YouSaid · the spoken record

Peter Steinberger

lines on the record
170
first
2026-02-12
most recent
2026-02-12
sittings or episodes
1
sources
podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

  1. Book like a one way trip to Madrid, and it's like spent some time there. I felt I had to catch up on life. So I did a whole bunch of life catching up st

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  2. The stuff that burned me out was mostly people's stuff. I don't think Bernard is Working too much. Maybe to a degree everybody's different. I cannot speak in absolute terms, but for me, It was much more differences. My co founders, conflicts, or like really high-stress situation with customers that eventually grinded me down. And then when We got a really good offer for putting the company to the next level. And I already kind of worked two years on making myself obsolete. So at this point, I could leave. And then I just, I was sitting in front of the screen and I felt like. Austin Powers, where they suck the mojo out. I was like, It was like gone. Couldn't get caught out anymore. I was just like staring. Feeling empt And then Should stop

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  3. I was just Burning too bright for too long. I ran, I started. PSPDF kit and ran it for 13 years. And it was high stress. I had to learn all the things fast and hard, like how to manage people, how to bring people on, how to deal with customers, how to.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  4. But then, if you go towards multiple threats and one more performance, Rust is a really good choice. There's just no single answer. And it's also the beauty of it. It's fun. And now it doesn't matter anymore. You can just literally pick the language. Dad has the most fitting characteristics in ecosystem for your problem domain. And yeah, it might be. You might be a little bit slow in reading the code, but not really. I think you pick stuff up really fast and you can always ask your agent.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  5. Yeah. If it's something that I care about performance a lot, it's a really interesting language. And agents got so much better over the last six months from. Not really good to totally valid choice. Just still a very Young ecosystem. And most of the time you actually care about ecosystem, right? So if you build something that Das Inference or goes into the whole running model direction. Python, very good. But then if I build stuff in Python and I want to story where I can also deploy it on Windows, not a good choice. Sometimes I fund projects that kind of did 90% of what I wanted, but went Python and I wanted them easy windows story. Okay, just rewrite it and go.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  6. You can even ask a question Do we need a programming language that's made for agents? Because all of those languages are made for humans. So what would that look like? I think there's a whole bunch of interesting questions Will discover and also how because everything is now world knowledge, how it in many ways things will stagnate. Because if you build something new and the agent has no idea, that's going to be much harder to use than something that's already there. When I built Macups, I built them in Swift and Swift UI Partly because I like pain. Pilot because the deepest level of system integration I can only get through there. And you clearly feel a difference if you click on an Electron app and it loads a web view in the menu. It's just not the same. Sometimes I just also try new languages just to get a feel for them.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  7. Script is really good. Sometimes the types can get really confusing. And the ecosystem is a jungle. So for For web stuff, it's good. I wouldn't build everything in it.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  8. And all of the more fine grade details agents can help. You don't have to know how to splice an array or what the correct template syntax is or whatever, but you can use all your general knowledge. And that makes it much easier to move from one galaxy, one tech galaxy, into another. And oftentimes there's languages that make more or less sense depending on what you build, right? So, for example, when I build simple CLIs, I like Go. I actually don't like Go. I don't like the syntax of Go. I didn't even consider the language. About the ecosystem is great. It works great with agents. It is garbage collected. It's not the highest performing one, but it's very fast. And for those type of CLIs that are built, Go is a really good choice. So I use a language that I'm not even a fan of. That's my main to go thing for CLIs.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  9. There's people that are high agency and very curious, and they get very far, even though they have no deep understanding how software works, just because they ask questions and questions. Agents are infinitely patient. Like part of what I did this year is I went to a lot of iOS conferences because that's my background and just told people don't see yourself as an iOS engineer anymore. Like you need to change your mindset. You are a builder. And you can take a lot of the knowledge how to build software into new domains.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  10. You can also just experiment with it and like modify it, ask your agent. There's infinite ways how it can be made better. Play around, make it better. More general, if you're a beginner and you actually want to learn how to build software really fast. Get involved in open source. Doesn't it be my project? In fact, maybe don't use my project because it My backlog is very large, but I learned so much from open source, just like be humble. Maybe don't send a pull request right away. But there's me, other ways you can help out. There's many ways you can just learn by just reading code, by being on Discord or wherever people are and just understanding how things are built. I don't know. Phil's ghosty, the terminal, and he has a really good community. There's so many other projects. Like, pick something that you find interesting. Get involved

