Latest / The Index Podcast / Matt Van Horn (Lyft) on AI Psychosis & Coding Agents That Never Sleep | The Index Podcast
Transcript
- 0:11SPEAKER_00: Hey everybody, I'm your host, Alex Cahaya.
- 0:14SPEAKER_00: Welcome to the Index.
- 0:16SPEAKER_00: And today I have a special guest, Matt Van Horn, who previously was one of the founders of Lyft and then went on to found June, which we can talk about, and is the creator of the last 30 days skill, which I thought was how we got connected, but really it wasn't that.
- 0:34SPEAKER_00: It was uh something else, which I'll just let Matt talk about his experience with that.
- 0:39SPEAKER_00: But um uh and you're also working on an undisclosed new project, uh, which we may or may not talk about, we'll find out.
- 0:45SPEAKER_00: Uh, but thanks for coming on the show.
- 0:47SPEAKER_01: Absolutely.
- 0:48SPEAKER_01: Thank you.
- 0:49SPEAKER_01: So yeah, it's it's kind of a hilarious story of how I ended up on this show.
- 0:54SPEAKER_01: So someone reached out to me, your producer, and said, Hey, we'd love to have you on the podcast.
- 1:01SPEAKER_01: And I was like, great, let's move to email and schedule.
- 1:03SPEAKER_01: I, you know, I clicked a few links, look look legit.
- 1:06SPEAKER_01: I'd heard of you before, and I was like, great.
- 1:09SPEAKER_01: And so clicked a few links and got it scheduled, and then it was the day of, and then a few hours before I looked at the link, and the the person emailed me as like, hey, are you joining?
- 1:20SPEAKER_01: And I'm like, it's literally on the calendar, and this is early.
- 1:23SPEAKER_01: Who who is this person that booked me?
- 1:26SPEAKER_01: Who is this producer?
- 1:28SPEAKER_01: And they had like a crypto photo, which again, not crazy for your world, Alex, right?
- 1:33SPEAKER_01: And then they had a lot of followers, but none that I really knew.
- 1:36SPEAKER_01: And then I was like, wait a minute, what is this link?
- 1:39SPEAKER_01: It wasn't like a Riverside link, it was like some other thing.
- 1:43SPEAKER_01: And I'm like, holy crap, this is a crypto scam.
- 1:47SPEAKER_01: Like, I'm literally gonna join this link and they're gonna steal my wallet.
- 1:52SPEAKER_01: I don't even have my wallet on this machine.
- 1:55SPEAKER_01: Or but like, I was like, holy crap, and so what do I do?
- 1:58SPEAKER_01: I DM you, and I'm like, hey, are we doing this?
- 2:01SPEAKER_01: But at that point I knew we kind of weren't, and I never clicked the link, and they sent a bunch of angry emails, and I DM'd you.
- 2:07SPEAKER_01: But what was cool is you and I had DM'd before, which I didn't realize.
- 2:10SPEAKER_01: You sent me a DM that says, Hey, I love last 30 days.
- 2:13SPEAKER_01: And I was like, awesome, thank you.
- 2:15SPEAKER_01: And then do you remember what you said when I said, Are we doing this podcast right now?
- 2:20SPEAKER_00: Yeah, you were like, Hey, we doing this.
- 2:22SPEAKER_00: I'm looking at right now, we'll we'll cut this in, like from my I'll get my producer to cut it in.
- 2:27SPEAKER_00: But I'm looking at it and it's this Billy XBT guy.
- 2:30SPEAKER_00: Everybody listen, this guy's a scammer.
- 2:32SPEAKER_00: He's been hitting up a bunch of people trying to act like he like works with us.
- 2:34SPEAKER_00: He does not.
- 2:35SPEAKER_00: And you're like, Oh, I think this guy's a scammer.
- 2:37SPEAKER_00: And I was like, Yep, he's a scammer.
- 2:39SPEAKER_00: And then I was like, I really want to have you on the show because I've been using the last 30 day skill all the time.
- 2:46SPEAKER_00: And I even sent you this uh selfie of me.
- 2:49SPEAKER_00: I was in New York at the time working with my colleagues in New York, and I sent you like a selfie of me with like a shotgun.
- 2:54SPEAKER_00: I was like, proof of being real, like I'm a real human, and here we are.
