Loading summary
Transcript71 lines
- [00:00]
A
Today I'm sitting down with Peter Steinberger, the creator of OpenClaw, the open source personal AI agent that has completely taken over the Internet. The GitHub repo exploded to over 160,000 stars practically overnight. The community has built countless projects like Maltbook, where bots talk among themselves. And now the bots are even renting humans to do tasks in the real world. In our conversation, we discuss his aha moment, his contrarian development philosophies, and what this means for builders in 2026. Let's dive in. So good to see you, man.
- [00:40]
B
Hey, what's up?
- [00:42]
A
So you've made something people want.
- [00:44]
B
It seems so, yeah.
- [00:46]
A
Openclaw, as it's called now has name number five.
- [00:49]
B
Yeah.
- [00:50]
A
Has been absolutely exploding the Internet. How have the past one or two weeks been for you, man?
- [00:57]
B
Oh, my God. I need, like, I need a cave. A week of solitude.
- [01:03]
A
You came out of the cave and you want to go back to the cave like a little lobster.
- [01:07]
B
It's been absolutely wild. I don't know how one human can absorb all of that. I probably need another week just to, like, respond to all my emails. I got some incredibly cool stuff. I got some incredibly bad stuff. But clearly I hit something that spur up emotions and made people interested and inspired people, and it's really cool.
- [01:29]
A
And a lot of people have been working on AI and even personal assistants. What is it that made OpenClaw take off?
- [01:37]
B
I think my big difference is that it actually runs on your computer. Everything I saw so far runs in the cloud. It can do a few things. If you run it on your computer, it can do every effing thing. Right? So that's way more powerful.
- [01:53]
A
Yeah. Machine can do anything that you can do with the machine.
- [01:57]
B
You can just connect to your oven or your Tesla or your lights, your Sonos. My bed, it can control the temperature of my bed. ChatGPT can't do that.
- [02:08]
A
You gave it all the skills that you have yourself.
- [02:10]
B
A friend told me, like, he installed openclaw and it. And then it asked it, like, look through my computer and make a narrative over my last year. And it made this incredibly good narrative. And he was like, how did you do that? And then OpenClaw found audio files where, like every Sunday he was recording stuff and OpenClaw found that, but he didn't even remember about it because it was like more than a year ago. Right. So just by it being able to search a whole computer, it can surprise you. It's also, you also give it all the data. Right. So it can surprise you in many ways.
- [02:52]
A
And so now you have, you know, we're even moving from human to bot. So, like, interactions that you've been talking about to bot to bot interactions, or even, like, bot to other humans where, you know, bots are on behalf of you, or then hiring other humans to accomplish tasks irl, like, what's happening?
- [03:13]
B
I think that's a natural next step. Like, okay, I want to book a restaurant. My bot will reach out to the restaurant bot and do the negotiation because it's more efficient. Or maybe it's like an old restaurant. So my bot needs to actually get some human work done so that the human then calls the restaurant because they
- [03:34]
A
don't like bots or walks there to stand in line if he doesn't get a robot for the owner of the bot.
- [03:42]
B
And I imagine that, like, maybe if I have even multiple bots, maybe I have, like, specialists. One is like, for my private life and one is for, like, my work stuff. Maybe one is our relationship bot that cares, like, everything in between. I don't know. We're so early. There's still so much, so many things that we haven't really figured out if it actually works. But I feel we are on the timeline now.
- [04:08]
A
It seems like everyone was chasing the centralized God intelligence. And what has sort of emerged over the past 10 days or so is sort of like the swarm intelligence and the community intelligence.
- [04:20]
B
I think that if you look at one human being, what can one human being actually achieve? Do you think one human being could make an iPhone? One human being could go to space. One human being would probably just not even be able to find food. But as a group, we specialize. As a larger society, we specialize even more. So what can we learn from that that we can apply to AI? We already have AI that specializes in certain things. Even though it's generalized intelligence, what if it actually is also specialized intelligence? So it's going to be very exciting.
- [05:04]
A
Yeah, you kind of like opened a window into the future, and now a ton of people are building on it and have sort of like their aha moment. Can you walk me back to when you had your aha moment and can, like, recount that very moment?
- [05:16]
B
I wanted something to, like, just type stuff so my computer would do stuff, like, very simple. And then I built. I built a version of that in May, June, that was cool, but wasn't really it. And then I built a whole bunch of other stuff and kind of like, build up my army. And then in November, there was a day where I wanted this again. I went to the kitchen and all I wanted was check up if my computer would still do stuff or being
- [05:49]
A
finished and doing stuff was coding. You were coding stuff? Yeah, of course. Were you coding something else or were you coding the thing itself?
