It's different than passing someone's Intellectual Turing Test
Epistemic status: Others have found my take here useful, but I haven’t engaged with the entire debate on the subject. Fairly confident that it’s useful to me, though see caveat.
Eliezer Yudkowsky recently disclaimed steelmanning on the Lex Fredman podcast. I haven't listened — I heard it second-hand. Here's a thorough roundup of people saying steelmanning is bad.
This post will argue that steelmanning is a useful tool.
A bit of obvious background that is nevertheless worth saying: I am not omniscient. Other people know things and have models that will help me achieve my goals better. I want to steal from them like a greedy man. I want to do this because I desperately want to be right about the world. Because it matters.
Yet it’s sometimes legitimately hard to steal from my interlocutor. They might be building their argument up from a different set of starting models than I am. This will make them explain their ideas in a way that’s kinda weird for me. Let’s see an example.
Dogs and cats
Suppose I don’t like dogs that much. My friend notices I’m feeling sad and tells me I should get a dog.
I might think their idea is crazy at first blush. If I try to pass their Intellectual Turing Test, I might focus on things they said, about how joyful dogs are. I will find this pretty uncompelling. But if instead I focus on the part of their argument that seemed strongest to me, about how their dog lies next to them and makes the house feel less empty, then suddenly I have an insight about how I should get a cat.
Steelmanning vs ITT passing
ITTs are basically easier to understand than steelmans. They are very pure. Your goal in an ITT is to understand your interlocutor. For someone on the other side, it feels very good to have someone pass your ITT. You feel understood, obviously. You may also have them ask you, “Does that capture everything?”
Steelmanning is a more subtle concept. Most sources define it as “constructing the strongest possible version of” a view you disagree with. This begs the question: according to who? I claim: according to you, the steelman-er. It is you who are trying to improve your models. If your interlocutor doesn’t like your steelman, that’s a sign you might have more to learn from them. But there’s no fundamental error going on.
Play with your steelman
I claim you should have fun with your interlocutor’s views. Make them your own, as much as possible. Worry less about fidelity, and simply follow the ideas where they take you. Think about what’s actually interesting to you about them. At the extreme, completely ignore the parts that seem bonkers to you, and focus on the parts that seem right. Ok, maybe do also see if there’s anything interesting in the bonkers parts, but only for as long as you feel like it.
Be clear about which thing is going on
If someone tells you they’re steelmanning you, don’t expect them to be perfectly representing your views. Think of it like a fanfiction of your idea. Feel free to talk to them about the ways you disagree with their steelman, but don’t expect an ITT.
When you’re talking to other people, if what you’ve done is a steelman, be clear that you can’t pass someone’s ITT, and instead this is your own spin.
When might you steelman vs ITT?
While live in your conversation, I would recommend adopting a more ITT frame of mind. Your interlocutor will appreciate it, and so you will get more out of the conversation. But afterwards, in the privacy of your own mind / google doc / conversation with your friend, go wild with your steelman. As a reviewer of this doc said, “ITT in the streets, steelmanning in the sheets.” Steelmans are what ultimately help me steal from people, if they’re starting from a different place from me.
That said, ITTs can lead to a stronger understanding of the other side’s position, which could help you steal from them as well. ITTs can make your eventual steelman stronger.
Because ITTs are harder, I often stop at doing a steelman. If I was less of a steelman-stan, maybe I would do more of the ITT mental motion, which would result in more overall stealing knowledge and models from people.
Thanks to Justis Mills for reviewing a draft of this post.
This is false, to be clear
Here this post switches from a fair descriptive to a fairly prescriptive frame. I’d like to have flagged that somewhere other than a footnote, but oh well.