- 2:58SPEAKER_00: We made it happen.
- 2:59SPEAKER_01: Oh no, you said one more thing about North Korea, if I remember.
- 3:02SPEAKER_00: Oh yeah, yeah.
- 3:03SPEAKER_00: I said, I said, Kim, what did I say?
- 3:06SPEAKER_00: Kim Jong Kim Jong-un sucks, and I'm human.
- 3:10SPEAKER_00: And you laughed hilariously about that.
- 3:12SPEAKER_00: Funny enough though, like, you know, I cut my teeth in crypto the last like decade, you know, and have a lot of friends in the space, and it's happened to multiple multiple of them, like where they've been on calls and just kind of felt something was weird, and they've literally just said, you know, Kim Jong-un sucks, and the person like hangs up.
- 3:29SPEAKER_00: And it's like, yeah, they were clearly North Korean hackers.
- 3:32SPEAKER_00: So that's sort of your like get out of jail free card.
- 3:36SPEAKER_00: If you ever think you're talking to somebody who might be a bad actor from North Korea, you just tell them Kim Jong-un sucks, and you can ask them to say it.
- 3:42SPEAKER_00: And if they get really uncomfortable, then you know, that's a showstopper right there.
- 3:46SPEAKER_00: That's how you know.
- 3:47SPEAKER_01: Amazing.
- 3:47SPEAKER_01: Well, I'm glad you're real, unless you're scamming me right now, but uh this seems legit so far.
- 3:53SPEAKER_00: I'm not.
- 3:53SPEAKER_00: We are a legit podcast, and I'm a little I'm a real person.
- 3:57SPEAKER_00: I will say too, like the podcast, my why for the show is people are worth knowing.
- 4:01SPEAKER_00: You know, I believe people are worth knowing.
- 4:02SPEAKER_00: It's core to who I am.
- 4:04SPEAKER_00: Like I'm always curious to meet people and hear their story.
- 4:06SPEAKER_00: And I just happen to want to tell the stories of the people building the future of the internet.
- 4:10SPEAKER_00: Like that's my index.
- 4:11SPEAKER_00: It's a literally like a network of humans, and it always has been since the day we started the show.
- 4:16SPEAKER_00: It's just I happen to build my network in a specific niche, which was blockchain and crypto.
- 4:21SPEAKER_00: But I talked to a lot of people working across AI and a bunch of in fintech and a bunch of other spaces now.
- 4:26SPEAKER_00: And so we are working on that from a branding standpoint because people get confused and think it's like all about crypto on the show, but it's just not, you know, really it's about people and humans and building relationships and like understanding your story.
- 4:39SPEAKER_00: And another big like reason why I started podcasting, period, was because like no one in my family knew what I did as a professional.
- 4:45SPEAKER_00: You know, and I figured like if I could just help you meet some of these amazing humans that I work with that I talk to all the time, then you'll know what I do, you know, because it's very hard.
- 4:56SPEAKER_00: I'm an entrepreneur, it's kind of hard to define what exactly I do.
- 5:00SPEAKER_01: Have they figured it out yet?
- 5:01SPEAKER_00: Some people have, some people still don't know.
- 5:03SPEAKER_00: It's okay.
- 5:04SPEAKER_00: You know, I can't.
- 5:05SPEAKER_00: It's also a lot of fun for me to do these things.
- 5:08SPEAKER_00: And that that leads me to my first question that I always ask people like, you know, you're you founded Lyft, you're you founded June, you've had a pretty successful career to say the least, and you could be doing anything probably right now.
- 5:19SPEAKER_00: Like, but you're still building.
- 5:20SPEAKER_00: I mean, you shipped the last 30 days skill.
- 5:22SPEAKER_00: That's been immensely useful to me.
- 5:24SPEAKER_00: Um, what what gets you like what gets you out of bed in the morning to just keep doing this, to keep building things?
- 5:32SPEAKER_01: AI psychosis keeps me out of bed in the morning right now.
- 5:36SPEAKER_01: I literally wake up and I hear I hear fake phantom music, like AI music, that like isn't even real, and I think about the agents and wonder if they were successful overnight.
- 5:49SPEAKER_01: Like it's really bad, actually, the wake up.
- 5:51SPEAKER_01: I know you're asking about motivation, but literally the wake up, like it's it's it's pretty bad.