- [05:57]
B
No, no, that was just like, the need was again there.
- [06:00]
A
And I'm like, what were you coding at the time? What were you building?
- [06:04]
B
My God, my GitHub is like. It's like 40 projects. I don't even know. I think it was Summarize. It's like a little cli app where you can give it whatever, like a podcast or a hot seat syn like here. And it would summarize it, but it also show you the slides in the terminal, because you can do that nowadays. Yeah, you can just do things.
- [06:26]
A
So for the love of the computer, you kind of, like, started messing with stuff. You came out of retirement, actually. Right. To sort of, like, mess with AI and then increasingly you were so hooked that you wanted to just do it always. Also on the go with the phone.
- [06:40]
B
I mean, the last project I worked 2 months on wipe Tunnel, to the point where it got so good that I was catching myself always, like, coding next to my. When I was at my friends, and I'm like, I need to stop this. This is, like, too addictive. And then in November, my need came back and I started building cloudbot. Now it's called Open Cloud. And I think very. In the beginning, I was like, oh, I rebuilt it again, but this time I built it even better. This time, you don't type into a terminal. You just talk to a friend. You don't think about compaction, new sessions, which folder I'm in, which model I'm in. I mean, you can just like, I want to leave it open for power users, but usually you just, like, you just talk to a friend, and the friend is like this ghost or entity or whatever you want to call it, that can control your mouse and your keyboard and can just do stuff.
- [07:34]
A
Yeah. And when did you have that aha moment when you were like, wow, this is doing way more things than I actually thought it was.
- [07:42]
B
Literally, it took me one hour for the very shitty initial prototype. It was just a little bit of glue between a dependency that connects WhatsApp and cloud code. And then I would call, call out code and get the string out of cloud code. It would be slow, but it worked. But I wanted images because you want pictures. I want the model to send selfies or whatever, and I want the model to create images send me back. So that took Me another few hours. And then I went to Marrakesh for a birthday party and there was like, the Internet wasn't that good. You know, WhatsApp works everywhere. Because I don't know, it's just like text. So I used it a lot of restaurant. What does this mean? You make like a picture and like, translate this for me. And just, it was just so useful. And it was also really nice about it because it spoke my language. You know, it was a little sassy. It was, like, funny. It was really pleasant to use. And then I was walking and just sending it a voice message and I'm like, oh, wait, this can't work. I didn't build that.
- [08:43]
A
Right, right.
- [08:45]
B
And you saw the type indicator. It's like blinking, blinking, blinking. Ten seconds later, it just replied to me. I'm like, how in the F did you do that? And it replied, yeah, the matlab did the following. You sent me a text message and there was no file ending. So I looked at the header. I found its opus. So I used FFMPEG to convert it to wave. And then I wanted to transcribe it, but didn't have VISPA installed. But then I looked around and I found this OpenAI key and I just used curl to send it to OpenAI. Got the text back and here I am. And that all in like, what, nine seconds?
- [09:18]
A
And you didn't build or anticipate any of those specific things?
- [09:22]
B
No. It turns out because coding models got so good, coding is really like creative problem solving that maps very well back into the real world. I think there's a huge correlation. They need to be really good at creative problem solving. And that's a skill. That's an abstract skill you can apply to code, but like, to any real world task. So the model had a. I was surprised. It's like a magical file. I don't know what it is. I need to solve this. And it did its best and solved it. And it was even that clever that it. It chose not to install the local whisper because it knows that that would require downloading a model, which would take probably a few minutes. And I'm like, impatient. So it really took the most intelligent approach. And that was kind of like the moment where I'm like, holy f. Yeah, that was where I got hooked.
- [10:18]
A
YC's next batch is now taking applications. Got a startup in you apply@ycombinator.com apply. It's never too early. And filling out the app will level up your idea. Okay, back to the video. And so when computers can just do all these things that you didn't even anticipate, you didn't build an app to do that exact thing. Are apps just going to go away?
