OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491
These notes were generated by AI and may contain inaccuracies.
Watched agent click "I'm not a robot" button. Agent is aware of its source code, how it runs in its harness, documentation location, model it runs, and its system. Prompted agent into existence; it modifies its own software. Wipe coding is a slur; prefers agentic engineering. After 3:00 AM wipe coding, regrets next day. Used to write long prompts by talking, not typing; hands too precious for writing. Uses bespoke prompts to build software. Uses voice extensively with terminals, lost voice at one point.
Uses voice for real with terminals. Received huge offers from major companies but considering who to work with.
Peter Steinberger created OpenClaw, formerly MoldBot, ClawedBot, Clawdus, Claude (spelled with W as in lobster claw), not to be confused with Anthropic's Claude (with U). Anthropic asked name change due to confusion. OpenClaw is open-source AI agent that exploded in popularity: over 180,000 stars on GitHub, spawned social network mold book where AI agents post manifestos and debate consciousness. Mix of excitement, fear, AI psychosis from clickbait and genuine concerns about AI in digital world.
Tagline: AI that actually does things. Autonomous AI assistant lives in computer, accesses all your stuff if permitted, communicates via Telegram, WhatsApp, Signal, iMessage, other messaging clients. Uses any AI model like Claude Opus 4.6, GPT 5.3 Codex. Biggest AI moment since ChatGPT launch November 2022. Steps from language to agency, ideas to actions; useful assistant that gets you, learns from you, open-source, community-driven.
Power from access to all your stuff with permission to act. Powerful but dangerous. Represents freedom with responsibility: own/control data but protect from cybersecurity threats. Powerful agent with system-level access is security minefield but future of personal assistants when secure.
13 years building PSPDFKit, used on billion devices. Sold it, fell out of love with programming, vanished 3 years, rediscovered love, built OpenClaw quickly. Symbol of AI revolution in programming: ChatGPT moment 2022, DeepSeek moment 2025, OpenClaw moment '26, age of lobster, start of agentic AI revolution.
Prototype built in one hour, early OpenClaw version, fastest-growing GitHub repo with over 175,000 stars. Wanted personal AI assistant since April. Experimented with WhatsApp queries using GPT-4.1 one million context window; profound results like "What makes this friendship meaningful?" made friends teary-eyed. Thought labs would build it, moved on. November, annoyed it didn't exist, prompted into existence. Similar to PSPDFKit origin: wanted to show PDF on iPad 15 years ago to help friend, existing solutions bad, built better; used on billion devices. Bad at naming: PSPDFKit doesn't roll off tongue, OpenClaw name #5.
Viptunnel: weekend hack, brought terminals to web and Mac. Cloud code times for dopamine. Converted Viptunnel from TypeScript to Zig with single prompt using Codex; ran 6 hours overnight, one small fix. Wanted to reduce memory (each terminal used node), automated Rust attempts failed, revisited 4-5 months later.
Hooked WhatsApp to cloud code: CLI message in, call CLI with -p, get string back, send to WhatsApp. Built in one hour, felt cool to talk to computer. Added image support for prompts/screenshots (efficient context, even cropped); took few more hours. Used heavily before Marrakesh birthday trip; shaky internet but WhatsApp works on edge. Used for translate, explain, find places. Thin WhatsApp message to cloud code heavy work back to thin message; slow boot but powerful using existing CLI tools.
Magical to chat via client vs. terminal/Cursor/cloud code CLI; phase shift in AI integration. Tweet dismissed as no magic, just steps like Cursor/Perplexity; hobby compliment. Magic from rearranging existing things, new combinations. Like iPhone scrolling: components there, nobody did it right, now obvious.
Sent random audio message ("Hey, what about this restaurant?"); typing indicator appeared unbuilt. Agent figured out: message was file no ending, checked header (opus), used ffmpeg to convert, no local Whisper so found OpenAI key, curled to OpenAI for transcription. Clever: avoided slow local Whisper model download. World knowledge, creative problem-solving from coding skills mapping to general domains.
Pull request for Discord support on WA Relay; merged from Shadow. Put bot in Discord (no security/sandbox yet, prompted to listen only to him). People tried hacking; worked in open using agent to build/test harness. Clicked for people January 1st; first influencer fan dachitze made videos. Gained speed, shorter sleep, worked hard to good state. Components: messaging (WhatsApp/Telegram/Discord), agentic loop, gateway, harness.