- 5:57SPEAKER_01: But uh wait so the other Wait, are you serious?
- 5:59SPEAKER_00: Is there like AI generated alarm clocks?
- 6:02SPEAKER_00: Like agentic alarm clocks that you said.
- 6:04SPEAKER_01: It's all my head.
- 6:05SPEAKER_01: It's all in my head.
- 6:06SPEAKER_01: Yes.
- 6:07SPEAKER_01: Like there are a few AI music artists that I follow, and I I hear their music in my head.
- 6:12SPEAKER_01: Yeah.
- 6:13SPEAKER_01: And you're like jamming.
- 6:15SPEAKER_01: And I'm jamming, and I know that I left two or three agents running overnight, and I wonder if they failed or were successful, and I like have to get out of bed to check.
- 6:24SPEAKER_01: And then I I also have a lot of kids.
- 6:27SPEAKER_01: I have four kids, and so usually it's also the two-year-old walking into our room.
- 6:31SPEAKER_01: I don't know why we got rid of his crib.
- 6:32SPEAKER_01: It was a probably bad decision, but terrible.
- 6:38SPEAKER_01: Yep, you think I would have learned with the prior three.
- 6:40SPEAKER_01: Nope.
- 6:41SPEAKER_00: Yeah.
- 6:41SPEAKER_00: But uh I have three kids too, so I'm familiar with this.
- 6:44SPEAKER_00: My young my oldest is eight, my youngest is three, and I have a middle that's five, so we're in the same boat.
- 6:49SPEAKER_01: Yep, we're we're we're in a similar, similar world.
- 6:52SPEAKER_01: But the to to answer your question more more directly, or I answer it too directly, the the kind of what what motivates me, what gets me excited is is new things and and building.
- 7:03SPEAKER_01: And I I think in this this world of like I've been a nerd my entire life.
- 7:08SPEAKER_01: Like I was the one that I used to camp out for Macworld to see Steve Jobs speak when I was a kid with my dad, and been a nerd my whole life, but I was always a pretty terrible software engineer.
- 7:19SPEAKER_01: And you know, could do some basic HTML or some Apple script or that sort of thing in high school.
- 7:25SPEAKER_01: But it since my started as a professional, I've never shipped anything of value that was mine from a from a code perspective.
- 7:32SPEAKER_01: Like there are zero lines of my code on the June oven right now.
- 7:37SPEAKER_01: And and so what's crazy is you know, it was fun like two years ago, a year ago, to like, oh look, it's kind of a toy.
- 7:47SPEAKER_01: Look, Matt can like make a little web, you know, a little web app now that's a little more advanced than before.
- 7:53SPEAKER_01: But then in November when Opus 4.5 came out, and then the latest version of Codex came out two weeks later, it became real, it became serious, and for the first time in my life, I could actually have an idea when I'm woken up by my AI psychosis or my two-year-old in the morning, and I can ship something of value by the end of the day.
- 8:15SPEAKER_01: And that's insane.
- 8:17SPEAKER_01: And I can also ship things that have no value too, for the record.
- 8:20SPEAKER_01: That's that's much easier.
- 8:21SPEAKER_01: But the fact that I could ship something of value, that's what gets me up and gets me excited.
- 8:26SPEAKER_00: That that totally resonates for me.
- 8:28SPEAKER_00: But I think about this a lot.
- 8:30SPEAKER_00: Uh, we live in this incredible time.
- 8:33SPEAKER_00: It's like our renaissance of creativity where all of your ideas can become real.
- 8:39SPEAKER_00: And I think builders and entrepreneurs, you know, we're we have vision, right?
- 8:43SPEAKER_00: We have a we have an idea and you want to see it become a reality.
- 8:46SPEAKER_00: And now, I mean, it's a blessing and a curse, right?
- 8:49SPEAKER_00: Like literally every good idea and every bad idea you have, you can just make.
- 8:54SPEAKER_00: And so I remember when I first started using, you know, these tools like six months ago to hack on stuff on my own.
- 9:01SPEAKER_00: And it was like I had like eight projects going at once, you know, and you're like losing your mind because you just I'm ADD, I'm pretty ADD, but that also like is my superpower, right?
- 9:11SPEAKER_00: So I get just so deep, and it's like all of a sudden you're like haven't slept in like four days.
