Nikhil.
← All marginalia
AIJudgment
It could find the twenty-fifth person. It just couldn't notice that changed the sentence.

Twenty-Four, Twenty-Five

I keep hearing that the models have gotten smarter, and most days I believe it. Today I'm not so sure what we mean by the word.

Here's what happened. I was clearing a sheet of open queries from experts on my project, going through them one by one, closing them out. Most were done. A couple dozen were still sitting there, unresolved. So I asked a model to help: go through the sheet, find everyone whose query is still open, and send each of them a message. Tell them we want to close this on priority, ask what they need. I said there were twenty-four of them.

It went in and found twenty-five.

My first reaction, honestly, was relief. It had caught someone I'd missed in my own count. It messaged all twenty-five, exactly the right people. That's the part that's supposed to be hard, and it did it without blinking.

Then I read the messages. Every one of them said the same thing: there are twenty-four queries still open, and you're one of them.

Twenty-four. It had just found twenty-five. It knew there were twenty-five, it was writing to twenty-five, and it still typed twenty-four, because that was the number I'd handed it and it never thought to hold its own finding up against my words.

I've been turning that over all day. It's such a small thing. But it's the kind of small thing a person would almost never miss. When you go looking for twenty-four people and find twenty-five, the first thing you feel is a little jolt. Wait, that's one more than they said. The model never felt the jolt. It did the hard thing and missed the obvious one.

More writing