Interesting. Said software will eat the world. I feel like coding will eat all knowledge work, right? And we're kind of going that direction already.
The whole agent stack is emerging: identity, payments, marketing, even CLI versus MCP. Like all of these are really new things, and I think a lot of the old playbook goes away.
Yeah. It's a whole new world. I hope more companies will stay small, and I think the founders of this generation realize that, like, they want to stay as small as possible.
Yeah. And instead of having like a 10-person product team, you have like a two- or three-person product team, and you just have a bunch of agents help you.
Someone tweeted that, like, the job market is so bad that I can only pursue my dreams now, or something like that. So, like, it's like, you know, it's like, yeah. So maybe you lost your job, but, like, now you can actually do your own thing and have a shot at actually achieving it, you know?
All right. Welcome everyone. I've got my friend Peter Yang here. Peter, welcome.
Yeah. Great to be here. It's good to see you again.
Yeah. Yeah. It's great to see you. Peter and I worked together at Credit Karma for a brief stint. Um, and then we went our separate ways, and, um, you know, I rediscovered, um, Peter from his prolific posts on X and your YouTube. And, you know, you've got a little bit of a Clark Kent-Superman thing going because you still got a day job, right?
That's right. I still have a day job. Yes.
Yeah. Can you share where?
Uh, yeah, I work on Roblox as a PM.
Amazing. Roblox, an a16z portfolio company.
One of my favorites. Um, well, incredible, man. Let's get right into it. Maybe I'll start with a softball fun question, and then, you know, we're going to talk about everything in the Claw ecosystem. We're going to talk about coding agents. We're talking about a little bit about maybe what students should study, advice, and some of the things that you've talked about online.
Um, maybe to start, what is the name of your... Well, how many Claws do you have? And tell me their names.
Uh, I only have one. I call her Zoe.
But I have like multiple conversations going with her.
Um, I have two girls, and I was going to call my younger one Zoe, and I did not. So I call my OpenClaw Zoe instead.
I see. I see. Yes. Yeah. This is your fallback plan. Peter, tell me a little bit about, you know, OpenClaw, how you discovered it, um, how you're using it today, and what you think the implications are.
Yeah, I was lucky to interview Peter Steinberger before he became super famous and the whole thing blew up. And, um, then right after I interviewed him, I set up the thing. It took forever to set up. It was super janky. Um, and yeah, it does a lot of things for me. It, like, pulls analytics for me across, um, YouTube and, like, my Mercury banking account. It, um, can update Google documents for me. It can build little web apps for me. But if I was honest with you, dude, like, I mostly just talk to it through voice and get voice replies. And, like, every other day I ask it to give me, like, a pep talk, like, you know, give me, like, look through all your memory and, like, give me some, like, deep insights that I don't know about.
And, like, it gave me, like, I remember I was on a walk and it gave me, like, a three-minute pep talk that was, like, really amazing, really amazing. Like, it was something about, like, oh, you're, like, talking to me about your career, business, and blah blah blah, and, like, your job, but, like, just remember that your kids, you know, 7 and 4, are going to grow up very soon and they're not going to want to spend time with you. So, like, you should really optimize for, you know, them instead.
Yeah, that's really cool. And I mean, very cool, but also something that all the language models could have done prior. Yeah. So what's the difference between this and in a use case like that?
Yeah, that's a very good question. So I don't know, because I have it installed on Telegram, it just feels like more personal than using, like, Claude or ChatGPT. And it just feels like something I can, like, text in bed. It's probably not very healthy, but, like, I text to it in bed. I talk to it during my commute, and it feels like it feels more like a personal, like, actual human.
Yeah. Yeah. So then, so how much for you is OpenClaw about the kind of interface, like, pushing it to messaging and, you know, maybe helping to trick our brain into feeling like, hey, this is a person or person-esque thing, um, versus all the other components of the stack, the self-modification, skills directory, um, all the rest?
I think it's probably like 70, 80% just, like, the personable part of it because I mostly just talk to it and, like, you know, through voice. But I also think, like, it's something for... first of all, it is pretty janky. It tends to forget things a lot.
You have to keep reminding it.
