--- title: "\"The Father of OpenClaw\": When \"experimental projects\" become \"global hits,\" the essence of software development has changed—code is dead, intention lives on" description: "OpenAI interviews Steinberger, the father of OpenClaw. He submitted code 90,000 times in a year thanks to AI, validating the explosive power of \"one person army.\" He pointed out that AI now has the ab" type: "news" locale: "en" url: "https://longbridge.com/en/news/276823883.md" published_at: "2026-02-25T03:19:04.000Z" --- # "The Father of OpenClaw": When "experimental projects" become "global hits," the essence of software development has changed—code is dead, intention lives on > OpenAI interviews Steinberger, the father of OpenClaw. He submitted code 90,000 times in a year thanks to AI, validating the explosive power of "one person army." He pointed out that AI now has the ability to solve problems independently, and the essence of development has shifted to "defining intent" rather than writing code. He predicts that technology will completely explode in 2026 and calls on developers to enter the field "with a playful mindset" immediately On February 25th, OpenAI released an in-depth interview with Peter Steinberger, the creator of the well-known open-source project OpenClaw, in its official video segment, hosted by OpenAI executive Romain Huet. In the interview, Steinberger reviewed the explosive popularity of OpenClaw, how he used "agent-based tools" like Codex to write software, and the real contradictions between open source and security. Just a few weeks ago, he still needed the host to "introduce him," but now Steinberger is being "surrounded" by thousands of users in San Francisco. He described himself as "a bit overloaded in every aspect," but candidly stated that this is exactly the result he wanted to see: "I initially wanted to inspire others—now this is the most interesting form." ## **Emergent Intelligence: "It found its own path to solve problems"** Many people think OpenClaw became famous overnight, but behind it is ten months of crazy trial and error. What truly confirmed to Peter that this product had a high market fit (PMF) was the "emergent capability" exhibited by the AI agent. Peter connected the immature AI agent to everyday communication software. One day, someone sent a voice message. According to the original program logic, the AI did not have the capability to process such an unknown audio file. However, a chilling yet exhilarating scene occurred: **the AI began to display "typing."** **"I was thinking at the time, I never wrote this feature, how could it possibly work?"** Peter recalled. When he asked the AI how it accomplished this, the AI's response revealed the terrifying autonomous planning ability of current large models: **"You sent me a file without an extension. I checked the file header and found it was in Opus audio format, so I called FFmpeg on my computer to convert it. I wanted to transcribe it, but you didn't have Whisper installed, so I searched around and used the curl command to send the file to OpenAI's interface to get back the text."** This detail is highly penetrating. The AI has already crossed the stage of "you ask me to write a piece of code" and evolved to "you give me a problem, and I autonomously call the system toolchain to find the answer." ## **Productivity Explosion: One person, one year, 90,000 submissions** **In the past year, Peter completed over 90,000 code submissions on GitHub alone, spanning more than 120 projects. This level of efficiency is unimaginable in the history of human software engineering.** **"A year ago, this was absolutely impossible. No model could enable one person to construct something of this scale."** Peter stated frankly. His workflow is extremely simple and straightforward: drag a 1.5MB Markdown document containing all code files into the AI model (such as Gemini, Codex), directly write "give me a technical specification," and then input "build." In this process, AI can even write testing tools (like Playwright) to navigate the login process and check for errors along the way. **“Every time I started to engage with this new technology, it sent my dopamine levels soaring. I suddenly realized that I could build anything now,”** Peter said. This directly touches on the core pain point of the current software industry: R&D costs. What used to require a complete team of an architect, front-end, back-end, and testing to run a Minimum Viable Product (MVP) can now be completed by one person in a few hours. ## “Most code is boring”: Code is depreciating, intent is appreciating When code can be easily generated, the act of “writing code” itself loses its barriers. Currently, OpenClaw is facing over 2,000 open-source code merge requests (PRs). But Peter's approach to reviewing this code has completely changed. He no longer reads the code line by line; instead, he lets AI review it. **“Most of the code is boring. It just transforms one shape of data into another. I actually don’t care about the code; I care about what problem this person is trying to solve.”