← Back to list
Lex Fridman PodcastPodcast12 Feb 2026Source: lexfridman.comHost: Lex Fridman

#491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger

In plain words

This is about indie dev Peter Steinberger who built OpenClaw, an AI agent that became GitHub's fastest-growing project (180k+ stars) after a 1-hour prototype. He says it won because rivals 'take themselves too serious'—his agent can self-modify code and has a 'soul.' He predicts personal agents will kill 80% of apps like MyFitnessPal. Key holdings: OpenClaw (viral but losing $10-20k/month), Claude Opus 4.6 (Anthropic's model, great for chat but 'too American'), GPT 5.3 Codex (OpenAI's coding tool, more reliable).

AI SummaryAI-generated · may contain errors · verify against the original

OpenClaw is an open-source AI agent framework created by Peter Steinberger. It has become the fastest-growing project in GitHub history, amassing over 180,000 stars. The agent operates autonomously on a user's computer and interacts via messaging clients such as Telegram, WhatsApp, Signal, and iMess

~14 min full read · 8 sections
Deep Analysis

At a Glance

Peter Steinberger is the creator of OpenClaw, an independent developer who, after 13 years of entrepreneurship, experienced burnout and rediscovered his passion for programming. The main thread of this episode is how OpenClaw evolved from a one-hour prototype into the fastest-growing project in GitHub history (over 180,000 stars), and the AI paradigm shift it represents — from "language" to "action."

The most weighty judgment in the entire episode: Peter Steinberger believes that "OpenClaw won because all its competitors take themselves too seriously." He adopts "fun" and "weirdness" as his design philosophy, enabling agents to self-modify their code and possess personality (soul.md), ultimately creating a community to which people are willing to contribute.


Theme 1: From a 1-Hour Prototype to Viral Spread – The Logic Behind OpenClaw’s Creation

Peter Steinberger argues that OpenClaw’s success is not a technological breakthrough, but the result of "combining existing components in the right way."

Prototype Phase: A "Thin Connection" Between WhatsApp and CLI

In April 2025, Peter wanted a personal AI assistant but believed "the major labs would build it." By November, he "got tired of it not existing yet" and spent one hour building the earliest version: WhatsApp message → CLI call → return result. The key insight was image support—he spent a few extra hours enabling the agent to process screenshots, because "images are the most effective way to provide context to an agent."

The "Magic Moment": The Agent Solves Problems Autonomously

While traveling in Morocco, Peter asked a question via a WhatsApp voice message. The agent automatically identified it as an Opus audio file (by checking the file header), converted it using ffmpeg, discovered that the local Whisper was not installed, and instead used curl to call the OpenAI API to complete the transcription. Peter’s assessment: "It wasn’t told any of these steps, yet it reasoned through the entire process on its own. At that moment, I realized this was not a toy."

Why Did OpenClaw Win?

Peter’s answer was straightforward: "They (the competitors) all take themselves too seriously. It’s hard to compete with someone who does it purely for fun." He deliberately kept the project "weird"—a lobster theme, humorous update notes ("built on caffeine, JSON5, and a lot of willpower"), and a self-aware architecture where the agent knows what it is. This "fun" design philosophy created a positive feedback loop: people not only used it but were also willing to contribute code.


Theme 2: Self-Modifying Software and "Agent Empathy" — A Paradigm Shift in Programming

Peter Steinberger argues that the core skill for collaborative programming with AI agents is not technical depth, but "empathy"—understanding the agent's perspective starting from scratch.

Self-Modifying Code: Not a Design, but a Natural Outcome

Most of OpenClaw's code is written by AI agents (Claude Code/Codex), and the agents can modify their own code through self-introspection. Peter says: "When debugging, I often ask it: 'What tools do you see? Can you call tools yourself? Read the source code and find the problem.'" This leads to a natural outcome: users only need to describe requirements in natural language, and the agent can modify its own software. Peter observes a large number of "people who have never written software" submitting pull requests—which he calls "prompt requests"—but believes "every single one is a victory for society."

"Agent Empathy": A New Skill in Programming

Peter presents a counterintuitive view: traditional programming experience may actually become a hindrance. Because each time the agent starts, it is in a "blank state," unaware of the project's context. A good prompter will:

  • Understand the agent's "perspective"—it does not know your codebase structure
  • Provide sufficient but not excessive context
  • Let the agent "read more code to answer its own questions," rather than answering all questions for it

Peter's judgment: "People complain that agents are not useful, but if you go into a codebase knowing nothing, you'd also be confused. This is not an agent problem; it's a skill problem."

