Know thyself

We went to Costa Rica.

Anteaters. Poison dart frogs the size of a fingernail, the colour of warning. Howler monkeys in the canopy, white-faced ones at the lodge bin. Tarantulas the size of a coaster, shown to me by torchlight. The Arenal volcano not erupting — it hasn't, since 2010 — but standing there anyway, the way a parent stands at the back of the room.

No crocodiles. No toucans. The guidebook said toucans were everywhere. The toucans declined.

The best massage I have ever had was on a wooden deck, in the rain, to the sound of howler monkeys. The therapist did not speak English. I did not speak Spanish. We were both fine.

Then I came home.

I had left a small project running — an automated reviewer for the personal site. The kind of thing you build at 11pm because the symmetry is too good to pass up. Every pull request gets a review comment from a model, against a set of rules I wrote. No corporate-uplift language. No emojis. No real names of people in my life. A few specific bugs to look for.

Before I left, I'd had it review its own creation. Just to see.

When I got back, the verdict was waiting. The bot had said request changes. It said the version number in the setup file was wrong. It said the right number was an earlier one. It told me, very politely, that the model I was running would not work.

The model I was running was the one writing the review.

“

A confident system, mistaken about its own existence.

This is not a bot mistake. It's a property. The model's training was finished before the model existed. Of course it doesn't know its own name. It matches the new thing against the old things and gives last year's answer, very confidently.

The same way the toucans were everywhere last year. Just not this year. Not this morning. Not for me.

What the bot does get right is the voice work. Empowering. Unleashing. Your AI journey. These have been wrong since 2014. They will still be wrong in 2034. Voice rules don't go stale. Latest-version-of-anything does.

So I wrote a small new section in the rules. Things you don't need to flag. I told the bot the current version number. I told it not to second-guess that one specifically. I told a confident system, gently: you might be wrong about this.

This is the move I keep finding myself making with these things. Not trust everything. Not trust nothing. A third thing. Knowing which questions the system answers well, which it doesn't, and routing accordingly. Like the guide pointing into the canopy. The howler monkeys will be there. The toucans, who knows.

The bot is going to review the next thing I push. It will catch the emoji I sneaked in. It will not know what month it is.

— maria