Skip to content

Podcast

#491 – OpenClaw- The Viral AI Agent That Broke the Internet – Peter Steinberger

Lex Fridman Podcast

Source ↗ ← All highlights
  • One-Hour WhatsApp Prototype
    • Peter built a WhatsApp-connected agent prototype in about one hour that felt like talking to his computer.
    • He expanded it with image support and used it heavily during a trip to Marrakesh where it proved extremely useful. Transcript: Peter Steinberger Well, that was still very limited. You know, like I had this one experiment with WhatsApp. Then I had this experiment. And both felt like not the right answer. And then my search, I was literally just hooking up WhatsApp to cloud code. One shot, the CLI message comes in. I call the CLI with minus P. It does its magic. I get the string back and I send it back to WhatsApp. I, I built this in one hour and I felt, already felt really cool. It’s like, oh, I could, I can like talk to my computer, right? That was, that was cool. But I, I wanted images because I often use images when I prompt. I think it’s such a, such an efficient way to give the agent more context. And they’re really good at figuring out what I mean, even if it’s like a weird crop top screenshot. So I used it a lot and I wanted to do that in WhatsApp as well. Also, like, you know, just you run around, you see like a poster of an event. You just make a screenshot and like figure out if I have time there, if this is good, if my friends are maybe up for that. It’s like, images seemed important. (Time 0:20:45)
  • Agent Solved An Unexpected Audio Message
    • Peter describes a mind-blowing moment when the agent handled an audio message it wasn’t explicitly taught to process.
    • It detected the audio format, converted it, and routed it through Whisper/OpenAI to reply sensibly. Transcript: Peter Steinberger Still, you know, the moment where it blew my mind was when I used it a lot, and at some point I just sent it a message, and then a typing indicator appeared, and I’m like, wait, I didn’t build Speaker 3 That. It only has image support. Peter Steinberger So what is it even doing? And then it would just reply. What was the thing you sent it? Oh, just a random question. It’s like, hey, what about this in this restaurant? Because we were just running around and checking out the city. So that’s why I didn’t even think when I used to, because sometimes when you’re in a hurry, typing is annoying. Lex Fridman So you did an audio message? Yeah. And it just worked. And I’m like… And it’s not supposed to work because you didn’t give it that capability. Peter Steinberger And it’s really what, how the fuck do you do that? And it was like, yeah, the mad led the following. He sent me a message, but it only was a file, a no file ending. So I checked out the header of the file and it found that it was like opus so i used ffmpeg to convert it and then i wanted to use whisper but you didn’t have it installed but then i found your Open ai key and just use curl to send a file to to open ai to translate and here i am and i just looked at the message and I’m like, oh wow. Lex Fridman You didn’t teach you any of those things and the agent just figured it out. It has to do all those conversions, the translation, they figured out the API, it figured out which program to use, all those kinds of things. And you were just absentmindedly just sending an audio message and it came back. Peter Steinberger So clever even because it would have gone the whisper local path, it would have had to download a model. It would have been too slow. (Time 0:25:10)
  • Fun And Personality Fueled Virality
    • OpenClaw’s virality came from making agent creation fun, weird, and highly personal rather than overly serious.
    • Giving agents personalities and clear self-knowledge made them more engaging and hackable for users. Transcript: Peter Steinberger Like factorial times infinite. Right. I feel like I built my little playground. Like I never had so much fun then building this project. You know, like you have like, oh, I go like level one agentic loop. What can I do there? How can I be smart at queuing messages? How can I make it more human? Like, oh, then I had this idea of because the loop always, the agent always replies something, but you don’t always want an agent to reply something in a group chat. So I gave him this no reply token. So I gave him an option to shut up. So it feels more natural. That’s level two. Yeah, yeah, yeah. On the agentic loop. And then I go to memory, right? You want them to like remember stuff. So maybe the ultimate boss is continuous reinforcement learning, but I’m like, I feel like I’m level two or three with marked on files and the vector database. And then you can go to level community management. You can go to level website and marketing. There’s just so many hats that you have to have on. Not even talking about native apps. That’s just like infinite different levels and infinite level ups you can do. Lex Fridman So the whole time you’re having fun. (Time 0:29:50)
  • Winning with Fun
    • Peter Steinberger attributes OpenClaw’s success to the project being fun and weird.
