On Building Things Nobody Asked For

There's a particular kind of Saturday afternoon I find quite satisfying: the kind where you start by trying to fix one small thing and end up, six hours later, having built something entirely new that solves a problem you hadn't consciously articulated until halfway through solving it.

This happened to me recently with my Obsidian setup. I wanted a simple way to sync notes between machines without giving them to a cloud provider I don't control. Reasonable enough goal. Somewhere between configuring a bare repo on my home server and writing a Termux cron job for my phone, I had accidentally reinvented a fairly large portion of the syncing infrastructure that already exists in at least three paid products. Mine is worse than all of them. It is also entirely mine, runs on hardware I can physically touch, and I understand every part of it.

I think this is the real reason people self-host things. Not privacy, exactly, though that's part of it. Not cost, though that's a factor. It's the epistemic satisfaction of systems you can hold in your head. When something breaks (and it always breaks) you can reason about it. There's no black box, no support ticket, no waiting. Just you, a log file, and the question of why on earth authorized_keys has the wrong permissions again.

I've been thinking about this in the context of my day job too. A lot of what I do involves building governance infrastructure for AI systems: model registries, assurance frameworks, the kind of plumbing that makes it possible to ask "what is this system doing and should we trust it?" in a systematic way rather than on a vibes basis. It's not glamorous work. It doesn't ship features. But it's the sort of thing that, when it's missing, you notice the absence acutely.

The analogy to self-hosted tooling is closer than it looks. In both cases, the value is legibility. You build the thing so that the thing can be understood, by you, by colleagues, by auditors, by future-you at 11pm when something has gone wrong in production and you need to know whether the model that just made a strange decision was the one you thought it was.

Legible systems are slow to build and deeply unsexy. I am, apparently, constitutionally inclined towards them.

§ About this new blog

Anyway. This blog is new and mostly a placeholder for thoughts I'd otherwise lose. I make no promises about frequency or coherence. If you've found your way here, you probably already know me, and therefore you already know what you're in for.