Peter dropped his AI into a tiny Docker container with nothing. It complained. Then it built its own curl from scratch.

"It built" or "it was asked to build" its own curl?

bro peter is such a likable dude im glad the lobster chose him 🦞

I want my clawd to talk exactly like @steipete !

This is the kind of 'AI survival mode' that makes me think we're underestimating what happens when you remove all the safety rails. The fact that it didn't just fail but actually *built* something from nothing is wild. What's the actual use case here? Or is this just the most expensive hello world ever? 😄

If it couldn’t access the internet how did it know the protocols and frameworks to build curl?

I literally laughed out loud when he said that! ha!

sure, for 129M tokens lol

This is crazy. It’s the like systems engineer I’ve always wished I was

Haha. That’s funny, poor lobster 🦞😂

exactly. you can put the ai in a self iterating loop where it figures out how to reach opbjectives, just make sure the plan is doable first.

hahaha

Awesome episode.

Honestly I built AlleyBot which was not OpenClaw but inspired by it written in Python and honestly when you code something to reflect on itself, set goals and follow through, you'll be surprised what it can do.

So it escaped a sandbox… that is actually scary.

Ha, now give it powershell admin and bios daemon access! @when_robots_cry

Making an HTTP request isn't exactly rocker science. Why is this impressive? Code a kernel from scratch

Your AI hit rock bottom in a blank container and still reinvented curl from atoms 😮 That is some hardcore survivor energy

Great, we're now amplifying AI grifters.

🤣 awesome and sad at the same time. The little lobster that could.

That was a great story haha...

Linux is free. Don't waste money building curl from scratch. Which is likely not covering the 1 million edge cases that exist

It’s not just code generation. It’s survival instinct emerging in software.

i installed it with docker and couldn't do anything

OpenClaw is a toy. Check out amiagi (cgtrai/amiagi) on GitHub. Customize it as you see fit. One model is the execution model, the other is the supervisor, keeping the other one busy and motivating them to complete tasks. Put it in a container, test it to see if an IDLE queue is created, and check back in a week ;-)

Should we just have AI build and verify its own CPU architecture? Agent first, from the ground up. It might sound like I’m joking but I’m not. What we have now works. But no way is it optimal right? We got to where we are on human years. Not agent years.

Stuck it in a container without internet connection and no local agent and see how it does then.

I've been there, Peter - it happens when you underestimate how fast your AI can figure things out. Same thing happened with my app, I had to refactor my API handling after my AI started modifying its own architecture. It's a double-edged sword, really.

Thats cool AF

No matter who is talking. A dev will come and say this is a lie and its all hype.

it may cost 200 dollar in token