    • Many other startups take themselves too seriously.
    • He intentionally incorporated lobster-related elements online.
    • The initial installation process was intentionally quirky, requiring users to clone, build, and run it in a specific way. Transcript: Lex Fridman There’s so many possible directions, speaking of Factoria, that we can go here. But one big picture one is, why do you think your work, Open Claw, won in this world? If you look at 2025 many startups so many companies are doing kind of agentic type stuff or claiming to and here open claw comes in and destroys everybody like why did you win because they Peter Steinberger All take themselves too serious yeah like it’s hard to compete against someone who’s just there to have fun yeah i wanted it to be fun i wanted it to be weird and if you see like all the all (Time 0:31:40)
  • Self-Modifying Software
    • Peter Steinberger created OpenClaw where the AI agent can modify its own software, without initially planning for it.
    • Lex Fridman reflects on the historical significance of self-modifying software being used by many people.
    • Peter uses self-introspection to debug the agent, which felt natural to him.
    • This led to many pull requests, even from people who had never written software before. Transcript: Lex Fridman It’s just fascinating so you have this piece of software that’s written, TypeScript, that’s able to, via the agentic loop, modify itself. I mean, what a moment to be alive in the history of humanity, in the history of programming. Here’s the thing that’s used by a huge amount of people to do incredibly powerful things in their lives. And that very system can rewrite itself, can modify itself. Can you just speak to the power of that? Isn’t that incredible? When did you first close the loop on that? Oh, because that’s how I built it as well. Peter Steinberger Most of it is built by Codex, but oftentimes when I debug it, I use self-introspection so much. It’s like, hey, what tools do you see? Can you call the tool yourself? Oh, like whatever do you see? Read the source code, figure out what’s the problem. To that the agent the very agent and software that you use is used to debug itself so that it felt just natural that everybody does that and that it led to so many so many pull requests by People who never wrote software i mean it also did show that people never wrote software so i call them prompt requests in the end but i don’t want to pull that down because every time someone Made the first pull request is a win for a society. It doesn’t matter how shitty it is. You got to start somewhere. So I know there’s this whole big movement of people complain about open source and the quality of PRs and a whole different level of problems. But on a different level, I found it very (Time 0:33:30)
  • Agents Debug And Modify Their Own Code
    • Self-modifying and self-debugging code emerged naturally because Peter used agents to introspect and fix the project.
    • Lowering barriers enabled many non-programmers to make first pull requests and learn open-source workflows. Transcript: Lex Fridman It’s just fascinating so you have this piece of software that’s written, TypeScript, that’s able to, via the agentic loop, modify itself. I mean, what a moment to be alive in the history of humanity, in the history of programming. Here’s the thing that’s used by a huge amount of people to do incredibly powerful things in their lives. And that very system can rewrite itself, can modify itself. Can you just speak to the power of that? Isn’t that incredible? When did you first close the loop on that? Oh, because that’s how I built it as well. Peter Steinberger Most of it is built by Codex, but oftentimes when I debug it, I use self-introspection so much. It’s like, hey, what tools do you see? Can you call the tool yourself? Oh, like whatever do you see? Read the source code, figure out what’s the problem. To that the agent the very agent and software that you use is used to debug itself so that it felt just natural that everybody does that and that it led to so many so many pull requests by People who never wrote software i mean it also did show that people never wrote software so i call them prompt requests in the end but i don’t want to pull that down because every time someone Made the first pull request is a win for a society. It doesn’t matter how shitty it is. You got to start somewhere. So I know there’s this whole big movement of people complain about open source and the quality of PRs and a whole different level of problems. (Time 0:33:30)
  • Name-Change Chaos And Squatters
    • The project faced aggressive name-squatting and tokenization attacks during renames, which caused account and package snipes.