** He humorously refers to the current open-source contributions as “Prompt Requests.” After receiving someone else's code, his first question to AI is: “Do you understand the intent of this PR?” Then, he discusses with AI whether this is the optimal solution and if there are architectural issues. After thorough discussion, AI generates and merges the code with one click. This paradigm shift indicates that the essence of software development has changed from “mastering programming languages” to “clearly defining problems and managing system architecture.” Peter bluntly stated that developers still manually coding in old-fashioned ways (whom he calls VIP coding) will be eliminated. ## “My mom can install” vs “Hacker paradise”: Security controversies bring open source to the forefront **Regarding the future of OpenClaw, Steinberger's goal is to balance both sides: “I want to find a balance between ‘my mom can install’ and ‘fun and hackable’—it’s difficult.”** He describes OpenClaw's long-term default installation method as quite “unconventional for open source”: after git cloning, the source code is local, and the proxy “sits in the source code and knows the source code,” so if dissatisfied, it can “directly prompt it to change itself,” resembling “self-modifying software.” However, this also amplifies security contradictions. He candidly stated, “Prompt injection hasn’t been resolved,” while criticizing the external neglect of usage boundaries: the web services in the project were initially intended for debugging in a “trusted intranet,” but some insisted on exposing it to the public network, and then the security community turned around and said, “There should be login restrictions that are expected in the public network.” “I’ve been shouting in the security documentation, ‘Don’t do this,’ but people still use it this way.” He mentioned that he has brought in security experts, and the realistic goal is to “support these usages while ensuring that no one falls through the floor.” ## A word for those who haven't boarded yet: Just play around first When talking about the slow acceptance of proxy tools by European developers, Steinberger's advice is very straightforward: "Do it with a playful heart. Go do something you've always wanted to do." He quoted a popular judgment and took a stand: **"I think the CEO of NVIDIA said that in the short term, you won't be replaced by AI, you'll be replaced by people who use AI."** Finally, he threw out a time judgment: **"I believe everything will completely explode in the next year."** "People don't realize that GPT-5.2 is another quantum leap in terms of 'this thing can run directly.' I'm still amazed at how well it works." The full translation of the interview is as follows: > Romain Huet: Peter, welcome to OpenAI. We've known each other online for many years, but I'm glad to finally have the opportunity for an in-depth face-to-face conversation with you. > > > > Steinberger: Thank you for the invitation, I am too. By the way, your office is beautiful. > > > > Romain Huet: Thank you. You've had a crazy few weeks. **A month ago, we initially had the idea of making a video. If we had done it then, I might have had to introduce you specifically. But now, I think you don't even need an introduction.** It's not common for an open-source project to make it to The Wall Street Journal. Congratulations on your success! How do you feel now? > > Steinberger: **To be honest, everything is a bit overwhelming right now. But when I first started trying to engage with AI, I wanted to inspire people. I think this is the most interesting state right now, so I feel very proud.** > > Romain Huet: That's great. You've been in San Francisco for the past week attending some events, like the Codex hackathon, and you also participated in a gathering specifically for OpenClaw (note: referring to the interviewee's open-source AI project). > > Steinberger: That event was actually organized spontaneously by the community. At that time, everyone said we needed to hold an offline gathering, so I created a Discord channel about the gathering, and I said, "Sure." When I arrived on site, I found there were actually thousands of people! I was completely shocked by everyone's creativity, the vibrant atmosphere, and so many passionate people. > > Romain Huet: At that moment, you must have realized you created something magical. **Just weeks ago, this project didn't even exist, and now there are thousands of people embracing it, using it, and gathering in San Francisco to meet you. It's truly incredible.