14:40 in you know good platforms for tooling 14:43 or something because companies actually 14:45 do invest a lot in tooling isn't they 14:46 they spend so much on maintaining their 14:48 Excel spreadsheets uh and their manual 14:52 human processes um and the uh you know 14:55 the the the emails that have to get sent 14:57 on the you know the end of the month and 14:59 and so forth. And you mean you kind of 15:00 look at those as internal tools. I mean 15:02 they're not good internal tools, but 15:03 they're internal tools nonetheless. And 15:05 uh my um my my brother John co-founder 15:08 uh also at Stripe, he started a podcast 15:11 uh recently um because he's the sort of 15:13 the funny and charming and uh 15:15 charismatic one. Uh and so so sorry 15:16 about bad news for you guys. You got the 15:18 other one. Uh but um so he started this 15:20 podcast uh called uh Cheeky Point. Um 15:23 and um it's you know in in a world where 15:25 nobody's drinking you know there has to 15:26 be a podcast where some people are. Um 15:29 and um and he recently had Mark and Dre 15:31 on the podcast and Mark said something 15:33 that I thought was very um Mark can fit 15:36 a lot of words into a two-hour podcast. 15:38 Uh so you know we sort of had a a war 15:39 and peace novela uh you know in two 15:42 hours. Um but uh marks something that 15:44 that really stuck with me which is 15:46 companies will uh companies are willing 15:48 to endure any amount of chronic pain uh 15:52 in order to avoid a short dose of acute 15:54 pain. And I think that is profoundly 15:57 true. It completely accords with my 16:00 experience. Uh it is uh maybe it's even 16:02 true of individuals. But um but I think 16:05 there's something here related to 16:06 tooling where all sorts of ad hoc and 16:08 formal processes arise and there's a 16:11 natural tendency to sustain them even in 16:13 the face of their manifest inadequacy 16:15 and inefficiency and so forth and it 16:17 takes some effort to say no this kind of 16:19 sucks and we're going to rebuild it and 16:21 do it you know the right way. Um and I 16:24 think companies are almost certainly uh 16:26 too hesitant uh to do that. H. So, 16:30 another question I was I was going to 16:32 ask later, but perhaps now is a good 16:33 time for it, is 16:35 why do companies have so many Excel 16:37 spreadsheets? 16:40 This is uh this is one of life's great 16:42 questions. And uh you know, when we um 16:44 maybe when we successfully define the 16:48 ineffable nature of progress, we will 16:50 then be able to turn to even more vexing 16:52 questions like why so many spreadsheets? 16:54 Um but um well look I think the um 16:59 I think the bull case on 17:03 on spreadsheets is that um 17:07 you know the the the software industry 17:09 and you know this is part of why we're 17:11 going full circle here. Part of why I 17:12 thought lisp and small talk were so 17:14 interesting is because the environment 17:17 was so dynamic and malleable. You could 17:19 poke at everything. You could inspect 17:21 everything. you knew exactly 17:23 what was happening at every moment and 17:25 and you know everything could always be 17:27 changed. Obviously in a spreadsheet uh 17:30 it it it has many of those same 17:31 characteristics like you see a lot 17:33 simultaneously. There are so many cells 17:35 and when something breaks it's typically 17:37 pretty clear. Uh you have some you know 17:40 spatial uh some some some spatial 17:42 awareness and hints. Um it's easy to 17:44 collaborate. Um you can uh you can build 17:48 very incrementally like you know if you 17:50 solved a problem that you might solve 17:51 with a spreadsheet with Java you know 17:52 you'd spend your first afternoon you 17:54 know defining your number factories and 17:57 you know dependency injection strategies 18:00 and so forth whereas the spreadsheet you 18:01 know you can just get going. So I think 18:02 spreadsheets in some sense they were 18:04 kind of miraculous technology and I 18:06 think you know it's it's um I think what 18:09 the the programming world has done uh 18:12 less well is pulled it all together uh 18:15 where you have some way to to see your 18:19 data alongside the code and to define a 18:23 sort of tight feedback loop between your 18:24 data and the code. And uh you know I um 18:28 uh zero speaker fees were you know 18:30 transferred uh uh for this event. But 18:33 part of why I was so compelled um by 18:36 retool when I first saw it was it felt 18:39 like some kind of missing piece between 18:43 the um the Lispen small talk uh you know 18:46 regions which I had you