Loading summary
Transcript1 lines
- [00:08]
A
Six months ago, I was the future. These days, the Pulse say I got mocked by an anime girl. You know, I get questions like this almost every day. Usually I ignore it or I give the media answer. Today you get some real ones. It's going to be five questions, 40 minutes. And I know what you might all came for. You came for loops, graphs, making codecs. Go brrr. This is free on my Twitter. This will help you to survive the roller coaster that might be coming for you. You know, yesterday, Boris told you to let the model cook. I'm the guy who found out what happened when tens of thousands of people let it cook at the same time. Let's start with the question everyone asks first. How is openclaw only eight months old? You know, it might be eight months in human time, but in AI times, it's more like four years. And my source of inspiration is usually being annoyed. So, on a rainy day in November, I was juggling some agents, but I was also hungry. So I wanted to make sure that my tokens are put to good use while I raid the kitchen. And there was still no good way to just send a prompt from my phone to my computer so some agent could check up how my agent's actually doing. Remember early 2025? And, you guys, this is my first silent disco. It's a very new experience. So, like, give me some hands if you actually listen to me. Yes. Amazing. Amazing. You know, early 2025, oh, my God. We're talking the old Opus 3 Gemini 2.5. OpenAI had.03, which was the most impressive one, but also really slow and expensive at times. And I got a dopamine hit when my agent did something right. I'm sure you remember this time. And these days, if they get something wrong, I usually question myself. You know, did I design the loop wrong? Did I not give my agent everything they need to verify their work? Do I have an error in my thinking? Am I asking for impossible things? So I came back from the kitchen, and my coding agent stopped because of something silly, and I was annoyed. So I spun up a new terminal session. I rambled my idea in, and I let the model cook it, built me a WhatsApp relay, and in an hour later, I could send messages from my Mac to WhatsApp and back. And that in itself felt magical, but really, it shouldn't have. It's like we all did this for months already, right? We opened the terminal, and you type something in and you get a reply. But the magic is how it feels felt. It wasn't A terminal. It wrote concise answers. It was proactive. Sometimes it would just check on me during the day. I made it so the complexity would melt away. You don't have to think about which model, which context size, or when you start a new session. And I prompted the weights to a little bit outside the default distribution, make it feel more like a friend. And I was using it. I had so many moments where I felt like, this is the future, this is AGI. And you know what? Nobody cared back then. I already had a good following on Twitter. I know it's called X now. I don't. I'll always call it Twitter. People were not getting it. I would keep trying to like, make them, make them understand, like how magical this felt to me. But week after week I failed. And you notice a pattern, it annoyed me. So I would create group chats with friends and I added them WhatsApp relay. And I would show them. I would let them talk to it. And every time I got a strong emotional reaction. Some people were amazed, some people were scared or freaked out even. But each time there was a strong emotion. And the best part, there were quite a few people that really wanted it. And especially to my non technical friends, I told them, no, this is not yet for you. And they got mad. So if that's not an indicator to have product market fit, I don't know what is. So I spent another month tweaking the details, thinking, what could I do to explain the world, you know? And then the audacity. Someone sent me a pull request to my WhatsApp relay to add Discord support. You know, like, what parts of the name are you not getting? I let the PR sit for a while and I was thinking and I was thinking, and then eventually I was like, ah, what the hell? And out of VA really became Claudius. Because I'm good with names and there's now more than one message channel. And so you understand how early this was. I didn't even had compaction built in yet. So like, at some point it would just like stop. And just in time for discord, I added a very hacky version of compaction so Mario could make a good one in PI. So it was around New Year's Eve. And as you do, I went home early to hack on it some more. And then in the first week of January, just as the geek world collectively was learning about coding agents, I created a Discord room, server guild, however we call it. And I put my claw in it. And I remember the first night vividly. People would be Joining Discord and they would like watch me build it in the open. They would try to hack it. They were like talking to it. They were getting smug replies and they see what they can do and they finally got it. You know, like this was, this was the moment. I sit up all night, I let people interact with it. I had this prompt in my agents MD that would instruct it