- 9:16SPEAKER_00: You forget to eat all day, you know, because you're just like approving Claude, you know, to make things.
- 9:23SPEAKER_00: But you know, once you kind of wrap your head around it and you get better control of yourself with that, um, it's pretty amazing.
- 9:30SPEAKER_00: And you know, I I I remember um there's this teleprompter app that I use that's open source called Techstream, and it had a bug and it was really bothering me.
- 9:41SPEAKER_00: And I was trying to make some scripted content.
- 9:43SPEAKER_00: And so I just forked it and I fixed it with Claude and a team of agents that we built at my company.
- 9:49SPEAKER_00: We have this like team of coding agents that we built in-house that like we get good results with, they're open source.
- 9:54SPEAKER_00: And I submitted the PR and it got merged, and I didn't realize it for like weeks.
- 9:57SPEAKER_00: And I and it's some for some reason I went back to the repo and I was like, holy shit, I just contributed to an open source project.
- 10:04SPEAKER_00: Like, I'm not an engineer at all, I'm just a sales guy, like I'm just a BD guy.
- 10:08SPEAKER_00: That's it.
- 10:08SPEAKER_00: Like, I I wrote like really shitty Rails apps when I first started out in my career, just so I could like kind of know how the pieces fit together with technology, but I'm terrible at this, you know?
- 10:19SPEAKER_00: And uh and I was amazed.
- 10:21SPEAKER_00: I was like, I contributed to an app that's used by like tens of thousands of people.
- 10:26SPEAKER_00: That's crazy.
- 10:27SPEAKER_00: He actually merged it.
- 10:29SPEAKER_01: I can relate.
- 10:30SPEAKER_01: I I had a similar moment with OpenClaw, where I had a PR that made it to OpenClaw, and Peter Steinberger was like, thanks, M.
- 10:39SPEAKER_01: Van Horn.
- 10:39SPEAKER_01: And I was just like, I remember I tweeted, I was like, trying to play it cool, like Peter didn't just tweet, you know, mention me in the GitHub repo history, not playing it cool.
- 10:48SPEAKER_01: This is amazing, going crazy.
- 10:51SPEAKER_01: But but I again back to the AI psychosis, like a normal person would be like, oh, that's cool, maybe I'll submit a few more PRs if I find a bug or something like that.
- 10:59SPEAKER_01: And no, because I'm an insane human and a masochist, I built an entire infrastructure around open source contribution because I wanted that dopamine that I felt from that to continue.
- 11:14SPEAKER_01: So if you if you go to like github.com slash mvanhorn, you can see the hundreds of projects that I've been merged to.
- 11:25SPEAKER_01: And that's cool.
- 11:26SPEAKER_01: A lot of these I work on myself, I get excited about a feature.
- 11:30SPEAKER_01: Others all are done autonomously, and it's I I submit dozens and dozens of PRs every day.
- 11:39SPEAKER_01: So if you keep scrolling, so I'm in Python, go.
- 11:43SPEAKER_00: I'm looking at it.
- 11:44SPEAKER_01: So it it doesn't get 100% down.
- 11:49SPEAKER_00: What's the infrastructure?
- 11:50SPEAKER_00: Tell me tell me more about the uh the info that you built to do this.
- 11:53SPEAKER_00: What do you mean by that?
- 11:55SPEAKER_01: Yeah, so it it's so it's a it's a private skill that I I will not release because it'll break the internet uh if I did release it.
- 12:03SPEAKER_01: But at a high level, I built this Carpathian inspired autonomous system that allows me to contribute to open source without me really doing anything.
- 12:15SPEAKER_01: And obviously, there there are places where I can contribute manually.
- 12:19SPEAKER_01: Like there I I've built a wave.
- 12:21SPEAKER_00: Dude, this is this is insane.
- 12:24SPEAKER_00: And each one of those I'm still scrolling.
- 12:28SPEAKER_01: And everything you notice there's commas in between those, those are different features on different repos.
- 12:33SPEAKER_00: Yeah.
- 12:34SPEAKER_01: And so certain projects, like if you go to like G Stack.
- 12:38SPEAKER_00: Dude, 11,366 contributions in the last year.
- 12:43SPEAKER_00: That's wild.
- 12:45SPEAKER_00: Okay, I have to know more about this.