know reluctantly 18:48 come to accept over the intervening 18:50 decade were not going to take over the 18:51 world. um uh but sort of the promise of 18:53 that power with the ease of use um and 18:56 the uh frictionlessness of the 18:58 spreadsheet uh and you know to some 19:01 extent I think retold success you know 19:02 since then has has borne you know 19:04 testament to that but uh I think I think 19:06 there's a way of you know looking at all 19:08 the spreadsheets uh that um that the the 19:11 uh the fact that our civilization now 19:13 runs on this proliferation of uh of 19:17 poorly um you know poorly structured 19:19 spreadsheets as a kind of um uh as a 19:24 kind of critique of how the programming 19:27 world's tools in some broad sense are 19:30 not quite fit for purpose. 19:32 >> Yeah. I'm trying to think what is 19:34 similar to a spreadsheet. Is spreadsheet 19:36 sort of the only 19:36 >> thing in that type? Yeah. 19:38 >> Yeah. No, it's a good question. Um I 19:40 mean I I I think 19:41 >> Photoshop is not quite Yeah. I think the 19:42 web inspector actually in browsers has 19:45 some of these characteristics for like 19:46 when you have the console and the DOM 19:49 and they're sort of like 19:50 >> yes 19:51 >> you know live updating together. Yeah. 19:53 Yeah. So so yeah I think those real time 19:56 environment I mean Mathematica I think 19:58 is kind of like this but you know there 19:59 are even fewer Mathematica users than 20:00 there were list users. So um yeah 20:03 there's not there's not many I mean 20:04 Brett Victor some of you guys um must 20:06 know his work. I mean I think he he he 20:08 did some really um is doing in fact some 20:11 some really interesting stuff in this 20:12 space but again 20:13 >> nothing he does catch on. What's the 20:16 issue? 20:17 >> Yeah. No it's very interesting so 20:18 popular. So 20:19 >> yes the most popular programming tool in 20:22 the world in history has these 20:23 characteristics and none of the others 20:25 do. 20:25 >> Yeah. Yeah. 20:26 >> Why do you think that is? 20:28 >> Do people just not want this? 20:30 >> Well evidently they do right for the 20:32 popularity of XL. I I mean there maybe 20:34 there's also something um where god I 20:36 feel like we could have a real you know 20:37 symposium discussion here on this and 20:39 solicit everyone's opinions on it but um 20:42 maybe there's something where the 20:44 spreadsheet and uh Excel um are are are 20:47 so good that maybe they they kind of 20:49 dominate uh where it's you know anything 20:52 that's spreadsheet adjacent is you know 20:53 it has to be so much better right than 20:55 than um but 20:56 >> but maybe it's going to be reachable. H 20:59 well uh maybe going back to uh internal 21:01 tools being a weapon to affect progress. 21:06 >> What has Stripe built internally that 21:09 has really affected progress? 21:12 >> The internal tools that matter for us? 21:14 >> Yeah. Gosh. 21:14 >> Are there any you're proud of? 21:16 >> Um yeah. No, we we we've invested a lot 21:18 here. Um, so hm the well look the one 21:22 that I'm excited about at the moment but 21:23 this is just some kind of recency bias 21:25 um is we've um we recently added to our 21:28 bug tracker um a um an end toend um fix 21:34 it button basically whereas we we we 21:36 we've spent you know gone to a lot of 21:38 effort of the last couple years to get 21:40 our our development environments fully 21:42 into the cloud. Um, and you know, often 21:45 engineers don't like that because it's 21:47 kind of annoying and there's some 21:48 latency and you know, blah blah blah. 21:49 Uh, you know, people like developing 21:50 locally and and to be clear, I get it. 21:51 you know uh you know I there there is 21:53 some attraction in that but having the 21:55 environment in the cloud and having the 21:56 development environments kind of fully 21:58 uh um you know standardized and all the 22:02 scaffolding uh provisionable you know 22:05 essentially instantly uh and atomically 22:08 I think is going to have huge benefits 22:10 as it comes to AIdriven development and 22:12 agentic development and so forth. And so 22:13 because this again was all kind of well 22:15 structured and defined um we added this 22:18 yeah fix it button in our bug tracker. 