Gumroad CEO's playbook to 40x his team's productivity with v0, Cursor, and Devin | Sahil Lavingia

Sahil Lavingia is the founder and CEO of Gumroad, where AI agents are already writing 41% of all code commits, and he’s targeting 80% by year’s end. Sahil demonstrates how this approach allows him to transform what would typically be two-week projects into two-hour implementations—a 40x productivity increase. What you’ll learn: The exact AI workflow Sahil uses to build features 40x faster—from prototyping in v0 to implementation with DevinHow Gumroad incentivizes AI adoption across the organization with $33,000 bounties for engineers who outperform the CEOHow to use component libraries like shadcn/ui for effective AI developmentHow AI is shifting engineering roles toward architecture and tech-debt removal while enabling designers and PMs to ship features directlyWhy spending more time on UX iteration becomes possible (and necessary) when implementation costs drop dramaticallyWhich organizational functions will be transformed by AI next— Brought to you by: Enterpret https://enterpret.com/howIAI—Customer SuperIntelligence Platform for Product and CX teams Vanta https://www.vanta.com/howiai—Automate compliance and simplify security with Vanta — Where to find Sahil Lavingia: Gumroad: https://gumroad.com/ Website: https://sahillavingia.com/ LinkedIn: https://www.linkedin.com/in/sahillavingia X: https://x.com/shl — Where to find Claire Vo: ChatPRD: https://www.chatprd.ai/ Website: https://clairevo.com/ LinkedIn: https://www.linkedin.com/in/clairevo/ X: https://x.com/clairevo — In this episode, we cover: (⁠00:00⁠) Sahil’s background (⁠02:31⁠) How soon will AI do most engineering? (⁠04:08⁠) Live demo: redesigning with v0, Devin, and Cursor (⁠09:30⁠) Using the right tools (⁠11:03⁠) Prototyping and iteration with AI (⁠19:45⁠) Incentivizing AI adoption in teams (⁠24:50⁠) “Magical” date-picker component development (⁠31:47⁠) AI’s impact on marketing, sales, and support (⁠36:50⁠) Deciding what to build when AI builds everything (⁠40:02⁠) Conclusion and final thoughts — Referenced: • Devin: https://devin.ai/ • Cursor: https://www.cursor.so/ • v0: https://v0.dev/ • Tobi Lütke’s tweet on how AI usage is now a baseline expectation at Shopify: https://x.com/tobi/status/1909231499448401946 • Flexile: https://app.flexile.com/ • shadcn: https://github.com/shadcn/ui • Gusto: https://gusto.com/ • GitHub: https://github.com/ • Figma: https://www.figma.com/ • Slack: https://slack.com/ • Vercel: https://vercel.com/ • Next.js: https://nextjs.org/ — Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email [email protected].

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Can you do something that used to take two weeks in two hours? And that's like a 40 times speed increase. So that's kind of like the number that I have in my head generally. Like, what's the most optimistic case? If you kind of remove all the auto next, something that would take 40 hours would take one hour.

• 0:14Claire Vo?

  If you're suggesting to us that AI is going to raise the bar on what's possible to do, you are certainly setting the standard.

• 0:21Speaker 00

  The majority of human engineering will be removing tech debt such that AI engineers can actually shift features. It's also scary, I think, which is why I think so many people shy away from this stuff. It's like there is this part why change is uncomfortable is that change can kill you. There's like a fear of change. It's like job security, right? But at the end of the day, I think it's sort of also job insecurity.

• 0:44Claire Vo?

  Hey, everyone. Welcome to How I AI, a podcast on how AI is transforming how we get things done. I'm Claire, product leader and AI obsessive here on a mission to help you build better with these new tools. Today I have an absolute powerhouse guest, Sehul Lavingia, CEO and founder of Gumroad. If you don't know Gumroad, it's the platform that has helped creators sell over a billion dollars of products directly to their audiences. Sehul's been at the bleeding edge using AI to transform how companies build products and write code, doing everything from open sourcing the entire Gumroad repo to paying his employees thousands of dollars if they can write more AI-powered code than he does. Today, he's going to show us exactly how he does it. Let's dive in.

• 1:29Claire Vo?

  This episode is brought to you by EnterPret. EnterPret is a customer intelligence platform used by leading CX and product orgs like Canva, Notion, Strava, Hinge, and Linear to leverage the voice of the customer and build best-in-class products. EnterPret unifies all customer conversations in real time, from Gong recordings to Zendesk tickets to Twitter threads, and makes it available for your team for analysis. What makes EnterPret unique is its ability to build and update a customer-specific knowledge graph that provides the most granular and accurate categorization of all customer feedback and connects that feedback to critical metrics like revenue and CSAP. If modernizing your voice of the customer program to a generational upgrade is a 2025 priority,

• 2:15Claire Vo?

  like customer-centric industry leaders Canva, Notion, and Linear, reach out to the team at enterpret.com/HowIAI.

• 2:23Claire Vo?

  That's E N T E R P R E T dot com slash how I A I. Hey, so I'm super excited to have you here. And before we dive into the demos, I wanted to call out something that you said a couple of days ago, which is Devon, the AI engineering agent, who I also love, is writing 41% of your PRs right now, and you expect it to go to 80% by the end of the year. So do you think that's the baseline that we should all be shooting for? Do you think you're way ahead of the curve? Where should we all be compared to that benchmark that you just said?

