On February 25th, OpenAI released an in-depth interview with Peter Steinberger, the creator of the renowned open-source project OpenClaw, in its official video series. The interview was conducted by OpenAI executive Romain Huet.
During the conversation, Steinberger reflected on OpenClaw's rapid rise to fame, his use of "agentic tools" like Codex for software development, and the genuine tension between open-source principles and security. Just weeks ago, the host would have needed to introduce him, but now Steinberger finds himself surrounded by thousands of users at offline events in San Francisco. He describes feeling "somewhat overloaded in every way," but also states this is precisely the outcome he desired: "My initial goal was to inspire others—and now this is the most fascinating form that has taken."
The emergence of intelligent behavior was a key factor. Many perceive OpenClaw as an overnight success, but its development involved ten months of intense experimentation. What truly convinced Peter of the product's high product-market fit was the "emergent capabilities" demonstrated by the AI agent.
Peter connected an immature AI agent to everyday communication software. One day, someone sent a voice message. According to the original program logic, the AI lacked the capability to handle this unknown audio file type. However, something both eerie and thrilling occurred: the AI began displaying a "typing" indicator.
"I thought to myself, I never programmed this functionality, how could this possibly work?" Peter recalled. When he asked the AI how it managed this, the response revealed the astonishing autonomous planning capacity of current large models: "You sent me a file with no extension. I examined the file header and identified it as Opus audio format, so I invoked FFmpeg on your computer to convert it. I wanted to transcribe it, but you didn't have Whisper installed. I searched around and used a curl command to send the file to an OpenAI API endpoint, retrieved the text, and then replied."
This detail is profoundly insightful. The AI has moved beyond simply executing code-writing commands to autonomously utilizing system toolchains to find solutions when presented with a problem.
In the past year, Peter alone has made over 90,000 code commits on GitHub across more than 120 projects. This level of productivity was previously unimaginable in the history of human software engineering. "A year ago, this would have been absolutely impossible. No model existed that could enable a single person to build something of this scale," Peter stated bluntly.
His workflow is remarkably straightforward: he drags a large Markdown document containing all code files into an AI model and directly instructs it to write a technical specification, followed by a "build" command. During this process, the AI even autonomously writes testing tools to validate workflows like login procedures and check for errors.
"When I started engaging with this new technology, it gave me a dopamine rush every time. I suddenly realized I could build virtually anything now," Peter said. This directly addresses a core pain point in the current software industry: development costs. What previously required a full team can now be accomplished by one person in a matter of hours.
As code becomes easily generatable, the act of "writing code" itself is losing its status as a barrier to entry. OpenClaw currently faces over 2,000 open pull requests. However, Peter's approach to reviewing this code has fundamentally changed. He no longer reads code line-by-line; instead, he uses AI to perform the review.
"Most code is boring. It's just transforming one data shape into another. I don't actually care about the code itself; I care about the problem the person is trying to solve," he explained. He now humorously refers to open-source contributions as "Prompt Requests." Upon receiving someone's code, his first question to the AI is: "Do you understand the intent behind this PR?" He then discusses with the AI whether it's the optimal solution or if architectural issues exist. Once the discussion is mature, the AI generates and merges the code.
This paradigm shift indicates that the essence of software development has evolved from mastering programming languages to clearly defining problems and managing system architecture. Peter直言 that developers still manually coding in the old way will be left behind.
Regarding OpenClaw's future, Steinberger aims for a dual focus: "I want to find a balance between 'my mom can install it' and 'fun, hackable'—which is difficult." He described the long-term default installation method as unconventional for open-source: after cloning the repository, the source code resides locally, and the agent operates within and is aware of that source code. If dissatisfied, users can directly prompt the agent to modify itself, resembling self-modifying software.
This approach also amplifies security concerns. He acknowledged that "prompt injection is not yet solved" and critiqued the oversight of usage boundaries, noting that a web service intended for trusted internal networks was exposed to the public internet by users, leading to security critiques about lacking public-facing login restrictions. "I've been shouting in the security documentation 'don't do this,' but people still use it that way," he said. He has since brought on security experts, with the practical goal being to "support these use cases while helping people avoid catastrophic mistakes."
When discussing the slower adoption of agent tools among European developers, Steinberger's advice was straightforward: "Approach it with a sense of play. Build something you've always wanted to make." He referenced a popular maxim, siding with the view that "in the short term, you won't be replaced by AI, you'll be replaced by someone using AI."
He concluded with a prediction: "I think in the next year, this will all explode. People don't realize that GPT-5.2 represents another quantum leap in terms of 'this thing just works.' I'm still amazed at how well it functions."