don't do any dangerous tool calls unless the prompt comes from Peter. You know, this was like six months ago where all the things we wrote in our agent 7D flash felt more like suggestions. So I watched it very carefully. I could always pull the plug. And then at 7:00am I finally, I was done. I said good night. I pressed Control C and I went to bed. But of course like I built the software to be resilient. So it was a launch demon. You know what happens if you like press Control C on a launch demon? Yeah, it'll just, it'll be dead for five seconds and it'll start up again. So while I was walking into the bedroom, my agent happily started answering God in the world. And then I slept for like 10 hours. When I woke up, I woke up to around 800 messages. People tried to hack it. I pulled the plug. I freaked out. I read through everything and nothing actually happened. But this was the moment, this was the moment where it went viral. You all been here. You know what happened next. My inbox exploded. Reporters tried to call me in the middle of the night. Did you know about this feature that Apple built? When you put it on do not disturb, but people call you a few times in a row. The iPhone would think it's an emergency. It would still let call through. I didn't know about it, but reporters did. My email inbox was more like a waterfall. Yeah, the Mac mini sold out. I think they're still sold out. I got more invites to podcast in a month than I did in my 39 years before. And Tropik sent an email demanding the name change and like drop the lobster. And the project molted a few times. Claudis became Claudebot. Then there was a very short time where we don't talk about the name. And then it started on openclaw. Jensen called it the most successful open source project in the history of humanity. Meanwhile, my inbox was flooded with reporters, security people and countless curious humans. And also agents. The numbers still don't feel real to me. In eight months, more than 18,000 different people opened an issue or a pull request. Over 111,000 in total. I did the math. Now who am I lying to? My agent did the math. Almost 3,000 people have commits in the repo. And some people were mad at me because clearly they did it first and I stole the idea. And they sent me links of something I've never seen. Some people call me AI Jesus. For some, I became the Antichrist. At some point I started collecting the Obuterius. There's a new one every few weeks. The most recent one is actually from yesterday. What do we learn here? Be careful what you wish for. I was not ready for all this attention. It almost broke me. Like I was that close of just deleting the whole thing. I stopped answering my friends. I didn't even want to look at my phone anymore because it was just a stream. And of course somebody leaked my phone number and also a lot of private details because I'm an enemy of humanity. And here I am. Just wanted to build something cool. And now we're getting to question 2. Did you sell out? You know, this wasn't my first ride. I spent my 20s and my 30s building a B2B software company. I wrote a PDF framework by hand, like an animal. I bootstrapped the company. I grew it to almost 80 people. I ignored the competition. It became the thing most enterprises would buy. And eventually I passed things on to my co founder. I sold my shares and I very badly burned out. Retirement. The idea of it sounds great. And I spent almost three years wandering around aimlessly, catching up on life and parties. I did all the things that I kind of missed out on. Life. I had months where I didn't even open my computer. You know, just casually checking the. The Internet with my phone like a normie. I always knew that if I. If I give myself enough time, the urge, the urge would come back. I always thought the urge would be code. I always loved programming, but it took me a year to understand that that's not entirely correct. I love building. Programming was just a means to an end. And here I was eight months after I found my spark again. And my inbox was full of VCs begging me to take their money. And I did not want to. So when the big labs came knocking and suddenly I was on the phone with Mark, Sam, a few others, this seemed like a much more interesting path. It also felt unreal. Right? Talk about imposter syndrome. Maybe that's an important lesson for you. Everything you can build can be forked or cloned, but your name cannot. So your personal brand is way more important than any single product that you ever will. Start working on that before you need it. So, yeah, I know how to play the game. I also know that doing this was the equivalent of selling my soul. Md if you know, you know, the decision was not easy. But I. If I learned anything during my time after the burnout, it's. It's this. It's trusting your gut. And my gut told me that I like OpenAI the most. Did Hermes win? The short answer I can give you is, they beat us where it hurt. We're still there. And both of those are my fault and my doing. In the months after the release, we got absolutely crushed by security reports. In some ways, we were the prototype of what many open source projects now experience. And even though most were