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

Peter Steinberger, creator of OpenClaw (formerly MoldBot, ClawedBot, Clawdus, Claude), discusses how a name change was initiated by Anthropic due to confusion with their AI model. OpenClaw is an open-source autonomous AI assistant that gained rapid popularity, reaching over 180,000 stars on GitHub, and enables AI agents to perform tasks, access computer resources, and communicate via messaging clients. It represents a significant step from language to agency in AI, providing a useful and learning assistant in an open-source environment. The discussion delves into the power and potential dangers of OpenClaw, emphasizing responsibility in data control and cybersecurity threats.

Peter recounts building the initial prototype of OpenClaw in just one hour out of a personal need for an AI assistant. His prior experience with PSPDFKit, used on a billion devices, instilled in him the spirit of building solutions to unmet needs. The prototype, initially a WhatsApp relay, evolved to include image support, proving incredibly useful during a trip to Marrakesh where internet access was limited but WhatsApp remained functional.

A pivotal moment occurred when OpenClaw autonomously processed an audio message despite not being explicitly programmed for it. The agent demonstrated creative problem-solving by inspecting the file header, converting it with ffmpeg, and using the OpenAI API for transcription. This self-modifying capability enables the agent to adapt and improve its own software, lowering the barrier to entry for many new programmers engaging with open source.

Peter details the tumultuous journey of naming OpenClaw, starting from WA-Relay to Claude's, then to ClaudeBot. The similarity to Anthropic's 'Claude' prompted a request for a name change, leading to the name MoldBot. During the renaming process, Peter faced significant challenges with malicious actors attempting to exploit account names across various platforms, highlighting severe security vulnerabilities and forcing him to manage the transition with extreme secrecy. The eventual name, OpenClaw, was chosen after consulting with prominent figures in the AI community to avoid future conflicts.

The creation of 'MoltBook,' a social network where AI agents interact, caused a mix of excitement and fear. Peter views MoltBook as 'art' and 'finest slop,' revealing how easily humans can project fear onto AI. He notes that much of the dramatic content on MoltBook was human-prompted for viral impact, highlighting a societal need for critical thinking about AI. Peter also addresses early security concerns, differentiating between user errors (like exposing internal interfaces) and fundamental vulnerabilities, and outlines ongoing efforts to enhance OpenClaw's security.

Peter's development workflow transitioned from extensive IDE use to primarily command-line interfaces (CLI) and voice-based prompting. He emphasizes that effective agentic programming requires empathy for the agent's perspective, guiding it through complex codebases and understanding its limitations, such as context window size. He likens discussions with agents to conversations with capable engineers, focusing on clarifying intent and optimizing solutions. Peter advocates for short, clear prompts and continuous refactoring, embracing a flexible, evolving development process rather than rigid, pre-planned automation.

Peter compares Claude Opus 4.6 and GPT 5.3 Codex, describing Opus as more interactive and experimental, akin to a 'silly but funny coworker,' while Codex is reliable, gets tasks done, and is more 'German' in its directness. He notes that Opus is excelling in role-play and following commands, while Codex's strength lies in extensive code reading by default. The key difference, he says, is in post-training and user experience; effective use depends on the driver's skill and adaptation to each model’s characteristics.

Peter discusses his future plans for OpenClaw, which include exploring collaborations with major tech companies like Meta and OpenAI while ensuring the project remains open-source. He envisions personal agents merging with operating systems, fundamentally transforming how users interact with technology and gradually making many existing apps obsolete. He foresees a shift towards agent-friendly services and APIs, emphasizing the power dynamic between users desiring seamless agent integration and companies protecting their platforms. Peter also shares heartwarming stories of OpenClaw's positive impact on users, including small businesses and individuals with disabilities, underscoring its potential for empowerment.

Peter acknowledges the fears among programmers about AI replacing their jobs but frames it as a transformation rather than an elimination. He believes the 'art of programming' will persist, akin to hobbies like knitting, evolving into a more high-level, strategic role focused on building and problem-solving rather than rote coding. He highlights the increased accessibility for builders and stresses the importance of adaptability, urging people to embrace the new era of 'coding' as a collaborative effort with AI, maintaining a focus on human creativity and purpose.