22:20 And so we we call these um these kind of 22:22 fully orchestrated dev boxes minions. Um 22:27 and now with a bug you can just like ask 22:29 a minion to fix it. Um and um like l 22:33 button like fix with minion. Uh and um 22:37 and I was reading an email uh two weeks 22:39 ago from so Atlas is one of our 22:41 products. uh it helps people incorporate 22:43 companies. Now I think 20% of all 22:46 startups in uh in the US is it more I 22:49 guess are are are started with with with 22:51 Atlas and there's kind of an Atlas fix 22:53 it week where you know we went to fix a 22:55 bunch of bugs and I think from memory 22:58 but 30% of the bugs were fixed by a 23:01 minion. Wow. 23:02 >> Um last two weeks ago 5% of all pull 23:07 requests at Stripe were generated by a 23:08 minion. Uh and again like this is I mean 23:11 everyone's using you know LLMs and 23:12 cursor and uh so forth. Uh but this is 23:15 this is like a human never logged into 23:18 the dev box. This is like completely 23:19 automated and now a human did approve 23:22 the pull request. So it was reviewed. 23:24 It's not just happening you know in a in 23:25 an unsupervised way. Our minions are 23:27 running wild in the asylum. But um uh 23:29 but uh so hey that that's just that 23:31 that's a recent internal tool that we're 23:33 very happy with. I mean the um the other 23:35 one that's been pretty I think impactful 23:39 for well uh I I should note uh that uh 23:42 we we've rebuilt a lot of and I'm not 23:44 going to name any names here but there 23:46 are a lot of third party tools um that 23:49 uh Stripe uses that you know they're 23:51 they're standard in their industries or 23:54 um they they have maybe very powerful 23:57 subterranean functionality be very 23:59 difficult to um to rebuild but that have 24:02 terrible interfaces um and that people 24:03 don't like using. Uh and so for many of 24:05 those, we've rebuilt them with retool. 24:07 Um and that has two benefits. One, 24:10 people are happy rather than sad when 24:12 using them. And you know, that's that's 24:14 no small thing. Um but it also gives us 24:17 um much more control and kind of ability 24:20 to customize and much more flexibility. 24:23 Uh but also we we can integrate 24:25 different tools and different data sets 24:26 together. So you know something that 24:28 might be siloed in one tool now 24:30 aggregates context from uh from multiple 24:32 tools and so if you work at stripe you 24:34 are a retool user of necessity. There 24:36 are many kind of very fundamental things 24:37 at stripe uh that that go through 24:39 retool. So in that sense retool is very 24:41 very important for us. Uh and then the 24:42 other one is um and I was a skeptic of 24:45 this one uh but um my um my skepticism 24:51 uh has been has been refuted. uh we've 24:54 built our own internal documentation 24:58 platform. Um so you know there's a whole 25:01 bunch of wiki systems and documentation 25:03 systems and so forth and for a variety 25:05 of reasons you know you want to have 25:07 some concept of maintainership and you 25:09 know who reviews edits to you know which 25:11 things and you want to there's a bunch 25:13 of other parts of Stripe we want the 25:14 documentation platform to integrate with 25:16 and so forth. We we we built our own 25:18 like I saw this project get started. I 25:19 thought it was a bad idea. Um but okay, 25:22 fine. Somebody feels passionately about 25:23 it, you know, let's do it. It was 25:24 definitely the right thing to do. Um and 25:27 uh and 25:28 >> why? 25:29 >> Well, I think 25:30 >> it feels why not just go buy Confluence, 25:33 go buy any other wiki solution. 25:35 >> Um it 25:40 well I don't I don't want to say 25:41 anything bad about other products. Um 25:42 but there were there were a bunch of 25:44 there there was a bunch of specific 25:45 functionality that we thought was 25:47 important for our use case. Um and you 25:51 know it it 25:54 documentation is a kind of tool in the 25:55 sense that it you know elevates the 25:57 productivity of everybody else. Um and 25:59 and I think documentation is actually 26:01 going to matter more and not less in the 26:03 next couple of years because obviously 26:05 the the the minions are going to want to 26:06 read the documentation. Um and the 26:09 minions can to ask the person next to 26:11 them uh in the way that you know like if 26:12 if you've shitty documentation uh you 26:15 know the company can still survive uh 26:17 you know in most cases and you just like 26:19 bother your colleagues more um but 26:21 that's that's tough for the AIS uh and 26:23 so I think the value of accurate 26:26 standardized up-to-date documentation is 26:28 actually get higher over time. 26:30 >> What percent of labor at Stripe do you 26:32 think will be done via large language 26:34 models in the next few years? Hm. Um, 26:39 and I asked this question because, um, 26:41 last night I was having dinner with, uh, 26:43 Jim, who's your head of corporate 26:44 engineering. 