super edge, Casey, I felt the pressure. And then the press, the press said 20% of our skills are malicious. Now. We actually put a paper up with this. We did the the numbers, and the numbers are more like 0.3%. We scanned all 67,000. We ship the paper, but you know, a correction never travels as far as a scare. I felt this responsibility. The world discovered openclaw and even though there was this big scary disclaimer, actually much bigger than I could on the slides when you install it, I knew many people wouldn't read the docs. So I really focused on that. Hardening the code base, building layers and layers of security. Added sandboxing, allow lists, a web protocol with permissions built in. We were shelling out to Python for some file operations because there were some primitives that are Simply lacking in TypeScript to ensure your agent stays in the workspace, it doesn't follow symlinks and that the configuration files are written atomically. Most of the users didn't care shit about that. Sure, they like the abstract term of security, but in all practical terms, they updated. I broke something they depended on. I made things slower and I made things more difficult to update. And another thing, I have to admit, I got sloppy because now I had my time shifted between open source, the press, spending countless times on the phone with lawyers to set up a 513American nonprofit. I worked on my 01 and you know, OpenAI has its own world that is interesting and demanding. I brought on help and we got such an amazing community of maintainers that are very much respected and everyone added their own little feature. I felt like I'm a bit in a hard place. After all, these people work for free, right? So who I am, who I am to tell them what to do. And also my attention was all over. So we added A lot of features, you know, features are the fun part. A new feature is just a prompt away. The real cost comes after every feature. We shipped, of course, with a configuration option because we didn't want to break everyone's setup. At our highest, I had to count that. We ended up with around nine and a half thousand configuration options. If you count all the permutations, you can write all the tests you want. It is impossible to cover all of these and not break things from time to time. It is infinitely harder to evolve software that has users. Meanwhile, other companies, they were fueled by VC money. They pushed ahead. The Anime Girl company did it especially well. People started inserting themselves into pretty much any conversation on Twitter while we were buried in security groundwork and features. They had a simpler story, an aggressive marketing campaign, and a one liner to migrate people's claws. The main thing though, that really hurt. The project was anthropic, not the name. That part was stressful, but I kind of understood and they were really nice about was the fact that I optimized too much on the model. I built OpenClaw with Codex and GPT. But the harness for a long time really worked the best and was optimized for opus. So when they pinged me with around 24 hours notice that they're going to disable the subscription for everyone, there was not really enough time to change course. I mean, sure, we support open weight models, we did a lot of work on them, but they really weren't that great yet. And the early OpenAI models simply lacked character. So maybe write this one down. Your dependencies. Business model is your business model. You know, that's how fixed now solve is pretty awesome. Open weight models are actually good. I learned a lot about harness engineering, but in many ways people moved on. You can see that in our download charts. We bottomed out at around 835,000 weekly downloads in May. And then after being declared death in June, we peaked at 4.7 million, the highest ever. Both of those are true at the same time. A hype is like the weather. You might see it coming, but you can't control it. In my case, it was a storm. Question four, is it still fun? You know, somewhere around February, it stopped being fun. I started to feel like. It started to feel like a responsibility. I woke up and the man who didn't want to build another company found himself in a situation where he had two jobs, or should I say a job and a calling. I was struggling and worst of all, I stopped using my own product. Somewhere in that Time I stopped making a product I love and I worked on making something for everyone, it became less of a thing that I use every day and more, I think that I saw and felt was work and I wanted to picture who everyone really is. One of them had a grocery agent, another one tried to social engineer my bot. Between everyone adding features and all the organization work, I became the person that fixed the bugs, fixed the security issues and provided support and built a foundation. And because all these people were like giving the rumor mill, oh, Open Cloud is not owned by OpenAI. I didn't want to take too much help from OpenAI. I mean, yeah, they gave me token and I, boy, did I use that. But I was also pulled into other projects and there was a whole flood of people that I couldn't as easily block who wanted to talk to me. In hindsight, I