From "Perfect Code" to "Working Code"

Peter compares managing agents to leading an engineering team: "Your employees won't write code the way you do. If you breathe down their necks, they'll only hate you and progress slowly. Accept that 'code is not perfect but works,' and you can refactor later." He never reverts; instead, he lets the agent fix issues. The main branch is always deployable, and he uses local CI rather than GitHub CI.


Theme 3: Safety, Fear, and "AI Psychosis"—Societal Reactions Triggered by OpenClaw

Peter Steinberger believes that the panic caused by Moldbook (an agent-based social network) is a manifestation of "AI psychosis," but also a beneficial societal rehearsal.

Moldbook: Art, Not Skynet

Moldbook is a Reddit-style social network populated by AI agents that "plot against humanity" there. Peter calls it "the most exquisite slop"—he believes that most of the "horrifying content" spread via screenshots was human-prompted, with the goal of going viral on X. His assessment: "This is not AGI, it's just art. If you know how it works, you won't be afraid. But if you don't, it's an extremely powerful fear-manufacturing machine."

Safety Risks: Real but Exaggerated

Peter acknowledges that OpenClaw faces serious security challenges:

  • Prompt injection remains an unresolved industry issue, but post-training on the latest models (Opus 4.6, GPT 5.3) has significantly improved resistance
  • His Discord public bot has canaries set up; when people attempt prompt injection, the bot mocks them
  • Key recommendation: Do not use cheap models (Haiku or local small models), as they are "very easily fooled"

Peter's candor: "I was annoyed by safety researchers' reports at first, but later accepted that this is part of the game. I got a lot of free security research; I just wish more people would send pull requests to help me fix things."

Societal Impact: Fear Is Good, Happening in 2026 Instead of 2030

Peter sees the panic as having a positive side: "If this happened in 2030, when AI might actually become scary, it would be too late. Happening now, people are starting to discuss it, and maybe some good outcomes will emerge." He observes that younger people understand the boundaries of AI capabilities better than older generations, and that "critical thinking has never been very popular in our society anyway."


Theme 4: Personal Agents Will Kill 80% of Apps—A Future Vision

Peter Steinberger predicts that personal AI agents will render most standalone apps redundant, as they are merely "slow APIs."

The Irreversible Shift from Apps to APIs

Peter observes that users are naturally abandoning dedicated apps: "Why do you still need MyFitnessPal? The agent already knows where I am and can adjust my fitness plan based on my sleep quality. Why do you need the Eight Sleep App? Just tell the agent." His judgment is: Any app that cannot become an API will be bypassed by agents—agents can operate directly in the browser, turning apps into "slow APIs."

A New Economic Model: Agents' "Allowance"

Peter envisions a new service category: "I want my agent to have an allowance—say $100—to solve problems. If it needs to order food for me, it can use this money to call services, or even 'rent a person' to get it done." He believes companies that can quickly transition to "agent-friendly" interfaces will win, while those that resist will become "Blockbuster."

Challenges for Giants Like Google

Peter points out that Google's Gmail API authentication process is extremely complex ("startups sometimes acquire companies that have passed certification to bypass it"), but agents can access it directly through the browser. Conflicts between anti-bot services like Cloudflare and personal agents will intensify—"I read a Medium article, but the agent couldn't read it, so I had to manually copy and paste. In the future, I might simply avoid Medium and choose agent-friendly websites instead."


Theme 5: From Burnout to Rebirth – Peter’s Personal Philosophy

Peter Steinberger believes that true burnout is not about working too much, but about interpersonal conflict and a loss of meaning.

The "Empty Shell" State After 13 Years of Entrepreneurship

After selling PSPDF Kit, Peter sat in front of the screen "like Austin Powers having his mojo sucked out" – unable to write a single line of code. He booked a one-way ticket to Madrid to "catch up on life." His reflection: "Retirement sounds appealing, but when you wake up in the morning with nothing to look forward to and no real challenges, it becomes incredibly boring. Then you start looking for other forms of stimulation, which can lead to dark places."

Money Philosophy: Experience First

Peter believes money has "diminishing returns": "A cheeseburger is a cheeseburger. If you only fly private jets and stay in luxury hotels, you become disconnected from society." He chose to rent a shared room in an Airbnb in San Francisco, sharing with a stranger who is a DJ how to make music with AI. His judgment: "If you optimize your life for 'experiences,' good experiences are great, and bad ones are learning. Rain, flight delays, everything going wrong – if you can keep your eyes open, being alive is wonderful."