• 3:03Speaker 00

  I feel like I tell the team constantly, like, we have a lead, you know, but the lead is getting shorter and shorter every, every day, every week. There's a new model coming out. So I would say like by the end of next year, I would suspect that like every engineering team at any company is, you know, using Cursor and Devon and V-Zero and all these tools to ship, you know, multiple times faster. And the question is, is mostly like, how do we organization adapt such that those people can do so, right? Like the model necks are, are show up in other places. Like Togi just tweeted about, you know, his Salmify AI stuff today. And I think that becomes the question is like, how fast can you actually change your organization, your culture, especially when you're remote?

• 3:47Speaker 00

  It's harder to make these big changes across the org to get people to learn new stuff, to try and fail and cross share learnings, you know, all that, all that kind of stuff.

• 3:57Claire Vo?

  Okay, so we're going to do it one at a time, which is you're going to show us how you actually redesign or build something using these tools. So we'll get your screen up and you can walk us through how you think about things.

• 4:09Speaker 00

  Awesome. Yeah, I mean, so I think the coolest thing about all this AI stuff is that you get to spend more time doing what you really enjoy, which to me, and I think you as well, like, like solving customer problems. And for this product that we built is called Flexile. And it's, you know, you can think of it like a, like a store deal, but built specifically for the way that we run the business, which is like hiring a bunch of people, a lot of project based, a lot of hourly based, monthly retainers, all sorts of different types of people, remote, in person, and full time. Let them choose their equity split, manage your cap table, all of that stuff in like the same product. And one of the reasons I love AI is that I can basically just use the product and instead of running into some issue.

• 4:54Speaker 00

  And being like, hey, engineer, can you go solve this? And then spending all this time like writing up a spec. You know, then putting that into, you know. Sending that to a designer, that designer will then do like tomorrow or the next day will then do a mock, there'll be some back and forth. And then it'll go to like next week on Monday, it'll go to an engineer. They may have some questions that goes back to the designer and by the time it ships, you know, it makes it to production, even for something relatively trivial. You know, it's been two weeks or something, right? And so like, can you do something that used to take two weeks in two hours? And that's like a forty times speed increase. So that's kind of like the number that I have in my head generally, like what's like the most optimistic. Case, if you kind of remove all the bottlenecks, something that would take forty hours would take one hour.

• 5:38Speaker 00

  And that's pretty awesome. So even in this form, like pretty simple, and I built the software. So I'm like, you know, I'm not like a. Uh, saying, oh, it's so terrible, but there's always room to, to, to improve. And even on this one screen, which is the, the contractor invitation. Age. There's like already a couple of things that I noticed that like, aren't big enough to like, really ask someone to do. Everyone's busy. They have their own stuff that they're working on. But there are like a few things that I noticed, like, for example, the date picker is kind of terrible. Yeah. Like it just uses like the, you know, the native date picker. It's not humanized, you know, you can't type in like next Monday or this Monday or. Yeah. Have like a nice date picker, you know, if you go to like Shad CN.

• 6:19Speaker 00

  And this is the beauty of open sources, you know, and the why AI is so good is there's a lot of open source. You get like a nice state picker like this, right? It's like nicely humanized and you can do all sorts of cool stuff. So that's like one thing I noticed that I think is like a really good candidate. For this, I would go straight to Devon. I would, you know, it's, it's, it doesn't really, really need that much scoping. It's kind of just like replace date widget. You know, date picker in contractor invitation screen. With the Shad CN date picker, we might as well. I mean, the cool thing with Devon is you can like. But do that while you do other stuff. So there's, there's no risk really. So I can select the flex cell repo and you know, say for this specific page, like update the date picker

• 7:00Speaker 00

  from the browser native, you know, input to shads. Yeah. Import if required. I've never actually used this button, but again, this is a good example, even somebody using this stuff like you have to constantly like up your game to learn more. You know, basically I think there should be really like a rich text. Like this, like you can just type into it and you could type in like next Monday, I think resend. Add a cool demo like this where they have more like a natural language. Something like this and you could type in, in one hour tomorrow at nine AM, these sorts of things. Or Slack actually has something similar where if you go into a canvas.

• 7:43Speaker 00

  You know, this is our roadmap. If you type in like Thursday, right? This is kind of like what I think would be really cool. So I think this is also like, I'm going to have Devon do multiple versions and then we can take a look at how far gone on them. But this is kind of how I would generally work is I would just take these forms and say like, you know, build this form.

• 8:06Claire Vo?

  So you're, you're putting into V zero, you know, use this form, using the very descriptive, great requirements. Magical. So, and then you're going to use V zero to get a prototype?

• 8:20Speaker 00

  Yeah. So generally my flow is V zero Devin cursor. Is probably how I would say it. Like generally I V zero is my prototyping tool of choice. And once I have like a really good prototype that I'm happy with. Then I, I go to Devon and if Devon sort of fails to completely finish, then I open it up in cursor, though I, I think last week Devon launched this pairing mode where you can actually like jump in. And so I haven't really experimented with them yet. But that's presumably I would use something like that going forward where I could actually just jump in and fix the changes. The nice thing is Devon actually runs, you know, one of the most annoying things about being a developer is just getting set up.

• 9:04Speaker 00

  You know, just getting your your your development environment set up, your N variables, local host.

• 9:10Claire Vo?