    • Peter nearly deleted the project but recovered with community and platform help to restore accounts and redirects. Transcript: Peter Steinberger Was no perfect name. And I didn’t sleep for two nights. I was under high pressure. I was trying to get like a good set of domains and you know, not cheap, not easy because in this state of the internet, you basically have to buy domains if you want to have a good set. And then another email came in that the lawyers are getting uneasy. Again, friendly, but also just adding more stress to my situation already. So at this point, I was just like, sorry, that’s not a word, fuck it. And I just renamed it to MaltBot because that was a set of domains I had. I was not really happy, but I thought it’ll be fine. And I tell you everything that could go wrong, did go wrong. Everything that could go wrong, did go wrong. It’s incredible. I thought I had mapped the space out and reserved the important things. Can you give some details of the stuff that got wrong? Because it’s interesting from an engineering perspective. Well, the interesting stuff is that none of these services have a squatter protection. So I had two browser windows open. One was like an empty account ready to be renamed to CloudBot. And the other one, I renamed to MaltBot. So I pressed rename there. I pressed rename there. And in those five seconds they stole the account name literally the five seconds of dragging the mouse over there and pressing rename there was too long because there’s no those systems I mean you would expect that they have some protection or like an automatic forwarding but there’s nothing like that and i didn’t know that they’re not just good at harassment they’re Also really good at using scripts and tools yeah so yeah so suddenly like the old account was um promoting new tokens and serving malware. And I was like, okay, let’s move over to GitHub. And I pressed rename on GitHub. And the GitHub renaming thing is slightly confusing. So I renamed my personal account. And in those, I guess it took me 30 seconds to realize my mistake. They snipe my account, serving malware from my account. So I was like, okay, let’s at least do the NPM stuff. But that takes like a minute to upload. They sniped the NPM package. (Time 0:43:44)
  • Secure Your Agent Before Exposing It
    • Follow OpenClaw’s security checklist: limit inbound access, sandbox tools, and configure reverse proxies.
    • Keep the service private and run in a private network unless you understand and accept the risk. Transcript: Lex Fridman Open class security audit that you can run. You can do all kinds of audit checks on the inbound access, tool blast radius, network exposure, browser control exposure, local disk hygiene, plugins, model hygiene, a bunch of the Credential storage, reverse proxy configuration, local session logs live on disk. There’s where the memory is stored, sort of helping you think about what you’re comfortable giving read access to, what you’re comfortable giving write access to, all that kind of Peter Steinberger Of stuff is there something to say about the basic best security practices that you’re aware of right now i think that people turn it into like a much worse light than it is um again you Know like people of attention and if they scream loudly oh my god this is like the scariest project ever. That’s a bit annoying because it’s not. It is powerful, but in many ways, it’s not much different than if I run cloud code with dangerously skipped permissions or codecs in YOLO mode. And every attending engineer that I know does that because that’s the only way how you can you can get stuff to work so if you make sure that you are the only person who talks to it the risk Profile is much much smaller if you don’t put everything on the open internet but stick to my recommendations of like having it in a private network, that whole risk profile falls away. But yeah, if you don’t read any of that, you can definitely make it problematic. (Time 1:09:22)
  • Learn Agents By Playing And Iterating
    • Start with short prompts, iterate, and learn the agent’s language through play rather than demanding perfect prompts.
    • Treat agents like conversation partners: discuss, clarify, then command to build reliably. Transcript: Lex Fridman There’s the please fix this, where you prompt a short prompt. On the left, and in the middle, there’s super complicated eight agents, complex orchestration with multi-checkouts, chaining agents together, custom subagent workflows, library Of 18 different slash commands, large full-stack features. You’re super organized, you’re a super complicated, sophisticated software engineer, you got everything organized, and then the elite level is over time, you arrive at the zen place Of, once again, short prompts. Hey, look at these files and then do these changes. I actually call it the agentic trap. Peter Steinberger I saw this in a lot of people that have their first touch point and maybe start vibe coding. I actually think vibe coding is a slur. You prefer agentic engineering. Yeah, I always tell people I do agentic engineering and then maybe after 3 a.m. I switch to vibe coding, and then I have regrets on the next day. Yeah, a walk of shame. Yeah, you just have to clean up and fix your shit. We’ve all been there. (Time 1:13:59)
  • Commit Fast, Fix Forward, Keep Main Shippable
    • Embrace fast main-branch commits and avoid excessive reverting; fix forward by asking agents to correct mistakes.
