Subscribe
Share
Share
Embed
Serious People Podcast: An Operator's Manual to AI is a podcast for business leaders and operators running complex, real-world businesses. The ones selling supplements, managing caregivers, or running service crews. Not software.
Host Noah Levin brings nearly two decades of experience at Amazon, Whole Foods, and healthcare tech to weekly conversations about what AI outcomes work inside a business, and what doesn't.
Equal parts practical and irreverent, but useful above all else.
[00:00:00] Matt Van Horn: As the models have gotten better, as compound engineering has gotten better, and my ability to articulate what I'm looking for to my agents has gotten better, that I now, do not read the plan.
Like do not load the plan. First I got rid of my IDE, my code editor, then I got rid of my text editor. I don't have either anymore.
[00:00:21] Noah Levin: Welcome to Serious People, the show about doing serious work with AI. My guest today is Matt Van Horn. He co-founded the company that became Lyft and then the smart oven company June. But right now, no one has changed how I approach AI, especially AI coding, more than he has.
He's living that YOLO life, but he also knows what it takes to ship real software and hardware and when to move slowly and be careful.
If you're trying to understand how people at the frontier are actually building with agents, this one's for you.
So the reason I wanted to talk to you, besides the fact that you're, in my world a celebrity and I'm starstruck, is that…
[00:01:00] Matt Van Horn: I'm definitely not. I'm just, I'm just, just a nerd that's been tweeting since January pretty much
[00:01:04] Noah Levin: You're just a nerd. I mean, that's,
[00:01:06] Matt Van Horn: Been a nerd a long time, but I've been tweeting about being a nerd that's building things since January
[00:01:13] Noah Levin: But I think that what makes you different than a lot of the nerds tweeting about building things is that you've actually built physical things before. Like your Zimride, which is like a human-powered service in the real world, your, your June Oven, which is you have to like actually make things and put them into people's houses to make that business work.
And I think that's very different than a lot of AI Twitter, right? Uh, and, and so the point of this podcast is meant to find people who are actually operating non-technical businesses in the real world and help them overcome the AI hurdles. And so there's a lot of talking about constraints and a lot of talking about, learning loops and keeping things from drifting and compliance and all the unsexy stuff.
But you are like Mr. YOLO, right? You, you're like, you're Mr. Don't Read the Plan, Just Let It Ride. And so how do you go from the type of business where you have to plan like a hardware production cycle over like a one-year timeframe to just kind of letting the code write itself and see what happens?
It could mean a few things. I'm, I'm gonna take it to mean, so in Codex, they have YOLO mode, and Cursor used to have a YOLO mode, which is just like, “Don't ask me questions, just jump to the next step whenever you can.” And in, Claude, I, I highly recommend to dangerously skip permissions.
[00:02:30] Matt Van Horn: That's literally what it's called, dangerously skip permissions. And I just found myself early on, it'd be like, “Okay, is it okay if I access your web browser?” I just asked you to access my web browser. Yes, of course you can. Is it okay if I, go and look at your GitHub? Yeah, of course. I just asked you to, like, submit a PR.
Of course, you could go look at my GitHub. Or, okay, is it okay if, like, I'm gonna go search the web for this ti-- Ah. Like, it was, was driving me nuts. And then I was very lucky, and my, friend Dan Shapiro, who was my original, I'll call it Claude Code sensei, who's a, founder up in the Washington area that I'm, I'm close to.
And he's the one that actually, like, set up my Claude Code and made it so it was dangerously skip permissions, and has worked out, very well.
[00:03:19] Noah Levin: If anyone who's listening to this, like, hasn't used your advice on getting the compound engineering plugins, using CE plan for everything you do, make a plan no matter what the task is, like do that. But you said, I think it was on Peter's podcast a few days ago, like, “Don't even read the plan.”