26:45 >> Yeah, he's big rutual fan. 26:48 >> And, uh, he was I asked him, "What 26:50 should I ask Patrick?" And he asked me 26:51 to ask you. 26:55 >> Uh, there's so much excitement around 26:57 AI, but what does Stripe look like with 26:59 AI in three, five years time? Is it 27:01 Stripe grows 10x, but the headcount 27:03 doesn't grow 10x? Is described as a 27:05 massive layoff today? What does 27:08 >> because there's so many minions now? 27:11 >> Well, okay. Um, well, f first look, I I 27:15 think, um, you know, there's the there's 27:17 the line that, you know, from the 27:18 economist that predictions are hard, 27:20 especially about the future. Um, and in 27:21 the domain of AI, I think that generally 27:24 speaking, predictions made over the last 27:25 five or 10 years have a a very poor 27:27 track record. You know, very, very few 27:28 people thought we'd be in this 27:30 particular world with these particular 27:31 characteristics at this moment. Um so 27:34 like you know even the AI labs in some 27:37 sense didn't think so right you know 27:38 they were working on other technologies 27:40 and you know it's it's it's only kind of 27:41 relatively recently that LLMs became the 27:44 the kind of the um coordination point of 27:46 choice. Um so so I think it's it's just 27:49 fundamentally hard to predict. Um, 27:50 having said that, uh, everybody jumps 27:53 immediately to 27:55 I I think naturally, um, to how can we 27:59 take what we're currently doing and do 28:00 it more efficiently, which yes, that is 28:03 kind of a relevant and an interesting 28:04 question. But I think in some ways the 28:06 much more interesting question is like 28:10 what what would we elect to do much more 28:13 of once the price declines 28:15 precipitously? And so, for example, for 28:18 everybody setting up a business, um it's 28:20 it's kind of at an interesting 28:22 intersection of um of characteristics 28:24 where it is it's hard, uh it is 28:29 complicated, it's important, and people 28:31 are doing it typically for the first 28:33 time. Um and you know, there aren't that 28:36 many things that one does that about 28:38 like maybe you know, booking a flight is 28:41 another task, but you know, you're 28:43 usually not doing it for the first time 28:44 and it's you know, of less import. But 28:46 when we survey Stripe users, they 28:47 typically tell us that their business is 28:49 the most important thing in their life 28:50 apart from their family and their 28:52 religion. And so they they really care 28:54 about getting it right. And so I think 28:55 that things like well what if um you 28:58 know for our um for our uh largest 29:01 customers uh we of course have account 29:02 teams that work with them full-time and 29:05 assist them and monitor their use of 29:06 Stripe and give them helpful suggestions 29:08 and so on. I think the obvious question 29:10 is well how can we take the account team 29:13 that Shopify or Amazon has at Stripe and 29:16 how can we make that available to every 29:19 single Stripe customer including the 29:20 person you know for whom it's just a 29:22 weekend side project uh and uh and to 29:25 kind of scale extensively um vastly you 29:28 know what it is that we do uh or you 29:31 know we um we today you know we 29:34 obviously uh we we you know work very 29:37 hard to protect our customers from fraud 29:39 But um again there are sort of natural 29:42 constraints over how much cognition or 29:45 computation we can do for any given 29:47 transaction. But as the price of 29:48 computation comes down you can in 29:50 principle conduct you know a a forensic 29:53 investigation over six hours to figure 29:54 out like what happened with this 29:55 transaction? How can we protect you know 29:57 a business from from a dispute or a 29:58 chargeback. So, uh, yeah, I I think the 30:01 I think the more interesting but also 30:04 speculative version of the question is 30:06 what what is Stripe going to do way more 30:09 of uh in the years to come and 30:12 recognizing the the speculative nature, 30:14 I think that as we look back in 5 years 30:17 over the biggest change or at the 30:18 biggest changes uh in Stripe or really 30:20 any organization over the next couple of 30:22 years, I think they're going to be more 30:24 those kind of type two changes, i.e. 30:26 doing more rather than type one of just 30:28 kind of straightforward substitution. 