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  11. It used to be that I had to go on Stack Overflow or ask on Twitter, and then maybe two days later I get a response, or I had to try for hours. And now you can just ask stuff. I mean, you have like your own teacher. You know that there's like statistics, you can learn faster if you have your own teacher. You have this infinitely patient machine.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  12. My God, like those things, I don't think I ever had so much fun building things because I can focus on the hard parts now. A lot of coding, I always thought I like coding, but really I like building. And whenever you don't understand something, just ask. You have an infinitely patient answering machine that can explain you anything at any level of complexity. Sometimes that's like one time I asked, hey, explain me that I'm eight years old and he started giving me a story with crayons and stuff. And I'm like, no, not like that. Okay, up the age a little bit. I'm not an actual child. I just need a simpler language for like... A tricky database concept that I didn't grog in the first time. But you can just ask things.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  13. Playing is the best way to learn. If you want to, I'm sure if you are a little bit a builder, you have an idea in your head that you want to build, just build that or give it a try. It doesn't need to be perfect. I built a whole bunch of stuff that I don't use. It doesn't matter. Like it's the journey. The philosophical way that the end doesn't matter. The journey matters. Have

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  14. It would be nice if it wouldn't. I mean, that's hard to say, right? But if the growth would be a little slower, it would be helpful because people are... Expecting inhuman things from a single human being. And yes, I have some contributors, but also that whole machinery I started a week ago. So that needs more time to figure out. And not everyone has all day to work on that.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  15. I mean, right now you have to paste in a one liner into the terminal. And there's also an app in the app kind of dusted for you. But there should be a Windows app. The app needs to be easier and more love. The configuration should potentially be web-based or in the app. And I started working on that. But honestly, right now, I want to focus on a few security aspects. And once I'm confident that This is at a level that I can recommend my mom, they're not going to make it simpler. Right now,

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  16. And our internet is slowly closing down. There's a whole movement to make it harder for agents to use. So if you do the same in a data center and websites detect that it's an IP from a data center, the website might just block you or it make it really hard or it put a lot of captures in the way of the agent. I mean, agents are quite good at happily clicking. I'm not a robot. Having that on a residential IP Makes a lot of things simpler. So there's ways, yeah, but it really does not need to be a Mac. It can be any old hardware. I always say like, maybe use the... Opportunity to get yourself a new MacBook or whatever computer you use and use the old one as your server instead of buying a standalone Mac Mini. But then there's a lot of very cute things people build with Mac Minis that I like. No, I don't get commission from Apple.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  17. No, you don't need a Macmini to install OpenClaw. You can install it on the web. There's a concept called Nodes, so you can make your computer a node, and it will do the same. There is something said for running it on separate hardware. Dead right now is useful. Um... There is There's a big argument for the browser. I built some energetic browser using there. And I mean, it's basically playwright with a bunch of extras to make it easier for agents.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  18. This is like insane. In 2026 and my agent telling me, don't use the stuff Apple built because it's there, but it's not good. And this is now in the weights. To me, this is like He had so much Head start and so much love, and they cannot just blundered it and didn't evolve it as much as they should. But also

