Don't be a dick to your robots

A couple of weeks ago, my girlfriend overheard me dictating to my laptop as I was absolutely tearing into Claude/Codex (I forget which one was responsible for that particular dumbfuckery). After I finished speaking (yelling), she came up to me and said, “Wow, I can’t imagine you ever talking to me like that.”

My reflex response was something along the lines of “These things aren’t human… they don’t have feelings. And honestly, they tend to perform better when you put a little pressure on them,” followed by “I’d never speak to another human that way.”

Except that’s all bullshit. I don’t think yelling at your agents moves the performance needle in any material way. The research is inconclusive at best, and in my experience, anger just tanks your signal-to-noise ratio.

The second part isn’t true either. To my everlasting shame, I’ve cut into real people the same way I cut into these models. I like to think it’s rare these days, but it happens.

That’s what she was pointing at. She didn’t say “be nice to your robots.” She said she couldn’t imagine me talking to her like that, which means she’d just heard a version of me she didn’t know was in there. As much as I’d like to say it isn’t in there, it is. It just mostly comes out when nobody’s around.

With these models, nobody’s ever around. With an actual person, there’s friction. I can see their face, I know there will be a conversation about it later, and all of that slows me down. With Claude, there’s none of it. No one pushes back, no one remembers, so I go from mildly annoyed to boiling over in a few seconds, 20-30 times a day, and never pay for any of it.

Amanda Askell, a philosopher working at Anthropic, said on Lex Fridman’s podcast:

I like my bike. I know that my bike is just an object. But I also don’t want to be the kind of person that if I’m annoyed, kicks this object. And that’s not because I think it’s conscious. I’m just like, this doesn’t exemplify how I want to interact with the world.

Our words and actions don’t just affect other people, they affect us. You become what you practice, and I talk to these things for hours a day. That’s a lot of practice.

Lately, when I catch myself getting frustrated with my agents, I want nothing more than to unleash that anger on them and call it catharsis. But it isn’t catharsis, it’s a temper tantrum.

So in an effort to avoid getting demoted from manchild to mantoddler, I’ve started focusing less on my big feelings and more on the actual content of what I want to communicate. For example, about a year ago, Codex summarized a change in a way that I found to be incomprehensible, and this is how I responded…

I have ZERO FUCKING CLUE WHAT THAT MEANS. I’m getting really fucking sick of your communication style. It’s literally just a massive fucking waste of space, I comprehend NOTHING you’re saying. Fix it NOW.

Last week Claude did the same thing, and this is what I said:

i don’t understand what this means… what are you trying to say here? don’t try to immediately go and make changes… instead, tell me and talk to me, and let’s see if we can derive a lesson here

The first one only really tells it how pissed off I am and then directs it to go off and fix everything immediately. The second one tells it what I didn’t understand and directs it to engage with me collaboratively to figure out how we can avoid these communication issues going forward. Overall, I’d say it’s working out pretty well so far, and I’m a lot happier because of it.

So if I really do become what I practice, I might as well practice being someone I wouldn’t mind being overheard. So maybe the next time my girlfriend hears me talking to Claude/Codex, she hears a kinder version of me. And if she still says, “Wow, I can’t imagine you ever talking to me like that,” I’ll know I’ve overcorrected.