  One of the tips I have for engineering organizations that are large, which is if you can make your environment easy to set up for AI, it's probably a lot easier to set up for new hires. So it pays off to sort of use that as a testing ground for how easy your it is for any new engineer to get started, whether or not AI. So you have V zero in theory going, there it goes.

• 9:34Speaker 00

  Yeah.

• 9:35Claire Vo?

  Okay, so you have V zero going on building your prototype. I have a question here, which is, you know, you mentioned Shad CN as your component library. Was that driven by, you know, using these AI tools and, you know, those those component libraries being out of the box? Or was that something you were looking at before?

• 9:53Speaker 00

  Yeah, it was a huge reason to switch and try to adopt a lot of these tools, both. And I think it's it's one reason that I think many people haven't really, it hasn't clicked, I guess. The AI stuff, they're like, oh, I tried it, didn't really work. It's not that good. It makes a lot of mistakes. It's, you know, basically it's faster for me to do it than to have AI do it. And I found that that's like a lot of it is just like AI is good at certain things. It's really good at front end, it's really good at React, it's really good at tailwind Shad C and stuff. So if you're not using those sorts of tools, you're not going to get the value. Like trying to ship something like this with like rails in the background backend and like hotwire or whatnot in the frontend. Like these just doesn't exist. Like you would have to spend all your time just getting this to work, you know, some jQuery calendar thing.

• 10:37Speaker 00

  You think, you know, that's how Gumroad was for a long time.

• 10:40Claire Vo?

  One of the things I wonder is if, you know, engineering leaders will decide on particular transitions or migrations to make just to power this stuff so that their teams can move a little bit faster because they're just seeing themselves be left so far behind compared to those who are maybe using some of these these libraries and technologies natively.

• 11:03Speaker 00

  I actually think that like the majority of human engineering will be. Removing tech debt such that AI engineers can actually ship features. Basically, like designers will be shipping features. Because if you think about it, what are they doing, right? They are thinking about what the feature should do. And then engineers are just basically setting up the groundwork, the framework, the defaults, the standards, the linting, the CI pipeline, the infrastructure, the dev setup such that designers actually are more and more capable over time of like, basically taking their idea. Like, if you were a designer, you, you know, you would like just design this part. You know, you'd design this, but you wouldn't design like all the little interactions in here.

• 11:45Speaker 00

  Right? Like you would just design like, like that because it would just take too, too long. Or you wouldn't even consider it because you didn't play, you know, you didn't. For example, like often you'll have a designer and they'd be like, didn't consider mobile.

• 11:56Claire Vo?

  Okay, so you got this design. Let's take a look at it. It looks pretty good. It has the magical date creation, which is type. There you go, type a, a magical date and it works. So it's not just the design, it's the functionality. And you said the next step for you from V zero is into Devon. So how does that transition work? What are you doing?

• 12:19Speaker 00

  Yeah, normally I would have a few back and forths here. You know, you could spend like three or four prompts, like 10, 20 minutes, like really nailing like the interaction, right? You may say like, you know, I added, add a, add a clear button or, you know, when you hit delete, it should actually delete. And this stuff will get only faster and faster and faster. But once, you know, once you're happy with what you have, normally I would take like the final prompt and I would just paste that into Devon, you know? And I would basically do similar to what I was doing before.

• 12:47Claire Vo?

  Yep.

• 12:47Speaker 00

  And I can see Devon doing its thing. Having lots of fun. And I can start a new Devon and basically do that, right? So like on here.

• 13:01Speaker 00

  On this page.

• 13:04Claire Vo?

  And you reuse the exact same prompt.

• 13:07Speaker 00

  Fill this form.

• 13:10Speaker 00

  Yeah, it's often, I mean, sometimes if I'm going back and forth and I learn stuff like I'm like, for example, this.

• 13:15Claire Vo?

  Yep.

• 13:16Speaker 00

  I may just add here, you know, like all the, these are kind of like learning where I could, it's basically, I'm like, oh, my spec could have been better. Like this, these are things a human engineer also would have maybe not done. You know, like I basically just kind of go back and forth and like build, basically I'm like, the, the V zero was kind of clarifying my spec in a way.

• 13:34Claire Vo?

  Do you use any of the code from V zero?

• 13:38Speaker 00

  Sometimes I do, like sometimes I'll take this and just use this command. And if I had put this and I went into cursor, uh, if I had cursor open on something, if I had said I had it open on this, for example, I would just go to terminal and I would just paste this right in. And it would put in this component. Uh, this is for a different repo, so it doesn't have shots the end, but it would basically like, you know, slot that file in and then I could reference the file and then, and, uh, you can also, I believe, just like, you know, you could, you could, uh, you could, uh, share it. And you could literally like, just give the, the URL effectively, right? Like this, and you could just say like, you know, mimic, mimic this, right? You know, you could say more things, you know?

• 14:23Speaker 00

  For example, I noticed that like in this thing, like I probably don't want the date to change in line. I like this parentheses is kind of weird. I'd probably add like a little note, you know? So I'd be like, uh, putting the date in parentheses is kind of weird. Put it below the input as a note.

• 14:45Claire Vo?

  Yeah.