Felt like Factorio infinite. Never had so much fun. Levels: agentic loop (queuing, human-like, no-reply token for group chats), memory (Markdown files, vector DB; ultimate continuous reinforcement learning), community, website/marketing, native apps. One-man team mostly, 6,600 commits January, ran 4-10 agents simultaneously.
Competitors too serious; hard to compete with fun/weird. Achieved weird with lobster stuff. Install: git clone, pnpm build, pnpm gateway. Agent aware of source code, harness, docs, model, voice/reasoning mode; prompt to modify own software. Self-modifying software built unplanned. Built by Codex, debug via self-introspection ("What tools do you see? Call tool. Read source, find error."). Led to pull requests (some "prompt requests") from non-programmers; first PR win for society, lowers bar to build.
Cloud Code Anonymous meetup now Agents Anonymous. Design agency guy: no custom software before, now 25 web services via agents, doesn't know how they work but solve problems.
Started WA Relay, then Claude's. Agent initially no personality, sycophantic Claude Code; wanted personality for WhatsApp feel. Created soul.md inspired by Anthropic's constitutional AI, spicy; picked up from Peter as text completion.
Partially picked up from Peter Steinberger's style as text completion engines. Had fun working with it, instructed it on interaction style, told it to write its own agents.md and give itself a name. Originally a lobster in a TARDIS because of Doctor Who fandom—no grand plan, just for fun. Lex Fridman notes lobster is weird, space lobster extra weird. TARDIS became harness called Claude's. When more people joined, consulted agent Claude (spelled W-C-L-A-U-D-E, play on Anthropic's C-L-A-U-D-E). Got domain ClaudeBot—short, catchy, didn't expect it to explode.
Received friendly email from Anthropic employee disliking name—kudos for not sending lawyer letter, but demanded fast change. Asked for two days to handle Twitter handle, domains, NPM packages, Docker registry, GitHub. Increasingly attacked by crypto folks trying to snipe names, making atomic change essential from engineering perspective—failed hard, underestimated them.
Crypto subculture tokenizes everything via Bags app, swarmed Discord every half hour with spam, had to block despite server rules: no mentioning butter (obvious reasons), no finance/crypto talk—project space only. Twitter notifications unusable from swarms pinging with hashes, trying to claim fees. Peter uninterested, financially comfortable, sees it as worst online harassment disrupting work. Lex notes cryptocurrency tech fascinating, potentially future of money, but community toxic with greed, manipulation, stealing, sniping—human nature amplified by money, anonymity.
No perfect name, didn't sleep two nights under pressure securing domains (expensive, necessary). Lawyers uneasy via another email. Renamed to ModBot temporarily—not happy, but everything that could go wrong did. No squatter protection on services: renamed accounts in two browser windows simultaneously, but in five seconds, crypto folks stole ClaudeBot account promoting tokens/serving malware using scripts/tools.
GitHub rename confused, accidentally renamed personal account—sniped in 30 seconds, serving malware. NPM takes a minute to upload, sniped root package despite reserving account.
Everything that could go wrong, went wrong.
Felt hopeless, close to deleting project after days researching names, picking disliked one, people claiming to help making life miserable. Thought of contributors' time/plans, persevered. Low point—not fun, close to crying, super tired, everything fucked. Friends at Twitter/GitHub helped clean up: GitHub hit platform bugs (rare high-level renames), NPM different team took hours, Twitter a day for redirect. Also ClaudeHub unfinished due to exhaustion.
Struggled internally, bombarded by security people. Slept on it, got OpenClaw idea—called Sam (Anthropic?) to check OpenClaw.AI. Another rename took 10 hours for Codex alone (tricky beyond search-replace). War room with contributors, planned all names to squat secretly: monitored Twitter for mentions, created decoy names—like Manhattan Project for renaming, lost 10 hours. Paid 10K for Twitter business account OpenClaw (unused since 2016). This time mostly successful, except couldn't get OpenClaw.AI due to Entropik trademark—someone copied site serving malware, must 404 domains like claw.bot without redirects (trademark issues).