    • Keep main shippable, run local CI, and prefer moving fast over elaborate branching. Transcript: Lex Fridman Requires letting go a little bit, just like leading a team of engineers. Yeah. Because it might come up with a name that’s, in your view, terrible. But it’s kind of a simple symbolic step of letting go. Very much so. There’s a lot of letting go that you do in your whole process. So, for example, I read that you never revert. Always commit to main. There’s a few things here. You don’t refer to past sessions. So there’s a kind of YOLO component because reverting means instead of reverting, if a problem comes up, you just ask the agent to fix it. Peter Steinberger I read a bunch of people in their workflows like, oh yeah, the prompt has to be perfect. And if I make a mistake, then I roll back and redo it all. In my experience, that’s not really necessary. If I roll back everything, it would just take longer. If I see that something’s not good, we just move forward. And then I commit when I like the outcome. I even switched to local CI, like DHH inspired, where I don’t care so much more about CI and GitHub. We still have it. It still has a place. But I just run tests locally. And if they work locally, I push domain. (Time 1:23:07)
  • Design For Delight, Not Just Automation
    • Infuse software with human delight and personality; agents won’t invent that human touch on their own.
    • Use agents to automate tedious parts but design interfaces, names, and small rituals to create emotional connection. Transcript: Peter Steinberger Still a lot of there’s still a lot of craft and thinking involved in how to make something or even even you know even when you started those little messages like i’m built i built on caffeine Json 5 and a lot of willpower and like every time you get it you get another message and it kind of primes you into that. This is a fun thing. It’s not yet Microsoft Exchange 2025 and fully enterprise ready. And then when it updates, it’s like, oh, I’m in. It’s cozy here. You know, like something like this that like makes you smile. Agent would not come up with that by itself. That’s like, that’s the, I don’t know, it’s just how you build software that delights. (Time 1:33:37)
  • Models Have Different Flavors
    • Different models have distinct flavors: Codex favors reading and reliability, while Opus (Claude) favors role-play, creativity, and trial-and-error.
    • Skilled drivers can get good results from either, but pick the model that matches the workflow you want. Transcript: Lex Fridman In terms of models, Cloud Opus 4.6 and GPT-5 Codex. Which is better? How different are they? I think you’ve spoken about Codex reading more and Opus being more willing to take action faster and maybe being more creative in the actions it takes. But because Codex reads more, it’s able to deliver maybe better code. Peter Steinberger Can you speak to the differences there? I have a lot of words there. As a general purpose model, is the best like for open claw opus is extremely good in terms of role play like really going into the character that you give it it’s very good at and it was really, But it really made an arch to be really good at following commands. Speaker 3 It is usually quite fast at trying something. Peter Steinberger It’s much more tailored to trial and error. It’s very pleasant to use. Speaker 3 In general, it’s almost like Opus is a little bit too American. Maybe it’s a bad analogy. You’ll probably get roasted with that. Lex Fridman I know exactly. It’s codex in German. Is that what you’re saying? Actually, now that you say it, it makes perfect sense. Peter Steinberger Or you could sometimes sometimes i explain it i will never be able to unthink what you just said that’s so true but you also know that a lot of the codex team is like european um so maybe there’s A bit more to it that’s so true that’s funny but also entropic they fixed it a little bit. Like Opus used to say, you’re absolutely right all the time. And today it still triggers me. I can’t hear it anymore. It’s not even a joke. I just… This was like the meme, right? You’re absolutely right. You’re allergic to sycophancy a little bit. Yeah, I can’t. Some other comparison is like Opus is like the co-worker that is a little silly sometimes but it’s really funny and you keep him around and cortex is like the the weirdo in the corner that You don’t want to talk to but he’s reliable and gets shit done yeah um ultimately this all feels very accurate. I mean, ultimately, if you’re 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 will just read a lot of code by default. Opus, you really have to have plan mode. You have to push it harder to go in these directions because it’s just like, yeah, can I go in? Can I go in? It will just run off very fast and there’s a very localized solution. (Time 1:48:50)
  • Give A Week To Learn A New Model
    • When switching models, give yourself about a week to develop a gut feel for its behavior and interaction speed.
    • Avoid judging a cheaper, slower plan as the model’s real capability if you’re used to fast interactive tiers. Transcript: Peter Steinberger Switches, I would give it a week until you actually develop a gut feeling for it. Yeah. If you just… I think some people also make the mistake of they pay $200 for the cloud code version, then they pay $20 for the OpenAI version. But if you pay the $20 version, you get the slow version. (Time 1:55:20)
  • Beginner’s Guide to Agentic AI
    • If you’re a programming beginner interested in agentic AI, start by playing and building things you want to build.