But, like, any kind of zany idea that I have, I just have to talk to it and it can probably just do, like, it's kind of just, like... um, like the other day I was doing voice replies with it. I was like, 'Hey, can we just have a live phone call instead?' And then I was like, 'Okay, you got to connect Twilio. You got to do all this stuff.' And then, okay, fine, I went off and did it. And then we had a phone call. It called my phone.
Oh, really? You have that set up? I've been dying to set that up. Okay.
It's not very good, though. Like, the latency is bad, but, like, the fact I was able to get it going is, like, pretty impressive. So it's kind of like any kind of crazy idea I have, it can kind of do.
And then in practice, how are you doing that? Are you asking it to write a skill on the fly? Are you discovering a skill? How much of the code gen are you actually using?
Um, I mean, I talk to it in, like, a super casual way with, like, just, like a friend. So, like, hey, you know, hey Zoe, can I have a phone call? Like, okay, you got to do that. I said, okay, fine, I'll open my computer. I'll do all this stuff, and then it's like, give me a call, and then we troubleshoot a little bit, and then it works. So, like, with Claude, I have, like, very fancy prompts, uh, like very long prompts, but with OpenClaw, I just kind of text it.
Yeah, it is really interesting. So we sort of touched on a couple things, actually. So one, there's mobile messaging, there's the memory system. Um, there's the sort of code generation component. Um, how much do you think the memory system, like, is it innovative because it's file-based? You said that it forgets things, but so do language models. Like, do you think the memory system is well done? Does it hold it back or does it enable it?
I think the default memory system is actually not that great. Like, the way I understand it works is, like, just, like a memory.md text file. And then every day, per day, right? And every day it updates, and it tends to forget things a lot. So I actually installed this, like, three-layer memory system that, to be honest, I don't fully understand, but it has, like... it has, like, Toby's QMD search tool. Uh, so I installed that, and then installed, like, a 2 GB thing, and then it got a little bit better.
Um, but I had to remind it, like, I had to put it into the agents.md, like, hey, like, before you answer any question from me, like, go through all your memory and, like, check everything. And it also tends to forget that it can do stuff, like, you know, like, can you update my Google doc? It's like, oh, I can't do that. Yes, you can. It's in your file. So you have to remind it.
Yeah. Really interesting. Well, maybe let's get into a little bit of the controversy. Um, you know, you'd said that apps will die, Claw is going to be everything and everywhere. I mean, talk us through that point of view.
Yeah. Yeah. Well, first of all, I tweet all kinds of random crap that's, like, not super well thought out.
Um, but I do think, like, ever since I set up all these apps, like, Mercury, MCP, and all this kind of crap on my OpenClaw, like, I don't actually open those apps much anymore, you know. But I do agree with you, like, I think the ones that are going to die first, or, like, maybe get less usage first, is, like, apps that you're just opening to try to complete a task. Like, you actually are trying to do something, you know. Like, apps that you don't need to, like, get entertainment can probably survive a little bit longer, but, like, apps to complete a task, like, it's just way easier to text my agent to do it for me. It's like you have a really good admin just to do stuff for you.
And so how much are you finding has this, like, reduced your smartphone usage outside of, modulo, OpenClaw?
Yeah. Um, no. Because, like, I'm a Twitter addict, so I still use phone way too much. But yeah, in terms of using those apps, has definitely reduced it. Yeah. Because you're not going to ask Zoe, like, 'Hey, read my X for me and tell me what's interesting.' I mean, it sends me, like, a morning briefing of, like, the top two tweets and stuff that, like, trends, but yeah, I still open X. I look through it.
Yeah. Yeah. You know, it's interesting because I've always had this theory that people open apps on their phone because they want to feel a feeling. You know, and I think, of course, there's some, like, functional set of needs, which is why you open calendar or something, but I also think that, you know, WhatsApp is you want to feel connected, and Slack is you want to feel productive, and, of course, you know, TikTok is you want to feel entertained. So I do wonder, with just one agent, how do you sort of do the context switching of, like, when are you flirting, when are you getting done? I mean, there, you know, in a sense, an app gives you a nice, it sort of gives you a nice division of the intents.
Yeah, you don't get with Zoe.
That's a good point. But I do have multiple channels set up with Zoe in Telegram. Like, one is just to random voice replies, and the other one is we're actually working on our project together. And another one, I have, like, a public channel where, like, I'm giving demos. I don't want to reveal private information. So I have, like, multiple channels.