30:30 And I think that in part because um if 30:33 you if you're just substitutionoriented 30:36 uh i.e. not improving the product, I 30:39 think you will be you know you you will 30:42 um you will suffer at the hands of 30:43 somebody who is using it to improve the 30:45 product. Uh and so um that's that's 30:47 where I find myself spending more of my 30:49 time. 30:50 >> Yeah. So it feels like the value of LMS 30:53 will really acrue to the consumers. I 30:55 think so. I mean again it's so hard to 30:57 predict the value chain here but um but 31:00 um 31:01 >> rather than Nvidia actually in fact 31:03 >> well Nvidia seems to be doing okay but 31:05 um but um I well no look I I I tr I 31:10 think it is true that a very small 31:12 percentage of the value will acrue to 31:15 the the providers of the I mean it's the 31:18 same with with financial services with 31:19 with you know with payments right like 31:22 Stripe fortunately is a profitable 31:23 business and you know we're we're we're 31:25 we're happy with the business we have 31:26 but the the value of all the act you 31:28 know accrs far more sort of to the rest 31:30 of society into the businesses built on 31:31 the platform I think that'll be true of 31:32 AI and uh whatever you know number of 31:36 you know whatever the market cap is um 31:38 of of AI companies in aggregate in five 31:39 years uh I think it'll be 100fold more 31:43 um in terms of consumer and societal 31:45 surplus 31:46 >> yeah I was watching this interview with 31:48 you where I think um someone asked you 31:50 how you use LLM 31:52 >> and uh I think the answer you gave was 31:54 around uh finding factual knowledge 31:56 online and coding. 31:59 What are some weirder ways Patrick uses 32:01 LMS? 32:02 >> Um 32:04 gosh. Um 32:05 >> we have minions running around. That's 32:07 one. 32:07 >> Yeah, we have minions running around at 32:08 the office. Um so um what the weird ways 32:11 to use LMS? Um 32:15 >> I mean look, I'm um 32:18 I do love deep research, right? Um, and 32:23 it's just there's kind of a a um a 32:26 perspective shift when you realize you 32:28 can deep research anything, right? Like 32:31 there's no 32:32 >> deep research police uh who first 32:35 adjudicate whether your question is 32:37 worthy of deep research. Um, and so you 32:40 can pass an interesting building and ask 32:42 for like a deep research report on what 32:45 is this building, why was it built, um, 32:48 uh, who who was the visionary behind it, 32:50 like why why is our world the way it is? 32:52 And so I feel like I'm still um, you 32:55 know, back propagating to fully realize 32:58 that uh, that um, you can I I haven't 33:02 hit the usage caps um on on the 33:04 providers yet in terms of how much deep 33:06 research one can do. Uh, And there's a 33:08 lot of puzzling things about the world. 33:10 So I'm um I um I'm commissioning a lot 33:14 of reports. 33:15 >> Yeah. 33:17 >> One thing I've been struggling with a 33:18 bit is thinking about how much latency 33:20 matters because I I myself also use 5 33:23 Pro 03 Pro quite a bit and I have you 33:25 know I launch multiple research projects 33:28 at once but you know it takes 15 20 33:30 minutes sometimes. 33:32 >> How much do you think latency matters? 33:34 >> Well I I think it makes the product much 33:36 worse. Um although I um you know I also 33:38 kind of love it, right? Because it's uh 33:40 you know I'll often forget about the 33:41 reports I'm commissioning and then it's 33:43 like a a gift for my past self. Um so um 33:47 it cuts both ways. Uh but yeah, no I 33:49 should be arrested by the police 33:51 commissioning wasting GPU cycles. Well, 33:54 I'm supporting Nvidia market cap. Uh but 33:56 um but um uh yeah, I often think about 34:01 what some of these product experiences 34:02 would feel like if they were type ahead 34:04 instant, right? um in that I mean 34:07 everyone here has you know you used 34:09 products where um uh you know the the 34:11 the the experience when something is 34:13 updating and you know providing an an 34:15 sort of an instant answer you know 34:17 spotlight style keystroke by keystroke 34:20 is very different to one where yeah it's 34:22 a it's a you know multisecond or 34:24 multi-minut uh latency uh and I mean is 34:27 it possible to generate sophisticated 34:29 deep research reports in under a second 34:31 I mean I guess it depends a bit on I 34:34 don't know the nature of how much 34:34 internet you can cache or something. Uh 34:36 but um but I think that I think that 34:39 would be very cool. Um and you know what 34:43 practically speaking what it would mean 34:44 is you would end up refining what you're 34:46 doing you know many times as opposed to 34:47 