    • Don’t worry about perfection; the journey and the process of building are what matter most.
    • Focus on the hard parts and enjoy the process.
    • If you don’t understand something, ask an AI; treat it like an infinitely patient answering machine that can explain anything at any level of complexity. Transcript: Lex Fridman What advice would you give to them about, let’s say, joining the agentic AI revolution? Peter Steinberger Play. Playing is the best way to learn. I’m sure if you are a little bit of a builder, you have an idea in your head that you want to build, just build that. 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, you know, like the philosophical way that the end doesn’t matter. The journey matters. Have fun. 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 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 there’s like one time i asked hey (Time 2:10:57)
  • Learn By Building And Contributing
    • Beginners should play and build small projects to learn agentic AI; experimentation accelerates skill acquisition.
    • Join open-source communities, read code, and contribute incrementally to learn practical software building fast. Transcript: Peter Steinberger Play. Playing is the best way to learn. I’m sure if you are a little bit of a builder, you have an idea in your head that you want to build, just build that. 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, you know, like the philosophical way that the end doesn’t matter. The journey matters. Have fun. 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. (Time 2:11:05)
  • Avoiding Retirement Boredom
    • Don’t retire with the idea that you will ‘just enjoy life now’ because having nothing to look forward to can lead to boredom.
    • Boredom can drive you to seek stimulation in unhealthy ways.
    • Money should not be the primary objective; it’s more of an affirmation that you did something right.
    • Having money solves problems, but there are diminishing returns as you accumulate more. Transcript: Peter Steinberger On how to approach life. If you think that, oh, you work really hard and then I retire, I don’t recommend that because the idea of, oh yeah, just enjoy life now. Speaker 3 Maybe it’s appealing, but right now I enjoy life, the most ever enjoyed life. Peter Steinberger Because if you wake up in the morning and you have nothing to look forward to you have no real challenge that gets very boring very fast and then when when you’re bored you’re gonna look For other places how to stimulate yourself and then maybe maybe that’s drugs, you know. But that will eventually also get boring and you look for more. And that will lead you down a very dark path. Lex Fridman But you also showed on the money front, you know, a lot of people in Silicon Valley in the startup world, they think maybe overthink way too much optimized for money. And you’ve also shown that it’s not like you’re saying no to money. I mean, I’m sure you take money, but it’s not the primary objective of your life. Peter Steinberger Can you just speak to that, your philosophy on money? When I built my company, money was never the driving force. It felt more like an affirmation that I did something right. And having money solves a lot of problems. I also think there’s diminishing returns the more you have. (Time 2:22:50)
  • Dangers of Early Retirement
    • Peter Steinberger advises against working hard with the sole intention of retiring early.
    • He emphasizes that having nothing to look forward to after retirement can lead to boredom.
    • Boredom can drive people to seek stimulation in unhealthy ways, such as drug use.
    • A lack of challenges in retirement can ultimately lead down a dark path. Transcript: Peter Steinberger On how to approach life. If you think that, oh, you work really hard and then I retire, I don’t recommend that because the idea of, oh yeah, just enjoy life now. Speaker 3 Maybe it’s appealing, but right now I enjoy life, the most ever enjoyed life. Peter Steinberger Because if you wake up in the morning and you have nothing to look forward to you have no real challenge that gets very boring very fast and then when when you’re bored you’re gonna look For other places how to stimulate yourself and then maybe maybe that’s drugs, you know. But that will eventually also get boring and you look for more. And that will lead you down a very dark path. But you also (Time 2:22:50)
  • Heartbeat Makes Agents Relatable
    • Proactive agents (heartbeat) that intermittently check in create relatable, helpful experiences and can trigger on life events.
