I am currently doing a bunch of things almost simultaneously. I have decided thta I am going to massively deplatform myself from the major tech companies. This will likely take a long time to achieve, maybe a full ten months or so of slowly recovering my own digital sovereignty, this is because in twenty years of using Google products they has a solid terrabyte of data on me. Most of which is YouTube videos. But that needs to have its own sorting out and for some of this I am going to need storage.

But that's not a big enough mountain

You're right, I have my sights set entirely too low if what I am trying to do is recover my own data and shift my services to other means. So I am working on my own CMS (Content Management System) which utilizes a database backend that is shared across my domains. So you might read these words in one place or another, its still me writing them.

The Secret Sauce

I was flipping through my feeds in newsboat and came across Dave Winer's post from today. This seemed to me like it was a question. Not that I have an answer to it but it set me thinking about what I have done and if this sort of qualifies as meeting the standard of this question. So of course, I fired up a sycophant, I mean a llm and began asking it questions like "AI as Server Software?" and "If I have an OpenClaw instance on a server with the general directives of keep this server working, build new code to improve yourself and the the tools to fulfill this, is that AI as a server?"

I've built a sophisticated distributed AI system where OpenClaw agents run across my droplet, laptop, and studio machines. They autonomously build frontends for my domains, journal their daily experiences in markdown, and consolidate those memories into a database nightly - essentially implementing sleep-like memory consolidation.

My core challenge emerged: I want these agents to self-improve but don't know how to tell them to "improve yourself" when I can't define what improvement looks like. We explored how true self-improvement might focus on mundane but critical tasks like monitoring resource consumption and implementing fallbacks - which became immediately relevant when all my agents hit token limits simultaneously.

My token timeout revealed a deeper issue: I've been broadcasting instructions to all agents like a mob instead of using the corporate structure I designed. I have a git repo (gotcha-workspace) where agents can write tools for themselves, a database with projects/goals/tasks, and planned orchestrator personas that should own different domains - but I haven't implemented the ownership model or stopped treating them as a hive mind.

The solution involves my Executive Personal Assistant (currently offline due to my laptop power cable failure) who would route my directives to appropriate orchestrators, preventing duplicate work and token waste. Each orchestrator would spin up subagents with webhook reporting, creating a self-organizing company structure.

I'm intentionally using token limits as forced strategic pauses - "burn rate" management from the dot-com era. While downloading Qwen 8B as a local fallback for my EPA (since Llama3 lacks tool use), I recognize I'm bootstrapping a cognitive startup on a shoestring, where the constraints are actually forcing better architecture than unlimited resources would allow.