just you know dispatching it once and 34:49 then being maybe a little bit 34:50 dissatisfied with the result. I mean, a 34:51 thing that I really want, if you're 34:52 asking for my weird usage of LLMs, is I 34:55 it's um LLMs are the world's um you 34:58 know, like they're they're um it's like 35:02 having an LLM is at least when it comes 35:05 to 35:07 I often use an an LLM when I'm I'm 35:09 reading some book and I want to better 35:10 understand it or something, but it's a 35:12 bit like having a 150 IQ grad student 35:16 who's trying to sabotage your career. um 35:19 in that there are so many hallucinations 35:22 and subtle errors and so on, but they're 35:24 so plausible, right? Um but like you 35:26 know I'll ask I'll ask for an example or 35:28 a citation or a quote or you know 35:30 whatever and like oh that's interesting. 35:32 I I totally missed that and then I you 35:34 know pull in the thread and hunt it down 35:35 and then you know it's you know 35:36 eventually you get to you're absolutely 35:38 right. Uh that did not happen. Um and so 35:42 um that that that I look I I I haven't 35:44 learned my lesson. That is still a weird 35:46 usage that I uh engage in substantially, 35:48 but um you have to be uh you know 35:50 permanently mad moody sort of 35:52 perpetually on your guard. 35:54 >> Yeah. Yeah. The other day I was asking 35:56 an alm about a public company and it 35:58 told me it was in the board meeting and 35:59 it told me about the minutes that had 36:02 written. I was great. Of course it's 36:04 hallucination. Uh 36:05 >> you never know. 36:08 >> Shifting gears a bit. Um you've written 36:11 a bit about how software today is almost 36:14 behind bulletproof glass. So to go back 36:16 to the Excel analogy with Excel you can 36:18 play around with the software. In fact 36:19 you create it yourself in many cases. 36:21 >> Today most apps that we use you can't 36:23 really modify it. 36:25 >> Yeah. 36:26 >> Why do you think that is? And it wasn't 36:27 always this way. 36:28 >> Yeah. Yeah. Yeah. 36:29 >> What what changed? Why? 36:31 >> Yeah. Yeah. No, I I think um I think 36:34 this is 36:38 I don't quite know why. So um like 36:42 people here used desktop versions of 36:44 Microsoft Office, I assume, right? Um 36:47 yeah, and you know, Office has has VBA, 36:50 you know, Visual Basic for applications 36:52 and or maybe that's evolved. It's VB.NET 36:54 at now or something. But um but you know 36:56 fundamentally you could I mean you could 36:59 script um uh you know your your office 37:02 application right and if you want to do 37:04 some I don't know super sophisticated 37:06 mail merge or wire up I don't know Excel 37:08 and Word in some like weird way 37:09 fundamentally you could do that right 37:11 and it's interesting to me that with SAS 37:13 applications by and large you can't 37:16 right and and it's strange because we've 37:18 moved to these more dynamic programming 37:19 languages and 37:21 >> I don't know HTML is in some sense much 37:23 more mutable and so forth. But uh if I 37:26 want to um you know 37:29 >> there more programmers than ever before 37:31 as well. There are more programmers than 37:32 ever before. 37:33 >> Exactly. 37:35 Generate the code. So there all these 37:36 reasons I think as to why we should all 37:38 be writing little customizations to and 37:41 extensions to um uh the the the SAS 37:44 applications we're using. Um but by and 37:46 large it just isn't. So now now again 37:48 obviously there are exceptions you know 37:49 for some companies like Salesforce or 37:52 Service Now etc. they've obviously 37:53 really leaned into this customization 37:55 strategy and indeed that has been you 37:56 know enormously to their benefit but for 37:58 like the median SAS application is just 38:00 not that customizable and if it is it's 38:01 maybe in like very tightly narrowly 38:04 circumscribed boxes um and and and by 38:08 the way this is uh you know Stripe is 38:11 you know I think an offender here um 38:13 where uh you know Stripe is immensely 38:17 flexible in obviously in terms of what 38:18 you can do with it um but if you want to 38:20 customize your dashboard storage. We 38:23 have some facilities for doing that. You 38:25 know, we we stripe applications, but 38:27 it's not as customiz customizable as, 38:30 you know, I I I would like it to be or 38:32 as customizable as it will be in the 38:34 future. Um, so stay tuned. Uh, but um I 38:37 think we've just gotten it I mean I 38:39 maybe there is a fundamental structural 38:40 reason here, but I think we've just 38:42 gotten it culturally wrong. I think it 38:44 could be otherwise. I think the wrong 38:46 patterns got established. I think a lot 38:49 of things in programming are just norms. 