Description Transcript
Watch this exclusive 45-minute fireside chat between Patrick Collison, co-founder and CEO of Stripe, and David Hsu, CEO of Retool, recorded live at Retool Summit 2025 on October 7, 2025. The two industry leaders explore the opportunities and challenges that AI presents for companies of all sizes in this intimate conversation from SFJAZZ Center in San Francisco.
This intimate conversation covers:
How large language models are reshaping enterprise software
The role of AI in financial infrastructure and internal tooling
Scaling AI-powered solutions across different company sizes
Real-world challenges and successes in AI implementation
Future predictions for AI in business software
Read more 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
