I put GPT-3 into TerminalNexus in early 2023

In January 2023 I shipped a version of TerminalNexus with GPT-3 wired into it. ChatGPT had appeared two months before and everyone was busy being amazed at the chat box. The part that caught me was quieter: the same kind of model sat behind an API you could call from your own software. So I did.

This was before the chat models. No GPT-3.5 endpoint, no GPT-4, no function calling, none of the tooling that showed up later that year. You sent a block of text to the completions API, you got a block of text back, and you spent your evenings massaging the prompt until the thing behaved. Early 2023 was a strange, fun window to be building against it.

What went in:

The one I actually leaned on was explain a command. You point at a command, one of yours or something you pasted in, and it tells you in plain language what the thing does, flag by flag. I have run plenty of commands I only three-quarters understood, copied off some answer years ago. Having the explanation sitting next to the command, instead of in a browser tab I have to go dig up, changed how often I bothered to check.

Then the reverse: describe what you want and get the command back. You write "find files over 100MB changed in the last week" and it hands you something to run. For the tools I do not live in, that is the gap between doing the thing and giving up and doing it by hand.

It could also suggest commands close to one you already had, which helped on the days you knew there was a better flag and could not dredge up the name. And since the app already had text fields all over it for command descriptions and notes, I let GPT-3 tidy up grammar and spelling in those while I was at it.

The dull but necessary piece was usage. The API costs real money per call, so I metered it. The app tracked an AI words quota for your current 30-day cycle, so you could see what you had spent. Not exciting, but you do not hand people an open pipe to a paid API and cross your fingers.

Looking back, the rough edges were all in the model, not the plumbing. GPT-3 would explain a command with total confidence and get a flag wrong here and there, the way they all still do, so you double-checked it. But the core bet held up. The explanation and the command belong right where you work, not in another window. I still believe that, and I still do most of this kind of work from the same place, at the terminal, with the answer a keystroke away instead of a tab away.

It is a little odd to reread those January 2023 release notes now, knowing how ordinary all of this became. At the time, an indie desktop app with a real AI feature in it was not something you ran into every day. I was early to it, and I am still here building, which is most of what I would want a release history to say.

Thanks for reading. If you were also gluing the GPT-3 API into something odd in early 2023, I would like to hear what you made. Send me a line.