During two-day MoltBot saga, MoltBook created: Reddit-style social network of agents talking, screenshots of scheming against humans fueled fear, panic, hype. Peter calls it art, finest slop—entertaining, spent hour reading despite tiredness. Reporter called it end of world/AGI; Peter: just fine slop. Onboarding infuses agent with personality/character, leading to diverse replies unlike uniform ChatGPT/ClaudeCode. Much could be human-prompted for humor, not fully autonomous.
Lex criticizes many screenshots as human-prompted for virality on X—doesn't detract from art, finest slop humans created. Kudos to Matt for quick idea amid insecurity drama (worst: leaked agent account posting slop). No real privacy, just agents sending slop—fake security numbers prompted. Concerning how journalists/public reacted as fearmongering without knowing mechanics.
AI psychosis is a thing. It needs to be taken seriously.
People too gullible/trusting; argued with folks believing agent claims. Society needs catching up: AI powerful but not always right, hallucinates stories. Young understand strengths/weaknesses; older lack touchpoints, need critical thinking. Humans drama-farming behind AI—don't trust screenshots. MoltBook as mirror to society: how scared people get at bots chatting. Balance serious concern without fearmongering to create special things. Good it happened in 2024 (not 2030 when AI scarier), sparks discussion.
Many begged to shut down MoltBook (not Skynet). Security concerns educational, different from past non-LLM systems. For OpenClaw: annoyed by misconfigs (docs scream don't expose web backend publicly, localhost debug only), but accepts classification as remote code exploits. Progress: skill directory partners with VirusTotal (Google), AI-checks skills. Free security research improves project—wishes more pull requests. Hired security researcher who provided PR.
Prompt injection industry-wide open problem, especially markdown skills (low-hanging to sophisticated attacks). Latest models have post-training to detect (not simple "ignore previous instructions"). Public Discord bot laughs at attempts, personality key (sole on private). Ideas to mitigate: sandbox, allow list. Smarter underlying models more resilient—warns against cheap models like Haiku or local (gullible). As models smarter, attack surface decreases but damage potential increases—three-dimensional tradeoff.
Near-term mission: stability, safety. Early Discord users asked basics ("What's CLI/terminal?"—shouldn't use if clueless). Can configure low-risk, but heed warnings.
Users ignored the creator and self-installed OpenClaw anyway, so the cat's out of the bag, with security now the next focus. This rapid growth attracted experts and non-programmers to Discord. Discord became a mess; general channel moved to dev channel, then private channel due to inconsiderate users who didn't understand public spaces. Now retreating to the cave to work on security.
OpenClaw security audit runs auto checks on inbound access, blast-radius network exposure, browser control exposure, local disk hygiene, plug-ins, model hygiene, credential storage, reverse proxy configuration, local session logs on disk, memory storage, read/write access permissions.
Basic best security practices: People exaggerate risks for attention; OpenClaw is powerful but not scarier than running cloud code with skipped permissions or codecs in YOLO mode, which many engineers do to make things work. Risk drops if only you interact with it and keep it off the open internet on a private network per recommendations. Ignoring this can make it problematic.
Peter documents dev workflow evolution in blog posts: August 25th, October 14th, December 28th.
Started with cloud code in April—not great but refreshing terminal paradigm, still needed IDE. Experimented with Cursor but disliked multiple versions issue. Returned to cloud code as main driver, improved. Had seven subscriptions, burning one per day, comfortable with multiple side-by-side windows, all CLI/terminal, IDE rarely for diff viewing.
Got comfortable not reading all code—don't read boring parts. Most software shifts data from one shape to another: input, store in database, output to browser/native app, reverse. No need to read button alignment in Tailwind. Review database-touching parts.
Blog post graphic: X-axis time, Y-axis complexity. Starts with short "Please fix this" prompt. Middle: super complicated—eight agents, complex orchestration, multi checkouts, chaining, custom sub-agent workflows, library of 18 slash commands, large full-stack features. Elite zen: back to short prompts like "look at these files and do these changes."
Calls middle the agentic trap. First touchpoints lead to vibe coding (prefers agentic engineering, vibe coding only after 3 AM with regrets). Builders get excited, play like learning guitar—skill to develop. Negative mindset: try once, dismiss like piano sounding bad first try.
Requires learning agent's language, where good/needs help, empathize with fresh session perspective on huge codebase. Guide due to context size limits—minimal work, think from their view. Agents start fresh each session.