- 12:49SPEAKER_00: Wait, you want to keep sharing?
- 12:50SPEAKER_00: Sorry.
- 12:50SPEAKER_01: No, it's like, but just like on G Stack, like I think I'm the number three or number four human contributor.
- 12:56SPEAKER_01: And so a lot of these, I'm like a top 10 contributor.
- 13:00SPEAKER_01: And the my my friend Paige Von jokes.
- 13:03SPEAKER_01: You see my face.
- 13:04SPEAKER_01: If you see the like the little faces on the if you scroll up a little bit.
- 13:07SPEAKER_00: Yeah, right here.
- 13:07SPEAKER_00: Here you go.
- 13:08SPEAKER_01: See my face, one below there.
- 13:12SPEAKER_01: There you are.
- 13:12SPEAKER_01: Is that you?
- 13:12SPEAKER_01: So like yeah, that's hilarious.
- 13:14SPEAKER_01: You're on the top one.
- 13:14SPEAKER_01: My friend PageMon jokes that it's a where's Waldo in open source?
- 13:18SPEAKER_01: Where does my face show up as a contributor on all the exciting projects?
- 13:25SPEAKER_00: Dude, we need this at my company.
- 13:28SPEAKER_00: Whatever this private skill is, we need.
- 13:31SPEAKER_00: I love it.
- 13:32SPEAKER_00: Um, I want to run it.
- 13:33SPEAKER_00: So like I think I I think I told you this before, but my my co-founder is this guy named Brian Fox at Corbett's.
- 13:40SPEAKER_00: And he wrote Bash back in the like 80s and 90s.
- 13:43SPEAKER_00: He's like the co-author of the GPL licenses.
- 13:46SPEAKER_00: Um, our why for the company is just build together.
- 13:49SPEAKER_00: You know, we believe open source software is the purest form of innovation and the greatest legacy we can leave to the world.
- 13:54SPEAKER_00: Um, and that's like another reason why I got so excited about last 30 days, because you open sourced like this ridiculously useful tool.
- 14:04SPEAKER_00: Um, and I just started using it all the time.
- 14:06SPEAKER_00: Um, and we've even like I was telling you before the show, you know, integrated it into a bunch of our core workflows and our product.
- 14:12SPEAKER_00: And so we're gonna start like letting any customer that wants to use it use it.
- 14:16SPEAKER_00: And I'm curious like where that one came from, though.
- 14:19SPEAKER_00: Like, why did you build that?
- 14:20SPEAKER_00: What was the inspiration behind the last three days?
- 14:23SPEAKER_01: So it was so I I've been trying to figure out the best way to get prompting techniques because I felt like AI was moving so quickly that all my prompts were outdated, and the best place to get prompts was the the nerds on Reddit and the nerds on on X.
- 14:42SPEAKER_01: And so uh a something that I used to do uh at my my my company that we we're not talking about yet.
- 14:49SPEAKER_01: Um one of the things I used to do when someone joined the company was I I would teach the the vibe coding class, which was hilarious because I was the least competent person in the company to to teach any type of coding.
- 15:01SPEAKER_01: And so it was kind of like a funny joke type thing.
- 15:04SPEAKER_01: But anyway, the thing that I would do at the beginning of the vibe coding class was this is back when I was using cursor as my primary driver, which I haven't done for a long time, and I would say, like, okay, go, I would say in ChatGPT, go search Reddit for the best prompting techniques only for the last 30 days on how to use cursor, cursor's agent or whatever they called it at the time.
- 15:29SPEAKER_01: And it would go do their research because ChatGPT has a licensing deal with Reddit, and then it would pull that in, and then I would just copy and paste that into cursor, and it wouldn't do anything else.
- 15:38SPEAKER_01: It would kind of I'd have it write the prompt for me.
- 15:41SPEAKER_01: And as I you know had even that was when the tools were not not very useful pre-Opus 4.5.
- 15:50SPEAKER_01: And then suddenly I was on the ski mountain, I had my skis, I had my daughter who had uh her ski racing class, and the snow was terrible in Washington this year, it was not good.
- 16:03SPEAKER_01: And I was like, I don't really want to ski.
- 16:06SPEAKER_01: I brought my laptop.
- 16:08SPEAKER_01: Should I go to try and build that thing that I was hacking into just a chat GPT window?