38:53 Um, and sometimes on 38:54 >> the wrong branch now. 38:55 >> Exactly. And sometimes you just you end 38:57 up in a in an inefficient branch of the 38:58 multiverse. 38:59 >> Yeah. 39:01 >> As you were saying that I was thinking, 39:03 >> but again, you guys are going to fix it, 39:04 right? 39:04 >> Indeed, we will. Yes. But 39:06 >> well, I mean, it's actually a question, 39:07 right? I mean, obviously you're building 39:08 retool, you know, the the platform that 39:10 people can build their own applications 39:11 on, but could it be like the re retool 39:14 meta platform where you like 39:16 >> we see people doing this actually? I'm 39:18 building a SAS application and somehow I 39:20 use the retool meta platform to make my 39:22 SAS application customizable. 39:24 >> Yeah, we see people doing that actually. 39:26 So uh many builders uh they maybe they 39:29 are an IT for example they or 39:30 engineering they build the application 39:32 and then we see uh others actually 39:34 forking that application and saying I'm 39:36 going to go make a modification or 39:37 something like that and so we do see 39:38 that a bit. Um, but as you maybe should 39:40 be a pillar of your company strategy, 39:42 >> maybe it should be. Yeah, maybe vote of 39:44 hands. 39:46 >> Uh, as you were saying the stripe thing, 39:47 you know, I was thinking 39:48 >> I think you're launching today, Dawson. 39:49 >> Thank you. Yeah. Um, as you were saying 39:52 the stripe thing, you know, I remember 39:54 >> I think I implemented the checkout page 39:56 actually for retool way back when and we 39:58 considered I think it was called 39:59 elements then or stripe checkout and we 40:01 obviously chose checkout. It was way 40:03 faster 40:04 >> and so I guess I'm also an offender here 40:06 in the sense that you gave me two 40:07 options actually. I could actually go 40:09 build my own form. 40:10 >> Yeah. Right. So you 40:11 >> worry about it. But I was like you chose 40:12 high converting. It's 40:14 >> Yeah. Exactly. You chose the the the the 40:15 easy to integrate high converting you 40:17 know whatever option. But now if you 40:19 want to go tweak that and customize I 40:20 mean we give you certain ways in which 40:22 you can do it but it's not as 40:22 customizable as it should be in my 40:24 opinion at least. Uh so we're we're um 40:26 and and like that that dichotomy is too 40:28 dichomous. Uh so we're going to turn it 40:30 into more of a continuum. 40:31 >> Okay. We're excited for it. 40:32 >> Great. Um so uh with the rise of LLMs 40:38 and with uh LM driven software 40:39 development you saw our launch today. 40:41 >> What do you think uh the future holds 40:44 for us uh in terms of end user 40:46 customizability of applications? Do you 40:48 think that end users so I suppose part 40:50 of the reason I as the question is do 40:52 you think they want to customize it but 40:53 cannot and now they can or is it that 40:55 they don't even want to and just want to 40:57 be told just use this software and be 40:59 happy. I think um 41:03 >> so so specifically 41:04 >> I I I think I think much more the first 41:07 I think people really do want to 41:08 customize their software um and I mean I 41:12 believe this partly from my partly from 41:14 my own direct experience partly from my 41:16 interaction with others um and partly 41:19 because it's what I want to believe 41:21 about the species um in that uh you know 41:24 I I it's a 41:27 it's a grimmer world uh when you imagine 41:30 that certain disconnected architects of 41:33 tools are going to make all the relevant 41:35 decisions about them and then we are all 41:36 just um you know uh uh um uh downstream 41:42 you know in in the in the information 41:44 propagation you know flow we are all 41:46 downstream we have no agency and we are 41:49 at the mercy of their seen little 41:51 Britain you know the computer says no 41:53 skit uh like that vision of computer 41:56 says no as the future of our sector uh 41:58 seems uh seems rather bleak to me. So I 42:02 think I think it is the if you go way 42:05 back to the beginning uh so the long now 42:08 uh um has held many events here at the 42:12 at the Jazz Center. Um and I see the 42:14 long now as uh carrying the torch to 42:18 some extent for kind of California 42:21 humanism. And a lot of the early 42:23 thinking around the computer industry 42:25 was of course about how um computers 42:29 would not be you know vessels for 42:32 entertainment or even efficiency uh for 42:36 humans um but um but levers for us and 42:41 things that would enable us to create in 42:43 new ways and to you know fundamentally 42:45 have more agency. 