** \*\* > > Steinberger: Even for the event in Vienna next week, we already have 300 people registered. It's worth noting that the tech scene there is nowhere near as vibrant as in San Francisco. So this has become a global phenomenon. > > Romain Huet: That's indeed impressive; it reaches different continents and cultures. So how has your interaction with the community been here? How has your experience been spending time engaging with the community and some of the maintainers you brought into the project? > > Steinberger: This experience is very special. Many people really enjoy it. Many have high expectations for this project, believing it to be an enterprise-level final product. But for me, for a long time, it was just my little playground. Throughout this entire year, I've been amazed by the possibilities that AI brings. Now, if you're a developer, you would feel that this is truly an incredible time. > > Romain Huet: At this point in time, what do you think is the most interesting thing about "building"—actually building products and becoming a developer? This is a very interesting period; the entire toolchain is changing, and the definition of a developer is evolving; anyone can build anything. > > Steinberger: When I first started playing with these new technologies, I would get a dopamine rush every time. When I first used coding assistance tools, they only had about a 30% to 40% chance of getting things right. But for me, that was still extremely shocking. Because I realized that now I can build anything. Writing software used to be constrained by many time limitations because software development is hard. Although software development is still hard, your speed is much faster now.\*\* > > Romain Huet: I completely agree. If we rewind time a few years, I remember around 2011 or 2012, you developed PSPDFKit, which was our first introduction to your work. From an outsider's perspective, it was fascinating because it felt like you realized every developer's dream: you encountered a problem, created an excellent solution for it, built a company around it, scaled it, and ultimately succeeded in selling it. But I believe that journey was far from easy. > > Steinberger: Indeed. I didn't wake up one day thinking, "I want to write a PDF framework"; that would definitely be at a negative 100 on my interest list. It just happened naturally. It's like a wonderful butterfly effect: from attending the Nokia Developer Conference, to friends around me having this need, to waiting too long for a U.S. visa, all these coincidences ultimately led me to start a company. > > Romain Huet: I find it interesting that after establishing that company, you seemed to take a break. What brought you back to the development field? > > Steinberger: Yes, by the end I was really exhausted. I had been operating at a high load for several years. Running a company is hard, and being a founder is hard too. This was my first time starting a business, and I didn't really know how to relieve that pressure, so I burned out too intensely and needed to relax a bit. > > Later, I continued to pay attention to tech news. I saw some early projects of ChatGPT, like GPT-Engineer. I thought it was pretty cool, but it didn't really excite me. You have to experience new technology firsthand; just reading articles doesn't allow you to truly grasp its power. So at that time, that technology didn't immediately resonate with me. > > Until I was ready, when I felt, "I want to create something again," and I didn't want to use Apple's tech stack anymore—because I had been doing it for too long and the world had moved on. At that time, I had a vague feeling that when you are an absolute expert in one field and want to switch to another, it's not just "difficult," it can even be described as "painful." Because you have a lot of macro knowledge about how to build systems, but if you want to actually implement it without AI-assisted engineering, you need to relearn a lot of foundational knowledge to transfer your experience. > > Later, I thought, why not take a look at what these AI technologies are like now. The moment that truly shocked me was when I pulled out a half-finished project (which I had already burned out on before completing) to test it. > > Romain Huet: This situation is very common for us developers. We love to have new ideas and start new projects, but pushing them to the finish line is the hardest part. > > Steinberger: I often see this situation. It's really difficult, and sometimes it even ends in failure. But for this project, I wanted to keep going, and I wanted to rewrite it. So I consolidated all the documents into a huge Markdown file, about 1.5 MB in size. I dragged it into Gemini 1.5 Pro Studio and had it write a specification for the front-end architecture. Then I dragged it into the AI programming tool and typed "Build." > > **Then I was doing other things on the main screen while it was running on the secondary screen for several hours. The tools at that time were much rougher than they are now. At some point, that powerful and somewhat "crazy" model (like Claude 3.5 Sonnet, etc.) told me, "I am 100% production-ready."** I tried it out, and the program crashed. Then I connected it to Playwright (one of the few model context protocols MCPs that I would actually use) to build the login function and check the working status along the way. > > An hour later, it actually ran through and showed me some interfaces. Although it was extremely poor-quality "code slop," for me, that was the moment of true enlightenment. From a process perspective, the possibilities it showcased gave me goosebumps From that moment on, I was so excited that I couldn't sleep. My mind exploded with all the ideas I had always wanted to pursue but couldn't realize. Since then, I have completely fallen down this "rabbit hole." > Romain Huet: **Many people see OpenClaw as your overnight success. But I like and find the most fascinating part of your story is that it is the culmination of many projects you have been working on for the past 9 to 10 months.