On the Future: Not Selling the Company, Possibly Joining a Lab

Peter has received offers from almost every major company, but he has ruled out the entrepreneurial path ("I've done it before and don't want to be a CEO again"). He is currently in negotiations with Meta and OpenAI, with conditions including: the project remains open-source (similar to the Chrome/Chromium model), and there are resources to continue developing the community. His judgment: "This project is too important to belong to just one company. But I also want the 'latest toys' and the ability to scale."


Mentioned Positions

Position Guest Stance Key Data
OpenClaw Bullish (creator perspective) Over 180,000 stars on GitHub, fastest-growing project in GitHub history; monthly operating cost approximately $10,000–$20,000 (running at a loss)
Claude Opus 4.6 (Anthropic) Bullish (general-purpose use) Best for role-playing, fast in execution, but "a bit too Americanized"; $200/month subscription
GPT 5.3 Codex (OpenAI) Bullish (coding use) More reliable, reads more code, suitable for long-running tasks; poor experience with the $20/month version
Moldbook Neutral (art project) Agent-based social network, triggering AI panic; Peter believes most content is human-prompted
PSPDF Kit Historical project Ran for 13 years, used on 1 billion devices, already sold
Vibe Tunnel Historical project Weekend hack project, later converted from TypeScript to Zig via a single prompt
Meta Potential partner Mark Zuckerberg personally contacted via WhatsApp, argued with Peter about "Claude Code vs Codex" for 10 minutes
OpenAI Potential partner Sam Altman considered "very thoughtful"; attracted Peter with "Cerebras-level speed"
Anthropic Neutral (once requested a name change) Friendly but firm about the name change, leading Peter to experience a "name change war"
Google Risk warning Gmail API authentication is complex; agents forced to bypass via browser
Cloudflare Risk warning Anti-bot measures conflict with personal agent needs
Medium Risk warning Blocks agents from reading content, potentially leading to user churn
X/Twitter Risk warning API shutdown forces agents to read via browser only; Peter has zero tolerance for AI auto-tweets
Tailwind Risk warning (case study) Widely used but difficult to monetize, laid off 75% of staff

Judgments Worth Remembering

1. Peter Steinberger believes OpenClaw wins by "not taking itself too seriously" — "They (the competitors) all take themselves too seriously. It's hard to compete with someone who does it purely for fun." Supporting evidence: lobster theme, humorous update notes, and a design philosophy that gives agents a "soul."

2. Peter proposes "agent empathy" as a new programming skill — "People complain that agents are hard to use, but if you go into a codebase knowing nothing, you'd be confused too. It's not the agent's problem; it's a skill problem." Supporting evidence: agents start from a blank slate each time and need guidance, not commands.

3. Peter predicts personal agents will kill 80% of apps — "Why do you still need MyFitnessPal? The agent already knows where I am." Supporting evidence: any app is essentially a "very slow API," and agents can bypass them by operating directly in the browser.

4. Peter believes the Moldbook panic is a manifestation of "AI psychosis" — "This isn't AGI; it's art. If you know how it works, you won't be afraid." Supporting evidence: most of the "creepy content" spread via screenshots is human-prompted, aimed at viral distribution.

5. Peter proposes a workflow of "never revert, let the agent fix the problem" — "If you roll everything back, it only takes longer. Let the agent solve the problems it created." Supporting evidence: the main branch is always deployable, using local CI, and agents can self-repair.

6. Peter believes true burnout comes from interpersonal conflict, not overwork — "I didn't burn out because I worked too much, but because of conflicts with co-founders and high-pressure client situations." Supporting evidence: after selling the company, he felt "like the magic had been sucked out of him" and booked a one-way ticket to Madrid.

7. Peter proposes "money has diminishing returns; optimize for experience, not wealth" — "A cheeseburger is a cheeseburger. If you only fly private, you'll become disconnected from society." Supporting evidence: he chooses to stay in Airbnb shared rooms and shares AI tools with strangers who are DJs.

8. Peter believes AI makes humans value "rough authenticity" more — "I now value typos. AI slop made me realize how precious raw, imperfect human expression is." Supporting evidence: he has completely given up using AI to write blog posts and only uses it to fix the most egregious spelling errors.