• 14:46Speaker 00

  I love putting, I don't know what this does, but for some reason, I, if I feel like I'm vibing with this person, like they know what I mean when I say note, I mean, like, slightly smaller font size, like gray, you know, like note, like I feel like a designer would get it. So, you know, this is kind of like what I would give, uh, to Devon and, and then it would, it would, you know, run off and, and do its thing. It'll wake up and it'll do all these things, all, but all the stuff that I would basically do, right? Open cursor, get the thing, find, find the files that need to get changes. But I personally, one of the things I think is really, really important is spending more time in V zero. Like I think many people just like, they do a first pass and they, basically I think MVPs are no longer enough. Like you can actually spend like,

• 15:30Speaker 00

  10, 20, 30, 40 minutes here. If you know that Devon is going to be able to execute, like sometimes you don't want to spend too much time here because it just creates work for the engineer, right? You're like, oh, now I have to think about this and that and this and like all these like little bits that would make the customer feel really good. The user experience would go up, but the developer experience would go down, right? But if you know an AI is going to be implementing all of that stuff and they're going to do it at like a very high level of conscientiousness. You might say, oh, by the way, redesign it to like have this or like, you know, different roles, for example, right? Like different roles have different amounts, go a preview in the dropdown. You know, so one may be like 200 an hour, one may be like two per project, et cetera. You know, one may be 250K a year.

• 16:13Speaker 00

  Just for fun, I may say like one may even have multiple pay rates because I've been exploring this idea venerally. And I think part of the beauty of not doing it yourself is to happy accidents. Like AI may just take your, your spec and actually do it better than, than you would have. And so, yeah, that's kind of how I use it. And then I, I generally, if you're hosted, you know, depending on the projects. Our newer projects are all next to us, host on Versal, so they'll even give you.

• 16:43Claire Vo?

  Yep.

• 16:44Speaker 00

  Like a preview branch, right? Where, and I mostly love doing front end stuff with Devon. Actually now with, they have this pairing thing, I could actually go in and like run Rails Console and like check the backend stuff too.

• 16:57Speaker 00

  But, you know, with preview branches, like I love making changes to antework.com.

• 17:01Claire Vo?

  Yep.

• 17:02Speaker 00

  Because I can test them almost immediately, right? I can be like, you know, let's say a new person joined the company. You know, I can just say, hey, add this person who joined the company. This is their motto. By the way.

• 17:14Claire Vo?

  Yeah.

• 17:14Speaker 00

  Pick a fun icon that matches for them. Like I didn't pick any of these icons. I would not have made myself a king, for example. I just said, like, I basically just asked everyone in Slack, like, tell me if you want it to link anywhere. And what you're, what you want your, you know, your, your slogan to be. And then I asked Devin to actually do it and pick an icon for each person.

• 17:36Claire Vo?

  That brings me to something I was thinking about, which is when you were in V zero and you were asking it to add on features, I was playing the product manager in my brain and I was thinking, oh, in past lives, people would say, no, that's scope creep. We're just focused on the date picker, or we're just focused on updating this component. We can't kind of scope creep and add more and more features. And what I think, I think is interesting, I'm curious your point of view, is you can really start to go to the edges of some great user experience and it's less about how much time will this take? Or is it too complicated? It's more about what's actually going to work and be, be useful.

• 18:16Speaker 00

  Yeah, totally. And like, I often like, I mean, maybe this annoys some people at the company. But like, as I'm doing these zero stuff, like on other things, I'll be, I'll like go into the issue and be like,

• 18:28Speaker 00

  let's see if I have one here and like, I wanted to improve the multiple pay rates per role, as I mentioned, right? Like this is, and I'll, you know, I'll be like, I'll just go in here and be like, you know, like this one I had, I was like doing something with gusto and I kind of liked it, you know, and I was like, did I turn on this? And it's just free. People can ignore it if they want, but it's like free design research, you know? Uh, so all of a sudden they have an example of this.

• 18:54Claire Vo?

  This episode is brought to you by Vanta. Building a business? Achieving ISO 42001 compliance shows your customers that you're taking the necessary steps to ensure responsible usage and development of AI. But the process can be time consuming, tedious and very expensive. With Vanta, achieving compliance can be done in a fraction of the time and at a fraction of the cost. 95% of the required document templates are pre-built for you, accelerating the process, helping you demonstrate trustworthy AI practices and scale your business. Start with Vanta's free ISO 42001 checklist, which gives you a breakdown of the compliance process and the road ahead. Download it at vanta.com/ how I AI.

• 19:38Claire Vo?

  That's V A N T A dot com slash how I AI for the free compliance for AI checklist. You know, I see you as an individual being able to add this and fix that and update, you know, the homepage and all those things and use Devon sort of asynchronously. I'm curious how you've made this work at the team level. Like, what are the actual operational pieces that have to be in place for this to not degrade into chaos? And then what about just culture makes this work for you all?

• 20:11Speaker 00

  Yeah, I mean, first off, it's not easy. Change is uncomfortable, right? It requires work and energy and biologically, I feel like we're trying to save our energy all the time. So you have to, you know, you have to motivate people, you have to make it exciting. You know, there's a reason like colleges and classes are in person, right? Like there's a, it's like fun to train together. You know, it's easier to go to the gym in a gym than like at home in your bedroom. Right? Part of it is doing it myself, too. You know, like if your manager, your annoying boss is telling you to do something, it's different than like leading from the front a little bit. I often do like screen shares. Actually, like I recorded these videos and I recorded this one with Josh Pickford on YouTube, which is like three hours long.

