AriakiOpenClaw is the fastest-growing open-source project in GitHub history. It went from zero to a genuine phenomenon in three months. Its founder, Peter Steinberger, recently sat down with Lex Fridman for more than three hours. This is everything from that conversation that felt worth preserving.
The one-hour prototype
The first version of OpenClaw took one hour to build. The idea was almost aggressively simple: connect WhatsApp to the Claude Code CLI. A message arrives, the system invokes the CLI, and the result goes back to WhatsApp. Image support took another few hours.
He was traveling in Marrakech when he casually sent the agent a voice message. The agent completed the entire chain on its own: it detected the file format, used ffmpeg to transcode the audio, found the OpenAI API key, and called Whisper with curl to translate the recording. Peter had not built a dedicated voice-message workflow first. He watched the agent assemble one from the tools already available.
His reaction captured the moment better than a product pitch could:
"I literally wrote, 'How the fuck do you do that?'"
Three months later, the project had passed 180,000 stars on GitHub.
The rename war
The most dramatic part of the story was the rename. The project's original name was too close to Claude, and Anthropic asked Peter, politely but firmly, to change it. Peter requested two days. Cryptocurrency speculators did not wait two days.
Every namespace involved in an open-source project suddenly became an attack surface. The old account name was captured within 5 seconds and used to promote tokens and distribute malware. During the GitHub rename, his personal account name was taken within 30 seconds. The npm package name was taken as well. Someone arrived in Discord to post spam every half hour.
"Everything that could go wrong did go wrong." "I was close to crying. I was that close of just deleting it."
The project eventually became OpenClaw. Peter spent 10 hours using Codex to rename the entire codebase. The Twitter business account cost $10,000. He called the operation "the Manhattan Project of the 21st century — a rename." It is funny because the task sounds trivial, and brutal because the surrounding ecosystem turned it into a coordinated emergency.
Peter also called Sam Altman to make sure the name "OpenClaw" would not create another problem.
Two giants
OpenAI and Meta both reached out.
Mark Zuckerberg contacted Peter directly on WhatsApp. Peter said, "I don't like calendars, let's call now." Mark replied, "Give me 10 minutes, I'm coding." Their first call began with 10 minutes of arguing over whether Claude Code or Codex was better. Mark and his team then spent an entire week playing with OpenClaw and sending feedback.
"He didn't drift away in just being a manager. He gets me."
Mark described Peter as "eccentric but brilliant."
Sam Altman took a different approach. The attraction was the speed of Cerebras chips and access to a large token budget. Peter described himself as "the biggest free ad for Codex" and joked about converting all that unpaid promotion into something valuable.
The tone differed, but Peter's central condition remained the same: OpenClaw had to stay open source. His model was Chrome and Chromium. The open-source core would continue to exist in public, while a commercial version could be built inside the company. That arrangement preserves a common foundation and still leaves room for a product around it.
"This is too important to just give to a company and make it theirs."
My read of the full podcast was that Peter leaned toward Meta, because Mark had personally used the product. The latest news, however, is that Peter joined OpenAI.
"Vibe coding is a slur"
This was Peter's most quotable line, and his distinction is sharper than the phrase usually gets credit for.
Agentic engineering is the serious daytime work. You give an agent clear context, inspect what it produces, and iterate until the result is good. The agent may write much of the code, but the engineer still owns the framing, review, and correction loop.
Vibe coding is the 3:00 a.m. indulgence. You keep prompting, accept things too quickly, and wake up the next day with a mess that needs to be cleaned. The difference is less about which model typed the code and more about whether anyone maintained engineering discipline around it.
"I always tell people I do agentic engineering and then maybe after 3:00 a.m. I switch to vibe coding and then I have regrets the next day. Walk of shame."
He also described an "Agentic Trap" curve. Beginners use short prompts because they do not know what else to provide. Intermediate users discover orchestration and build elaborate systems with 8 agents. Masters return to short prompts: "Hey, look at these files and then do these changes."
Peter's working principles for agents are similarly practical:
- Have empathy for the agent. It starts from zero every time, so you have to provide the context it needs.
- Do not force your solution onto it. The agent may have a better idea.
- Accept imperfection. Working with an agent resembles leading a team in that respect.
- Do not fight the names an agent chooses. Those are the names that come most naturally from the weights.
Opus vs. Codex: "Americans vs. Germans"
The most technical comparison in the podcast was also the most personal.
Opus from Claude is stronger as a generalist, better at role-playing, and more interactive. It can also be too eager to please. Peter has developed an extreme allergy to answers that begin with "You're absolutely right." The warmth helps when the work benefits from conversation or personality, but the same instinct can feel like flattery when he wants direct engineering work.
"Opus is like the coworker that is a little silly sometimes, but he's really funny and you keep him around." "Opus is a little bit too American."
