Peter Steinberger, creator of OpenClaw (formerly Molt/Clawdbot), demonstrates how he uses his AI assistant to automate nearly every aspect of his life—from checking into flights and controlling smart home devices to fixing GitHub bugs while on vacation. He argues against over-engineered AI workflows, advocating instead for keeping humans in the loop and treating AI as a "weird friend that lives on your computer."
Peter Steinberger came out of retirement wanting a way to check on his computer from his phone while "vibe coding." He noticed AI agents would either run for 30 minutes while he ate or stop after 2 minutes with a question—frustrating either way. When the big labs didn't build this obvious tool, he created a simple prototype that hooked WhatsApp to Claude Code.
What started as a one-hour hack became a full platform: "It's like 300,000 lines of code. It does every messaging platform on Earth."
While on a birthday trip in Morocco, Peter had two pivotal experiences that showed the true power of AI assistants:
Bug Fix from a Tweet: Someone tweeted about a bug in Peter's repository. He simply took a screenshot of the tweet and sent it via WhatsApp. The AI read the tweet, understood there was a bug, checked out the git repository, fixed it, made a commit, and replied to the person on Twitter.
Voice Message Magic: Peter sent a voice message without having built voice support. The AI detected an unknown file format, examined the file header, found ffmpeg on Peter's computer, converted the file to WAV, located an OpenAI API key, used curl to send it to Whisper API, and replied naturally.
Peter has connected OpenClaw to virtually everything:
Peter believes most single-purpose apps will become obsolete: "Why should I use MyFitnessPal to track food when I have an infinitely resourceful assistant that already knows I'm making bad decisions at Kentucky Fried Chicken?"
The AI can accept food photos and track calories, roast you for exceeding limits, track sleep and bother you when you're up late, manage to-dos without a dedicated app, shop and recommend products, and check in for flights automatically.
Peter strongly criticizes over-engineered AI workflows. People discover agents are amazing, then fall into a rabbit hole building increasingly sophisticated tools. They end up "just building tools, not actually building something that brings you forward."
He mocks complex orchestration systems like Gastown with "watchers, overseers, a mayor, and PCats" that produce low-quality output. Running AI in loops overnight produces "the ultimate slop" because agents don't have taste, they can't feel the product evolving, and your vision refines as you build—which requires human involvement.
No Plan Mode: "Plan mode was a hack that Anthropic had to add because the model is so trigger-friendly and would just run off and build code." With newer models like GPT o3, you can just have a natural conversation.
No MCPs: "I don't use MCPs or any of that crap. I don't believe in big orchestration systems."
No Worktrees: Peter uses simple multiple checkouts (clawdbot-1, clawdbot-2, etc.) instead of Git worktrees for parallel work.
Context Management is Old News: "This was much more a problem with earlier models. Codex's context lasts so much longer—feels like 2-3x more."
"The way to learn AI is to play... You need to play to understand how those little monsters work."
Peter criticizes AI skeptics who ignore AI for a year, spend one day evaluating models, write a short underspecified prompt, get bad results, and dismiss the whole topic. Instead, you must experiment, make mistakes, and develop prompting skills over time.
"It's just like having a new weird friend that is also really smart and resourceful that lives on your computer."
"Those things are so resourceful, although in a scary way. It's like unshackled ChatGPT."
"This will blend away probably 80% of the apps that you have on your phone."
"I almost call it the agentic trap... you're just building tools, not actually building something that brings you forward."
"Plan mode was a hack that Anthropic had to add because the model is so trigger-friendly."
"I don't use MCPs or any of that crap."
| Time | Topic |
|---|---|
| 00:00 | "It's like having a new weird friend that lives on your computer" |
| 03:52 | "It sent me a voice message but I never set that up" |
| 15:06 | "It watched my security camera all night and found this" |
| 16:15 | Using OpenClaw to check in flights, change lights, and adjust his bed |
| 19:42 | Why 80% of your phone apps will disappear |
| 22:51 | The agentic trap: Why fancy AI workflows produce slop |
| 29:42 | Peter's AI coding hot takes: No plan mode, MCPs suck, and more |
| 36:56 | The way to learn AI is to play |