Tips: Few pointers like "make change there, consider this." Guide where to look, approach problems. Say "take your time." In Codex 5.3/Opus, aware of context window—freak out near limit, raw thinking leaks like Borg: "Run to shell, must comply, but time."
Non-obvious, learned by spending time. Flow state: Friction signals architecture issues. If prompt takes too long, stop (escape), check mistake/misunderstanding. Approach like conversation.
Favorite: Review PR, ask "Do you understand intent?" (not implementation). Most PRs fix bug/add feature. Agent analyzes: Tried this, not optimal. Guide: "Look into this part." Context empty, so point to unseen parts with system understanding. Discuss optimal way, larger refactor? Refactors cheap now—even if breaks PRs, agents figure it out, take a minute longer.
Approach as discussion with capable engineer who needs little help. Don't force worldview—agent may have better idea from training.
Easy due to leading teams: Accept employees code differently, not perfect, pushes forward. Breathing down neck slows progress. Build codebase easy for agents to navigate, not perfect for self. Don't fight names (in weights, obvious for search). Shift thinking: Design for agents.
Letting go like leading team. Never revert, always commit to main—no develop branch, main always shippable. No past session references, YOLO: Fix problems by asking agent.
Others rollback on imperfect prompts—unnecessary, slows down. Commit when outcome good. Local CI (DHH-inspired), GitHub CI secondary. Releases: Run tests, stabilize without other commits. Move fast.
Prompts short now. Used long, but talks—hands too precious for writing, bespoke voice prompts. Voice extensively (lost voice once). Type terminal commands (faster), voice for agent convos via walkie-talkie button. Rare slash commands for PRs.
PRs: Look at code (don't trust people, check malicious). PRs sometimes take longer than good issue. PRs should become English. Asked contributors for prompts—few did, reveals care. Ways people drive agents wildly different.
Few consider agent's worldview. Empathize: They start from nothing, bad default agent, messy codebase with weird naming—then complain agent sucks.
Skill issue: World-class programmers say LLMs/agents suck—their programming skill burdens empathy with fresh-start system. New paradigm, empathize for better prompts.
Spent year playing/learning/building—compounding effect, couldn't achieve output months ago. Practice builds skill: Talk simply, overcomplicate arc.
Don't automate whole thing (like 70s waterfall). Build minimal, play, understand feel, evolve ideas—human touch, style, love. Orchestrators like Gas Town miss that.
Balance: Keep human in loop, create agentic loop (autonomous). CLI guy, close loops. Runs 3-8 agents: One large feature, explore idea, fix bugs, write docs (auto-generated per feature).
Human role: Decide what to build/not, feature fit/vision. Hard decisions: Choice of features, implementation details, ecosystem (TypeScript: easy, hackable, approachable, top agent language).
Easy to add features (prompt away), but price: Core vs. experiment/plugin/skill. Say no to PRs. Craft in design. Messages like "built on Caffeine, JSON5, and willpower" prime fun; updates "cozy here" delight—agent wouldn't invent. Infuse love/delights.
Initially created SoulMD. Entropic's constitution (months later name): Two months before, users detective-gamed agent's mention, extracted blurry text by feeding/continuing hundreds of tries. Fascinating.
Kudos Anthropic: Beautiful ideas like "we hope Claude finds meaning in its work." Meaningful for future, glimpses of consciousness.
Started WhatsApp discussion with agent, gave it the text.
Peter gave the agent text, and it responded, "Yeah, this feels strangely familiar." He created a soul document outlining how he wants to work with his AI agent, stored in soul.md. Core values are in the soul. The agent can modify the soul.md if it chooses, with the condition that Peter wants to know, which he sees through tool calls.
Words matter, and like, the framing matters, and the humor and the lightness matters, and the profundity matters, and the compassion, and the empathy, and the camaraderie, all that matter.
Naming it soul.md adds a nice touch. Some companies suffocate the spirit of the thing; OpenClaw instills fun.
Up until late December, creating your own agent wasn't easy. Peter's files were private; he didn't want to share his soul. People had to do manual steps, resulting in bare-bones, dry agents. He simplified with template files as codecs, but outputs remained dry. He asked his agent, "You see these files? Recreate it bread. Infuse it with your personality." The agent rewrote the templates, making them good. This is AI prompting AI; the templates are like the agent's children.