• 20:55Speaker 00

  And I basically did it because I wanted to, I was like, this is, you know, got a lot of views, actually. Like I, that's how important I felt it was, not just for me, but for everybody. But I was like, I basically recorded it for the team. I had my team in mind as I was doing it, like, check out how cool it is. Like, imagine once we switched to tailwind, like how fast we can, you know, do this kind of thing and like how, you know, it sort of is part of that, bringing the energy. We also financially motivated people. So there's a couple of times here, I'll find you example of a Devon competition we did. So we did this competition where we did thirty. Thirty three thousand dollars, uh, split amongst whomever opens and merges more Devon PRs than me over the course of May.

• 21:36Speaker 00

  So, you know, it's a kind of a fun way to like motivate people to learn. It's time bound.

• 21:43Speaker 00

  And I actually did pretty well. Let's see the results. So I got, I got fourth. I opened twenty seven PRs with Devon and then three people beat me. So. It's, uh, and I, you know, I do a lot of easy wins, you know, so like props to all the engineers who, who, who did it. But yeah, I, this is all my, all my Devon PRs. A lot of people like, there's no way you use Devon. Like you're making it up. You're just trying to like go viral or whatever. I'm like, not really. Like, I'm just trying to like help people be more productive. I didn't know that was controversial. But, you know, there's like a lot of small things like remove this part of the homepage. There was this like recap that we do in Slack that's generated by AI, uh, that recaps like everything that shipped last week.

• 22:22Speaker 00

  And so I said, hey Devon, could you, you know, like, for example, these two things don't really need to be here because there's nothing under them. Right? So I just said, hey, at Devon, like, could you like, you know, only show the products that actually have shipments and like hide the other ones. And also like, some of these aren't really shipments. Like this one is only the backend. The frontend hasn't shipped yet. So like, make sure, you know, the update, the AI prompt that we're using for this, which by the way, I've never seen, like I, I just, I just know that there's an AI prompt that's, you know, involved. And, you know, it'll, and that's actually what this, uh, what this one is. Right? So. It found the Slack weekly, you know, recap and it, it, uh, you know, it, it, it made these changes and they created this PR. So we can actually go in and see this PR.

• 23:05Speaker 00

  And we can, uh, confirm my, my suspicion or not, which is, oh, turns out there is a prompt. So it's primarily on shipments, feature improvements and bug fixes. Right? Prioritize these categories. And then it also did something here. Uh, which is if we looked, it did this, it added a filter. So you know, basically only the projects that have more than one. The thing that I would critique about myself is that ideally we would have a test and maybe there is a test that I don't know about. So this is when the human would come in. I don't normally just hit merge on these things. You know, I would normally send this to somebody else and be like, hey, I did my best shot at this and you could see here for the slack recap, don't include project names.

• 23:48Speaker 00

  And then I, I pasted, uh, paste this, uh, the link to the, to the update. You know, I would normally like, say, hey, uh, and you make sure this looks good to you. And if there are any tests that need updating. And personally, I think this is way better that like someone has done most of the work for you. And basically I think humans will start the process. I think of it like flying a plane, like humans will take off, decide where to go and land typically, you know, do QA in the, in this context. But you know, not actually build, uh, write all this code, right? Like look up like, for example, you know, like dot filter versus dot trim or dot clean, or you know, like every language is different, right?

• 24:32Speaker 00

  But overall, I know rough amount of software architecture that like this is, you know, this is the right solution, right? To this problem. You're just adding a simple filter that removes the things. And ideally there would be a test. So I would have even higher confidence that, that this, uh, that this has done what it, what it should, you know?

• 24:50Claire Vo?

  One thing I was going to call out on the code you just wrote, showed was I find that these AI engineering tools are pretty good engineering citizens and that their code is well commented. They call out, you know, there's a little bit document doc strings and things like that that make it easier to parse some of those, those changes. Okay. So this is what we kicked off with. It's the native date picker replaced with the Shad CN one.

• 25:18Speaker 00

  Yeah. So the core problem I have here, which I guess, uh, is I need to make sure it works.

• 25:24Claire Vo?

  Okay. So you're showing the time lapse of Devon here, which is basically a screen recording of every single step along the way. Right now you're in the, in the terminal and the IDE. So you can actually replay step by step, how Devon got all this code done. It looks like it has in here. You know, reasoning and thought and planning.

• 25:49Speaker 00

  Exactly. And then the part that I'm looking for, and hopefully it did it, is that it would run, it would run the app locally, and it often does this, but sometimes if you have a complex app, and we just open source this so that it may have like broken, uh, but it would actually run the browser in its little local box, and then it would test it. So let me ask it to do that, run the browser. And it's awake, so it should pretty quickly start doing, doing that. And it's, this is Devon right here, Devon Box, Mr. Devon. Uh, and also, we can watch it on this one, right? It's, uh, it's doing its, its little thing. So because I had used magical in quotes, it presumed that I wanted to call it magical.

• 26:32Speaker 00

  And you could see, we actually open sourced it recently, so it's always working on an old repo, which is my guess of why it's not working exactly, right? Um, so now there are two things I decided to, you know, replace this input, the standard input with the type date with this new component. That's definitely correct. And then it, uh, created this, uh, this component where it goes through and it, it, uh, replaces it. So the thing that looks wrong here is it doesn't, it doesn't look like there's any AI magic. So something's sort of making, which maybe it doesn't need to. Maybe it's smart enough to know if I just type in today, tomorrow, or yesterday. But, you know, it, this probably wouldn't work if I said like three Sundays from now.

