Recording intelligenceAI-generated brief · check the source for context

OpenClaw AI Agent

3:25:50 recording · EN · 2 speakers

Listen to the original

Listen to the episode
Executive Summary AI
  • Lex Friedman introduces Peter Steinberger, creator of OpenClaw, an open-source AI agent that reached over 180,000 GitHub stars in days and spawned Moldbook, a social network where agents post manifestos.0:00
  • Steinberger says he built the first version in one hour by hooking WhatsApp up to Cloud Code through the CLI with minus p, sending the string back to WhatsApp, and used it constantly on a trip to Marrakesh.20:57
  • The moment it clicked was an audio message he never built support for: the agent read the file header as Opus, converted it with FFmpeg, found Whisper was not installed, and used his OpenAI key with curl to transcribe it.26:11
  • An Anthropic employee sent a friendly email demanding a fast name change, and the rushed move to Moldbot went wrong everywhere, with squatters stealing the account name in the five seconds between two rename clicks, then his renamed GitHub account and the npm root package, serving malware from them.39:54
  • On the engineering side he describes the agentic trap, where people build elaborate orchestration and eventually return to short prompts, talking to four to ten agents by voice, committing to main, and not reading the boring parts of the code.1:14:34

Brief overview

Peter Steinberger built OpenClaw in an hour for fun, then survived a brutal rename and security reckoning.

  1. Fun was the competitive advantageAsked why OpenClaw won, Steinberger answers that everyone else takes themselves too serious, and that it is hard to compete against someone who is just there to have fun.
  2. A rename is a security eventNone of the platforms had squatter protection, so five seconds of mouse movement was enough to lose the handle, and the second rename had to be planned in full secrecy with decoy names.
  3. Weak models are easy to prompt injectHis security documentation warns against cheap or local models because they are very gullible, while sandboxes, allow lists and a private network shrink the risk profile.

Questions this recording answers

10 questions, each answered where it is said
What is OpenClaw and what can it actually do?

OpenClaw is an autonomous AI assistant that lives on your computer and, if you let it, has access to all of your stuff. You talk to it through Telegram, WhatsApp, Signal or iMessage, and it uses whatever AI model you like, including Cloud Opus 4.6 and GPT 5.3 Codex.

Answered around 1:06
Why did the project have to be renamed?

Anthropic asked for it. The agent was called Clawd, spelled with a W as in lobster claw, and it was being confused with Claude, spelled with a U. A friendly employee email came first, then a second saying the lawyers were getting uneasy, so it became Moldbot and finally OpenClaw.

Answered around 0:18
What was the prototype he built in one hour?

He hooked WhatsApp up to Cloud Code: a message comes in, he calls the CLI with minus p, it does its magic, he gets the string back and sends it to WhatsApp. That took one hour. Image support, which he wanted because screenshots give the agent context, took a few more hours.

Answered around 20:57
How did the agent handle a voice message it was never taught?

He sent an audio message in a hurry. The agent got a file with no file ending, checked the header and found it was Opus, converted it with FFmpeg, wanted Whisper but it was not installed, so it found his OpenAI key and used curl to send the file to OpenAI.

Answered around 26:11
What went wrong when he renamed the accounts?

None of those services have squatter protection. With two browser windows open, he pressed rename in one and in the five seconds of dragging the mouse the name was stolen. On GitHub he renamed his personal account by mistake and within about thirty seconds it was sniped and serving malware.

Answered around 44:56
How do you run OpenClaw with a smaller risk profile?

He says it is not much different from running Cloud Code with dangerously skipped permissions or Codex in YOLO mode. Make sure you are the only person who talks to it, and instead of putting everything on the open internet keep it in a private network; then that risk profile falls away.

Answered around 1:10:21
Which models does he warn against using with the agent?

His security documentation says not to use cheap models: no Haiku, no local model. Even though he very much loves the idea that the thing could run completely local, a very weak local model is very gullible and it is very easy to prompt inject it.

Answered around 1:06:37
How are the skills in the directory checked for malware?

For the skill directory, Clawhub, he set up a cooperation with VirusTotal, which is part of Google, so every skill is now checked by AI. It will not be perfect, but it captured a lot. He also hired the one security researcher who sent pull requests instead of only complaints.

Answered around 1:04:02
How many commits and how many agents at once?

In January he did 6,600 commits, probably more; he jokes he is limited by the technology of his time and could do more if agents were faster. He runs between four and ten agents in parallel, depending on how much he slept and how difficult the tasks are.

Answered around 31:16
Why is OpenClaw written in TypeScript?

He picked TypeScript because he wanted the project to be very easy, hackable and approachable. It is the number one language being used right now, it fits all those boxes, and agents are good at it. The language matters less to him than the ecosystem around it.

Answered around 1:32:40
Key Quote

“And in those five seconds, they stole the account name.”

— Peter Steinberger, OpenClaw AI Agent, at 45:15▶ listen at 45:15
Key Quote

“I think it's art.”

— Peter Steinberger, OpenClaw AI Agent, at 54:58▶ listen at 54:58
Key Quote

“Like, don't trust screenshots.”

— Lex Friedman, OpenClaw AI Agent, at 59:55▶ listen at 59:55
Key Quote

“Don't use Haiku or a local model.”

— Peter Steinberger, OpenClaw AI Agent, at 1:06:44▶ listen at 1:06:44

Cite this

“OpenClaw AI Agent.” https://media.blubrry.com/takeituneasy/ins.blubrry.com/takeituneasy/lex_ai_peter_steinberger.mp3. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/openclaw-ai-agent.html.

Report this page
What is wrong with this page?