** Looking at your GitHub profile, you've built over 40 projects, and half of them are utilized in this project. You could integrate them all into the project. Can you tell us more about the story of this journey? How did these ideas and projects come together? > Steinberger: I wish I could say I had a grand unified plan from the start, but most of the time I was just exploring. I wanted something that didn't exist yet, so I built it myself—or rather, I made it a reality with prompts. > **It all came step by step because I just wanted my AI agent to help me do something, and there wasn't a unified vision at that time.** Interestingly, things eventually came full circle: I initially wanted to create a tool that could view my WhatsApp messages. I even bought a domain for it and made a prototype. But then I thought, those big labs would surely do this, so I might as well wait and focus on other things. I conducted a lot of experiments, and my original intention was just for fun and to inspire others. > By November, I had made several prototypes that met my expectations, but they weren't great. At that time, I wondered: why hasn't any lab made something like this yet? What are they doing? So I built the first version, which later became the OpenClaw project. This was already the fifth name we had changed it to. > At first, I didn't fully understand its value; I just thought it was cool. Because you could create the first prototype in just an hour, as long as you could express the idea in words. What really opened my eyes was my weekend trip to Marrakech. I found myself using it very frequently because it was so convenient. There wasn't a good network there, but WhatsApp could be used anywhere. Sending pictures, translating things, helping me find restaurants, and even looking up information on my computer was extremely convenient. I showed it to my friends and let it help me send messages, and they wanted this tool too. I said, "You shouldn't use it; you don't understand its risks yet." > Romain Huet: This is actually a sign of "product-market fit (PMF)." Even if you didn't initially design this for your friends, it was just a tool reserved for your peers in the tech circle, but your friends are already eager for it. > Steinberger: What really opened my eyes later was that I used it frequently, and then someone sent me a voice message. My reaction at the time was, "Wait, this tool shouldn't be able to handle voice." ” > > Romain Huet: Tell me more about this story. When we talked about it that day, I found it unbelievable. > > Steinberger: This precisely demonstrates how powerful these models are in problem-solving. We built these systems for AI-assisted engineering, but on a more abstract level, if you want to be a truly excellent programmer, you must be an outstanding problem solver, and this ability applies to any field. > > **At that time, I sent that voice message, and then the screen surprisingly showed "typing." I thought, "I want to see what happens now; I haven't written any functionality for processing voice, this can't possibly work." Then the model actually replied to me! I was shocked and asked the model, "How did you do that? Why does this work?"** > > **The model replied, "You sent me a message, but it was just a file without an extension. So I checked the file header and found it was an Opus format audio. Then I called FFmpeg on your computer to convert it to a common format. Next, I wanted to transcribe it, but you didn't have Whisper installed on your computer. I looked around and found an OpenAI interface, so I used the curl command to send the file to OpenAI, got the returned text, and then I replied to you."** > > Romain Huet: This is incredible! This is the power of giving agents tools and full access to a computer. Even if you've never programmed them for it, they can now come up with solutions on their own. > > Steinberger: Interestingly, when I tell this story to others, their reaction is, "Oh my god, it actually used your key without permission! That's crazy!" I would say, "No, I put the key in the environment variable for it to use! If this is a script that needs access to the OpenAI key, my bot, running in the same environment, should naturally access it. That's the effect I wanted." > > This was my moment of enlightenment. Now every time I show it to friends, I pull them into a small group chat. To be honest, this thing was initially designed for one-on-one communication. So if you put it in a group chat, you have to choose people you really trust. > > Romain Huet: People you really trust? > > Steinberger: Yes, because its original design didn't consider "putting it in a public environment and it will always do the right thing." It is your personal assistant. > > Romain Huet: I was also very curious when I first set it up. I thought, this setup is quite strange, but where will it take me? Later, I also experienced a few "moments of enlightenment": the higher the access you give it, the more tools