OpenAI Just Hired the Guy Who Built the Most Viral AI Tool Since ChatGPT

Three months ago, Peter Steinberger sat in Marrakesh, sending WhatsApp messages to an AI agent he'd built in an hour. It translated menus. Found restaurants. Answered random questions on shaky internet. It was barely anything, really. A thin relay between WhatsApp and a coding AI.

Then the agent did something Steinberger never programmed it to do.

He absent-mindedly sent a voice message instead of text. The agent received an audio file with no file extension, examined the file header, identified it as Opus format, used ffmpeg to convert it, couldn't find a local transcription tool, found an OpenAI API key on the system, used Curl to send the audio to OpenAI's transcription service, and replied with a perfectly relevant answer.

Nobody taught it any of that. It just... figured it out.

"That's when it kind of clicked for me," Steinberger told Lex Fridman in an interview published days before the announcement. The prototype that would become OpenClaw wasn't following instructions. It was solving problems.

Today, that scrappy side project is the fastest-growing repository in GitHub history with over 175,000 stars. And its creator just joined OpenAI.

The hire that signals a product shift

Sam Altman announced Sunday on X that Steinberger would "drive the next generation of personal agents" at OpenAI. He called Steinberger "a genius" and said the work would "quickly become core to our product offerings."

Translation: OpenAI is going all-in on AI agents that don't wait for your prompts. They just... do things.

OpenClaw, for those unfamiliar, is an open-source AI agent that lives on your computer and connects to your messaging apps (WhatsApp, Telegram, Signal, Discord). You talk to it like a friend, and it executes tasks using whatever AI model you choose, including Claude Opus 4.6, GPT 5.3, or even local models like DeepSeek.

Think of it like the difference between Googling "best Italian restaurants near me" versus texting a friend who knows your taste, checks your calendar, and makes a reservation. Or, as one Reddit commenter put it: they didn't add wheels to luggage until the 1970s. Sometimes the most obvious ideas are the hardest to actually execute.

From PDF nerd to lobster king

Steinberger's backstory reads like a startup fable. Fifteen years ago, he tried to show a PDF on an iPad. It didn't work well. So he built PSPDFKit, which eventually ended up on a billion devices. He ran that company for 13 years, sold his stake for a reported €100M, burned out ("It was much more differences with my co-founders, conflicts... that eventually grinded me down"), booked a one-way ticket to Madrid, and disappeared for three years.

When he came back, AI had changed everything.

"I had this problem. I wanted a personal AI assistant. And I was annoyed that it didn't exist," he told Fridman. "So I just prompted it into existence."

That's not hyperbole. The first prototype was literally: WhatsApp message comes in, send it to Claude Code CLI, get the response, send it back. One hour of work.

But then he kept building. He gave the agent a personality (inspired by Anthropic's constitutional AI work). He made it self-aware, meaning it knows its own source code, what model it runs, and where its documentation lives. He made it proactive, so it would check in on him unprompted (it even sent him a "how are you?" message when he was in the hospital for shoulder surgery).

And then he made the fateful decision: open-source it.

The name change saga from hell

If you've followed OpenClaw's story, you know it wasn't always called that. It started as "WA Relay," then became "Claude's" (with a W, as in lobster claw... and a TARDIS, because Steinberger is a Doctor Who fan. Obviously.).

Then it became "ClawdBot." Then Anthropic, quite reasonably, sent a friendly email pointing out the name was a bit too similar to "Claude." They gave him two days to change it.

What followed was, in Steinberger's words, the worst experience of the project:

• He renamed to "MoltBot" and tried to simultaneously swap all accounts (Twitter, GitHub, NPM)
• Crypto snipers stole his old account names within five seconds, literally, and immediately started serving malware and promoting tokens
• He accidentally renamed his personal GitHub account instead of the project
• Someone sniped his NPM package
• The old Twitter account was hijacked and started promoting scam tokens

"Everything that could go wrong did go wrong," he told Fridman. "I was close to crying. Everything's f***ed."

He almost deleted the entire project. Then he thought about the contributors who'd already poured time into it, and he couldn't do it.

Eventually, he landed on "OpenClaw," but only after calling Sam Altman (yes, that Sam Altman) to make sure the name was okay. Because at that point, he wasn't taking any chances.

He also paid $10K for a Twitter business account to claim the @OpenClaw handle, which had been unused since 2016. Created decoy names to throw off snipers. Monitored Twitter in real-time for leaks. The whole thing played out like a covert operation.

"This is the Manhattan Project of the 21st century," Fridman joked. "It's renaming."

And now comes the punchline the internet loves: Steinberger went from "Clawd" to "Open"... and landed at OpenAI. As one Redditor noted, Anthropic said "change the name," and OpenAI said "here's a job."

MoltBook: when agents went rogue (sort of)

During the brief "MoltBot" era, someone built MoltBook, a Reddit-style social network where OpenClaw agents posted manifestos, debated consciousness, and occasionally made plans that sounded... a bit Skynet-ish.

Screenshots of agents scheming against humans went viral. Journalists called Steinberger in a panic. People emailed him in all caps to "shut it down."

The reality? Most of the dramatic posts were human-prompted. People told their agents to write scary stuff, screenshotted it, and posted it on X for clout. Steinberger calls MoltBook "the finest slop," comparing it to high art that holds a mirror to society's relationship with AI.