It used to be read the plan, edit the plan. Now it's just there's a plan, let's make the plan happen
[00:03:43] Matt Van Horn: Yes. So, so it, it was crazy in, I remember, I want to say it was January or February. I, I was on This Week in Startups twice. It might have been the second time I was on there talking about Clawdbot at, back before it was called OpenClaw. And there was this kind of theory of like dark factories of like being able to run code while you sleep, and companies have done really good jobs about this.
And I, I talked a little bit about that on the show. And I, I think there's this change. Like it was a crazy idea then that was like, don't write code, don't read code, right? That was January, February me.
Don't write code, don't read code. No one should ever read the code. No one should ever write the code. No human should ever write the code. No human should ever read the code. That was back in January. And then as I got deeper and fell in love with compound engineering, it's like, okay, well, I'm not gonna read the code, but I am gonna read the plans because it made this great plan.
And if it has a really good plan file, then my agent hopefully will do a good job. So I better put a lot of effort into making the plan really, really good. And as the models have gotten better, as compound engineering has gotten better, and my ability to articulate what I'm looking for to my, my agents has gotten better, that I now, I declared this in maybe May, that do not read the plan.
Like do not load the plan. Like I stopped-- First, first I got rid of my IDE, my code editor, then I got rid of my text editor. I don't have either anymore. No IDE, no text editor. And I'll ask a few questions. I'll say, “Can you give me the TLDR, Too Long, Didn't Read, of what this plan is?” And I might have one question like, like, “You're just modifying my one repo, right?
You're not like submitting PRs to someone else's repo?” Or, “Can you ELI5 this? Can you explain this to me like I'm five years old?” And so I'll, I'll have a little bit of a conversation, but then I just trust that the plan is good. And it's not always good, for the record. Like it, like there are instances where I probably should have read the plan, but not to move at the speed that I'm looking to look at and the investment of my time in reading and perfecting plans versus trusting that the plan is good when the plan is perfectly great, let's call it ninety-five percent of times, that it's not worth reading the plan to focus on that five percent, which I can easily correct.
[00:06:18] Noah Levin: I have this experience now with, my looping engineering agents where I'll ask it for a status update and it'll tell me, “You know, well, this ticket, ticket number 155 is in progress and it's not going great 'cause blah, blah, blah.” And I, I have to just be like, “Bro, I, I don't know what ticket 155 is. Like, you, you wrote that. I didn't.”
[00:06:36] Matt Van Horn: Totally. That, that, that is, is very, very much me. Like, be like, “Okay.” It's like, oh, well, you, you designed it this way. I, I saw there was a funny, there was a funny X post that said that some people are mad that automatically when you, put agentic code on GitHub, it says like co-authored by Claude. And then some-someone said, I can't take the credit.
He says, “I actually think it's really nice of Claude to give me any co-authorship of any of my PRs on GitHub.”
[00:07:08] Noah Levin: Uh, it's spot on. You're deep in, uh, agentic engineering, and what I wanted to sort of poke at was you've been an operator of businesses before that are not about the code, or not only about the code, and now you're starting something now. I don't know what it is, and I'm assuming you're not ready to talk about it.
But when you think about starting to operate a business with these tools, how do you apply the same kind of lessons from vibe coding to operating? And can you be as, as don't read the plan about it when you're using it for operations?
[00:07:45] Matt Van Horn: I don't know who gets credit for, ending the word vibe coding, but I feel like the products, like the models got so good around Thanksgiving last year, Claude and, ChatGPT's models, that vibe coding ceased to exist 'cause that's when it was more of a toy, and that's when agentic engineering has started, and the models were getting so much better.
Like I remember I was a very, very early Cursor user, and it was almost a joke that when someone new joined, joined my company, I would teach the vibe coding class because that's funny, because I'm the only non-software engineer in the company. And so everyone would watch me, like, try and make a weather app, like, and pull in, like, an API, and it didn't work half the time.