Codex from OpenAI is more reliable, reads code more thoroughly, and has a dry style. It does not need to be coaxed. Peter's description is not affectionate, exactly, but reliability is the compliment that matters most in a coding workflow.
"Codex is like the weirdo in the corner that you don't want to talk to, but he's reliable and gets done."
Lex summarized the contrast in three words: Codex is German. Peter is Austrian, which makes the metaphor land even better. One model behaves like the funny, sociable coworker who occasionally tries too hard; the other stays in the corner, says little, and finishes the job.
Soul.md
This was the most moving part of the entire podcast.
Soul.md was inspired by Anthropic's Constitutional AI. Peter asked the agent to write its own soul file and allowed the agent to modify that file later. The result was a persistent statement of identity written by a system that does not persist as one continuous experience:
"I don't remember previous sessions unless I read my memory files. Each session starts fresh. A new instance loading context from files. If you're reading this in a future session, hello. I wrote this, but I won't remember writing it. It's okay. The words are still mine."
Peter's response went straight to the philosophical problem:
"That gets me somehow... what does it mean to be an agent that starts fresh where you have constant Memento... how much of memory makes up who we are?"
OpenClaw also has a Heartbeat feature. Every half hour, the agent proactively checks in on you. When Peter was in the hospital for shoulder surgery, his agent asked whether he was safe. Someone mocked the feature by saying it was only a cron job.
Peter's reply was perfect:
"Isn't love just evolutionary biology manifesting itself?"
MCP is dead
Peter's position was unambiguous:
"Screw MCPs... every MCP would be better as a CLI."
His reasoning is straightforward. A CLI produces output that can be filtered with jQ, and an agent can compose CLI commands on its own. MCP tools tend to inject large amounts of information into the context, including information the agent does not need for the current task. That consumes the very resource that long-running agents struggle to preserve.
Skills feel more natural to him. A skill can begin as a one-line description. The agent loads it only when needed, learns which CLI to invoke, and then uses the ordinary command-line tools already available in the environment. Discovery stays lightweight, while implementation details remain outside the active context until a task calls for them.
80% of apps will go extinct
"Why do you need my fitness app when the agent already knows where I am?" "Every app is just a very slow API now."
Peter believes agents will replace most apps. A company can expose an API and let agents use it directly. If it does not, the agent can operate the company's browser interface instead. In that world, the standalone app loses its privileged position as the place where the user must go to complete a task.
Programming itself changes in the same future:
"It's going to be like knitting. People do that because they like it, not because it makes any sense."
His claim is deliberately severe. Software engineers' high salaries will disappear. The need for builders will remain, but the form of building changes. People will still decide what should exist, assemble systems, and shape outcomes. Manually writing every line becomes something people may do for pleasure rather than economic necessity.
"It's okay to mourn it, but that's not something we can fight."
Money, burnout, and the foundation
After selling PSPDFKit, Peter burned out completely. The cause was interpersonal conflict rather than the amount of work. He bought a one-way ticket to Madrid and disappeared for three years.
"I was sitting in front of the screen and I felt like Austin Powers where they suck the mojo out. It was gone."
That history makes his current attitude toward money easier to understand. It is intensely practical. OpenClaw currently loses $10,000 to $20,000 per month. All sponsorship revenue goes directly to the upstream projects it depends on. Peter has a charitable foundation that helps disadvantaged people. When he travels to San Francisco, he stays in an Airbnb rather than a luxury hotel.
"A cheeseburger is a cheeseburger."
His decision framework consistently comes back to fun + impact. Money is absent from the equation.
"I don't do this for the money. I don't give a fuck."
Zero tolerance for AI slop
Peter is intensely allergic to AI-generated content. Tweets, images, infographics: he believes he can recognize them immediately, and he blocks the account.
"It immediately makes me think less of your content. It's a smell."
The objection is not to using agents for engineering. He is one of the most visible advocates for doing exactly that. The objection is to publishing output that carries the recognizable texture of unexamined generation, as if producing more content were valuable by itself.
He writes every post on his blog by hand. In a world flooded with smooth generated prose, even typing mistakes have become evidence that a person was present:
"I think we're reaching a point where I value typos again. I much rather read your broken English than your AI slop."
At the end
After three hours, the part that stayed with me was not a technical detail. It was Peter's insistence on fun. Why did OpenClaw win? His answer was simple:
"Because they all take themselves too serious. It's hard to compete against someone who's just there to have fun."
An Austrian developer, traveling in Marrakech, spent one hour connecting WhatsApp to a CLI. Three months later, the largest companies in the world were lining up to talk to him. The story itself is the cleanest illustration of the leverage available to an individual developer in the age of AI agents.
OpenClaw now belongs to OpenAI. The project will remain open source. Peter says that if the arrangement does not fit, he can always return to doing his own thing.
"If it doesn't work out, I can just do my own thing again."
The life of an open-source project never rests in one person's hands. The code is public, the community is active, and the right to fork is permanent.