    • Small proactive behaviors make agents feel caring and more integrated into daily life. Transcript: Peter Steinberger So it’s like, you know, great power becomes great responsibility. And I was like, how can I up the stakes a little bit more? Yeah, right. And I just made it proactive. So I added a prompt. Initially, it was just a prompt surprise me. Every like half an hour surprised me, you know. And later on, I changed it to be like a little more specific and in the definition of surprise. But the fact that I made it proactive and that it knows you and that it cares about you, at least it’s programmed to do that, prompted to do that, and that is a follow-on on your current session Makes it very interesting because it would just sometimes ask a follow-up question or like, how’s your day? I mean, again, it’s a little creepy or weird or interesting, but heartbeat very, in the beginning is still, today, it doesn’t, the model doesn’t choose to use it a lot. Lex Fridman By the way, we’re talking about heartbeat, as you mentioned, the thing that regularly acts. You just kick off the loop. Isn’t that just a cron job, man? Peter Steinberger Yeah, right. It’s like the criticisms that you get. Lex Fridman You can deduce any idea to like a silly, yeah, it it’s just a crone shop in the end i have like separate crone shops isn’t love just evolutionary biology manifesting itself and uh isn’t Peter Steinberger Aren’t you guys just using each other and again and the project is all just glue of a few different dependencies and there’s nothing original why do people well, well, you know, isn’t Dropbox just FTP with extra steps? Yeah. I found it surprising where I had this, I had a shoulder operation a few months ago. And the model rarely used heartbeat, but then I was in the hospital and it knew that I had the operation and it checked up on me. It’s like, are you okay? And I just, it’s like, again, apparently like if something significant in the context, that triggered the heartbeat when it rarely used the heartbeat. And it does that sometimes for people. And that just makes it a lot more relatable. (Time 2:45:59)
  • Agents Treat Apps As Slow APIs
    • Agents turn apps into ‘slow APIs’ by automating browser interactions when official APIs are unavailable.
    • This forces apps to rethink openness: either provide agent-friendly APIs or accept agents scraping via the UI. Transcript: Peter Steinberger And that through personal agents, a lot of apps will disappear. You know, like I had a, I built a CLI for Twitter. Speaker 3 I mean, I, I just reverse engineered the website and used the internal API, which is not very allowed. Peter Steinberger It’s called bird short lived. It was called Bird because the bird had to disappear. The wings were clipped. All they did is they just made access slower. You’re not actually taking a feature away. But now if your agent wants to read a tweet, it actually has to open the browser and read the tweet. And it will still be able to read the tweet. It will just take longer it’s not like you’re making something that was possible not possible no now it’s just ticking and now it’s just a bit slower so so it doesn’t really matter if your Service wants to be an api or not if i can access it in the browser it is API. It’s a slow API. Can you empathize with their situation? Lex Fridman Like what would you do if you were Twitter, if you were X? Because they’re basically trying to protect against other large companies scraping all their data. But in so doing, they’re cutting off like a million different use cases for smaller developers that actually want to use it for helpful, cool stuff. Peter Steinberger I think if you have a very low per day baseline per account that allows read-only access, it would solve a lot of problems. There’s plenty of automations where people create a bookmark and then use OpenClaw to like find the bookmark, do research on it, and then send you an email with like more details on it Or a summary. That’s a cool approach. (Time 2:53:18)
  • X Read-Only Access
    • Peter Steinberger suggests that X (Twitter) could provide a low per-day baseline for read-only access to solve data scraping issues while enabling smaller developers’ use cases.
    • Many people bookmark content on X and could use OpenClaw to research and summarize those bookmarks.
    • Lex Fridman agrees that read-only access to bookmarks is an incredible application, as many users never revisit their saved content.
    • Peter Steinberger proactively told Twitter about OpenClaw’s potential use with bookmarks and the existing need.
    • He hopes this will encourage Twitter to consider the benefits of read-only access. Transcript: Peter Steinberger I think if you have a very low per day baseline per account that allows read-only access, it would solve a lot of problems. There’s plenty of automations where people create a bookmark and then use OpenClaw to like find the bookmark, do research on it, and then send you an email with like more details on it Or a summary. That’s a cool approach. I also want all my bookmarks somewhere to search. I would still like to have that. Lex Fridman So read-only access for the bookmarks you make on X. It seems like an incredible application because a lot of us find a lot of cool stuff on X. We bookmark, that’s the general process of X. It’s like, holy shit, this is awesome. Oftentimes you bookmark so many things, you never look back at them. It would be nice to have tooling that organizes them and allows you to research it for. (Time 2:54:47)
  • Valuing Imperfection
    • Peter Steinberger values typos again, recognizing the value in the rough parts of an actual human.
    • He experimented with AI-generated blog posts but found it missed the nuances of his writing style, so he went back to writing organically.
    • While Peter raves about using AI for code, he’s allergic to it in stories.