42:47 um and people like Ted Nelson who um you 42:51 wrote Computer Lib and um live not far 42:54 from here. Um or Alan K and Dana Park 42:59 and so forth. Um and so I actually think 43:02 there's something very fundamental here 43:03 about what we want computers and 43:06 computing to be in society. And again, I 43:09 don't think it's for ordained which 43:11 direction we go. And 43:14 um 43:15 I mean no pressure. 43:18 >> I was Yeah, I hope we can change it 43:20 because I think that's one of the big um 43:22 motivators behind our launch of 43:24 application generation earlier today is 43:27 that we think that by uh allowing 43:30 tomorrow's developers to go generate 43:32 these applications that they can modify 43:34 themselves, they're closer to the 43:36 problem, they're closer to the domain. 43:37 And so we suspect that software quality 43:39 will increase as we do that because 43:41 there's less of a gap between the 43:42 builder and the user if you will. 43:43 >> And by the way, one one micro story on 43:45 this front um from Stripe is uh we um we 43:50 maybe it's another category of you know 43:51 internal tool. We have a little um agent 43:54 builder framework and platform where you 43:56 can define agents, you can wire them 43:58 together, you can have workflows, you 44:00 know, etc. That came from our um 44:04 financial operations team. didn't come 44:07 from some like you know super empowered 44:09 AI lab thing. It came from people to 44:12 your point very close to the work who 44:14 realized that if they built some 44:16 slightly superior tools and in this case 44:17 dramatically superior tools that it 44:20 would be you know directly relevant and 44:21 applicable and and a big improvement in 44:23 that domain but in fact it has ended up 44:25 yielding dividends uh across Stripe more 44:27 broadly and so what you're describing in 44:30 terms of franchising people in that 44:32 respect I mean at least from the Stripe 44:34 standpoint it's it's not just some uh 44:36 speculative gain you know we we've seen 44:37 it play out very tangibly over the last 44:39 six months. 44:39 >> Yeah. For you personally, what is the 44:42 minimum amount of malleability that 44:44 you're happy with in a software product? 44:47 >> Um, 44:48 >> iOS, I I presume you use an iPhone. 44:50 >> Yeah, 44:51 >> super malleable. 44:52 >> Well, I I I think the challenge is and 44:54 and this is again why like again I I I 44:57 speak positively about lisp and small 44:59 talk and mathematic and these things. I 45:00 think the challenge with infinite 45:02 malleability is it's easy to paint 45:03 yourself into a box, right? or into a 45:05 corner um and to end up with something 45:08 unmaintainable and that nobody else can 45:10 understand and that is really flaky and 45:14 flimsy and yeah just like it's kind of 45:16 malleability and maintainability and the 45:19 ability to collaborate with others and 45:20 so forth like reconciling those is I 45:22 think no no trivial thing uh and I I I 45:25 you know I don't understand your problem 45:27 domain nearly as well as you do but I 45:28 imagine that a lot of what retail has to 45:30 do is figure out well when somebody 45:31 breaks the application you know what's 45:33 the what's the roll back functionality 45:35 and how do you make sure that nobody 45:37 just blows everything up by accident and 45:38 so forth and and so I think there's I 45:40 think there's a lot of nuance and 45:41 subtlety there subject to be able to 45:43 solve all of those then I think you want 45:44 infinite malleability h 45:47 yeah we have a almost belief actually 45:50 that 45:52 uh governance enables freedom which is 45:55 that if you can define the security 45:56 guard rails and say hey you know here 45:58 are the building blocks that you can 45:59 assemble then actually people can go 46:01 crazy it's kind of like how I think I 46:04 don't watch Formula 1 but I think 46:05 Formula 1 drivers I think wear seat 46:06 belts and seat belts enable them to go 46:08 faster and so we think that 46:11 with governance comes freedom in fact so 46:14 that's the reach platform 46:16 >> we can workshop it in a bit 46:19 >> great well thank you Patrick it was 46:20 really great having you um thanks for 46:22 being a customer too 46:23 >> congratulations on all the progress um 46:25 and uh yeah it's congratulations on a 46:27 wonderful event 46:28 >> thank you