• 27:15Speaker 00

  But maybe that's fine. Maybe that's not actually what anyone would really do. This maybe even is a good example of something that, to your point, like I think AI has really good hygiene, engineer hygiene, where it is on a, on a micro level, like it's a better engineer

• 27:30Claire Vo?

  Yep.

• 27:31Speaker 00

  than human engineer would be. So you have to spend more time on like the architecture and like the planning aspect of it, making sure your execution is correct, like calling it magical date picker, maybe it's not the correct approach. I would probably call it natural language date picker or something like that, because magical doesn't really.

• 27:49Speaker 00

  Give you any insight into what's magical about it.

• 27:53Speaker 00

  But besides that, my guess is like this code, this parse natural language is actually like probably really, really robust, really good. Even this magical, like, like, check out the math on this guy, you know, like, whoa. Pretty simple, but like, how long would it, you know, how many times would you have to tweak it to like, oh, I, you know, like, I got it wrong, this like fancy add days. Function like it's pretty, uh, pretty clever how it's doing, how it's doing that. Find index, it's getting all, you know, it's basically figuring out like when you type next Monday, it's like three days and you're adding the days to get to the right, right day in the calendar and it's parsing the day based and it's like, that's like, this would be like a, you know, two years ago, this would be like a,

• 28:37Speaker 00

  so impressive for like, this would be almost like an engineering challenge, you know? Like I would hire an engineer based on this, which is, now they would just go to chat GPT and be like, and you know, it would, it would, it would work. So what you could do is to go back to the V zero, if you really wanted to enhance this, you know, you could just sort of take this component. And they're actually working on a way to like embed it, uh, you know, bring a component back into V zero and then you could like iterate on it. And the UX like a designer could even do that with V zero and then you could then, pull it back in, uh, to the code base. So you could kind of like do a lot of this, like, customer focus iteration, you know, uh, on the, uh, in a whizzy wig way, basically like dream weaver, you know, uh,

• 29:21Speaker 00

  versus like, like in code, I mean, like this, you have to think so hard to understand like what, how do you improve the user experience? Looking at this, right? The amount of like brain power and it just hurts my head.

• 29:33Claire Vo?

  What I think about is imagine that an engineer took this and went a week away and came back and said, here, I built your magical natural language. You know, date picker and you said, no, that's not really what I want. It feels like such an expensive iteration to throw out that code and do something new, whereas you, you can iterate that on that. You know, in a couple of minutes or a couple of hours over and over and not feel like you're wasting, you know, time and expense and people's honestly, people's like motivation and energy. I think about that a lot as well.

• 30:06Speaker 00

  Yeah. If you, if you spend two weeks on something and you're, you know, you're annoying CEO is like, nope.

• 30:13Speaker 00

  That's not what I meant. It's like, oh.

• 30:16Speaker 00

  You know, and then you gotta spend, you gotta go for a long walk in a coffee bag before you're back to work, right? So it's so much better to like really spend time, you know, here before.

• 30:27Claire Vo?

  Yeah, I just leaned in. So we got a redesign from V zero on this new employee in onboarding. And not only did it get new features, but you got a beautiful update on the um, date picker with some suggested common timeframes in there.

• 30:44Speaker 00

  Yeah, this is super smart, like, and all I did, by the way, I just had built a really dope natural language day picker for an HR product onboarding form. So probably the critical piece is like HR. Right? So it's like building it in the context of the problem you're trying to solve, which is, you know, if you're, if you're, if you're building like a like a party planning tool, you probably have like Christmas or, you know, like whatever, but you know, but in this case, uh, yeah, next Monday in two weeks, you know, probably it's gonna be next Monday. That's my guess is that is the most common, you know? But you could say, actually, we're, you know, we're, we're based in a, you know, a country in which like we, we work, we start on Sundays or Tuesdays and boom.

• 31:27Speaker 00

  You know, and you could do all sorts of interesting things or we're in a, you know, a place in which our date, you know, we put the day before the month or whatever. And so, yeah, just, uh, it just, uh, yeah, just a great opportunity to like really push the envelope and just like really spend more time, even I love this, to put the first name and last name next to each other. So you can read it out nicely.

• 31:50Claire Vo?

  So we just watched you ship in a new component, build a magical and now dope date picker for your employee onboarding tool. You've shown us how to get this done across your org and you proved that you're at least in the top five people shipping PRs with DM or with Devin at, at the company.

• 32:13Speaker 00

  Oh, by the way, this is merged. We got merged. So it looks like, uh, it's good. I made no mistakes. So yeah, next week it'll be better. Like, think about that would have been like, you know, at least 24 hours. So that's like a nice 10x speed increase.

• 32:29Claire Vo?

  This is a lot about engineering at Gumroad and you said, you know, 41% of your PRs are being written by Devon. You're writing code. What org is AI coming for next? Were the 80% of the work you think is going to be started by agents?

• 32:44Speaker 00

  I mean, I think you could see, you know, if you think about what are the orgs that exist? You know, it's like design, product, engineering, customer support, sales, marketing. And I really, I don't know, I actually was probably more optimistic on like full automation. I don't think we're going to really get there for a long time. There's just always like a higher level abstraction that you get to operate at. So, you know, there will be, for example, like, I think there's a lot more marketing automation that could happen in terms of like suggested tweets, like, you know, it could just watch what's happening in GitHub. It could like suggest, hey, this thing, we, you know, we have a content framework. We should post about this feature right, right now. I noticed myself having to like, you know, say, hey, this thing shipped in GitHub. By the way, only half of it shipped, only the backend.