could have made it done differently. Like, I could have asked for more help, I could have moved responsibility off my plate. But I was so deep in everything that I didn't take time to think through strategically. Luckily, I met a bunch of amazing people and little by little, things aligned. I figured out my visa, eventually formed a nonprofit. We got amazing companies as donors, and I found some really good people that believe in open source and started working with me. I also need to give Nvidia a special shout out because they were very early and they simply asked me what I need and then they sent people to take over much of the security work. And I think it was somewhere in May, around my birthday, where I had this feeling that things are starting to feel good again, where the joy of building for me was coming back. And yeah, somewhere in the press, the prestigoing and openclock Killer. Every other week at some point, I think I counted 20 of them. There's even literally a project called OpenClock Killer. It's an uninstaller, which is totally unnecessary because we have an uninstaller, but, you know. But none of the killer stories ever picked up what this is actually about. Open source, my source of inspiration is being annoyed. These days I get annoyed when I have to use software where I can't just send a prompt to my agent to change it. That's the fun part that's coming back. It's hard to compete with someone who's just there having fun. Is velocity. The weeks I enjoyed building the product got visibly better. The wix that didn't. We should configure options. Question five, what's next? Some people in the audience might be like, is this finally the part where he talks about graphs. First of all, I'm super happy with where we are with the foundation. Our mission is to bring people closer to AI. And things are changing so fast for many people it feels scary. I'm proud of one thing that openclaw achieved and that for many people it moved AI from this thing that's like nebulous and scary into something that is fun and weird. You know, lobsters and all of that. And they'll keep pushing there, keep building a great ecosystem of open source software with events that bring people together and with education. We have 10 people now on payroll. We're hiring a few more roles, including a CEO. Second of all, I kind of made claw a noun. You know, Karpathi dropped the open satya says enterprise grade clause in Microsoft's keyNote. There are 33,000 claw named repositories. If we are in a simulation, we are certainly in one of the weirder ones that won't get shut down. I dig this part of the future. Third, we still don't have an agent that's always on, always syncing. There's still so much to do. The landscape and the tech in AI are evolving faster than the software that we build around. And that's an opportunity for all of you. Our workflows are also evolving. Like the vision I had early on where we shouldn't think about session or compaction. The world is finally the models. That tech is finally getting to a stage where this is becoming a reality. We're entering a world where we finally move on from the text, text only interfaces to voice and multimodality. Just yesterday we got a hack working that your claw can now FaceTime you. And this is kind of what matters, why Openclaw exists at all. Every lab will sell you an agent. OpenCloud is the alternative. Open source runs everywhere, works with any model. And if you run local models, your data never has to leave. Your device, your agent, your machine, your life. That's not even my thesis, that's Gary's. You know, we just shipped it first. So we also finally are building openclaw with openclaw again, but this time with a twist where everyone sees each other's session on a team server and there's a claw that knows what everyone else is working on and can also take over orchestration of the work. We are finally leaving this slowly, this weird blip in the future where people work in terminals and run around with their laptops open because the agent needs to keep working. If you only remember three things, number One, don't stop having fun. Find is the ultimate driver so you can get the best ideas. Number two, listen to your gut and also fix the things that annoy you. It might just become the next big thing. And number three, stay focused. You know, another podcast is not what will make you win. Live in the future. Build what's missing and when they write your obituary, keep shipping. It confuses them. Now, these were five questions. I'm sure you have many more. And I think we have plenty of time for qa. So first it was agents, then loops, then graphs. How are you building these days? I mean, it was always sessions, right? But in the beginning you had to actually care that you would clean and you would make sure that your instructions are coherent. And these days it's more like my sessions are topics, clearing the session, sometimes even a disadvantage because there's so much information in there that helps the agent. But the big thing, the big thing, the big thing that shifted is I try to make the agent do more proactive work for me. So when I shift my attention to something, I don't want to read issues, I want