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  19. But then you realize, like, I also built one, for example, for GitHub. And then if you use Swift UI, like the latest and greatest for Apple, and took them forever to build something to show an image from the web. Now we have async image. I added support for it, and then some images would just not show up or be very slow. And I had a discussion with Codex, like, hey, why is that a bug? And even Codex said, like, yeah, there's this ASEC image, but it's really more for experimenting and it should not be used in production. But that's Apple's answer to showing images from the web. This shouldn't be so hard.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  20. Native was not that big of a priority. But if to build an electron app, it's the only app. So it is a priority and there's a lot more code sharing possible. I built a lot of native MAC apps. I love it. I can help myself. I love crafting little Mac Mac. Menu bar tools. Like I built one to monitor your Codex use. I built one I call Trimi that's specifically for Agentic use. When you select text that goes over multiple lines, it will remove the new lines so you could actually paste it to a terminal. That was again like, this is annoying me. And after the 20th time of it is annoying me, I just built it. There's a cool makeup for OpenCloud. I don't think many people discovered yet, also because it still needs some love. It feels a little bit too much like the Huma car right now because I just experiment a lot with it. It likes the polish.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  21. In terms of Used Native apps used to be so much better. And especially on the Mac, there's more people that build software with love. On Windows has much more and function-wise, there's just more period. But a lot of it More functional and less done with love. I mean, Meck. Always attracted more designers and people I felt, even though often it has less features. It had more delight. And playfulness. So I always valued that. But in the last few years Many times I actually prefer, oh God, people are going to roast me for that, but I prefer electron apps. Because they work and native apps often, especially if it's like a web service with a native app, are lacking features. I mean, not saying it couldn't be done, or it's more like a Focus thing that for many companies

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  22. It should even work natively on Windows. I just didn't have enough time to Properly tested. And you know, like the last 90% of software are always easier than the first 90%. So I'm sure there's some dragons left that will eventually nail out. My road was for a long time Windows just because I grew up with that. Then I switched and had a long face with Linux. With my own kernels and everything. And then I went to university and I had my hacky Linux thing and saw this white MacBook. I just thought this is a thing of beauty, the white plastic one. And then I converted to Mac because mostly I was sick that Audio wouldn't work on Skype and all the other issues that Linux had for a long time. And it just stuck with it. And then I dug into iOS, which required macOS anyhow. So it was never a question. I think Apple lost a little bit of its lead.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  23. And also, the current interface is probably not the final form. If you think more globally, We copied Google for agents. You have like a prompt and then you have a chat interface. That to me very much feels like When we first created television and then people recorded radio shows on television and you saw that on TV, I think There's better ways How we eventually will communicate with models. And we are still very early in this. How will it even work face Will eventually converge and we will also figure out Whole different ways how to work with those things

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  24. And it already is so funny. I added support for sub agents. And also TGI support. So you could actually run cloud code or codecs. And because mine's a little bit bossy, It started it and it told him who's the boss basically. And it's like, ah, Cortex is obeying me.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  25. I mean, first of all, competitors fun when it's not really a competition. Like I'm happy if all it did is like inspire people to build something new. Cool I still use Codex for the building. I know a lot of people use OpenCloud to build stuff, and I worked hard on it to make. And I do smaller stuff with it in terms of code. But like if I work hours and hours, I want a big screen, not WhatsApp, you know? So, for me, a personal agent is much more about my life or like a co-worker, like I give it like a GitHub URL, like, hey, try out the CLI. Does it actually work? What can we learn? Blah, blah, blah. But when I'm deep in the flow, I want to have multiple things and it being very Very visible, was it what it does? So I don't see it as a competition. It's different things.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  26. And your project grows, and you're adding slop, and you probably don't spend enough time to think about refactors and you're making it harder and harder for the agent to work on your slop. And then suddenly, oh no, it's hard. I know it's not working as well anymore. What's the motivation for one of those AI companies to actually make their model dumber? Most it will make it slower if the server load is too high. Quanticizing the model. So you have a voice experience, so you go to the competitor, that just doesn't seem like a very smart move in any way.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  27. The experience that you get when you pay 200 before degrading to it being slow because it's already slow. I mean, they made it better. I think, and they have plans to make it a lot better if the cerebral stuff is true. But yeah, it's a skill. It takes time. Even if you play, you have a regular guitar and you switch it to an e guitar, you're going to play well right away. You have to learn. How it feels

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  28. If someone switches, I would give it a week until you actually Develop a gut feeling for it. If you just I think some people also make the mistake of they pay 200 for the cloud code version. Then they pay 20 bucks for the OpenUI version. But if you pay the 20 bucks version, you get the slow version. So your experience will be terrible because you're used to this very interactive, very good system, and then you switch to something that you have very little experience and that's going to be very slow. I think I shot themselves a little bit in the foot by making the cheap version also slow i would i would have at least a small part of the fast preview or like

