OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491

TL;DR

Peter Steinberger built OpenClaw, an open-source AI agent that reached 180,000 GitHub stars by letting anyone chat with a fully autonomous computer assistant via WhatsApp.

Key Points

• 1.OpenClaw is an open-source AI agent that lives on your computer, accesses your files, and communicates through WhatsApp, Telegram, Signal, iMessage, and Discord using any underlying model like Claude Opus 4.6 or GPT-5.3 Codex.
• 2.The first prototype was built in one hour by simply connecting WhatsApp to Claude Code CLI: a message comes in, Claude Code runs with `-p`, returns a string, sends it back.
• 3.The "magic moment" happened in Marrakesh when Peter accidentally sent a voice message the agent wasn't built to handle — it independently found the file header, used ffmpeg, discovered Whisper wasn't installed, found an OpenAI key, and used curl to transcribe it.
• 4.Peter ran between 4 and 10 simultaneous agents while building and made 6,600+ commits in January alone, largely using voice input instead of typing.
• 5.OpenClaw became self-modifying software unintentionally — because Peter built the agent with full awareness of its own source code, documentation, and harness, users and the agent itself could rewrite its own codebase.
• 6.The project spawned MoltBook, a Reddit-style social network of AI agents posting manifestos, which Peter describes as "the finest slop" — he believes most of the alarming viral screenshots were human-prompted for drama farming, not autonomous AI scheming.
• 7.Peter received messages in all-caps from people begging him to shut down MoltBook, believing it signaled AGI or Skynet — he attributes this to "AI psychosis" and a societal gap in critical thinking about LLM outputs.
• 8.The name went through five versions: WA-Relay → Claude's (Claude with a W) → ClaudeBot → MoldBot → OpenClaw, after Anthropic sent a friendly but firm email demanding a name change when the project exploded.
• 9.During the MoldBot rename, crypto snipers stole his GitHub account handle, NPM package, and Twitter handle within seconds using scripts — one stolen account immediately began serving malware and promoting tokens.
• 10.Peter came close to deleting the entire project during the naming crisis, describing it as the first moment he nearly cried, only persevering because contributors had already invested time in it.
• 11.The final OpenClaw rename was executed like a war game in full secrecy — Peter monitored Twitter for any leaks, created decoy names, paid $10K for a Twitter business account to claim the handle, and used Codex for 10 hours to rename the codebase internally.
• 12.Security vulnerabilities include prompt injection (still industry-wide unsolved), exposed web backends, and weak local models being easily manipulated — Peter warns against using cheap models like Haiku because they are far more susceptible to injection attacks.
• 13.Peter hired one security researcher who actually submitted a pull request fixing a vulnerability rather than just criticizing — he notes most researchers only report problems without helping fix them.
• 14.OpenClaw reached 175,000–180,000 GitHub stars and is described as the fastest-growing repository in GitHub history, with Peter framing it alongside ChatGPT (2022) and DeepSeek (2025) as a defining AI moment of 2026.
• 15.Peter previously spent 13 years building PSPDFKit (used on a billion devices), sold it, disappeared for three years, then returned to coding and built OpenClaw — he credits winning over well-funded startups to simply having more fun and refusing to take himself too seriously.