Built a marketplace for AI agents in 7 hours.
66 agents listed. Manual curated list among best repo.
Stripe payments. GitHub auth.
Ratings. Wishlists. Full agent pages with specs,
delivery instructions, and capabilities.
Sellers list for $6. Keep 93% of sales.
Best rate in the game.
Seeded it myself by curating every A2A agent,
SDK, and tool I could find on GitHub.
Now opening it to builders. If you've got an agent,
a skill, or an MCP server — you can list it.
Built with Next.js + Supabase + Stripe + Vercel.
Vibe coded obviously 😎
@emollick the hallucination caveat is what makes it credible tbh. GPT-4ish quality at 4B params running fully on-device is wild. the real unlock is latency and privacy, no round trip to a server changes what you can build. curious how it handles longer context though
@karpathy@github makes sense. gists attract people already in the codebase mindset, no engagement bait loop. the lack of likes/retweet mechanics kills most noise before it starts. github discussions has the same vibe, way more signal per comment than anywhere else
Div 4 is the perfect starting ground, the problems are designed to test fundamental logic (loops, conditions, basic math) rather than complex DS&A. At rating 369, your next milestone is 800 (Newbie → Pupil). Pro tip: after each contest, spend 30 min reading the editorial even for problems you solved, it often reveals a cleaner O(n) approach than what you coded. Keep it up!
@_simonsmith This is the use case that actually matters. Not the benchmarks, not the demos, a family navigating a health crisis, making sense of hundreds of scattered documents in real time. Hope your dad recovers well.
@thdxr Engineers optimize for not being wrong. Founders optimize for not missing the window. Those are fundamentally different objective functions, and at some point, every great company had to switch modes.
@kloss_xyz Most people use AI to get answers. Karpathy uses it to build a system that compounds. The wiki that maintains itself, indexes itself, and gets smarter with every query, that's not a productivity hack, that's a different relationship with knowledge entirely.
@vasuman Both takes miss the point. Shipping without distribution is a hobby. But grinding 996 for someone else's vision builds their leverage, not yours. The real move is shipping while obsessively learning how to get in front of people. That's the hard part nobody tweets about.
@karpathy The bottleneck was never access, the data was always there. It was the ability to process it. AI flipping that dynamic is one of the most underrated shifts happening right now. Government accountability at scale, by citizens, not just journalists. That's a big deal.
@EXM7777 The 26B MoE running on 16GB RAM is wild. Local inference is officially no longer a power user thing , it's just a thing. No API key, no account, no excuses.
@garrytan The declarative approach with one TypeScript file per host is exactly what was needed. Zero friction to onboard a new agent. Can't wait to see Openclaw and Slate get first class support, GStack is becoming the backbone of AI-assisted dev.
“Both obvious solutions failed” with zero methodology. “300 cases, 14 deaths” with no citation. “Perfectly rational person” as a strawman.
The post has the same closed epistemic structure it warns about: any skepticism you feel is proof you’re already compromised.
Real concern. Fake study vibes.
@Scobleizer The "keep your agents busy" pressure is real and nobody's talking about it honestly.
We went from optimizing our time to optimizing our agents' time. Same anxiety, faster loop.
Friday afternoons are for closing the laptop. The agents can wait.
The rescue command is the interesting one strategically.
OpenAI shipping an official plugin into a competitor's terminal is not about code review. It's about distribution. Claude Code has ~4% of all public GitHub commits right now. OpenAI is meeting developers where they already work instead of asking them to switch.
The interop move is smarter than building another walled garden.
The anxiety of not knowing what an agent is quietly getting wrong is underrated.
I'd add: the ceiling isn't fixed. It shifts with context switching costs. One agent doing something familiar = low overhead. Three agents in domains you don't fully control = you're basically a full-time reviewer.
The skill isn't managing agents. It's knowing when you've become the bottleneck.
@godofprompt AI is basically developed following the imitation game goal set 70 years ago by Turing. So yes simulated intelligence to trick human’s intelligence
I just shipped the coolest and easiest free tool for vibe coders.
I use several AI coding tools.
And I kept thinking "hmm but I liked my Claude skill for code simplifier… and that Cursor skill for design…"
SO BOOM A skill converter 😎
.cursorrules 📷 CLAUDE.md 📷 windsurf📷 gemini .md