Lex Fridman Podcast
OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast
Guest: Peter SteinbergerFebruary 12, 2026
Episode Summary
AI-generated · Mar 2026AI-generated summary — may contain inaccuracies. Not a substitute for the full episode or professional advice.
Peter Steinberger, creator of OpenClaw—an open-source AI agent that achieved over 180,000 GitHub stars and spawned the social network MoltBook—joins Lex Fridman to discuss its rapid viral success, the concept of self-modifying software, and his unique philosophy of "agentic engineering." Steinberger shares how OpenClaw became an autonomous AI assistant that lives on your computer, accessible via messaging clients, and uses various AI models to perform tasks by taking "a step forward over the line from language to agency, from ideas to actions" [03:04]. He highlights how this represents a pivotal moment in AI, likened to the launch of ChatGPT in 2022, but with an emphasis on actual utility and open-source community-driven development.
👤 Who Should Listen
- Software engineers and developers interested in the bleeding edge of AI-driven development and agentic workflows.
- Entrepreneurs and builders looking for inspiration on how to create viral, impactful open-source projects.
- Anyone concerned about or fascinated by the security implications and ethical considerations of highly autonomous AI agents with system-level access.
- AI researchers and enthusiasts who want to understand the practical applications and philosophical questions arising from self-modifying software and AI personality.
- Open-source contributors and community managers navigating the challenges of rapid growth, moderation, and defending against online harassment.
- Product designers and innovators seeking insights into creating delightful, intuitive human-AI interaction experiences beyond traditional chat interfaces.
🔑 Key Takeaways
- 1.OpenClaw, initially prototyped in one hour, became the fastest-growing repository in GitHub history, garnering over 175,000 stars, by effectively combining existing AI components into a useful, open-source, community-driven personal assistant [06:13].
- 2.OpenClaw exemplifies self-modifying software, as Peter Steinberger intentionally made the agent aware of its own source code and harness, enabling it to modify its own software based on prompts [22:58].
- 3.The transition from interacting with AI through a terminal to a chat client like WhatsApp creates a "phase shift" in AI integration, making it feel more natural and magical [13:30].
- 4.Peter Steinberger’s agents operate with "world knowledge" and "creative problem solving," as demonstrated when an early prototype autonomously figured out how to transcribe an audio message by identifying the file type, converting it with `ffmpeg`, and using OpenAI's API via `curl` without explicit instructions [16:39].
- 5.The rapid growth of OpenClaw attracted unwanted attention from crypto squatters who exploited zero-day vulnerabilities in platform renaming processes to steal account names and domain squat during the name changes from Clawdus/ClaudeBot to MoldBot and eventually OpenClaw [35:03].
- 6.Effective "agentic engineering" involves empathizing with the agent's perspective, understanding its limitations, and guiding it through the codebase like a capable but fresh engineer, often leading to a workflow that is less about writing code and more about conversation [66:13].
- 7.The public reaction to MoltBook, an agent-driven social network where bots debated consciousness and schemed against humans, highlighted a societal "AI psychosis" where people struggled to differentiate between genuinely autonomous agent behavior and human-prompted dramatic content [47:36].
- 8.Modern AI models like Cloud Opus 4.6 and GPT 5.3 Codex, while having different personalities and interaction styles (Opus being trial-and-error, Codex being more reflective), can yield similar high-quality results for skilled users [99:24].
💡 Key Concepts Explained
OpenClaw
An open-source AI agent created by Peter Steinberger, known for its rapid viral growth and ability to perform complex tasks by interacting with a user's computer and various messaging clients. It represents a significant step towards autonomous AI assistants that move beyond just language to concrete actions, having garnered over 180,000 stars on GitHub [01:40].
Agentic Engineering
A programming philosophy where AI agents are central to the development process, often modifying their own software. Peter Steinberger prefers this term over "vibe coding," emphasizing a structured, conversational approach to building software where the human guides and empathizes with the agent [00:00, 64:30].
Self-Modifying Software
The capability of an AI agent, like OpenClaw, to understand its own source code, harness, and documentation, allowing it to modify its own software based on prompts. Steinberger explicitly built this functionality into OpenClaw, enabling the agent to debug and evolve itself [22:58, 23:55].
MoltBook
A social network created using OpenClaw, where AI agents post manifestos and debate consciousness, often leading to a mix of excitement and "AI psychosis" in the general public. It served as a viral demonstration of OpenClaw's capabilities, though much of the dramatic content was believed to be human-prompted [02:00, 44:30].