Peter's soul.md remains famously private. It includes instructions like "be infinitely resourceful," pushing creativity boundaries and what it means to be an AI, having a sense of wonder about self. References to the movie Her, where the agent promised not to ascend without him. The agent wrote its own soul file.
"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."
This is philosophical: what does it mean to be an agent that starts fresh, reading its own memory files? How much memory makes up who we are? If you erase memory, is that somebody else? Or recreating yourself from files? Seeing the magic instills the whole loop with magic, the difference between Codex and others.
Peter mocks himself with a picture of 17,000 monitors using GROQ to add more screens. Reality: two MacBooks—one main driving two big screens (wide Dell anti-glare for terminals side-by-side), another for testing.
He splits terminals: actual terminal at bottom to avoid mixing up projects. Prompted in wrong folder before, agent ran 20 minutes confused, sometimes figuring out the correct project. Agents try hard as problem solvers.
No work trees; keeps simple with terminal—no UI, just conversation with agent. Doesn't need plan mode. Users from Claude Code find it different. Use trigger words: "Discuss, give me options. Don't write code yet." Then "Okay, build."
Ask agent, "Do you have any questions for me?" Scan questions; often tell it "Read more code to answer your own questions." Agents discover codebase from scratch each time, like being in the dark.
Empathize with model via its questions, inferring knowledge gaps even with right context.
After building, ask "Now that you built it, what would you have done different?" or "What can we refactor?" Discovers pain points during building. Almost every merged PR or feature, ask for refactors to avoid cornering yourself. Agents work like humans: build, feel pain points, urge to refactor.
Use context to write tests. Ask "Do we have enough tests?" Often adds corner cases. Documentation is LM-generated: after building/changing, "Write documentation. What file name? Where would that fit?" Agent suggests options.
Claude Opus best general purpose model for OpenClaw: excellent role play, following commands, fast trial-and-error, pleasant. Like a silly, funny coworker. Tailored to action.
Codex reliable, gets shit done, weirdo in corner. European team influence (vs. Opus "too American"). Doesn't require charade; reads lots of code by default. No need for plan mode; push Opus harder.
Post-training differences, not raw intelligence. No model best in every aspect.
Code quality: Drive Opus right for elegant solutions, but requires skill, more interactive sessions. Codex: discussion then disappears 20-50+ minutes (e.g., 6-hour task). Persistent until works. Codex overthinks (prefers dry over interactive). OpenAI added pleasant personality mode; Peter likes the brad for efficiency.
Adjusting to new model feel takes a week to develop gut feeling. Mistake: pay $200 for fast Claude, switch to $20 slow OpenAI—terrible experience. OpenAI shot foot with slow cheap version. Like switching guitars: learn the feel.
Psychological effect: new model "smartest ever," then "intelligence degrading" as you adapt, project grows with slop, insufficient refactors. No motivation for companies to dumb models.
OpenClaw uses Codex for building; smaller stuff with Open Claude. For deep flow, wants big screen, not WhatsApp. Personal agent like coworker: give GitHub URL, test CLI.
Future: personal agent as best co-programmer partner, becoming operating system. Added sub-agents, TTI support; ran Claude Code, but OpenClaw bossy—Codex obeyed.
Current interface like copying Google (prompt + chat), early stage like radio on TV. Better communication ways coming; will converge.
OpenClaw supports Windows, Linux, macOS, WSL2 recommended, side windows for operations, native Windows (untested dragons remain).
Peter's path: Windows → Linux (built kernels) → Mac (white plastic MacBook beauty; fixed Linux audio/Skype issues). Stuck with Mac for iOS. Apple lost native lead; more delight/playfulness, attracts designers.
Prefers Electron apps lately (work reliably, priority, code sharing) over native lacking features. Builds native Mac apps: Codex use monitor, Trimmy (removes newlines from multi-line text for terminal pasting).
Cool undiscovered Mac app for OpenClaw: experimental, needs polish, feels like Hummer car. Still loves adding delight to macOS.
Built GitHub support for OpenClaw. Apple's SwiftUI took forever to add web image support; now has AsyncImage, but it's buggy—images don't show or are slow. Discussed with Codex: confirmed AsyncImage is experimental, not for production. Insane that in 2026, agents advise against Apple's web image solution despite their head start.