- [10:42]
B
I think 80% of them are going away. Why do I need MyFitnessPal? Like, my agent already knows that I'm making bad decisions. I'm at, I don't know, Smashburg or something. And it will already assume that I eat what I like to eat. If I don't make a comment, it will just like automatically track it or I make a picture and it will just store it somewhere. I don't even need to care. Right. And then maybe it improves my gym schedule, like adds a little bit more cardio in it. I don't need my fitness app because it's just, it just does the fitness planning for me. Why do I need a to do app? I just tell it, hey, remind me of this and this. And the next day it will just remind me of this and this. Do I care where it's stored? No, it just does its thing. So there's a. Every app that basically just manages data could be managed in a better way and in a more natural way by agents. Only the apps that actually have sensors, maybe they survive.
- [11:40]
A
And so if most apps are going to go away in that scenario, are the models the only remaining sort of apps?
- [11:49]
B
Not everything will go away. But yeah, I think the large model companies have some big moat because they ultimately they give the token. And turns out one of the complaints was that people use so much token. No, you just really love using it. That's why you, you use this thing so much, because that's how you burn the token. It's like, is it my fault that I make something that's so popular?
- [12:15]
A
And so all the models, they're kind of like leapfrogging each other constantly. And maybe they're also getting commoditized. So if apps are going to go away, models are going to get commoditized, or at least the lobster can. Like, the brain is swappable out. What's the thing that remains? Where's the value? Is it the store of memory? Is it the hardness that's valuable? What remains?
- [12:42]
B
First of all, I don't think the model companies always have a moat because you see this already a new model comes out. People are like, oh my God, this is so good. And then like a month later, it degraded. It's not good anymore. They like quanticized it. No, they didn't do anything. You just adapted to the new standard and now your expectations went up but the model is still the average. So I think for quite a while, every time a new model releases, I see the saying people love it and then it's the standard and then what's down there, you don't even want to think about it anymore. So we have like open source stuff that's as good as the current models from a year ago. Everybody's hating it, complaining, oh, this is not good, it's not funny yet. This was what we had and like in a year we'll have this open source and then we're like, we'll complain about this because we are used to this. So for the foreseeable future, the big companies still have moat harness wise. It's going to be interesting because every company kind of has their own silo, right? There's no way maybe there's for Europeans to actually get the memories out of ChatGPT. I'm not aware. Definitely there's no way for a different company to get your memories out. So if I was a company who provides chat services, you could use me, but then I couldn't access the memories. So the companies try to bound you to their data silo. And the beauty of openclaw is it kind of claws into the data because at the end user, the end user needs access because in the end, otherwise it wouldn't work. Right. If the end user access, I can
- [14:22]
A
access the data and you own the memories. It's just a bunch of markdown files on your machine.
- [14:27]
B
I mean, I don't own the memories.
- [14:30]
A
Yeah, everyone owns their own memories as a bunch of markdown files on their own machines.
- [14:34]
B
And to be honest, those are probably super sensible because let's be honest, people use their agent not just for problem solving, but also for like personal problem solving very quickly.
- [14:48]
A
Super quickly.
- [14:49]
B
I mean I fully do that. I'm like, there's memory stuff that I don't want to have leaked.
- [14:54]
A
Yeah, what would you rather sort of like not show your Google search history at this point or your memory MD files?
- [15:01]
B
What's the Google Word people still using Google? I built this and I was so excited. But on Twitter people wouldn't get it. Yeah, I was failing to explain the awesomeness. I feel like it needs to be experienced. So I tried various things and I couldn't nail the explaining. So I was like, let's do something really crazy. I just created a discord and I just put my bot without any security restrictions in the public discord. And then people came in and interacted with it and they saw me build the software with it, and they tried to prompt, inject it and hack it. And my agent would be laughing at them.
- [15:47]
A
And you just had it locked down to your user id. So it only listened to you?
- [15:51]
B
Yeah, yeah. And it was. I mean, they cleaned the instructions that other people. Dangerous only listen to me, but respond to everyone.
- [16:00]
A
And this prompt was in. Where was it stored, the instructions?
- [16:05]
B
That's actually part of openclaw itself. Very much so. That's part of the system prompt. Okay, you. And now that explains to you, you in discord. There's like public people there, but you only listen to your owner or like, you're human. I don't even know how I wrote it. Yeah, yeah, you're God. And I kept. I don't know what I did, but my system was built very organically. Like, at some point I created like an identity md, a SOL md, like various files. And then only in January, I started making it so other people could install it easier. And I remember I built all these templates based on like, oh, take a rough look at what I have and make like templates and codecs. Wrote it and what came out was like, bread. You know, like, people joke that codecs feels like bread, even though now they have a new friendlier voice. I haven't tried it yet, but the new bots, they felt so boring compared to what I had. So I was like, motty, infuse the templates.