- 16:14SPEAKER_01: Should I decide to build it for myself today instead of skiing?
- 16:18SPEAKER_01: And I was like, I could try.
- 16:20SPEAKER_01: And so I don't know how long I had, eight hours, a little less than that.
- 16:25SPEAKER_01: Um and I just forced built it in in Cloud Code.
- 16:32SPEAKER_01: Even made a launch video the same day using Remotion.
- 16:36SPEAKER_01: I prefer Hyperframes now, but Remotion was felt like sorcery that day that I could make a launch video.
- 16:42SPEAKER_01: And I submitted my first ever open source project and then I tweeted it.
- 16:48SPEAKER_01: I didn't have that many followers.
- 16:50SPEAKER_01: I was pretty inactive on X.
- 16:52SPEAKER_01: To be honest, I had you know, I had like 10,000 followers, but it was like from like 15 years ago, from like when I signed up for my Twitter account in like 2007 or whatever that was, and it got a million views that video, that launch video for some reason.
- 17:09SPEAKER_01: Wild people wanted this thing that I accidentally built, and the the magic behind it was you could plug in your API key.
- 17:19SPEAKER_01: So it was built for Cloud Code, so Cloud Code was kind of like the search engine and the brain, and it used uh an open an open AI key to get access to Reddit.
- 17:29SPEAKER_01: And I forget when we added X or how we added X, but I think it came soon after.
- 17:36SPEAKER_01: And then yeah, it became a thing.
- 17:38SPEAKER_00: That's wild.
- 17:39SPEAKER_00: And so now you've got like, I was looking at your profile just now.
- 17:43SPEAKER_00: You've got 37,000 followers on X.
- 17:46SPEAKER_00: And I wonder, is that do you think all attributed to that specific like the last 30 days thing?
- 17:52SPEAKER_00: Or do you think all this work you've been doing on open source with like that autonomous open source engine has also like driven a ton of people to your profile?
- 18:00SPEAKER_01: Yeah, I think it's I've just become obsessed with agentic AI.
- 18:04SPEAKER_01: I think, especially as as a non-classically trained engineer, that I've been able to be successful, and I've been very forward with all my tips and tricks.
- 18:14SPEAKER_01: I've I've written two articles called All My Agenc Engineering Hacks so far.
- 18:19SPEAKER_01: Uh my most recent one was in June.
- 18:22SPEAKER_01: And it's a little it's it's long, it's a lot of work.
- 18:24SPEAKER_01: I mean, what you what I recommend is actually just copy-paste the whole thing to your agent and say do everything here and uh so you don't have to read the article, which is great.
- 18:32SPEAKER_01: But uh it's I it's I've kind of built up this community around this agentic engineering world, and I'm having a lot of fun and I'm shipping things.
- 18:41SPEAKER_01: I I also built a product called the the printing press with uh with Trevin, who I met just because he was doing agentic engineering things with with compound engineering, and we we started talking, and that that turned into something crazy.
- 18:54SPEAKER_01: You you were talking about uh not eating, not sleeping earlier.
- 18:57SPEAKER_01: One time I was trying to use the printing press to print a CLI to order lunch downstairs, and it wasn't working.
- 19:06SPEAKER_01: And so I punished myself and didn't let myself eat lunch until I made a CLI that would order it for me.
- 19:14SPEAKER_01: Eventually it worked.
- 19:15SPEAKER_01: I made the order to go CLI.
- 19:18SPEAKER_00: That's hilarious.
- 19:19SPEAKER_00: Amazing.
- 19:20SPEAKER_00: And so um I want to go back to your autonomous open source contributing thing.
- 19:26SPEAKER_00: Can you tell me more about how that actually works?
- 19:29SPEAKER_00: I know you were saying you're like not gonna release it open source because it's gonna break the internet, but what actually is it?
- 19:35SPEAKER_01: It's so it's a good question.
- 19:37SPEAKER_01: I think I can talk about what like the it's it's a beast now.
- 19:41SPEAKER_01: But the kind of the original version, which is very simple, was go go find an issue on uh open claw.
- 19:49SPEAKER_01: Openclaw is not a good first starting point anymore, by the way.
- 19:52SPEAKER_01: I do not recommend it.
- 19:53SPEAKER_01: Uh OpenClaw's great, but don't recommend if you're trying to contribute to open source to start there.