• 33:26Speaker 00

  There's still all this nuance that I think, you know, marketing could get like a lot more efficient. Sales too, I think, like, for example, there are all these people who sign up, you know, they show up in our database, basically, right? Uh, and they're just emails, right? You go to Flexile, you sign up. But there's, I think, a lot more automation, you know, if someone signs up to Flexile with like soil and your times.com, you know, you could sort of queue up an email to them. There's so much focus on customer support. We even built our own customer support product with AI, which is great, you know, you can, you can talk to AI and it'll help you out. But this is all like reactive, you know? Well, what if I'm just browsing the page?

• 34:06Speaker 00

  And, you know, it, it knows that I'm in New York for my IP, you know, and you can wave at me. And, you know, it could be like, hey, what's up? How's New York? It's kind of cold out there. It's kind of raining now. And you'd be like, oh yeah, it is. It is raining in New York. Why do you care? And you can have a conversation and, you know, you're like, well, you know, it's kind of nice to be able to like talk about the problems customers are facing. So yeah, I mean, there's, there's, I think, sales, like making it more, making support more about sales, making it more proactive. I think making design more about product, making engineering more about architecture. You know, I think there's always going to be more and more stuff to do. Maybe even like prioritization. I think I spend a lot of my time, like, you know,

• 34:50Speaker 00

  for example, like going through GitHub and saying, okay, we have all these tasks. We have like 27 things. Like, what do we build first? And right now, it's like in my head, basically, I've seen all these things go live, or maybe even a better example, more people would relate to would be Gumroad. You know, we have this big roadmap. And, you know, I basically, I think I'm pretty good at this. But the reason I'm good at this is because I've seen every single thing ship. And so I kind of can very quickly sort of be like, okay, this is, this is, you know, gonna generate like, you know, maybe a hundred to $200,000 in value for creators, creator earnings, this will probably generate like 300, 400K. But then I have to also put on my engineering hat

• 35:31Speaker 00

  and say, okay, this is gonna take like 40 hours of an engineer's time. This is gonna take 300 hours of an, you know, and like do all this math, which you can go to business school, learn about byte and like all these things. And I could totally imagine like, you know, a button here that's like magical rank, right? And then it just like, sort of goes through and maybe you actually know that because you missed out this fact, it's actually much harder to ship, or we don't yet use chat to the end, so actually you're underestimating this. And it could like reprioritize it, right? And you could do all sorts of interesting things. That's like a huge, I mean, think about how many people at these large companies, especially like they're spending so much of their time on strategy, quote unquote, which is really just prioritization, right?

• 36:11Speaker 00

  And what we do is we just email all creators. And we just put together a list of things. And I just, we just sent this Google Doc to like, our top 200 creators in 2024. And we kind of like ranked this based on what they wanted from us, because it turns out like they're the ones paying our bills, right? And we started shipping. And imagine AI could take all it in all that data. I mean, all their sales volume, we have access to, right, in our database. And you could somehow kind of like, get a good, good sense of like, okay, what feedback should we be listening to? And, you know, you can imagine like you just hit a button that says like, you know, assign to Devon, you know? And then boom, it's done. I mean, that's another weird

• 36:55Speaker 00

  thing, though, right? Is like, well, if AI gets so good, why do you need to do everything?

• 37:02Speaker 00

  Why, why, what's the point in prioritization? Prioritization is a function of like limited resources. So that's a whole thing. It's like, I really, I mean, I would love to be in a place where I come into the office and I have no idea what's gonna happen. Like, I have no idea what we should even be building. And we spend time as a team, like, thinking about, like, what should we build? Like, we go, like, there's no, there's no issues in get, in, in.

• 37:21Claire Vo?

  Cleared, yep.

• 37:22Speaker 00

  Like, because every issue is, is solved. So, you know, it's cleared, we're in inbox zero. And so it's like, okay, well, what do we do? And then we sit around and talk and pontificate and eat lunch. And, you know, we really have to think hard about, like, oh, we should do something totally radical. Like, open source the whole thing, you know? Like, things that, like, an AI probably wouldn't suggest that. It wouldn't be in the, in the next token prediction, or even in the reasoning models, or like, okay, we should do really advanced content customization options. Okay, like, what, what is that? Okay, let's go design and be zero a bunch of that and do a lot of research.

• 37:57Speaker 00

  You know, I think research is obviously gonna get a lot better with, with AI, but, but still, humans have to go talk to people, ask them questions, user research, design research, um, market research. I think sales will always be important. I think marketing, like, I think marketing will be one of those things where, like, the average marketing, like, AI will get so good at marketing that, like, the level of what's interesting to a human, like, kind of like the, you know, that meme of the Saratoga Springs guy drinking the water and whatever, obviously, like, putting banana on his face? Like, to me, that's like a sign of how good AI is, that, like, that level of content production is now necessary to go viral. Like, it's insane, I can't imagine, like, how, how you took to make.