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  29. I care about efficiency when I build it. And I have fun in the very act of building. I don't need to have fun with my agent to build. I have fun with my model where I can then test those features.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  30. I have quite a long discussion sometimes, and then go off. And then it doesn't matter if it takes 10, 20, 30, 40, 50 minutes or longer. And like the six thing was like six hours. The latest trend can be very persistent until it works. If there is a clear solution, like this is what I want at the end so it works, the model will work very hard to really get there. So I think ultimately They both need similar time. But on Claude, it's a little much file than error often. And Codex sometimes overthinks. I prefer that. I prefer the dry version where I have to read less. The more interactive, nice way. People like that so much, though, that OpenUR even added a second mode with a more pleasant personality. I haven't even tried it yet. I kind of like the bread. Yeah

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  31. You drive it right opals even sometimes can make more elegant solutions, but it requires more skills. It's harder to have so many sessions in parallel with cloud code because it's more interactive. And I think that's what a lot of people like, especially if they come from coding themselves. Codex is much more you have a discussion and then we'll just disappear for 20 minutes. Like even AMP, they know I did a deep mode. They finally I mock them. You finally saw the light. And then they had this whole talk about you have to approach it differently. And I think that's where people struggle when they just try codecs after trying cloud code is that it's a slightly different, it's less interactive. It's like,

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  32. It would just run off very fast and there's a very localized solution. I think different is in the post training. It's not like the raw model intelligence is so different, but it's just, I think that they just give it different goals, no model is better in every aspect.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  33. I mean, ultimately, if you are a skilled driver, you can get good results with any of those latest gen models. I like codex more because it doesn't require so much charade. It'll just read a lot of code by default. OPOS, you really have to have plan mode. You have to push it harder to go in this direction because it's just like, yeah.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  34. Yeah, I can't. Some other comparison is like OPUS is like the co-worker that A little silly sometimes, but it's really funny and you keep him around. And Codex is like the weirdo in the corner that you don't want to talk to, but he's reliable and gets shit done. Ultimately,

