OpenAI Just Acquired the Developer Who Built 180K GitHub Stars in 3 Months Alone
10 lessons from the Lex Fridman interview that will change how you build with agents.
Peter Steinberger spent 13 years building PSPDFKit. Sold it. Retired to Madrid. Got bored.
Came back and built OpenClaw in 3 months by having conversations with an AI agent.
180,000 GitHub stars. The fastest-growing repo in GitHub history.
Last week, Sam Altman announced that Peter is joining OpenAI. He called him “a genius with a lot of amazing ideas about the future of very smart agents interacting with each other.”
If you want to see how OpenAI itself learned to explain its vision, I broke down their early pitch deck here.
Lex Fridman just released a 3-hour interview. Here are the 10 moments that matter:
1. His Agent Solved a Problem It Was Never Programmed to Solve
Peter sent an audio message to his agent in Marrakesh. The agent had never been given audio capability.
“A typing indicator appeared. And I’m like, wait, I didn’t build that. It only has image support. So what is it even doing? And then it would just reply.”
The agent checked the file header, identified Opus audio, converted with FFmpeg, found an OpenAI key on the machine, hit the Whisper API, and replied.
Zero instructions. It reasoned through the entire chain.
The takeaway: Architecture that lets agents improvise is worth more than any feature you ship.
2. OpenClaw Beat Funded Competitors With One Advantage: Fun
“They all take themselves too seriously. It’s hard to compete against someone who’s just there to have fun.”
The lobster mascot. The Soul.md file. Terminal-only install for months. Personality was the product.
The takeaway: Personality is a moat. You cannot copy it. For more examples of how AI startups are differentiating themselves, I curated the best recent AI decks.
3. Users Who Never Coded Started Submitting Pull Requests
Peter made the agent aware of its own source code. He used it to debug itself. Users copied the behavior.
“Every time someone made their first pull request is a win for society. Doesn’t matter how shitty it is.”
This is the same pattern happening with Claude Cowork. People who never thought of themselves as builders start automating their own workflows.
The takeaway: Agents that understand their own architecture create builders, not users.
4. Crypto Mobs Stole His GitHub Handle in 5 Seconds
Anthropic asked him to rename the project. He had two days.
“I pressed rename there. I pressed rename there. And in those five seconds, they stole the account name.”
Recovery required decoy names, a secret war room, $10K for a dormant Twitter handle, and a call to Sam Altman.
The takeaway: When you go viral, you become a target the same day.
5. Most Developers Fall Into the Agentic Trap
Short prompts → eight-agent orchestration → short prompts again. The middle phase is the trap.
“It’s the same way as you have to play with a guitar before you can make good music. It’s a skill.”
Basic prompting is dead. The skill now is knowing how to structure your requests. MIT recently dropped a technique that changes how AI reasons, and it connects directly to what Peter is describing.
The fix: After every feature, ask the agent “what should we refactor?” It just felt the friction. It knows where the debt is.
For a complete breakdown, I wrote a guide to modern prompt engineering that covers all of this.
6. Claude Opus vs Codex: The Honest Breakdown
"Opus is like the co-worker that is a little silly sometimes, but it's really funny and you keep him around. Codex is like the weirdo in the corner that you don't want to talk to, but he's reliable and gets shit done."
Opus: fast, pleasant, needs steering.
Codex: disappears for 50 minutes, comes back thorough and correct.
I wrote a full breakdown of Opus 4.6 if you want to understand how to get the most out of it.
The takeaway: Give any new model a full week before judging it.
7. Cheap Models Are a Security Liability
“Don’t use cheap models. Don’t use Haiku or a local model. They are very gullible. It is very easy to prompt inject them.”
Three things that change your risk profile: public web backend, weak models, ignoring documentation.
The takeaway: The psychological framing of your prompts also matters more than most people realize.
8. MCPs Are Mostly Unnecessary
OpenClaw does not use Model Context Protocol in core. Nobody has complained.
“Every MCP would be better as a CLI. And now this stuff doesn’t even have MCP support, and nobody’s complaining.”
CLIs let agents filter and pipe. MCPs dump everything into context.
This aligns with how serious builders are using Claude. The simpler the interface, the better the results
The rule: Stop adding MCP servers. Add CLIs instead.
9. He Sold a Company, Retired, Got Bored, and Built the Fastest-Growing Repo in GitHub History
Peter sold PSPDFKit after 13 years. Software on a billion devices. Financially set.
“If you wake up in the morning and you have nothing to look forward to, you have no real challenge, that gets very boring very fast.”
He booked a one-way flight to Madrid and spent three years catching up on life. Then came back and built OpenClaw in three months.
The burnout was not from code. It was from people:
Co-founder conflicts
Customer friction
The kind of grind that has nothing to do with what you are actually building
He routes all sponsorship revenue to OpenClaw’s dependencies and is losing $10,000 to $20,000 a month on the project without concern.
The takeaway: The work is the reward. The economic shifts happening right now are forcing everyone to rethink what work even means.)
10. 80% of Apps Will Be Replaced by Personal Agents
This is not a 2030 prediction. Peter thinks the displacement is already underway.
“Why do I still need an app to do that? Why should I pay another subscription for something the agent can just do now?”
A personal agent with full system access, persistent memory, and browser control already knows your location, calendar, and stress. No single-purpose app competes with that context.
Peter watched his agent click the “I’m not a robot” button.
The question: What does your product look like when the interface is an agent?
The Playbook
Build what frustrates you. The best products start with “why doesn’t this exist?” If you need ideas, here are 70 startup ideas YC wants funded.
Empathize with the agent. It starts from nothing every session. Context before tasks. This prompt structure helps.
The journey compounds. Peter could not have built OpenClaw six months earlier.
If you want to start building with agents today:
Every agent you build tomorrow will be more capable than the one you built today. The window to get ahead of this is open right now.
Start building.
Watch the full podcast here
If this saved you 3 hours, share it with one builder still opening five apps to do one thing.
The "MCPs are unnecessary, just use CLIs" take is interesting but I think it misses something.
CLIs work great when the agent already knows the tool exists.
MCPs give discoverability that CLIs don't. Both have a place tbh.