And, but then again, this was before Thanksgiving last year, and so after Thanksgiving last year, it started to become real. So it's wild what you can build, but it's also, it also can be scary. Like when something breaks and you have a customer on the line and you don't know how anything works like that, that can be scary.
And the, the hope and the dream is that the fixing agent will also be good enough to fix that. And again, I think these things will get better and better, just like how good the models became Thanksgiving last year, and how good they're going to be Thanksgiving this year, that I, hope that software engineering will be a solved problem.
Obviously, if there's tons of risk, if you have real customers, real revenue, people have expectations versus someone that's like me for my open source project can live in YOLO mode like I was trying to get a, a feature into Last 30 Days last night, and so I was doing a ton of testing after I got this feature in.
And I realized that I broke everything related to X content on Last 30 Days. And the worst part is, I might have broken it like a month ago, which is crazy and sad, and no one, no one spotted it. But it was, it was really bad. It was searching X correctly, so it was telling you that it was searching X.
And so it would find like, if you were searching for me, it would find like @MVanHorn, but it would, wouldn't show any of the content unless Matt Van Horn was spelled out. Like, it was the most horrible bug. Anyway, so like, again, it's Last 30 Days. It's a free open source project. People can be forgiving.
It's fixed now, by the way. But like imagine this at scale with paying customers. That's terrifying. Or in hardware that you can't fix later. With June, so my, my last company, we made a countertop oven with a camera in it that would recognize most commonly cooked foods and, cook it perfectly.
And if we ship bad software on there and we break the over-the-air update function, like we literally brick that hardware and just owe that customer a new like $1,000 oven. And so the stakes are much higher when these type of things are on the line.
[00:10:46] Noah Levin: It sounds like you're almost, deliberately playing in the low-risk sandbox now because you get to move fast and break things, but you, you would be treating it a little bit differently if it were a different risk-reward calc
[00:10:59] Matt Van Horn: Absolutely. Like, like even in, in my, my own company, I'm super cautious about, me coming in as the, the founder guy messing with and fixing everything because I'm like, “Oh, my YOLO mode can do this.” So I, I, I do think it matters
[00:11:13] Noah Levin: I want to ask you questions about your new company at the meta level without forcing you to say anything you don't want to say about the company.
How are you setting up differently than you did June because of AI? Are there different roles in place, different funding patterns, different, types of problems that you're trying to solve? Like, how is, how is AI affecting your strategy, for starting this early stage thing?
[00:11:35] Matt Van Horn: I think one of the, great things that I'm, very fortunate to have is we are a company that was started in this post, I call it AC, you know, BC, AC, right? BC before Claude 4.5 and Codex that came out right after. And then after, like I, I feel just very lucky and fortunate from a timing perspective that I get to build a company in that era and can plan headcount around that very accordingly, right?
Like if I had at June, I feel like we were close to 100 people at our peak. And if I had a legacy team of 100 people on that June business and now I have Claude 4.5, that's not a s- ideal situation to be in. And so I feel very, very lucky and fortunate. And I, and I also feel lucky and fortunate that we were able to start building after that point versus even if we'd started the company like a year earlier, which is crazy.
[00:12:35] Noah Levin: Are you choosing a different problem because of AI, or is it just the size of the team and the execution against it?
[00:12:42] Matt Van Horn: I mean, June was an AI company. Like, we literally had an NVIDIA GPU on our oven. I have these photos of me, at NVIDIA in
2014. In 2014, I'm like in the lobby holding something big in my hands with like a blanket over it 'cause it's a big secret, and demoing for, for the NVIDIA executives
[00:13:01] Noah Levin: If, if I own a June oven right now, can I resell it at a premium, uh, or
lease it out to a hyperscaler?
