Overlords

I, For One, Welcome our new Robot Overlords - a widely attributed meme

I’ll admit it, for a very, very, long time, I was an AI skeptic.

I used popular AI tools for dumb questions, got it to do trivial tasks, but never really trusted it.

I used a variety of models from a variety of providers, and tinkered a bit, and told myself I was being cautious and correct in thinking carefully before trusting these incredibly speedy and broadly incompetent robot idiots.

“It’s like an incredibly enthusiastic junior” I’d joke to work colleagues. We’d laugh as I recounted my most recent story of the latest and greatest AI assistant getting something wrong, and I felt lordy, and accomplished, and high and mighty.

I was doing it wrong.

In my long journey of being more keyboard-centric, I have taken the odd route of forcing myself to use keyboard tools more often. My mac, that bastion of modern interface design by the gods that are Apple (setting aside the hideous atrophied mess that is Mac OS Tahoe, may it burn in the fires of hell forever), is now… corrupted.

• I do more work in the terminal.
• I write these posts in Neovim.
• I installed and use Ghostty for all my terminal work - and I’m doing a lot lately.

I write posts in Neovim, save and quit using the common commands that are burned into my brain now (:wq), git add, git commit -m, git push, and let the magical little demons at cloudflare update my site. I even have a cron job on my server that pings the cloudflare api to trigger a rebuild at 9am every friday morning, so as long as i’ve pushed pages up to cloudflare, they will magically publish at the right time.

All done from the keyboard. (a bit of clicking in the cloudflare interface to set it up, but I don’t have to go back there now).

But today was my big milestone.

For years i’ve been telling myself that the AI tools are so untrustworthy that I can’t use them to code, and i’d never understand what they produced. I took courses, instead. Spent real money on courses, and - as is so frequent, and a reason why course sellers love them so much - never really built anything.

I promised myself I’d learn to code properly, and build a first step, before trusting the AI. Like it was an incompetent black box.

But all those videos by Theo and The Primeagen were chipping away at my resolve. I saw them ripping through code, building tools, using them every day, and I just couldn’t hold off any more. Hell, I’m an avid user of t3.chat now (Theo’s AI chat tool), with 3-5 long threads a day on anything that comes to mind.

I had ideas for tools. Like most old farts, i’d got used to my manual processes, and when called upon to produce a report, or catch up with what I was doing a few days ago, I’d rely on crappy spreadsheets and remembering where they were, keeping them up to date, and rigid collections of notes and file structures so I didn’t get lost.

Hint: I always got lost.

ADHD brain / short attention span / long embedded procrastination habits - call them what you will, but every morning was spent re-learning my complex structures of work. Asked to create a summary report? I’d be cobbling together data from all over the place and manually assembling it like some sort of neanderthal into reports that never looked the same, and which - if truth be told - I could never say were completely accurate.

I kept promising myself I’d build an app.

Well - I kept promising myself I’d build loads of apps - I’d started a list of them, and I never built any because I was still ’learning to code’. Finally the dam burst, and I created the nice, complex combination that unlocked enough interest to get me started.

• I installed tmux.
• I installed opencode.
• I archived all my ’testing code’ folders, and created a new one.

Then, I got cracking. Opencode used Claude Sonnet 4.5, and started building an app for me. Four hours later, I had an app - it wasn’t all that good (bad prompting on my part, to be fair), didn’t look all that good (personal taste, of course), and considering how long it took, didn’t have that many features, either.

These are mainly my own architecture problems - I forced it into using certain tools, and it fumbled. But I also just didn’t enjoy the experience. It asked me waaay too many questions, took ages because it kept stopping, and I just lost interest.

So I started again. deleted everything, simplified my decisions, and handed the reins to GPT Codex 5.2. Goddamn.

I spent one hour, four dollars, and I had an app with lots more features. I was absolutely blown away with the whole experience. Codex didn’t ask questions, didn’t bother with niceties, and just got stuck in. I wanted to think and plan everything through, but GPT always built a plan in one stage and was itching to implement it.

I’m absolutely going to use this again, and I am going to be building a lot more tools now; unshackled from learning to code, and instead letting a tool that is way more experienced than me just get on with it.

I’m also likely to be a tiresome bore in the future about this - but…

I, for one, welcome our new robot overloads.