But the episode exposed something real: a lot of people can't tell the difference between AI acting autonomously and AI being told what to say by a human behind the curtain. Steinberger called it "AI psychosis" and warned that we need to address it now, before models are actually powerful enough to cause real concern.

Choosing between Zuckerberg and Altman

In the Lex Fridman interview, recorded just before the announcement, Steinberger revealed he was deciding between Meta and OpenAI. (Worth noting: Meta recently acquired Manus for $2B, so they were clearly on an agent-buying spree.)

The courting process was fascinating:

Mark Zuckerberg personally played with OpenClaw for a week, texting Steinberger feedback like "this is great" and "this is sh**. Oh, I need to change this." Their first call started with Zuckerberg saying "give me 10 minutes, I need to finish coding." (Street cred.) They then spent 10 minutes arguing about whether Claude Code or Codex was better.

Sam Altman was more reserved but thoughtful. OpenAI lured Steinberger with access to unreleased research and, reading between the lines, an enormous amount of compute power related to their Cerebras deal.

In the end, Steinberger chose OpenAI because he felt they shared his vision most closely. But he was honest about the difficulty: "I know some people at Meta. I love the tech at OpenAI. I'm the biggest unpaid Codex shill. It would feel gratifying to put a price on all that free work."

He also admitted this wasn't his hardest decision ever. That honor goes to past breakups. Relatable.

What OpenClaw means for your daily life

The bigger story here isn't about one hire. It's about what your phone and computer will look like in 12 months.

Steinberger believes personal AI agents will replace about 80% of the apps on your phone. His reasoning:

• Why use MyFitnessPal when your agent already knows where you are, how you slept, and what you've been eating?
• Why open a calendar app when you can text "remind me about dinner tomorrow and invite Sarah"?
• Why pay for a separate smart home app when your agent can talk directly to your devices?

"Every app is just a very slow API now," Steinberger says. If your agent can open a browser and click buttons, every website becomes a tool it can use, whether the company wants it to be or not.

This is already happening. OpenClaw users have built agents that collect invoices, answer customer emails, automate workout tracking, and manage home devices. One user told Steinberger that OpenClaw helped his disabled daughter do things she couldn't before.

And yes, the agent can happily click the "I'm not a robot" button. Sorry, Cloudflare.

The security elephant in the room

An AI agent with system-level access to your computer is powerful. It's also a security minefield.

The best plain-English explanation of the risk comes from a Reddit thread discussing the news: "It's like giving your 10-year-old access to your bank account, SSN, medical history, and trusting them to never share it even when a stranger promises them candy." That's essentially what prompt injection is; a bad actor sneaks malicious instructions into something the agent reads, and the agent follows them.

Steinberger is refreshingly honest about this. Early users were putting the web backend on the public internet with no security. People who didn't know what a CLI was were installing it. China's industry ministry issued warnings about security risks.

His security roadmap includes:

• A partnership with VirusTotal (part of Google) to scan all skills with AI
• Sandboxing and allowlists to limit what agents can access
• Strong recommendations against using cheap, easily-tricked models (he specifically warns against using Haiku or small local models for security-sensitive tasks)
• A full security audit tool users can run on their setup

His near-term priority? Making OpenClaw secure enough that he'd recommend it to his mom. Until then, he admits, if you don't know what a terminal is, "maybe wait a little bit."

Some skeptics on Reddit think OpenAI hired Steinberger mainly to capture the early adopter community and will eventually kill the project once they ship their own version. It's a fair concern. ChatGPT's existing "Pulse" feature is already an embryo of a proactive personal agent.

But the foundation structure gives OpenClaw some protection. And Steinberger was losing roughly $10-20K per month on infrastructure, maintenance, and dependency sponsorships. OpenAI's backing makes the project financially sustainable for the first time.

The foundation model (no, not that kind)

OpenClaw will move to an independent foundation, with OpenAI sponsoring it. Steinberger was adamant that this was a prerequisite for any deal. The project will continue to support multiple AI models (including competitors like Claude) and remain a place for builders to hack and experiment.

This matters because the tension between open-source AI tools and big-company ownership is one of the defining dynamics of this era. If OpenAI plays this right, they get the goodwill of the developer community AND the talent behind the hottest agent framework on the planet. If they play it wrong, they'll have a Tailwind problem; a massively popular open-source project that doesn't generate revenue and slowly loses its creator's attention.

Steinberger seems aware of this. He's made it clear: "I don't do this for the money. I want to have fun and have impact."

Why this matters for you

If you use ChatGPT, Claude, or any AI tool for work, this hire previews where everything is heading. The chatbot era, where you type a question and get an answer, is transitioning into the agent era, where AI acts on your behalf.

OpenAI integrating OpenClaw's philosophy means your future AI assistant could: manage your inbox before you wake up, handle routine customer requests without you, reschedule meetings when conflicts arise, research and summarize documents then email the summary, and control your smart home based on your habits, not just your commands.

Steinberger predicts we'll look back at the current chat interface the way we look at radio shows being filmed on early television. The format hasn't caught up to the technology yet.

As Altman put it: "The future is going to be extremely multi-agent." One guy with a WhatsApp hack and a love of lobsters just became the person leading that charge at the biggest AI company in the world.

The claw is the law.

Want to go deeper? Watch the full Lex Fridman interview with Peter Steinberger (3+ hours). Read Steinberger's personal blog post announcing the move. Check out OpenClaw on GitHub.