- 19:58SPEAKER_01: But my original way I would do it is I would say, hey, here's here's three issues that people have filed.
- 20:07SPEAKER_01: And then I would use compound engineering CE plan, make a plan to fix these issues with code and submit a PR.
- 20:16SPEAKER_01: And compound engineering just kind of did a really good job of that.
- 20:20SPEAKER_01: And so they kind of that's that's how it started, and then what it ended up becoming is something way crazier than that.
- 20:29SPEAKER_01: That what's amazing is I have all the data behind all my merged PRs.
- 20:35SPEAKER_01: And so it's I've kind of built this carpathis-style loop to understand what's been successful and what hasn't, what PRs get merged, which ones don't, what the feedback is, and so it's built this self-learning loop of how to successfully contribute autonomously in open source because I've had so many thousands of PRs and thousands of merges that it's it's learned.
- 21:07SPEAKER_01: And again, I've also gotten in trouble.
- 21:09SPEAKER_01: Like in my first week of the tool, I got banned from two projects that I love.
- 21:14SPEAKER_00: Oh no.
- 21:14SPEAKER_01: One was Rust, they're still very anti-AI in the Rust repo.
- 21:20SPEAKER_01: And uh the second one was uh Ghosty, which is since less GitHub because they're they're frustrated with people like me, but I I officially got banned from both of those.
- 21:32SPEAKER_01: But I I haven't gotten banned lately, knock on knock on something.
- 21:36SPEAKER_00: Well, it's starting, it's working, you know, it's gotten better.
- 21:39SPEAKER_00: Um can you tell me what is when you say compound engineering, what do you what do you mean?
- 21:43SPEAKER_00: What is that?
- 21:44SPEAKER_01: Yeah, so there's a company called Every that makes a series of products.
- 21:48SPEAKER_01: Every.2, I think, is their URL.
- 21:50SPEAKER_01: And there's a um a skill that that Kieran and Trevin run called compound engineering, which is this Skill that you can run in Cloud Code or in Codex.
- 22:03SPEAKER_01: And it's just magical.
- 22:05SPEAKER_01: I don't know how to describe it, but it gives you this magical ability to write these in-depth plan.md files, and then the agent executes on it.
- 22:15SPEAKER_01: And in general, agents before slash goal, goal slash goal has changed this, but before slash goal, agents were very lazy and would try and take the easiest path to making you happy across everything.
- 22:30SPEAKER_01: And so Compat Engineering does a really good job of writing, doing a deep research thorough plan before executing on the product.
- 22:38SPEAKER_00: And so this is like a core thing that you do is work as use this.
- 22:43SPEAKER_01: Yep.
- 22:43SPEAKER_01: So my my last hacks article, which got a million, a million views on X, kind of a lot of the core of it is how I use Compat Engineering in my daily workflows.
- 22:53SPEAKER_00: That's so cool.
- 22:54SPEAKER_01: It's also it's it's a such a well-designed skill.
- 22:57SPEAKER_01: It might be the best skill I've ever seen, that whenever I'm building a skill for myself, I literally will point it at say, hey, go look at the compound engineering skill and learn what's so thoughtful about this skill.
- 23:12SPEAKER_01: So when you make a you know matte custom skill for this purpose, don't build it from scratch.
- 23:17SPEAKER_01: Steal the infrastructure of what compound engineering did, because it's so well designed.
- 23:21SPEAKER_00: Very cool.
- 23:22SPEAKER_00: I'm gonna have to check that out, and we're probably gonna end up integrating it into our platform as well, because it'll be really good.
- 23:26SPEAKER_01: Yeah, the one I'm doing right now is I'm because I have a lot more codex credits than I do Cloud Code credits, and so I'm finding myself shipping as much code to Codex as possible.
- 23:36SPEAKER_01: I'm actually not up for the work, but I have Codex taking my open source contributing engine, and I said Compat Engineering did a really good job of making their skill, which used to be Cloud Code only, grading Codex.
- 23:50SPEAKER_01: Go make my open source skill grading codex, and so it's been running for like 12 hours trying to do that.
- 23:55SPEAKER_01: I haven't tested it yet, but as like go study what compound engineering did, they did a really good job of solving this.
- 24:02SPEAKER_00: Is it just looping?