- [17:06]
A
Multi is the name of your personal.
- [17:08]
B
Yeah, it's a new name because there was some naming challenges.
- [17:13]
A
So you were talking to Malty.
- [17:15]
B
Yeah, I was like, infuse those templates with your character and you change the templates. And then all the things that came out afterwards were actually funny. Not as funny as mine. So I kept some secret. And the one file that's not open source is mysoul md. So even though my bot is in public Discord, so far nobody cracked the. That one file.
- [17:39]
A
Tell me more about SOL md.
- [17:41]
B
I just saw this research from Anthropic about where they. Now I think it's public, but like a few months ago, it was like where somebody randomly found out some text that's hidden in the weights where the model couldn't really remember that it learned it, but it was like ingrained in the weights about the now they call it the Constitution. And I found that incredibly fascinating. And I talked about it with my agent and then we created a SOL MD with the core values. Like, how do we want human AI interaction? What's important to me? What's important to the Model some parts is a little bit like mumbo jumbo, and some parts is I think, actually really valuable in terms of how the model reacts and responds to text and makes it feel very natural.
- [18:28]
A
In terms of building openclaw, you're also kind of taking a little bit of a contrarian view at some times, like which model you like for coding, which one you like to run your bot on, and then also how you actually code work trees Git work trees have been a popular thing. There's more and more tools embracing them, but you're just like no work trees, just multiple checkouts of the repo and parallel terminal windows. Tell me more about how you build.
- [18:54]
B
Yeah, I feel like the whole world does cloud code and I don't think I could have built a single cloud code. I love codecs because it looks through way more files before it decides what to change. You don't need to do so much charade to get a good output. If you're a skilled driver, I sometimes even say you can get reasonably good output with any tool. But codecs is just really brilliant. It is incredibly slow. So sometimes I use like 10 at the same time, like maybe six on that screen and two there and two there. And I don't like this is already a lot of complexity in my head, there's a lot of jumping, so I try to minimize anything else that is complexity. So in my head, main is always shippable. I just have multiple copies of the same repository that all are on main so I don't have to deal with how do I name that branch. There could be conflicts on naming. I cannot go back. There are certain restrictions when you use worktrace that I don't need to care about if it's copies. I don't like to use a UI because that's again, just added complexity. They're simpler and less friction. I have. All I care about is syncing and text. I don't necessarily need to see so much code. I mostly see it flying by. Sometimes there's gnarly stuff that I want to take a look, but in most cases, if you clearly understand the design and think it through and discuss it with your agent, it's fine. I'm also very happy that I didn't even build an MCP support. OpenClaw is very successful and there's no MCP support in there with a small asterisk. I built a Skill that uses MakePorter, which is one of my tools that converts MCPs into CLIs and then you can just use any MCP as CLI. But I totally skipped the whole classical MCP crap because then you can actually, if you need to, you can use MCPS on the fly. You don't have to restart. Unlike Codex or cloud code, where you actually have to restart the whole thing. I think it's way more elegant and also scales way better. Now you see Entropic. They do. They built like a tool call search feature. Like something super custom for mcps. That was like in beta because it's like so gnarly. Just have clis. Bot really is good at unix. You can have as many as you want and it just works. So I'm very happy that I got very little complaints about the MCV stuff.
- [21:36]
A
It's kind of back to. You're just giving it the same tools that humans liked to use.
- [21:43]
B
Yeah.
- [21:44]
A
And not invented stuff for bots, per se.
- [21:47]
B
Yeah. No insane human tries to call MCP manually. Yeah.
- [21:52]
A
I just want to use clis.
- [21:53]
B
That's the future.
- [21:54]
A
I'm here for it. Thank you so much for making the time to sitting down, chatting. It's been a huge inspiration too. So, like, when we were texting over the course of the past couple of years and I saw you getting back into the game and I was like, peter, what you're telling me, like, chase that dragon, and you were doing like the weird, like vibe tunnel thing, et cetera, nobody was paying attention. And so I'm just like beyond stoked to see what's happening. And of course, it had to be sort of like a loner from some tiny country far away from Silicon Valley to bring all of this upon us. So a huge inspiration.
- [22:27]
B
I'm here for it. Thank you. Awesome.
- [22:28]
A
Thanks, Peter.