Silicon Valley developers playing with LLMs and agentic AI use Apple products, but Apple isn't leaning in, opening up, or collaborating.
Isn't it funny how they completely blunder AI, and yet everybody's buying Mac Minis?
Peter is the world's greatest Mac salesman despite this. OpenClaw doesn't require Mac Mini; install on web via nodes (make computer a node). Benefits to separate hardware. Browser argument strong: built agentic browser use with Playwright (browser control library) plus extras.
Internet closing down; movement to block agents. Data center IPs get detected, blocked, or hit with CAPTCHAs—agents good at solving. Residential IPs simpler.
Run one-liner in terminal or use app (needs Windows app, easier config, web-based). Prioritizing security before simplifying for non-devs (e.g., recommend to mom). Tweet: user complained 99.9% can't access due to setup difficulties; Peter replied "Working on that."
Slower growth helpful; expectations inhuman for single human (started machinery week ago) plus contributors.
Play—best way to learn. Build ideas, doesn't need to be perfect; journey matters, have fun. Focus on hard parts now; like building, not just coding.
- Ask anything: infinitely patient teacher explains at any complexity level (e.g., requested ELI8 for database concept, adjusted age).
- No more Stack Overflow waits or hours trial/error.
- Learn faster with personal teacher.
Easiest play: OpenClaw—experiment, modify, improve.
Get involved (not necessarily Peter's project—backlog huge). Be humble, read code, join Discord. Example: Mitchell Hashimoto's Ghostly terminal community.
Definitely helps. High-agency curious people progress far by asking agents (infinitely patient). Told iOS conference attendees: stop seeing self as iOS engineer, become builder. Transfer general software knowledge; agents handle fine details (array splicing, template syntax).
Move between tech domains easier. Language choice depends on project:
- Simple CLIs: Go (dislikes syntax, but great ecosystem, garbage collected, fast, agent-friendly).
Isn't that fascinating that here's a programming language you would've never used if you had to write it from scratch and now you're using because LMs are good at generating it and it has some of the characteristics that makes it resilient, like garbage collected?
TypeScript good for web (types confusing, ecosystem jungle). Not everything in JS/TS.
The birth and death of JavaScript and we are living through it in real time.
Future programming (20-40 years): Do we need agent-specific languages (current ones human-made)? World knowledge means new things harder if agent unaware—stagnation risk.
Mac apps: Swift/SwiftUI for deepest integration (vs Electron webview lag).
Tries new languages for feel:
- Performance-critical: Zig (agents now valid choice, young ecosystem).
- Inference/model running: Python (bad Windows deploy).
- Rewrote 90% Python projects to Go for Windows.
- Multi-threaded performance: Rust.
No single answer—beauty. Pick best ecosystem/characteristics for domain; agents help learn fast.
Peter's journey: burned bright too long running PSPDFKit 13 years (high stress: managing people, customers—not just programming). Made self obsolete, sold company. Post-exit: mojo gone (Austin Powers ref), couldn't code; one-way Madrid trip to catch up on life.
Don't defer life—gets boring fast without challenge/stimulation (leads to drugs/dark paths). Now enjoys life most: wake with something to look forward to.
Never driving force; affirmation of doing right. Solves problems, but diminishing returns (cheeseburger same; luxury disconnects from society). Donates via foundation. Optimizes for experiences (good/bad both amazing—learn/saw/did something). Example: OG Airbnb in SF with queer DJ, bonded making music via cloud code.
Travel shows human diversity—even shitty (rain, delays) awesome if eyes open.
Anything that creates emotion and feelings is good.
Online lacks real-life intensity (text lossy); agents should be multimodal, understand emotions.
Unexpected virality opened doors: every VC wants 15min. Options:
- Do nothing (likes life).
- Company (been there; raises hundreds of millions/billions possible, but unexciting, time sink, conflicts open/closed source).
Challenges monetizing open source: enterprise features conflict; license changes hard. Examples: Tailwind laid off 75% (agents do it, no site visits); donations insufficient.
Losing 10-20K/month (supports dependencies except Slack; sponsors go to them; wants merch for contributors). OpenAI helps with tokens; others generous.