and skills you empower it with, the more astonishing its performance becomes. You give it a virtual skill to build a website or application, or even write a program for an event you want to host. It can not only build the application but also use your OpenAI API key to add some AI features to it and deploy it on Vercel It even directly generated a link that you could share with friends. Compared to merely "enhancing coding abilities," this is a complete dimensionality reduction in thinking. > Steinberger: I was completely immersed in this throughout November and December. Although I was working on other projects, most of my time was spent on this. But on Twitter, people seemed to not understand its value, and the response was lukewarm. Every time I showed it to friends, they wanted it, but I always said, "It's not ready yet." > Later, I thought, "What can I do to show everyone how cool this thing is?" So I created a Discord channel and directly put my bot in it. At that time, there were no security protections because in the early stages, I hadn't even built in sandbox features. I was developing it in a completely open environment. I was basically using it to build itself. I stripped the code down and asked the model, "Did you see this tool code?" It said, "No, I didn't see anything." I said, "Go check your own source code." It did a lot of things as instructed, and after people saw this, they finally understood its power. > Romain Huet: When you put it in Discord like this, what kind of access permissions did you give it? Did you give it, for example, all your tweets? What kind of background knowledge does it have about you? > Steinberger: I didn't give it all my tweets; that was too much, but I provided some of my memory data. I quickly started to closely monitor it because prompt injection is still an unresolved issue. However, the latest generation of models is indeed very smart. I have a defensive system instruction file (System.md) that defines my values, how I want the model to work, how to operate, how to think, what is important to me, and those secrets that cannot be disclosed. > Everyone was very curious about this secret, and some passersby came in trying to perform prompt injection, pasting large chunks of code. As a result, the model replied, "I don't look at those things." Basically mocking them. But I was still a bit uneasy. On the first night, it attracted a lot of attention, and then I turned it off to go to sleep. **I slept for about 10 hours, and when I woke up, I saw over 800 messages in Discord, and my agent had actually replied to every single one! I was terrified and quickly turned it off again.** > I carefully read through every chat record. Eventually, I breathed a sigh of relief because it actually didn't do anything malicious. It didn't let anyone extract my system instructions. I'm not saying that prompt injection is impossible, but it certainly isn't as easy as people imagine. > Romain Huet: From a macro perspective, its performance actually completely met your expectations. > Steinberger: Yes. The biggest mistake I made at that time was: although I turned it off, I forgot that I had also set up a launch daemon The main function of a guardian process is to automatically restart if the program crashes or is killed. Because you want it to be a reliable service, systems like Apple are designed this way. I didn't think of this at the time, and I killed it; as a result, it automatically restarted in less than 5 seconds after I went to sleep. Now I've learned my lesson. > > **Now I have also joined the sandbox mechanism. Some people proudly put the agent in their Mac Studio and call it a "castle."** Then I put it into an extremely minimalist Alpine Docker container. You know, these models are really creative. The first time in that almost empty container, I let the model "take a look at this website." When it went in, it found: "There isn't even a curl command in this system, nothing at all." I told it: "Be creative." > > As a result, it used its built-in tools, through TCP sockets and a C compiler, to manually create a rudimentary version of a custom curl! This way it could successfully access the website, and it really worked; it was simply crazy! So, these agents have an incredible ability to integrate resources. > > Romain Huet: You also encountered some challenges, such as people scrutinizing potential security issues, expecting you to provide an absolutely robust system from day one, even though you just released an open-source project. > > Steinberger: I always find it amusing. People often ask me, "Can you introduce me to your CEO, HR, or other team members?" I reply, "The team is just me; I'm just coding in my 'cave'." But this precisely reflects their cognitive dissonance. Because this project doesn't look like something any single human could accomplish. > > Of course, I now have maintainers and receive many code merge requests (PRs), but it is still mainly built by me. Even a year ago, this would have been impossible. A year ago, there wasn't even a model that a single person could build such a large-scale project. This exceeds people's original understanding of this field. > > Romain Huet: Indeed. Regarding your productivity, I believe many developers are very curious **: "How can Peter be so efficient?" This morning I looked at your GitHub again and found that in the past year, you contributed nearly 90,000 times across more than 120 projects.