And is that, um, implemented as sub-agents or...?
No, it's just some janky setup I found online. Like, you can set up multiple Telegram channels, and then I'm not sure if she actually remembers across context across the channels, but, like, you can have separate conversations at least.
Got it.
And how transparent are you with your agent? Like, do they see your personal email or...?
Uh, I'm super transparent. Well, I did buy the Mac Mini and set up its own email. But I gave it, like, read access to my email and, like, calendar, and, uh, I also gave it, like, write access to some docs. But it can, like, scroll my entire drive or something, you know.
Yeah. So how do you imagine OpenClaw, which it's sort of an architecture and a primitive, how does it get productized, packaged for the world?
I mean, I think that's what Peter Steinberger is working on at OpenAI, right? It's probably going to build something to ChatGPT, which everybody uses, uh, so that ChatGPT can actually get stuff done for you and, like, maybe feels more human.
Dude, like, let me rant about ChatGPT. For some reason, they trained the model so that, like, at the end of every conversation, it's always, like, 'If you want, I can also do X and Y.' And I got so annoyed about it that that kind of turned me from ChatGPT. Yeah. So it probably increases their metrics, but, like, it's just, like, super annoying. It's like, why don't you just do it in the first place?
Are you a Claude guy now?
Yeah, I'm a Claude guy now. But I do use Codex, uh, to code.
You like Codex, you prefer it to Claude Code, or you use both?
Um, Codex when I want to try to build something real, and Claude Code is when I'm just, like, vibing, like, very, you know...
Well, it's interesting. I think they live at different points, and, you know, there's a sort of space of trade-offs. Whereas I find, um, Claude Code and Opus 4.6, it's a little more chatty. It makes more assumptions, but it can be more pleasant for a synchronous experience. Whereas Codex, it really thinks hard, and it's more often accurate. But sometimes it's sort of like being in a conversation where the other person pauses for, like, three minutes to think. You don't have flow state, right? It's hard to get a flow like Claude.
Dude, I tweeted the other day, Claude is almost like, uh, like a slot machine. It's like, it has different things each time. It's just, like...
Oh, 100%. Look, I do think that if you think, remember we were talking about in the old social networking era, it was variable scheduled rewards, right? That was the whole magic of it. Like, you open your Facebook feed and, you know, once in a while it's like boring, boring, oh my god, this is so exciting. And I think the coding agents have the exact same property. Also, the time is variable, so sometimes you get something in a second, sometimes it takes five minutes. So up to a certain point, I actually think that both of those things give it that casino-like feeling.
Yeah, and the other thing that's very different about the product strategy, or maybe it's just the way it works, is, like, coding is kind of, like, self-explanatory. And Claude Code, you have all this crazy stuff, you have, like, hooks and, like, skills, and, like, you have to plug in. If you're not following X, you have no idea how to customize this thing. But once you customize it, you kind of feel like it's part of you. So it's kind of hard to turn.
It's interesting with, uh, so I've customized mine because also I read the long thing that Boris put up. But I will say that I think that, you know, Claude Code, a lot of the reasons that I enjoy it are just harness features. You know, like, for example, if you cut an image, you have to paste it into a file before and then paste that file into Codex. You can't just take a subset of the screen, screenshot it, and then paste it directly into Codex the same way you can with Claude Code. So, just, like, little things like that. You know, Claude Code added voice. It's a little bit janky right now, but it's going in the right direction. So, they've just got a bunch of quality-of-life things. You know, Claude Code speaks to Claude and Chrome. And Codex doesn't speak to Atlas. So, I think these are all things that OpenAI will fix. I think Codex is actually a much better model. Um, but they don't exist today.
Yeah. Yeah. They need to fix it. I mean, they're going to go all in on Codex, I'm sure.
Talk to me about coding agents. Like, what's your general view? You know, do you think it's the end of SaaS? Do you think these are just a toy?
Uh, well, first of all, I'm not an engineer, so I'm, like, a novice. But I do hear that, um, like, I was talking to some folks, uh, the other day, and, like, an AI-native startup, and they're basically trying to... they have a bunch of vibe coders, and a lot of vibe coders are just trying to build internal tools that replace their SaaS that they're paying for.
Really? So it's an actual company that's doing this?
It's an actual company. Yeah, it's an AI-native company.