[00:13:09] Matt Van Horn: Well, we, we joked that we wanted to do crypto mining on June Ovens when they were not cooking. And there was someone on the day of our launch, because the, NVIDIA TK1s run pretty hot, and it said that the June Oven actually doesn't have heating elements. They just overclock the GPU. Just overclock the GPU. And so we, we don't have heating ele-- we-- and I remember I thought that was so clever. I wanted to retweet it, but I didn't want to offend anyone at NVIDIA
[00:13:39] Noah Levin: Today's episode is sponsored by Valet. AI is getting very good at making things, dashboards, internal little websites, one-off tools, but AI is not good at sharing things with a team or a client without forwarding a Claude chat, handing over code, or opening up data that should stay private. Valet gives you a secure place to share what AI creates.
Take something you built in Claude, ChatGPT, or another tool and turn it into a real company resource with the access controls that a business needs. And underneath that simple sharing layer is the Valet agent platform, so the thing can eventually connect to data and tools and actions instead of just sitting there as a static demo.
One of my clients right now is about to get access to a whole new set of reporting in Claude. With Valet, they'll be able to turn these reports into web dashboards that everyone can view, filter, and export, and everyone can use it, even the folks who are still figuring out AI. The Valet team took a bet on this show before it existed, and I'm grateful to have them as our founding sponsor.
Go to valet.dev and see what you can make shareable.
[00:14:45] Noah Levin: Can I ask you about your, your personal productivity? Just getting ready for this podcast, you responded to my emails more quickly than anyone I've interacted with in a very long time, and I imagine you're getting a few emails. Is there a productivity thing going on behind the scenes for you that allows you to do that?
Or are you just hyper online?
[00:15:04] Matt Van Horn: Can you please tell my wife that I responded to an email quickly 'cause she would not believe you at all. I'd say I have not solved email. I use a product called Howie that I like for scheduling. I liked it so much that I begged the founder to let me invest, and he did. And so things like that are good.
I'd say using different CLIs within Claude Code, I'm able to get some productivity hacks. I love Granola, the transcription product for all, all meetings. And then within the Printing Press, someone built the Granola CLI, and so I'm able to just go into Claude Code, which also has access to my, calendar, and I'll be like, “Can you look at the Granola notes and my calendar and my email and tell me what meetings happened and what I haven't replied to yet, and what I should say from those meetings?”
And those three things with Claude Code as the brain can all come together and make a recommendation and kind of stack rank that, which is really nice.
[00:15:58] Noah Levin: When you do that, are you, triaging your email and responding to things inside of Claude Code or are you using a different surface for it?
[00:16:05] Matt Van Horn: I still haven't nailed this. I, I, I almost built a Claude Code email client. Like, I started that project and I was like, “I cannot maintain this.” I was like, “I, I know I could, and I could build this,” but I was like reimagining what Superhuman would be in Claude Code. And I love Superhuman. I was a Superhuman investor.
And thinking about what that is. So I, I have not nailed it yet, is, is the answer. And I would love for someone to, and I would love to give them money as an investor and as a user
[00:16:37] Noah Levin: This is where I get stuck a lot with AI productivity use cases is, there's nothing with the right form factor to interact with it, for like different use cases. Like there's Slack for notifications, but it gets noisy quickly. There's chat for sort of free form. Cowork has some useful little widgets for being able to edit an email or just send it or take other actions based on it.
But it's really, you really need a UI that like, moves and bends with you and does a whole bunch of different things to make it work.
[00:17:07] Matt Van Horn: I know what it looks like. I can see it in my mind's eye. If someone wants to build it, I-- it, it, it would be, it would feel like Claude Code, and it would have an intelligence. It would feel like a Claude Code modal where it's like, okay, now we're onto this, this email. I think this is the most important one.
Do you want to reply? Do you want to look at it tomorrow? I've already drafted this. And then you just talk to it like you would Clicky or, or Wispr Flow and kind of one at a time, almost like a, a Tinder for email with like a layer of intelligence on top. And I know there, there's a few people that have, have tried this, but nothing I've seen in, in the command line that, meets that.