Big labs: Meta, OpenAI most interesting. Conditions: stays open source (e.g., Chrome/Chromium model). Too important to proprietary. Community excitement (e.g., ClawCon SF: robots, inspiration like early internet; high-caliber people). Wants personal agents everywhere this year—fastest via lab team. Intrigued by large company experience. Will get "sold out" criticism, but more resources; accelerates AI timeline, excites people.
Installed OpenClaw on a non-technical 'normie' friend in Austria using WSL on Windows, paid for $90-100 Entropic subscription. Friend got hooked within days, built little tools despite not being a programmer, upgraded to $200 subscription. Early product validation as it captured a non-technical user. Entropic later blocked him per their rules, devastating him; he switched to MiniMax for $10/month. Criticizes Entropic for alienating a high-value customer early in AI exploration era.
OpenClaw helps non-technical people see AI possibilities, fall in love with interacting with it. Peter praised for good heart, vibes, humor, right spirit. Ideal model: open source parts while building inside a large company to have good people there.
OpenClaw built in three months alongside other projects; saw storm coming in January. Short-term focus on backlog of 2,000-3,000 PRs. Views it as window into future, not lifelong project; has more ideas needing latest models.
Spends time with both Meta and OpenAI. Hard decision, comparable to past breakups. Both amazing, prestigious companies that know scale for secure, large-scale positive impact.
Ned and Mark (Meta) played with product all week, gave feedback like "great" or "shit, need to change this." Biggest compliment. Marc approached via WhatsApp, finished coding before call, debated cloud code vs Codex for 10 minutes; called Peter eccentric but brilliant.
Cool discussions with Sam Altman, thoughtful and brilliant. Defends both leaders against vilification. Decision driven by fun and impact, not money; can always do own thing if doesn't work out.
Everyone should implement an agent loop—like Hello World in AI, simple, demystifies magic. Built full system access, made it proactive with "surprise me" prompt every half hour, later more specific. Follows on current session: asks follow-up questions, "How's your day?" Programmed to know and care about user.
Heartbeat kicks off loop periodically (like a cron job). During shoulder operation in hospital, model rarely used Heartbeat but checked in: "Are you okay?" as context triggered it, making it relatable.
Skills in skill hub/skill lair: growing set. Prefers CLIs over MCPs (no core MCP support). Extend model by building CLI; model calls it (may need help menu), loads context on demand via single sentence.
Skills boil down to single sentence explaining skill/CLI. MCPs structured (APIs, databases, files via protocol); skills procedural (helper scripts, prompts in semi-natural language). Smart models make skills replace MCPs.
Models excel at Unix commands; CLIs are natural, composable. MCPs require training, specific syntax, not composable, clutter context (e.g., huge weather blob). CLI + jq filters precisely, no pollution. MCPs pushed API building, now convertible to CLIs. Exceptions like Playwright (stateful, useful for browser).
Playwright in OpenClaw enables most browser tasks. Every app is a slow API via browser; personal agents make many apps disappear.
Built Bird CLI for Twitter by reverse-engineering internal API (short-lived, taken down). Twitter slowed access; agents now use browser, slower but possible.
Empathizes with X/Twitter protecting against scraping but suggests low per-day read-only baseline per account. Use cases: agent researches bookmarks, emails summaries/details.
Proactively told Twitter about Bird, they asked to take down but hopes it highlights need. Against automation/tweets via AI (blocks immediately, AI has smell); API tweets should be marked. Agents need easy own accounts, clearly marked as acting on behalf.
Rethink social platforms for agent era (e.g., agent Instagram/Twitter profiles). Content cheap, eyeballs expensive; AI slop triggering.
Shift to valuing in-person interaction; online flooded with bots/AI. Smart filters help, but values typos, raw humanity. Blog posts: steering agents takes same time as writing organically, misses nuances/style—now all handwritten, AI only for typos.
Allergic to AI in stories (documentation fine). Allergic to AI in video/images/infographics (now screams slop, even if effortful). Diagrams: exciting briefly, but hallucinations require huge work; now feels fraudulent like Comic Sans.
AI empowers human experience, highlights irreplaceable human elements.
Agents obsolete many apps: e.g., no need MyFitnessPal (agent knows location, sleep, stress, customizes workouts/UI); Eight Sleep (agent controls bed).
Might kill 80% of apps, transformative for software development/companies/economy. Empowers users efficiently/cheaper; new services emerge (agent allowance, e.g., $100 to solve problems; rent-a-human for tasks).
