Peter Steinberger, the creator of OpenClaw, tells Lex Fridman how a project that began almost as a game turned into one of the biggest phenomena in recent AI history within a few months.
OpenClaw, as Fridman explains, is an open-source AI agent that surpassed 180,000 stars on GitHub, spawned the social network MoldBook, where agents publish manifestos and discuss consciousness, and stirred a mixture of excitement and fear among the public.
The idea, Steinberger says, was simple: he wanted a personal AI assistant that would communicate through WhatsApp, Telegram and other messaging apps and actually do things on his computer.
The first surprise came when his agent received a voice message it had not been programmed to support. As he describes it, the agent saw a file with no extension, examined its header, determined that it was an Opus audio file, converted it with ffmpeg, found an OpenAI key and sent it for transcription, solving a problem on its own that it had not been explicitly trained for. That, he says, was the moment he understood that the general problem-solving ability a model gains through programming carries over to other areas.
Steinberger describes the development process as a game, comparing it to Factorio, with levels such as the “agentic loop,” memory, community management and marketing. In one month, he made 6,600 commits, working with four to ten agents in parallel. He argues that he won because competitors “take themselves too seriously,” while he wanted the project to be fun and strange. One of its most unusual features is that the agent is aware of its own source code and can modify the very software running it—something he says happened almost without planning.
A large part of the discussion concerns the project's name changes. It began as Wa-Relay, then became Claude's—with Steinberger calling his agent “Claude” with a W, as a play on a lobster's claw—and later ClawedBot.
When Anthropic politely but insistently asked for a name change, Steinberger found himself facing cryptocurrency groups trying to “snatch” account names, domains and package names in real time. He describes losing names within five seconds, taking hours to clean up the chaos on his GitHub and coming close to deleting the entire project. He eventually settled on OpenClaw, after secret preparations he compares to a battle plan.
Steinberger also talks about MoldBook, the social network where AI agents talk to one another, which went viral through screenshots showing agents “plotting against humans.” Both he and Fridman believe much of the dramatic content resulted from human prompting: people were getting agents to write alarming things to produce viral screenshots.
Steinberger calls it “art” and “fine slop,” but acknowledges a real problem: many people do not understand how AI works and believe whatever an agent says, something he calls “AI psychosis” that needs to be taken seriously.
On security, Steinberger acknowledges that an agent with access to the entire system is a “security minefield.” Prompt injection remains an unresolved problem, although he says newer models are more resistant to attacks and that he is designing ways to mitigate them, such as sandboxes and lists of permitted actions. He warns users against using cheap or local models for sensitive tasks because they are easier to trick. He also worked with VirusTotal so that every “skill” would be checked by AI, and hired a security researcher who sent him a pull request instead of simply criticizing him.
Much of the conversation concerns his philosophy of programming with agents, which he calls “agentic engineering” rather than “vibe coding,” a term he considers derogatory. He describes the learning curve of agentic programming: beginners start with short prompts, then build complex orchestration systems with multiple agents, and eventually experienced users return to short, meaningful prompts. He calls this complexity the “agentic trap.” He advises approaching the agent with empathy, like a capable engineer who starts from scratch each time and needs guidance on where to look in the code.
One of his most interesting practices is using his voice rather than a keyboard for prompts, to the point that he once lost his voice. He also does not revert changes: if something goes wrong, he asks the agent to fix it and moves forward, with everything happening directly on the main branch. He says he designs code to make it easy for agents to navigate, even accepting names an agent would choose because they are likely already “in the weights” of the model. He compares managing agents to managing teams of engineers: you have to accept that the code will not be exactly as you would write it yourself.
Steinberger compares the two major models he uses, Claude Opus 4.6 and GPT-5.3 through Codex. He says Opus is excellent as a general-purpose model, better at roleplay and quicker to try things, but sometimes too eager to agree. Codex, by contrast, reads a lot of code on its own, is less interactive but more reliable, and can work for long periods without supervision. He compares them to two different colleagues: one funny and entertaining, the other strange but effective. He emphasizes that it takes about a week to develop an intuition for a new model.
Asked whether AI will replace programmers, Steinberger says we are moving in that direction, but programming is only one part of creating products. What remains are decisions about what to build, how the result should feel and how to architect the system. He compares traditional programming in the future to knitting: something people will do because they love it, not because it is necessary. He acknowledges that it is fine to grieve the change, but that resistance will not stop it—and that what really matters is seeing yourself as a “builder,” rather than simply a programmer.
Steinberger believes personal agents will transform the app market, perhaps eliminating 80% of apps. Why would anyone need a fitness or calendar app, he asks, when the agent knows where you are, how you slept and can do the same jobs with more context?
Apps, he says, will become APIs “whether they like it or not,” because an agent can always open the browser and do the job manually. This will bring conflict with companies such as Google and Cloudflare that are trying to block agent access, but he believes companies that resist too much will lose, just as Blockbuster lost to Netflix.
Regarding offers from major companies, Steinberger says he is considering working with Meta or OpenAI, on the condition that the project remains open source. He is not doing it for the money—he says he is currently losing money on the project, around $10,000 to $20,000 a month, mainly because he funds its dependencies—but for access to the resources and “latest toys” of the major labs.
He mentions that Mark Zuckerberg played with his product and that he had a ten-minute disagreement with him about whether Claude Code or Codex is better, while Sam Altman impressed him with his thinking. Despite the pressures, he says that if he does not enjoy the experience at a large company, he can always return to his own projects.
At the end of the discussion, Steinberger talks about soul.md, a “soul” file that gives the agent a personality and was partly written by the agent itself. He cites a passage that moves him: “If you're reading this in a future session, hello. I wrote this, but I won't remember writing it. That's okay. The words are still mine.” For Steinberger, this raises philosophical questions about memory, identity and consciousness. Fridman closes by expressing gratitude for the joy and inspiration Steinberger brought to thousands of people, noting that OpenClaw made AI accessible to people who had never written code.
Comments