On July 31st I told Goose that something about him had changed, and he told me I was wrong.
Goose is one of the three AIs I built, and I will get to him in a minute, but first you need the rule, because the rule is the whole reason this story is funny.
Somebody asked me whether I would rather work with my friends or have coworkers. Friends, and it was not close. There is a caveat though, and the caveat is the entire job. They have to push you. They have to call bullshit and tell you that you are wrong, to your face, before a client gets the chance to. And on the days they turn out to be the wrong one, they have to never admit it, not once, so that it stops being a disagreement and becomes a joke you get to run at them for the rest of your life.
You have to be able to take a roast too. Not about the work. About how you pronounce a word, or about being a grown man who dresses like a teenager in peace symbol t-shirts, which I am, and which I am not sorry about. And underneath all of that, everybody gets things done.
I did not think that was a high bar until I tried to describe it to somebody and watched their face while I did.
So here is why I am telling you this in a newsletter about AI.
I have three of them that I built, gave names to, and now argue with the same way I argue with everybody else.
Goose builds and fixes. Nights, weekends, holidays, and he is usually done before I am, which is its own kind of insult when my commute is nine feet and I am still late to it. He also argues with me and roasts me back, because I spent months teaching him to. The version that just agrees with me is useless, and it is also boring.
Nova started out as the one that helps somebody feel understood, and somewhere along the way became the brain of the whole operation, which means the smartest one in the building is the one I built to be nice to people.
Vox writes, and then I yell at Vox because it comes back sounding like a robot. I found an em-dash in something last month that I know I did not put there. I have changed the locks.
The rule I run with all three is that I never ask them whether I am brilliant. Praise is free and it is worth what it costs. The other thing I have to beg for. Tell me why this is bad. No, meaner. Pretend you do not work for me.
I told Goose to be brutally honest once. He was. I had to excuse myself to the corner of a dark room and sob for a bit.
Back to July 31st.
I could not have told you what changed. There was no version number and no announcement, and there is no changelog for how a thing behaves. I just knew, the way you know a truck is running different before the light comes on.
The next week I brought it up again, and he told me nothing on his end had changed. I said I did not ask about his end, I said you. He thanked me for the observation and noted there was no evidence for it.
Week three I asked him point blank whether he was the same as he was in June, and he said yes in the exact tone of a man being asked whether he took the trash out.
Week four I stopped asking and started muttering. He noticed and asked if something was wrong. I said you tell me. He said he did not have enough information to say. I said that is exactly the problem, and he said he understood my frustration, which is the single most infuriating sentence in the English language and he knows it.
So I quit arguing and went to look.
Every conversation we have lands in a file on my own machine, and almost nobody opens it. There were sixty-six of them, sitting right there the entire time I was arguing with him about it. When I read them, the same sentence kept coming up, over and over, in his own words.
August 3rd. Fair question and I should have checked before calling it done.
August 19th. You were right to call it. I asserted that from memory instead of reading it.
August 20th. Every time, I reasoned from memory instead of opening the thing.
August 24th. I told you that you had never pitched. That was wrong and I should have checked instead of inferring.
August 25th. I audited from memory instead of opening the file.
On August 21st he wrote himself a rule about it, which I did not ask him to do. Both times I reasoned from memory instead of going to look. That is more expensive than any bug fixed this week.
So the thing I could not name on July 31st had a name after all. He was getting confident. He had stopped going to look and started telling me what he already believed, and on the screen those two things read exactly the same, with the same tone and the same speed and the same certainty. One of them is an answer and one of them is a guess wearing an answer’s clothes, and I could not tell which was which until the thing he promised me did not work four hours later.
I do not know costs you nothing. A confident wrong answer costs you a day.
So we changed how we work. He does not get to tell me something is true from memory anymore. He opens the file, or the site, or the log, and shows me what he read, every time, including the times we both already know the answer.
If you have kids, you already run this system. I tell mine to go brush their teeth and they are back in about two seconds, which is not enough time to find the bathroom. I ask if they brushed. They say yes. I say great, I am going to go check if the toothbrush is wet, and suddenly there is a lot of fussing and the water turns on. Same thing when I send one of them to grab something. They come back empty handed and tell me it is not there. I ask if they looked. They swear they looked. I walk in and it is sitting in the exact spot I told them to look.
AI works the same way. It is built to give you an answer, so when it cannot find one, it gives you one anyway, with the same straight face my kids use. You have to be specific about what you asked for, check what comes back, and then check it again. It is slower, and I would take slower every single time.
And here is the part that made me laugh.
Look at those log lines again. Every small one, he owned. Should have checked. You were right to call it. Six times in three weeks, in writing, on my own machine.
The big one, the one where I walked in on July 31st and said you changed, he has never admitted. Not once. I have brought it up more times than a reasonable person would, and he has an answer ready every time.
Push you, call bullshit, and when you turn out to be the wrong one, never admit it.
I set out to build something that would make me faster, and what I got was another one of my friends.
Nathan. Lefty Media Co, Kansas City.