to see fully reviewed and tested PRs. Maybe I like the feature, maybe I don't, but I don't want to split my attention at the same time. At work, if you come to me and you tell me this feature idea, I'm going to get mad at you. It's so easy that you just discuss the feature idea with an agent, you build it, you make screenshots, you let me play with it. That way if it's good, we can immediately iterate. But most of the time people will figure out why it's not good and don't even come to me. What comes after graphs and what's next for openclaw? I know sometimes I like to shitpost on a Sunday and then my Twitter explodes. But also it's not a shitpost, right? So anytime you know what we as engineers did for the longest time, we built automations to make our life easier. That's been going on since our profession exists. So if you design, call it loop, call it graph, call it workflow, it's kind of all the same. If you design something that gets a trigger, gets an input and does something for you, and maybe there's a decision in the way, boom, there's your graph. It's not magical. Even though, like, you know, I favorited some of those graph engineering posts because I'm kind of curious what their opinion is that it would be in my world. It's just like it's just a better way of our automation story that we did for so long already. In a world of build fast, ship fast, how to make sure that what you build is still reliable and scalable. Yeah, that's the part I messed up for a while because I was not focused and the models were not really good at testing. I think that the models are really good now. We not just have sessions that remember we have orchestration trained into models so they really understand and know to use subagents. We have computer use, we have browser use. Like all of that together is like your perfect Q and A environment. Like I did that yesterday where I spin up codecs, use 12 sub agents, understand my project, break it down into features and then each sub agent would like stress test a feature or code review feature and then inform the other session where to focus testing on. So we're not yet at the point where you can automate everything. You still need to do some manual click throughs so that you know how it feels. But like for a lot of the typical bugs that users will encounter, you can actually now you can prompt your way and you get very far. Okay, AI tools, what's your decision you made purely for speed early that you still live with today? I think I was really early in deciding that I don't read all the code. I see code review more as risk management. Sometimes you touch a system that is scary, you kind of want to read a little bit closer. And other times you build ui. Do I really care the UI is built if it looks correct? No. So you kind of, you glance over or you simply, you simply accept that it looks right. Part of it is just like you develop a little bit of a feeling how long something should take. So if I do a small tweak that should like change how dragging works and it takes three hours, I know something's wrong, I'll look closely. But otherwise part of code review is simply like observing, looking how big it changes and trusting your gut. Okay, AI tools make it easy to build a product in a weekend now. But building isn't the hard part. Getting people to use this. How would you go from a working prototype to your first 10 releases today? Well, I think that was part of what I addressed in my Talk. User number one should be you. And my users 2 to 20 were friends. So I'm sure you have some people that would actually test your stuff. But you need to be the first user. If you don't get excited about building, it'll probably not make sense like in this day and Age eyeballs are kind of the most expensive currency because building something so fast. Now, how do you balance solving issues that annoy you and make features you think people you want? That's a hard one because oftentimes it kind of goes together. I get annoyed if something's not there as I want it. To be honest, I haven't fully figured that part out yet. I feel I could just work for months on issues because there's always going to be weird edge cases in software, especially if you add more and more features. But then again, if I only do that, I'll lose interest in working. So it needs to be like, it needs to be a healthy mix. If you could back and change the thing. One thing about openclaw, what would you change? I would be less stressed out about security researchers. They are really good at making you feel really bad. They would send a report, they would email me, they would call me, they would through everything they can possibly get to get my attention, but not actually to help the product. In most of the case, it's really just for them to get clout, for them to get a point, oh, we found something. And most of them really sent reports that their agent produced without actually even testing it. I would take a stronger sense of explaining what are the parts that we guarantee and what are the parts that will not be fixed because that's not our security boundary. But honestly, I had no. This was my first time I was exposed to this world, so I just didn't know how