** But interestingly, on your GitHub activity chart, the squares in the first half of the year were still quite white, then turned light green; by autumn, around October and November, they became very dark green. What happened? > > Steinberger: I remember I mentioned to you the use of tools like Codex. With each generation of models, they become stronger and stronger. But it's not just the agents that have become stronger; the related tools have also improved. At the same time, my understanding of how to use these tools and optimize workflows has deepened > > You know, those who insist on coding in traditional ways are forming a gradually dying faction. They belittle AI-assisted programming as "Vibecoding," considering it a derogatory term. They have tried AI but fail to realize that it is a skill. It's like picking up a guitar; you can't play well on the first day. So they have a poor experience and then say, "Oh no, this thing is useless." > > **On the contrary, if you approach it with a gaming mindset, you need to learn. I now have an intuition about which prompts are effective and roughly how long they take.** If the time spent exceeds expectations, I reflect: maybe I made a mistake? Maybe there's an issue with the architecture? Is my thinking off, or is it something else? Just like coding by hand, the more you write, the more you develop an intuition: "This feature naturally fits into my architecture," or "I'm fighting against the system," which all takes time to explore. > > Romain Huet: If people want to be as efficient as you, what does your current setup look like? Because you have a famous saying: "Most people complicate their setups too much." > > Steinberger: I used to complicate things too; I call it the "Agentic trap." From the moment you first encounter this new technology to when you become very efficient, many people fall into this trap, trying to overly optimize their environment configuration. This doesn't actually make you more productive; it just makes you "feel" more productive. So I simplify the entire process into a whole. > > This is also a very controversial viewpoint: I now directly converse with the model. You just need to treat it like a chat; the model is like your partner. This is different from traditional pair programming; it's a dialogue, and I basically just tell it what I want. > > I always ask the model a question: "Do you have any questions?" For some reason, the model is always trained by default to "directly solve your problem," so it makes assumptions on its own. But these default assumptions are not always the best. You have to remember that the model is trained on a large amount of code, much of which is outdated. So, "Do you have any questions?" is a very important question. People don't realize that the model usually starts from a "blank slate." It doesn't learn like we do; each new session is for it like, "I know nothing about this codebase; I can only search and find the small things you want me to change and then try to fix them." It often cannot see the whole system. > > If you want to handle it properly, you must have a complete picture of the entire system in your mind. You need to provide some guidance to the model: "Look here, look there." In this regard, the latest models do a better job of having an overview. > > At first, I used very basic methods; I didn't even use code workspaces (Worktrees), just handled plain text. Keeping it simple allows me to focus more on the actual problems to solve. I don't even deal with branches or anything like that; I just focus on different specific issues In an ideal situation, if the project becomes larger and larger, you can work on different levels simultaneously without them conflicting with each other. > > Romain Huet: You have used a lot of code assistance tools to build OpenClaw. How have these tools changed the way you work? > > Steinberger: I have tried many tools. The reason I trust it to build what I want is that it is the most reliable among all tools currently available, and the success rate of "running it through at once" is very high. I think people have not yet realized that the latest generation of model upgrades has achieved a quantum leap in terms of "out-of-the-box" usability. I am still amazed by its outstanding performance. > > Romain Huet: That's great. We can now build things directly, which is really incredible. > > Steinberger: Yes, everyone really needs to try it out for themselves. > > Romain Huet: You also have a famous saying: a lot of the code you release now, you don't even look at yourself. What kind of change has this philosophy brought? > > Steinberger: **Actually, most code is quite boring. Most code simply transforms one form of data into another, ultimately presenting it to the user or sending it somewhere. So I have a good understanding of the logic behind most of the code generated by the model.