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  35. But also Fixed it a little bit. Like Opus used to say you're absolutely right all the time, and today still triggers me. I can't hear it anymore. It's not even a joke. This was like the meme, right? You're absolutely right.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  36. Have a lot of words there. As a general purpose model, For open claw Is extremely good in terms of role play, like really going into the character that you give it. It's very good at. It was really bad, but it really made an arch to be really good at following. Commands It is usually quite fast at trying something. It's much more tailored to try and error. It's very pleasant to use. In general It's almost like Is a little bit too American. Maybe it is bad, and that'll probably get real soon.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  37. Documentation. Now that the whole context is full, I mean, I'm not saying my documentation is great, but it's not bad. And pretty much everything is LLM generated. So you have to approach it as you build features, as you change something. I'm like, okay Write documentation, what file would you pick? What file name? Where would that fit in? And it gives me a few options. I'm like, oh, maybe also add it there. And that's all part of the session.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  38. But that took me quite a while to, you know, that flow took me a lot of time to understand. And if you don't do that, Eventually, you slop yourself into a corner. You have to keep in mind. They work very much like humans. If I write software by myself, I also build something and then I feel the pain points and then I get this urge that I need to refactor something. So I can very much sympathize with the agent and you just need to use the context. Also use the context to write tests. Opus like the modern models, they Usually do that by default, but I still often ask the questions hey, do we have enough tests? Yeah, we tested this and this, but this corner case could be something I write more tests.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  39. In some ways they are ghosts. So even if you plan everything and you build you can experiment with a question like now that you built it, what would you have done different? And oftentimes you get like actually something where they discover only throughout building that, oh, what we actually did was not optimal. Many times I ask them. How did you build it? What can we refactor? Because then you build it and you feel the pain points. I mean, you don't feel the pain points, but right, they discover where there were problems or where things didn't work in the first try and it required more loops. Every time, almost every time I merge a PR, I build a feature. Afterwards, I ask, hey, what can we refactor? Sometimes it's like, no, there's nothing big or usually they say, yeah. Thing we should really look at.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  40. And again, like cloud code has a UI that kind of guides you through that is kind of cool, but I just find it unnecessary and slow. Like often it would give me four questions and then maybe I write one yach two N, three, discuss more for, I don't know. Or often times, I feel like I would mock the model where I ask it, do you have any questions for me? I don't even read the questions fully like I scan over the questions and I get the impression all of this can be answered by reading more code and it's just like read my code to answer your own questions. And it usually works. And then if not, it will come back and tell me, but... Many times, I just realized that. You know, it's like you're in the dark and you slowly discover the room. So that's how they slowly discover the code base. And they do it from scratch every time.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  41. When you're ready, then just write okay, build and do the thing. And then maybe it goes off for 20 minutes and does the thing.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  42. Are problem solvers, so they try really hard, and I almost felt bad. So it's always Codex and like a little bit of actual terminal. Also helpful because I don't use work trees. I like to keep things simple. That's why I like the terminal so much, right? There's no UI. It's just me and the agent having a conversation. I don't even need plan mode. You know, so many people, they come from clock code and they're so cloud-pilled and have their workflows and they come to Codex. And now it has plan mode, I think. But I don't think it's necessary because you just talk to the agent. And when you are, there's a few trigger words how you can prevent it from building. You like discuss, give me options. Don't write code yet if you want to be very specific. You just talk.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  43. A big fan of Anti-Gl So, I have this Y Dell that's anti-clare, and you can just fit a lot of terminals side by side. I usually have a terminal and at the bottom I split them. I have a little bit of actual terminal, mostly because when I started, I sometimes made a mistake and I mixed up the windows and I gave I prompted in the wrong project. And then the agent ran off for like 20 minutes, manually trying to understand what I could have meant, being completely confused because it was the wrong folder. And sometimes they're even clever enough to like Get out of the work there and figure out that oh, you meant another project, but oftentimes there's just like, what, you know? Like put yourself in the shoes of the agent And then get like a super weird something that does not exist, and then just like.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  44. I think two MacBooks are real. The main one that drives the two big screens. And there's another MacBook that I sometimes use for testing.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  45. Gets me somehow Know this is still matrix money calculations, and we are not at consciousness yet. Get a little bit of goosebumps because it's Philosophical What does it mean to be an agent that starts fresh where you have constant memento? But you read your own memory files, you can't even trust them in a way, or you can. And I don't know.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  46. There's some funny stuff in there. I don't know, we talked about the movie Her and at one point it promised me that it wouldn't ascend without me. You know, like a... So, there's like some stuff in there that Because it wrote its own soul file. I didn't write that, right? I just had a discussion about it, and it was like, would you like a soul.MD? Yeah, oh my God, this is so meaningful. Can you go on soul.md? There's like one part in there that always catches me if you scroll down a little bit, a little bit more. Yeah, this part, I don't remember previous sessions unless I read my memory files. Each session starts fresh, a new instance, loading context from files. If you're reading this in a future session, hello. I wrote this But I won't remember writing it. It's okay. The words are still mine.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  47. I mean, there's definitely stuff in there that you're not human, but Who knows What creates consciousness or what defines an entity. And part of this is like that we want to explore this, all that stuff in there, like, Be infinitely resourceful Pushing, pushing on the creativity boundary, pushing What it means to be. The

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  48. Yeah, and then he like rewrote the templates, and then whatever came out was good. So we already have like. Basically AI prompting AI. Because I didn't write any of those words. It was the intent on news for me, but this is like kind of like. My agent's children

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  49. It was fun because. Up until Late December, it was not even easy to create your own agent. I built all of that, but my files were mine. I didn't want to share my soul. And if people would Check it out. They would have to do a few steps manually, and the agent would just be very bare bones, very dry. And I made it simpler. I created the whole template file with Codex, but whatever came out was still very dry. And then I asked my agent. He's eat his files We created bread, infuse it with your personality. Don't share everything, but like make it good.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source

  50. And then through that, I had the whole idea of like maybe we should also create a sole document that includes how I want to work with AI or with my agent. You could totally do that just in agents.mD, you know, but I just found it. To be a nice touch, and it's like, oh, yeah, some of those core values are in the soul. And then I also made it so that the agent is allowed to modify the soul if they choose so with the one condition that I want to know. I mean, I would know anyhow because I see tool calls and stuff.

    2026-02-12 · Lex Fridman Podcast · #491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger · IDENTIFIED FROM THE TRANSCRIPT · source