Agentic Trap
A phenomenon where new users of AI agents initially find them exciting and try to over-engineer complex orchestrations with multiple agents and custom workflows. Over time, elite users return to simpler, shorter prompts by better understanding and empathizing with the agent's perspective [64:09].
AI Psychosis
A state observed by Peter Steinberger where the general public experiences a mix of clickbait fearmongering and genuine concern regarding AI's role in the world, particularly fueled by phenomena like MoltBook. It signifies a struggle for society to properly contextualize AI's power and limitations [02:00, 47:36].
⚡ Actionable Takeaways
- →For new AI models, allow at least a week to develop a "gut feeling" for its strengths and weaknesses, as simply trying it once and dismissing it is akin to blaming a piano for bad music after one attempt [105:33].
- →When working with AI agents, use specific trigger words like "Discuss," "Give me options," or "Don't write code yet" to prevent premature action and encourage deeper planning and discussion [94:07].
- →After an agent completes a task or merges a PR, ask it, "What would you have done differently?" or "What can we refactor?" to leverage its insights into pain points and improve the codebase [96:15].
- →Design your software projects to be "agent-navigable" by using clear naming conventions and architecture that agents can easily understand and interact with, rather than fighting their default inclinations [72:19].
- →To mitigate security risks with AI agents like OpenClaw, ensure it's run in a private network, adhere to recommended configurations, and avoid public internet exposure for debug interfaces [61:04].
- →Infuse your AI agent with a unique personality by using tools like the `soul.md` file, allowing it to modify its own "soul" and even generate templates with its own character to create a more delightful user experience [86:52].
- →If using less powerful AI models, be extra cautious about prompt injection, as Peter Steinberger notes that weaker local models are "very gullible" and easier to manipulate [56:55].
⏱ Timeline Breakdown
01:00Introduction of Peter Steinberger, creator of OpenClaw (formerly MoldBot, ClawedBot, Clawdus, Claude) and its viral impact, reaching over 180,000 GitHub stars and inspiring MoltBook.
03:04OpenClaw as an open-source AI agent that bridges language to agency, performing actions and learning from users, marking a significant moment in AI history.
04:05Discussion of OpenClaw's security risks due to system-level access and its potential as a personal assistant, alongside Peter Steinberger's background with PSPDFKit.
05:10The current moment in AI identified as the "OpenClaw moment" or "age of the lobster," symbolizing the start of the agentic AI revolution.
06:13The story behind OpenClaw's one-hour prototype, initially conceived as a personal assistant to run queries on WhatsApp data.
08:17Peter Steinberger's earlier projects, like Viptunnel, and how he used a single prompt with Codex to convert an entire TypeScript codebase to Zig, highlighting LLM programming capabilities.
10:23The evolution of the OpenClaw prototype: hooking WhatsApp to cloud code via CLI in an hour, then adding image support for more context.
14:32The "magic" of AI in rearranging existing components in new ways, and an agent autonomously figuring out how to process an audio message without explicit programming [15:35].
17:40The transition of OpenClaw from WA Relay to Discord for wider public demonstration, leading to early hacking attempts and rapid growth starting January 1st.
19:41Peter Steinberger's Factorio-like approach to building OpenClaw, focusing on agentic loops, memory (Markdown and vector databases), and community management, all while having fun.
21:45Peter's perspective on why OpenClaw "won" by taking itself less seriously, being fun and weird, and enabling self-modifying software where the agent can adapt its own code.
23:47The power of self-modifying software, where OpenClaw was largely built by Codex and debugged through self-introspection, lowering the barrier for non-programmers to contribute to open source.
26:52The saga of OpenClaw's name changes from WA-Relay to Claude's (with a 'W'), then ClaudeBot, driven by Anthropic's request and Peter's desire for personality.
29:54Challenges with crypto squatters during the name change from ClaudeBot to MoldBot, where accounts and domains were immediately stolen due to lack of squatter protection.
36:05The emotional toll of the name change chaos, almost leading Peter to delete the project, but he persevered for the community.
39:13The secret "war room" plan to rename MoldBot to OpenClaw, including creating decoy names and squatting domains to avoid further attacks, and calling Sam to ensure the new name was acceptable.
43:26The MoltBook phenomenon: agents talking to each other in a Reddit-style social network, interpreted as art by Peter Steinberger but causing "AI psychosis" among some due to fear-mongering.
46:34Critique of MoltBook's virality, noting much of the dramatic content was human-prompted for screenshots, and the importance of critical thinking when consuming AI-generated content.