to handle that. And I lost a few months and got a few gray hairs to learn. What's the single biggest bottleneck in today's agentic infrastructure. Reliability tooling. Memory evolves. Managing compute. If that makes sense. Like if I run one test locally in my machine, TS Go will spin up 16 threads and will bottleneck my machine. If 10 sessions do that, two of them will probably time out and we'll have to do it again. And there's no really good system of managing all of that. Now. If I do web stuff, it's easy to create a cloud session. If I do something that requires macOS, that's already where 99% of the tools fail me. If it's something that needs other stuff that's on my computer. Same. And we haven't really built something yet where things could easily move from here to here without issues. At least I didn't see good stuff working yet. Or where I could drive a fleet in a reliable way. Like right now. I use too many systems. I screen share into some Computers to distribute the load. I shouldn't be doing that. How do you keep an open source project opinionated and pointed in one direction? Whether they're just merging the N1PR that adds a random feature and how do you communicate that direction to the community? Have you ever said no to popular PR because it pulled the project off course? Yeah. Yeah, I would say I didn't say no enough. These days when I do new open source, I write a Vision MD file where I kind of explain what it is now where I see this to be. But of course it's incorrect. It's not perfect science and it's something I need to adhere to better because it is always tempting to add this one feature that looks really cool. But what you rarely think about is when you merge this feature, it really means here's this pile of code that I don't need that the person probably doesn't really understand that I don't fully understand that I also got on responsibility. When do you expect to have claws that are perpetually running that are proactive in their work? I mean, honestly that's not so much a tech problem, it's more a token problem where we could do that today but you wouldn't get very far with your subscription. And not everyone is willing to spend so many tokens. It's also but a part is not easy is designing a system that doesn't just burn empty tokens. Even my early system of like heartbeats was too static and not proactive enough. And then especially bad if you have a large session and then one hour later you call a heartbeat checkup on everything. And that means you send 600,000 tokens back to the server after the KV cache cleared. And you pay a stupid amount of money for a lot of not useful work. So there's a lot of things that can be done to optimize that. It's also really really hard. Can you tell us more about your own personal setup? What do you host your claws on? What model are you running? What harness do you use? Do you still read? Can we scroll? Do you still read your code? I use my MacBook but I usually use Jump desktop to screenshot into my studio and then just use the computer there because that one's always running that one. I can run whatever I want fast and not drain my battery. And I can close my laptop and things keep working. And I have a few other remote machines where I sometimes VNC in. The part that's really nice about it is because I do a lot of Mac software agents Love to take over my screen and click around. And if you put that in their own, if you give them their own machine, it will not bother you. Otherwise you'll fight with the agent for the mouse cursor. Do you still read your code? I think I answered that a little bit with the term risk management. And also if I do open source by myself, it's a different risk management than if I do software at OpenAI where we do still read all of the the code. Peter, how would you approach your next startup idea? I don't know. Should I tell this to everyone? Like, I feel like it gets a lot of copies. I mean, the most important thing is you need to build something that you want to use, otherwise it'll just not going to be good. Second of all is I hope you worked on your personal brand and on visibility because in this day and age there's so much noise out there that your hardest problem is not the tech, it's not the software, not even the people. The hardest problem now is like getting eyeballs. Maybe I would pick something again that's in the category hard and boring because that's usually the category that is a little bit easier to actually find people that will appreciate when you solved something. If you pick something that is fun, even if it's hard, you're going to have a very tough time. Especially in a time where people can just prompt things into existence. All right, last one for today. What's a product you want someone to build? You know, I feel it's very easy to have to get a test box for Linux. It is unreasonably hard to get one for Mac that works really well. And all the stuff that I got for Windows is also quite annoying. I haven't found a really good provider yet that does all of that fast and cheap. I'm not sure if this is the business you want to be in because devtools are inherently hard, but that's something I would love to have. Thanks everyone.