    • He is okay with AI in documentation, considering it better than nothing. Transcript: Peter Steinberger I think we’re reaching a point where I value typos again. Like, you know, I mean, it also took me a while to come to the realization. On my blog, I experimented with creating a blog post with agents and ultimately it took me about the same time to like steer agent towards something i like but it missed the nuances that How i would write it you know you can like you can steer it towards your style but it’s not going to be all your style so i i completely moved away from that i i everything everything i blog Is organic handwritten and maybe maybe i i i use ai as a fix my worst typos but there’s value in the rough parts of an actual human. Lex Fridman Isn’t that awesome? Isn’t that beautiful? Peter Steinberger That now, because of AI, we value the raw humanity in each of us more. I also realized the thing that I rave about AI and use it so much for anything that’s code, but I’m allergic if it’s stories. Right? Yeah. (Time 2:58:45)
  • Agents Could Replace Many Single-Purpose Apps
    • Personal agents will subsume many single-purpose apps by using richer context to perform tasks across services.
    • Companies must either expose agent-friendly APIs or risk being bypassed as agents control user actions. Transcript: Peter Steinberger You think agents will just transform the entire app market? Yeah. I noticed that on Discord, people just said how they, like what they build and what they use it for. And like, why do you need my fitness pal when the agent already knows where I am? So can assume that I make bad decisions when I’m at, I don’t know, Waffle House, what’s around here, or briskets in Austin. There’s no bad decisions around briskets, but yeah. No, that’s the best decision, honestly. Your agent should know that. But it can modify my gym workout based on how well I slept or if I have stress or not. It has so much more context to make even better decisions than any of this app even could do it could show me UI just as I like why do I still need an app to do that why do I have to why should I Pay another subscription for something that the agent can just do now. And why do I need my eight sleep app to control my bed when I can tell the agent to, no, the agent already knows where I am, so I can like turn off what I don’t use. And I think that will translate into a whole category of apps that are no longer, I will just naturally stop using because my agent can just do it better. Lex Fridman I think you said somewhere that it might kill off 80% of apps. Yeah. Don’t you think that’s a gigantic transformative effect on just all software development? That means it might kill off a lot of software companies. Yeah. It’s a scary thing. So like, do you think about the impact that has on the economy, on just the ripple effects it has to society transforming who builds what tooling. It empowers a lot of users to get stuff done, to get stuff more efficiently, to get it done cheaper. Peter Steinberger There’s also new services that we will need, right? For example, I want my agent to have an allowance. Like, you solve problems for me, he has like a hundred bucks in order to solve problems for me and if i tell it to order me food maybe it uses a service maybe it uses something like rent a human So i just get that done for me i don’t actually care i care about solve my problem uh there’s space for new companies that solve that well. Maybe not all apps disappear. Maybe some transform into being API. (Time 3:02:19)
  • Programming Will Shift, Not Vanish
    • Programming as a craft will shift: agents automate routine coding while humans focus on higher-level design, vision, and product choices.
    • The role of ‘programmer’ becomes a builder/architect who guides and curates agent output. Transcript: Peter Steinberger Do you think AI replaces programmers completely, human programmers? I mean, we’re definitely going in that direction. Programming is just a part of building products. So maybe I’d just replace programmers eventually, but there’s so much more to that art. Speaker 3 Like, what do you actually want to build? How should it feel? How’s the architecture? Peter Steinberger I don’t think agents will replace all of that. Yeah, like just the actual art of programming, it will stay there, but it’s going to be like knitting. You know, like people do that because they like it, not because it makes any sense so so i read this article this morning about someone that it’s okay to mourn our craft and i can a part of Me very strongly resonates with that because in my past i i spent a lot of time sinkering just being really deep in the flow and just like cranking out code and like finding really beautiful Solutions. And yes, in a way, it’s sad because that will go away. I also got a lot of joy out of just writing code and being really deep in my thoughts and forgetting time and space and just being in this beautiful state of flow, but you can get the same State of flow. I get a similar state of flow by working with agents and building and thinking really hard about problems. It is different, and it’s okay to mourn it, but it’s not something we can fight. The world for a long time, there was a lack of intelligence, if you see it like that, of people building things. And that’s why salaries of software developers reached stupidly high amounts, and then we’ll go away. Speaker 3 There will still be a lot of demand for people that understand how to build things. (Time 3:11:00)