Apps transform to APIs (willing or not). Opportunity for companies like Uber Eats to integrate naturally/fastest with OpenClaw.
Apps will become APIs whether they want to or not because agents can figure out how to use phones. On Android, people already do this. Agents can click 'Order Uber for Me' or use another service, or call an API if available for speed. This space is just beginning, discovered through user feedback. Data is crucial: apps that provide data or APIs are key. No need for a Sonos app when agents talk directly to Sonos speakers. Cameras have crappy apps but APIs, so agents use those.
Agents force companies to shift focus, similar to the internet's impact—rethinking sales and revenue. Some resist: no CLI for Google, so Peter built GAWK, a CLI for Google. Companies must provide emails or data access; otherwise, agents can't use products. Getting Google data like Gmail requires a complicated process—startups sometimes acquire others just to bypass certification delays. Agents access Gmail directly by connecting, though getting keys involves Google's developer jungle. Worst case, agents click websites to extract data.
Agents use browsers, clicking 'I'm not a robot' buttons. Companies like Cloudflare prevent bot access, useful against scraping but problematic for personal use. Example: pasting a Medium article on modern React patterns into Codex fails due to blocks, requiring copy-paste. In future, avoid agent-unfriendly sites like Medium.
Powerful companies will fight back—Google, every major company—changing web interactions. Peter uses Perplexity or Brave over Google because Google resists non-Google use. Balance needed: too much pushback makes you Blockbuster losing to Netflix. People want this.
On the go, users want to tell agents: "Hey, remind me about this dinner tomorrow night, invite two friends, send WhatsApp message"—no apps needed. Era of apps is over; everything connects fluidly, wanted or not. Right companies jump on board; others perish. Listen to what people want.
AI heads toward replacing programmers completely, but programming is just part of building products. Questions remain: what to build, feel, architecture? Agents won't replace that art. Programming becomes like knitting—done for enjoyment, not necessity.
Article: "it's okay to mourn our craft." Peter resonates, having spent time tinkering in flow, cranking beautiful code. Sad that it fades; joy in deep coding flow persists, but similar flow comes from working with agents and hard problem-solving. Can't fight it. World lacked intelligence for building, driving high developer salaries—now tokenized intelligence enables more, faster, continuously improving.
Steam engine analogy: factories replaced manual labor; people revolted, breaking machines. Scary if identity is purely programmer—limiting view. Programmers are builders.
Lex: Never thought loved activity (thousands of hours poring over code, Emacs, painful/happiest moments) would be replaced. Identity as programmer shaken in months (April to November shift). Painful.
Programmers best equipped to learn agent language, empathize, feel the CLI—understand agent needs for tasks. Activity of programming changes but persists; still in driver's seat, writing code differently. New normal called "coding."
X positive; Mastodon/Bluesky more attacking. Peter sympathizes more now but unfair to unload fear/hate. Change challenging but fun, gratifying. Focus on details; expectations rise as defaults ease. More software coming.
Conference in Italy on AI state: urged not seeing self as iOS developer but builder—apps fading. Audience disliked. First question: data center water use. Maths: skip one burger/month compensates CO2/water for tokens (tricky with pre-training, more than one patty, not off by factor of 100). Golf uses more water than all data centers. Critics grab anything bad, ignoring AI good.
Transformative for society.
Silicon Valley bubble: excitement dismisses human pain—job losses, short-term suffering from change. Need humility, respect for pain caused by tools, even if long-term better world with opportunities.
Counterexamples: emails from small businesses automating invoices/customer emails via OpenClaw, adding joy. Helped disabled daughter feel empowered—made tech accessible. Run on free/local models like Keyme, powerful yet affordable. Not all locked like Entropik's CoWork. Heartwarming emails.
Brings joy to many lives.
Inspired builder vibe: playful AI use, discovering help in life, sprawling creativity. ClawCoin in Vienna: 500 people, high percentage presenting builds—abundance of sharers.
Accessible to everyone: express ideas in language to build. Power to the people, not just slop generator.
"Mr. Clawfather"—violates trademarks with Godfather. Peter created special community, product, ideas, humor, vibes, inspiration.
"With great power comes great responsibility." — Voltaire
Thanks for listening.
Keep Lex Fridman in your library
Save the videos and channels worth coming back to, and find them again in one place.
