Codex Micro and Knowing Where to Look

I bought a Codex Micro. It’s a little keyboard-like companion for Codex, the OpenAI app that does all sorts of things for you.

It has a knob you can twiddle, and when a thread finishes, its corresponding key lights up green. It’s exactly the sort of gadget that delights the boy in me. Best of all are the unusually large microphone button and the send button beside it. Hold down the microphone button and talk, and your words turn into text. Press send, and Codex springs reliably into action. Doing all of that with satisfyingly clicky physical buttons makes me feel like I’m sitting in the command room of the Ultra Guard.

There is something surprising about giving a physical form to an interaction that normally stays trapped behind a screen. It changes the whole experience. Expensive, but a good purchase.

So, to get used to Codex, I’ve been having it make all sorts of things I want. It has turned English technical articles I found on Twitter into clear Japanese reference pages for my own use. I have also had it build a small-scale shogi AI that can train and run locally.

When I already have some feel for a field, I can give reasonably good instructions. I can respond to what is taking shape along the way, and the result usually becomes something I’m fairly happy with. But when I know nothing about a field and only know what I want, things can get pretty grim. The shogi AI is the clearest example.

Codex gives me a work plan, but all I can do is bluff and say, “Just give it a try.” While it works, I stare at a screen full of what might as well be incantations and wait. Without any instinct for where to look, I have no choice but to go along with whatever it says. The result feels a bit off. But I don’t even know what to fix.

I don’t need to be an expert. It is enough to know where to give instructions and what to check, or to have enough experience in a neighboring field to think, “This should be possible.” That gives me some confidence as I go. If I don’t want the work to become a black box, I probably have no choice but to broaden and deepen the things I understand. That is disappointing, but also a little reassuring. It is a strange feeling.

In the not-too-distant future, AI will probably develop those instincts too. For now, though, experience and learning still seem likely to pay off. I prefer it that way.