The AI operating system has arrived

on

I've built my own orchestration framework for LLMs because I wanted something very simple and easy that I could tweak; I even started exploring programmatic tool calling with my framework. But for the last 10 days, my whole view on this endeavour has completely changed since I've been creating skills with OpenClaw and I must say, OpenClaw on its own seems to do a better job at nailing the processes that I've tried to implement with my framework.

I'm not saying everything is better per se, but it does one thing that I couldn't: it builds its own scripts, tries them, and fixes them until they work. But it does a lot more — it can install things on its own, it can read API documentation, it can create databases and so forth. It can do all that because it acts more like an operating system than a workflow tool. Just to give two simple examples that I've done just for myself:

a) I have a network audio player (Wiim Ultra) that can access my whole CD library (all FLAC ripped) via DLNA. However, I rarely use it because it's too cumbersome to search or browse for songs or albums, and it's too much hassle to create my own playlists etc. Yesterday, I was in my bed, about to wake up — and I thought to myself, wouldn't it be nice if I could ask OpenClaw to play something for me. So I asked, "Do you see any network devices called Wiim on the local network?" and it came back: "Yes, two devices ...." and so I just asked it to play me a song on one of the devices. It took a good 5 minutes, but in this time, it went to the Wiim website, got the documentation, wrote a Python script that can now interact with my DLNA server and the Wiims. Then I asked if it could put together a playlist, and in order to do that properly, it wrote a daemon process that acts as a webhook for the Wiims so that when the Wiims are finished playing a song, it would be notified and could play the next song — in other words, it created the whole playlist functionality on its own.

b) I took the train from my station to Oerlikon — that's about a 25-minute train ride — and during that time, I asked OpenClaw if it could enhance one of my web apps. The web app in question shows the next train, bus and tram schedules nearest to my home. So I thought, wouldn't it be nice that when I get out in Oerlikon, I could check what connections can bring me back home with a simple push of a button (on the web app). In that 25-minute train ride, I created a working version of exactly that - ON MY PHONE. It asks for GPS coordinates from my phone, looks for the nearest stations, knows the GPS coordinates of my home, and uses the public API to look for all connections that lead me home. Since then, I've "vibe-coded" this further and it's now really good. I didn't even do the deployment myself — OpenClaw deploys it.

What does it all mean?

  1. We are moving into the sphere of long-running AI tasks or processes; OpenClaw was the first salvo, and it appears everyone is now on board: Claude Code just released a loop function that can run at intervals with full prompts. Andrej Karpathy released autoresearch — it allows you to automatically train your own LLM and continuously improve on it while running overnight. Nvidia just announced the acquisition of nemoClaw, an OpenClaw competitor, and of course, we've known that the creator of OpenClaw is now working for OpenAI, so I would expect some of that functionality to make it into the next ChatGPT version.
  2. Everything you do at work can be much more efficient if you get access to these tools. In fact, I'm not the first to say this, but how we work will fundamentally change. Personally, I believe organisations will shift gears into more productivity by eliminating a lot of the overhead that comes with putting an idea into a workable product. Instead of presenting ideas, you can now present a prototype.
  3. The LLMs are good enough. I think I've made this point in the past — it's now all about scaffolding and harnesses and the right amount of context engineering. OpenClaw does a phenomenal job at this and I suspect there is a lot more still that we can get out of the existing models by being more efficient with context.

Food for thought. What's your opinion?

Comments

Post a Comment