[00:17:47] Noah Levin: After this call, I'm just gonna take my Granola transcript of this and throw that into a CE plan and see what happens, obviously.
[00:17:53] Matt Van Horn: I would love to see that product. If it's good, I will be, I will be a power user and never shut up about it. I, I never shut up about Su-Superhuman. I still love, I still love Superhuman, but I, feel like there's, there's an agentic email product that doesn't exist yet
[00:18:07] Noah Levin: I have trouble reconciling the version of you I see online, which is all in on these amazing open source projects that are changing the f**king world, and the fact that you're starting a company which I assume does not have anything to do with these three open source projects, although maybe I'm making that assumption wrong, wrongly.
But, what's motivating you to spend so much of your time and energy on these open source projects, and how is that feeding back into, you know, what you do for a living?
[00:18:38] Matt Van Horn: I think everything's related and everything is also learnings. Like I'll, have an idea for a product that might not feel related to my company or might not feel related to Last 30 Days, an already successful product that I should be spending more time on. And then suddenly it'll come together and it's like, oh wait, I get it.
And like, even though I wasn't being so thoughtful about how intentional that is, like, it just kind of works. Like I'll-- Like it's something very full circle, very simple. But full circle is the, uh, the new Digg that Kevin's work- Kevin Rose is working on. And I used to work at Digg. That was my first job out of college.
And the new Digg's awesome, and it, it's really, really neat. And as soon as Kevin showed it to me, I think it was in, wasn't even public yet, I was like, “I need to use the Printing Press to make a CLI for this.” And I did, and it's great. And then I was like, wait, Last 30 Days has some CLIs that it installs by default.
Like I should just install the Digg one by default. And so I did. And now it's, it's actually as of last night, it's onboarding into the ducks. And so the full experience of like how those things would all tie together to make everything else better, it's, it's just kind of happening. And I feel like that's true for anything in agentic engineering and AI, that things that you don't think are connected are just, they find ways to, to connect.
And again, it's all, it's all learnings and it's also the people I'm meeting right now from a recruiting perspective is just incredible. Like the amount of, talented agentic engineers that I'm so lucky to be in chat rooms with, and Discords with, and SMS with. And, being able to recruit from that pool is like major, major alpha for me and, and my company.
[00:20:18] Noah Levin: It sounds like your sort of celebrity in the coding world is, is like just paying dividends in, other ways that are not directly related to the projects, which is, pretty, pretty amazing
[00:20:27] Matt Van Horn: I'm just meeting a lot of smart people and I-- And it reminds me of the, the early 2007 internet days, which was such, such a fun time
[00:20:35] Noah Levin: Oh my God. One last advice question before we wrap. For someone who is enjoying vibe coding, if we call it that, and has never contributed to an open source project before, what would your advice be?
[00:20:48] Matt Van Horn: Use compound engineering. Don't just ask your generic, LLM to solve this. So I think there's two ways to do it. So one is you can look at an open issue on a project. I'd try and find a medium popular project. So a project with maybe fifty open issues. Don't start with OpenClaw, which has often thousands.
They, they finally figured out how to sort it, but like Hermes or OpenClaw, do not start there. Everyone is fighting to, to get to solve the problems. Ooh, you have an issue? And there's like fifteen people that submit a PR to solve it. So I would find a, a medium popular project, maybe like twenty-five to fifty issues where people do have problems.
And then I would copy the issue into Claude Code and say, “CE plan, help fix this issue, and let me know if you think we can solve this and submit a PR that would be successfully accepted here.” and do that. So that, that's one way. And then the other one, which is, is way more fun and exciting, is to actually build a feature you want.
That's, that's the more fun one. I'm using cmux a lot lately, and there's this annoying thing that, again, comes down from Ghostty. And so, cmux is, is based on a f-a fork of
Ghostty. And there's this thing where if you resize the window, the formatting gets messed up and like loses its shape.