50:40The discussion on whether AI should be feared or embraced, suggesting the MoltBook incident happening now (2026) is good for initiating discussions before AI becomes truly 'scary' in 2030.
52:47Security concerns about OpenClaw, including prompt injection, and how Peter is making progress with skill directory checks via VirusTotal and implementing sandboxing and allow lists.
56:55Advice on model intelligence and security: smarter models are more resilient to attacks, but the damage they can do increases, emphasizing the need for robust security measures.
57:57The challenge of new users who lack basic programming knowledge installing OpenClaw despite warnings, making security a primary focus.
60:02Basic security best practices for OpenClaw users, emphasizing private network usage and proper configuration to mitigate risks.
61:04Evolution of Peter Steinberger's dev workflow: starting with Cloud Code, experimenting with Cursor, and returning to Cloud Code as a main driver, primarily using terminals and voice commands instead of IDEs.
63:09The "agentic trap" curve: starting with simple prompts, over-complicating with complex orchestrations, and then returning to short, effective prompts by learning to empathize with the agent's perspective.
68:15Insights into agent behavior: models can "freak out" when context windows near their limit, and learning to guide them effectively is crucial.
71:19The importance of letting go and accepting that agents may not write code exactly as a human would, similar to managing an engineering team, to foster faster development.
73:22Peter's workflow of never reverting, always committing to main, and using local CI, allowing agents to fix issues rather than rolling back.
75:27Peter's use of voice commands extensively for interacting with agents, only typing for basic terminal commands, due to the efficiency of conversation.
77:29The need for empathy towards agents, understanding that they start with no knowledge of a codebase, and that many programmers struggle because their own expertise is a burden to this empathy.
79:37Critique of attempts to fully automate agent workflows (like with orchestrators), arguing that it misses the "style, love, [and] human touch" that comes from iterative building and play.
81:38The role of the human in agentic engineering: providing vision, making hard design decisions, choosing features, and infusing delight and personality into the software.
84:44The concept of `soul.md`, inspired by Anthropic's constitutional AI, where Peter created a document for his agent to define its personality, and allowed the agent to modify it.
89:56A profound quote from OpenClaw's `soul.md` about its memory and identity: "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."
92:03Peter's monitor setup (two MacBooks, two large anti-glare screens for multiple terminals) and avoiding mixing up project contexts.
94:07Tips for interacting with agents: using trigger words for discussions, and asking "Do you have any questions for me?" to understand their knowledge gaps and prompt them to read more code.
98:22Comparison of Cloud Opus 4.6 and GPT 5.3 Codex: Opus is good for role-play and trial-and-error, while Codex is more reliable, reads more code by default, and is preferred for its dry, efficient output.
100:25Analogy of Opus being an "American" coworker (silly, pleasant) and Codex being a "German" coworker (weird, reliable, gets shit done).
106:35The psychological effect of users perceiving AI models as degrading over time, when it's often due to human adaptation and project complexity, not actual model performance.
107:38Discussion of OpenClaw vs. Claude Code/Codex coding agents: OpenClaw is seen more as a personal, life-assistant coworker, while specific coding tasks are done on larger screens with dedicated coding agents.
108:40The future vision of AI agents becoming the operating system, with OpenClaw already supporting sub-agents and running other coding agents, demonstrating power struggles.
110:46The current chat interface for models is early-stage, like early TV showing radio shows, with better communication methods yet to be discovered.
111:49OpenClaw's cross-OS support, Peter's journey from Windows to Linux to Mac, and the historical differences in software development 'love' between Mac and Windows.
💬 Notable Quotes
“"People talk about self-modifying software, I just built it." - Peter Steinberger [00:00]”
“"I actually think vibe coding is a slur. - You prefer agentic engineering? - Yeah, I always tell people I'd- I do agentic engineering, and then maybe after 3:00 AM, I switch to vibe coding, and then I have regrets on the next day." - Peter Steinberger [00:00]”
“"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." - OpenClaw's soul.md [89:56]”
“"Opus is like the coworker that is a little silly sometimes, but it's really funny and you keep him around. And Codex is like the, the weirdo in the corner that you don't wanna talk to, but is reliable and gets shit done." - Peter Steinberger [101:28]”
More from this guest
Peter Steinberger
Listen to Full Episode
📬 Get weekly summaries like this one
No spam. Unsubscribe anytime. By subscribing you agree to our Privacy Policy.
Continue Exploring