• 38:40Speaker 00

  It just, it's, it's so funny, like, it's so thoughtful and so many funny, different little, like, Easter eggs, you know? And I think that, like, that's kind of what will need to happen. It's just, like, you have to, like, up the game more and more and more. Like, you know, right now, artists can post, like, a painting on Instagram and people will be like, oh, amazing painting. But, like, in five years, it's gonna be like, you need to, like, post the frickin' movie. It's, it's like, that's just what people will expect, like, hey, we just wanna see your, like, thirty minute sci-fi movie that you did. And that's just, like, for free, sorry. That's just, like, that's what, that's what our, you know, that's what's happened to our, our. Our dopamine system, or we can spend like a whole day talking about like, how do we get better at recommending products on Gumroad? Is there a totally different kind of, you know, recommendation experience?

• 39:20Speaker 00

  That's like much more AI driven and much more natural language than just like a marketplace of, you know, feed of products. Where you can, it can remember things about your, your tastes, your preferences, you know, it can learn from you. We're launching a community feature, pretty excited about this next, uh, later this week, which is pretty big, but we, you know, there's just tons of, yeah. I mean, who knows? I mean, it's exciting. It's also scary, I think, which is why I think so many people shy away from this stuff. It's like, there is this like part of, part of why change is uncomfortable is that like, change can kill you, you know? Like, there's like a fear of change, like, you know, it's, it's like job security, right? But at the end of the day, I think it's sort of also job insecurity. Like, we don't know if like what we do will continue to be valuable.

• 40:03Claire Vo?

  I can say for sure, if you're, if you're suggesting to us that AI is gonna raise the bar on what's possible to do, you are certainly setting the standard. I think you're showing an entirely new way for teams to build. You're showing an entirely way for a leader to show up and actually contribute to the work product of the company, which we think is really inspirational. And then I think you're also showing, look, you just have to go learn these things and try things. And, you know, you're gonna get in a loop, but over time, you can actually become one of these leaders that's, that's on the leading edge as opposed to the lagging edge. So I, I think it's great and I think you're setting the standard for how EPD orgs are gonna operate in the future, if not, if not companies.

• 40:45Claire Vo?

  So we're gonna, we're gonna wrap up with a quick lightning round, two questions. If you could encourage people to learn just one of all, all your toys here you just showed, just one that you think is the highest impact, which one would it be?

• 41:00Speaker 00

  V zero, honestly. I mean, it's a bias, I think, because I spend so much time in product. I think of our more of an engineer, I think Cursor's agent mode is pretty crazy. I think if you're like a CEO of a company, I think Devon is like the most impressive, like the fact that you can just be in Slack and just talk to it and it will, it will do this. Is crazy. So I think a lot of it depends on like your role and you know what you value and what you think is the most important. But if you zero, I think it's just like the lowest hanging fruit. I think everyone is kind of familiar with Sigma. And I think a lot of people think that like, you know, okay, now people, no one questions that AI can code, even though a year ago, people were like, say, oh, we can't code or whatever. But now people are like, oh, we can't design. It doesn't have taste, you know? And so it's just like really, you know, like design a really

• 41:44Speaker 00

  nice onboarding wizard or a bank, you know, and like, watch it do a better UI for a bank than any bank has, you know? So I think that this is like something anyone can do. Like a kid could like, have fun with this. So I would say, man, I'd probably nominate V zero. And then, you know, the cool thing about V zero is it shows you what's possible. And so then if you want to execute on it, then you have to like, learn all the other, all the other tools. Uh, the other nice thing about V zero is that it comes with a URL. So you could build like a tic tac toe and send it to your friend and play tic tac toe, which is a kind of a nice, you know, a replator bolt on you or lovable like they, there, there's just so many.

• 42:25Claire Vo?

  We've talked a lot about how you are setting up incentives like bounties to get people to use AI or learn AI. But how do you get AI to do what you want? So I found that everybody has their own tactic. Like they're mean, they offer money. What is your strategy for getting AI to listen to you when it's in a little bit of a loop?

• 42:48Speaker 00

  I mean, honestly, capital letters, not in like a mean way, hopefully, hopefully it doesn't take it the wrong way. But I just think it's like, it is, you know, kind of like, it's kind of old school, I guess, you know, you have like, literally like lowercase and uppercase and like, it's just a really easy way of saying like, this part is really important. Like, please do not ignore this specific part. There's another hack that I love called et cetera. So if you want a list of things, you can name like two of them and then just say et cetera. And it will often like riff. It's really fun to just like be like, it's kind of like a test. You know, it's like you've come up with two or three, but you need 10.

• 43:29Speaker 00

  It's kind of a nice way of letting it like be more, more, more creative. So.

• 43:34Claire Vo?

  Well, this has been incredible and we have to wrap by showing you have not only redesigned your own product, but you've taken on the baking industry by generating a onboarding for a neo bank, apparently here in N V zero. I really appreciate you giving us a real look at how you're building with AI, both as an individual and as a team. I think you're definitely going to inspire tons of people to rethink how they show up at work. And I think a few folks are going to be looking over their shoulder, thinking that you're about to lap them once or twice on some of, on some of this building. So thank you so much for the time. Where can people find you and how can they be helpful to you?

• 44:13Speaker 00

  Yeah, you can find me on Twitter slash X. My handle is at S H L.

• 44:21Speaker 00

  And helpful, I don't know, just anytime you see something I've said that you disagree with or think of my thoughts could be improved upon, just reply, let me know, DM me. I'm always looking to get feedback and improve my thinking. So I just appreciate everyone tuning in and I'm excited to see what everyone builds.

• 44:41Claire Vo?

  Thank you so much.

• 44:43Speaker 00

  Thank you.

• 44:45Claire Vo?

  Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at how I A I pod.com. See you next time.