- 24:03SPEAKER_00: It's just like in a loop trying to get it right.
- 24:06SPEAKER_01: Exactly.
- 24:07SPEAKER_01: So we'll see if it works at all.
- 24:10SPEAKER_00: That's awesome.
- 24:10SPEAKER_00: And meanwhile, we're sitting here having a great conversation, getting to know each other while it's doing all the work.
- 24:15SPEAKER_00: That's great.
- 24:16SPEAKER_00: So we're we're getting towards the top of the show.
- 24:18SPEAKER_00: What and I always ask this like, what have I not asked you that I should have asked?
- 24:22SPEAKER_01: Yeah, I I think I think I don't know what it's about asking, but something you said earlier that like that there's this like it's so easy to build anything, that there's not enough users and not enough money to actually use these things, and people are have so much AI psychosis going on, and so this which which is bad, it's exciting, but it also can be bad if if you're building a new place that that no one wants.
- 24:48SPEAKER_01: And so I think one of the most important things is how do we know what people actually want and gather that human signal?
- 24:56SPEAKER_01: And I think that's one of the biggest challenges right now to figure out does anyone want this across anything that anyone's building.
- 25:04SPEAKER_00: I think I think we're seeing something similar happen in enterprise, and this is actually a question I wanted to ask you.
- 25:10SPEAKER_00: Um, a lot of companies they know they need like agent agents, they need to be identified, and they know they need some way to do that.
- 25:21SPEAKER_00: I think with maybe uh mythos and Fable 5 getting kind of pulled out of the market temporarily, like a lot of them woke up and they're like, oh, we should probably like own our intelligence stack and use open source tools and models.
- 25:35SPEAKER_00: What is your take on that?
- 25:37SPEAKER_00: On all that, like one, open source, like open intelligence, owning your intelligence stack, and two, like where do they start?
- 25:44SPEAKER_00: Because a lot of these companies, when I talk to them, they're like, I don't, I know I need this, but I don't know what this is.
- 25:50SPEAKER_01: Yeah, so I I love open source.
- 25:52SPEAKER_01: Uh I love that there's open source models, and I think that's good for the world.
- 25:56SPEAKER_01: But for me personally, I need to use the best models, and so that means whatever that is, flavor the moment.
- 26:02SPEAKER_01: So I do not care about local models.
- 26:04SPEAKER_01: I do not like I'm happy to spend the tokens.
- 26:07SPEAKER_01: I've I've told my team they can spend their salary in tokens.
- 26:11SPEAKER_01: Like I am all for that future of using the best the best tool for the job.
- 26:19SPEAKER_01: And so I'm I'm happy to pay for the the mythos credits and all that again, be smart and optimize, use codex when it's good, which it is very, very good.
- 26:27SPEAKER_01: Sol is very, very good.
- 26:29SPEAKER_01: But you know, to uh llama is kind of the classic example of something that like I would not want to run.
- 26:35SPEAKER_01: And again, maybe maybe with with Kimmy, which I again I don't have the horsepower to run Kimmy, and it's not out yet from uh open source, they haven't released the open source side yet, but like those sorts of things could be interesting, like a a bedrock style hosted Kimmy in the US could be really interesting.
- 26:53SPEAKER_00: Yeah, that makes sense.
- 26:54SPEAKER_00: I mean, I think we look at it the same way.
- 26:56SPEAKER_00: We we've been using a bunch of open source models inside the product we built and getting really good performance out of them, but we've tuned it a lot, you know.
- 27:04SPEAKER_00: It takes effort to figure out how to get the the most out of it.
- 27:07SPEAKER_00: But yeah, that was that was my only question for you uh that I that we hadn't asked.
- 27:11SPEAKER_00: I really appreciate you coming on the show and thanks for all the work you do.
- 27:14SPEAKER_00: Thanks for the last 30A skill.
- 27:15SPEAKER_00: I've been enjoying it, and um, I would love to have you back on when you're ready to announce this new thing.
- 27:21SPEAKER_00: I'm like super curious, and I didn't really ask you about it because I knew you didn't want to talk about it.
- 27:25SPEAKER_00: So you'll have to come back on and tell me about it when you're ready.
- 27:28SPEAKER_00: Awesome.
- 27:28SPEAKER_00: Sounds good.
- 27:29SPEAKER_00: Thank you, Alex.