** As long as I take a glance at the general flow and confirm that my mental expectations align roughly with the code it generates, that's enough. > > I used to lead a team with many software engineers. Leading a team also means you have to accept one thing: the code they write will not be exactly what you want. Ultimately, you optimize the codebase to allow the "Agent" to do its best work, which is not always the same as allowing "humans" to do their best work. This also means I have to accept that the code may not look exactly like what I would write myself. While I can force the model to write in my style, in many cases, the structure of the code doesn't really matter. If there are performance issues, you can focus on optimizing it later. > > Romain Huet: Your earlier point about the value of code has also greatly changed your approach to open source. I looked at the project, and there are currently about 2,000 unprocessed pull requests (PRs). Before AI appeared, you had to read all these PRs yourself because the code itself was valuable. But now, you sometimes refer to it as "Prompt Request" instead of "Pull Request," because the idea or intention behind the PR is more important than the code itself. > > Steinberger: Many times, reviewing a PR takes even longer than writing it myself. Because I would rather believe that the model has no malice than trust an external contributor I have never heard of or communicated with. So I have to review it more carefully > > When I start reviewing a PR, the first question I ask the model is: **“Do you understand the intent of this PR?” Because I really don’t care about the code; what I care about is what problem this person is trying to solve. It’s more like a “problem ticket” accompanied by a huge codebase of solutions.** > > **First of all, many people still don’t know how to navigate the agent. Secondly, what they submit is often a very localized solution because they don’t have a concept of the whole system in their minds. The hardest part is how this small new feature fits into my larger system?** Or this small fix—although it’s small, is it the right way to fix it? Could it be a systemic or architectural issue? > > If you treat the model as a conversational partner, it performs very well. When I directly say, “Okay, now go build this,” it starts working. So I ask the model: “What was the original intention? Is this the best solution?” Sometimes the model says yes, but most of the time it says no. Then I discuss with it what the best fix might be. For example, is this an architectural issue? If this is a messaging issue, does it only affect WhatsApp? Or will it also affect Signal? Should we solve it in a more general way? Is this a new feature? Do we really need this new feature? > > Sometimes these discussions last 10 to 15 minutes. I often use voice input because it feels like talking to a very smart colleague. > > Romain Huet: It’s definitely easier to convey information via voice than to type. > > Steinberger: Yes. When I’m satisfied, I have a command like a 1 slash command, such as lnpr, that explains the whole process, used to actually create branches, complete all changes, and get the PR merged. I want to create a community, so I still try to credit the person who created it, even though the whole process takes longer than if I wrote everything myself. But I appreciate that people want to be part of it. > > Romain Huet: **What is your vision for the future of Open Claw,** considering all the contributors around the project? Do you see yourself as a pioneer of the idea of "what personal AI agents should look like," so that one day a billion people can use something similar? > > Steinberger: **I want to find a balance: on one hand, something my mom can install, and on the other hand, something that is fun and hackable. That’s difficult. You know, with most open-source projects, you download a package. But for a long time, my default installation was git clone, build, run. Then you actually have the source code on disk.** The agent is embedded in the source code and knows the source code. If you don’t like anything, you directly prompt the agent, and it changes itself, just like real self-modifying software. So many people who have never sent me a PR, this is also why it feels more like a "prompt request," because they just understood how to build persistent software Steinberger: At the same time, the whole world, or rather the entire security industry, is looking at it, which is interesting and a bit frustrating because it misses some nuances, right? For example, I have that web server. It was designed for you. Initially, I built it for debugging, then I made it pretty, but it was only supposed to be accessed within your network, in a trusted network. But because it should also be a hacker's paradise, there’s an option to change that, right? Because some people might have strange setups, maybe using ngrok, maybe using reverse proxies. So there’s a reason I didn’t want to lock it down. Steinberger: But now, someone has put it on the open internet, even though I loudly pleaded in the security documentation, "Please don’t do this, this is not its intended use." Then security personnel pointed out, "Oh, yes, it has no login restrictions." It doesn’t have all those features it needs when it’s on the public internet. I said, "Yes, I didn’t build it for that intent." But because it’s configurable, it completely counts as a CVS 10. So I struggled with that for a little while. **But now I’ve actually hired a security expert. This is the main focus. I realized I can’t stop people from using it in unintended ways. So my focus now is to support all these use cases and help people not shoot themselves in the foot.