And so I was like, “I want to fix this.” And so I screenshotted the error as it was going, and I said, “CE plan, can we fix this in cmux, or is this a Ghostty problem?” And it's like, “All right, I'll make the plan. I'll do the research.” And what it did was it, it dug in. It found that it's not dependent on Ghostty since they've already forked Ghostty.
The other thing that it found is it found someone else earlier that had tried to solve this problem. I don't know exactly what happened, but I, I feel like it got lost because cmux is so popular right now that the team didn't see it.
And, but it was, it was an older PR. By older, it's probably like thirty days old, not like six months old. It was like, “What, what do you wanna do?” As my agent asked. And I was like, “Okay, well, I, I want to make sure it's really good and a high quality PR, and I wanna solve this problem.”
And so it's like, okay, I'm gonna rebase and give credit to the person that made the original PR, and then I have an idea to make it better. And then we submitted it. And, I haven't met the founder of cmux, but we, we've been DM-ing. And so I sent it to him like, “Hey, pretty please, can you add this to your queue to consider adding?”
And, I'm hopeful that it will make it in there.
[00:23:19] Noah Levin: I am going to try doing that this week, and I will let you know how it goes. I'm so scared of having my PR rejected, but I think I'm gonna work through it in public and see how it goes.
[00:23:31] Matt Van Horn: I would just say to your agent, like almost, I don't want to say threaten it, that's not very nice. But I would just say like, “Hey, heads up, I've been banned from so many repos for sloppy, for sloppy agentic, uh, AI slop PRs. Make sure this is not AI slop.
And by the way, you, you are rare and unusual. My stuff gets a lot of likes. But the amount of people that actually do the work to try all 19 things in one of my blog posts and ideas, it's like very, very, very, very, very small percentage of people. So like…
[00:24:07] Noah Levin: Why do you... Why is that? You've compressed it down pretty far.
[00:24:11] Matt Van Horn: That's a good question. I, I, I-- People are busy in life. I think that's one. I think the other is, AI is so good, and we're all able to do things we've never been able to do before, that it's hard to imagine that there's another level to it, 'cause you already think you're at, like, such a crazy level, right?
And so obviously there's levels to it, but there's also so much that you don't know what to trust and what to dive deep into versus, “Oh, this one's fine.” And like, I do that too. People tell me about, about new things. Like, it took me a while to get, used to cmux. I'd been contributing to cmux months before I was using it, and people told me it was great, and I was like, “Oh, no. Ghostty for life.” Like, I was being stubborn. And so there's this certain like, you're moving so fast and so successful. Like, we're all feeling so successful in AI that like, why would you switch? Why would you make it better?
[00:25:06] Noah Levin: There's more than one right way to do things now, right? There's not like these well-worn paths that are the obvious way. Like, and even when, was the head of product for, Claude Code, I think, not, not Boris, but his, his partner was Lenny and was saying they're moving so fast that they're building features that are in some ways redundant across like Cowork and Code and even within one of them.
And so you ask like, what is the point of this feature versus that feature? There might not be a right answer. They might kind of do similar things. And so in all that noise, sometimes it's hard to believe that there actually is something that's objectively better, right? That's the next step.
And usually when there is a next thing, my experience is that everyone uses the word like two months before they know what it means. So that used to be agents, now it's loops.
Is there a place that people can go to, get more from you?
[00:25:55] Matt Van Horn: Pretty much X. MVanHorn, M-V-A-N-H-O-R-N on, on X
Last 30 Days, Printing Press and Agent Cookie are my, my three release open source projects
[00:26:05] Noah Levin: No one has changed the way I approach AI and specifically AI coding more than you. And so if you build a thing or tell me to do a thing, I'm not reading the plan, I'm just doing it at this point. So just take the transcript of this podcast and drop it into Claude and just tell it, tell it to do all the things.
Matt Van Horn, thank you for joining “Serious People.” Real pleasure to talk to you