I feel like most of the AI-curious world has dived head-first into autonomous agent madness with Clawdbot Moltbot OpenClaw. I have been doing something much less ambitious, spending my weekend attempting to wrestle Codex into setting up AdGuardHome on a Raspberry Pi 5 and using Nix to make the reproducible builds.

This was really an excuse to have a play with all of these things, although I still would have liked a modern equivalent of a Pi-hole at the end of it. As it was, I ended up with nothing but learnings. This pretty much boils down to “if you find yourself waiting hours for Nix to compile things from scratch, maybe stop and have a think”.

The first place I went wrong was that I didn’t really think about how to spend my innovation tokens. I should have focussed on learning about Nix, or spent time finding the edges of working with Codex, or done the most basic thing to get AdGuardHome up and running. Not all three.

The second place stems from doing something I would never do in client work: I yolo’d straight into it without doing any upfront research or planning, trusting that the agent would see me right. I think even a little bit of thought would have saved my weekend, but the prompt/feedback/prompt loop was just too tempting and I ended up swallowing just one more spider to catch a fly from earlier in the session.

The contrast between what people are talking about in public and what I could achieve in private is pretty stark. Are people really so many miles ahead? Should I be… afraid? Maybe a little! I haven’t had a lot of time to play with AI on the code side. My excuse is that I’ve been too busy trying to wrangle humans. Frustrating, but maybe job security in this AI timeline comes from both coordination skills and an ability to reason about systems rather than code.

The other positive side is that this was so painful that I ended up rethinking the whole “put a DNS sinkhole on a box I look after” plan. Instead, my subconscious prodded me into simpler solutions and I ended up setting up NextDNS and using that as the DNS for my Tailnet. I have comprehensive adblocking on all the devices I care about, both in my house and when I’m out and about. It took me all of 15 minutes and I don’t have to babysit a Raspberry Pi. Yay, technology.

Anyway, OpenClaw. It does feel somewhat closer to the TNG-era Star Trek computer, which is pretty much all I’ve wanted from an assistant: solid natural language parsing and a set of defined (maybe very widely defined) actions that can be taken as a result. I’d settle for a version of Siri that did this, though.