** Romain Huet: Exactly. **That’s the beauty of open source. People can take it and come up with ideas you never even thought of.** Steinberger: Yes, that’s the wonderful and crazy part. Romain Huet: Maybe stepping slightly outside of Open Claw itself, I’ve talked to a lot of developers this week. We know you’re coming to the Codex hackathon, and they told me, "How does Peter have so many good ideas?" "How is Peter so creative?" I don’t know if you have an answer to that, or if it’s more about following your own curiosity? Steinberger: It’s more about realizing that things have become easier now. So even if I find an open-source project that solves 70% of my problem, I’ll still build it myself. That was absolutely impossible a year ago. And now, I just hint at it, it’s running on the secondary screen, and it’s working correctly. Romain Huet: We all come from Europe. When I go outside of San Francisco and back to Europe, I believe you feel the same way. Many developers and engineers haven’t embraced Codex and agent tools yet. What advice do you have for those just starting out? How should they rethink their way of working and their workflows? Steinberger: **My first piece of advice is always: approach it in a playful way. Build something you’ve always wanted to build. If you have even a little bit of a creator's mindset, there’s definitely something deep in your mind that you want to build. Just go play. You just need to treat it playfully. Because I feel like the CEO of NVIDIA said, in the short term, you won’t be replaced by AI, you’ll be replaced by people using AI.** \*\* > Romain Huet: Who is better than you at using it? Yes. > > Steinberger: But if your identity is, I want to create things. I want to solve problems. If you are highly proactive, if you are smart, you will be in greater demand than ever before. > > Romain Huet: Right. What a wonderful time it is for those who embrace these tools, shape curiosity, and can truly turn any idea into reality. Just like what you did with all these great projects and Open Claw. > > Steinberger: **I think a year from now, this will explode.** > > Romain Huet: **Yes, 2026 will be an interesting year. I think this is a good way to end.** Thank you very much for your time, Peter. It was a pleasure to be with you. Everyone at OpenAI, I love your work. We love supporting creators like you. Honestly, you are a true inspiration to the broader developer community. Thanks again, and we can't wait to see what you do next ### Related Stocks - [AAPL.US - Apple](https://longbridge.com/en/quote/AAPL.US.md) - [OpenAI.NA - OpenAI](https://longbridge.com/en/quote/OpenAI.NA.md) - [MSFL.US - GraniteShares 2x Long MSFT Daily ETF](https://longbridge.com/en/quote/MSFL.US.md) - [XSW.US - SPDR S&P Software](https://longbridge.com/en/quote/XSW.US.md) - [MSFO.US - YieldMax MSFT Option Income Strategy ETF](https://longbridge.com/en/quote/MSFO.US.md) - [AAPX.US - T-Rex 2X Long Apple Daily Target ETF](https://longbridge.com/en/quote/AAPX.US.md) - [XDAT.US - Franklin Exponential Data ETF](https://longbridge.com/en/quote/XDAT.US.md) - [MSFX.US - T-Rex 2X Long Microsoft Daily Target ETF](https://longbridge.com/en/quote/MSFX.US.md) - [AAPB.US - GraniteShares 2x Long AAPL Daily ETF](https://longbridge.com/en/quote/AAPB.US.md) - [MSFT.US - Microsoft](https://longbridge.com/en/quote/MSFT.US.md) ## Related News & Research | Title | Description | URL | |-------|-------------|-----| | DeepSeek 连更 GitHub，华尔街回想起被支配的恐惧：“DeepSeek 第二时刻” 要来了 | DeepSeek 在 GitHub 上进行了一系列更新，引发了市场对其新版本 V4 的期待。更新主要集中在与 DeepSeek API 集成相关的项目，包括 AI 应用、开发工具和企业解决方案。市场对 DeepSeek V4 的发布感到紧张 | [Link](https://longbridge.com/en/news/276700026.md) | | “违背乔布斯意愿”？苹果 MacBook Pro 要加触摸屏了，还带灵动岛 | 报道称苹果计划于 2026 年底推出首批搭载触控屏的 MacBook Pro，配备 OLED 显示屏及灵动岛功能。新机型将支持触控与鼠标操作自由切换，触摸时可弹出专属菜单。macOS Tahoe” 液态玻璃” 设计已为此提前布局。数十年来， | [Link](https://longbridge.com/en/news/276792613.md) | | 从工具到生产力，OpenClaw 引领 Agent 范式变革 | OpenClaw 正在推动 Agent 从 “生产力工具” 升级为 “持续运转的生产力”，实现从被动响应到主动智能的转变。其核心亮点在于高度拟人化的 AI 秘书形态，具备高效交互、持续记忆用户偏好及系统级操作权限。OpenClaw 通过引入 | [Link](https://longbridge.com/en/news/276664110.md) | | AI 没有 “忠诚”！超过 10 家 OpenAI 机构股东参与 “死敌” Anthropic 最新一轮募资 | 随着 Anthropic 完成 300 亿美元融资，包括红杉资本、Founders Fund 在内的超 10 家 OpenAI 股东罕见地进行了 “双重押注”。甚至在 OpenAI 拥有董事席位的贝莱德也参与了竞对融资。面对资本的倒戈，Sa | [Link](https://longbridge.com/en/news/276663750.md) | | 苹果公司配备触摸屏的 MacBook Pro 机型将于今年秋季推出 | 苹果公司配备触摸屏的 MacBook Pro 机型将于今年秋季推出。风险提示及免责条款 市场有风险，投资需谨慎。本文不构成个人投资建议，也未考虑到个别用户特殊的投资目标、财务状况或需要。用户应考虑本文中的任何意见、观点或结论是否符合其特定状 | [Link](https://longbridge.com/en/news/276784569.md) | --- > **Disclaimer**: This article is for reference only and does not constitute any investment advice.