FILE · 05.002 NOTE · 3MIN --:--:--
// writing / 002 · note
05file · 0202
typenote
read3 min
words580
filedapr 2026
revv1.0

Notes after reading one AI paper twice.

The first pass is for showing off. The second pass is for the thing the paper is actually saying. They are rarely the same paper.

I read the same paper twice this week, four days apart, and the second one was a different paper. Not because the words changed — because I had, slightly, in the interval. I want to write down what shifted, before it un-shifts.

01First pass.

On the first read I was looking, mostly, for the parts I could quote. The diagram on page 4. The number in the abstract. The one-liner I could drop in a Discord thread and feel slightly older than I am. I underlined a lot. I felt smart, in that specific way that means I haven't actually understood anything yet — the way you feel after watching a film with the subtitles on, when you've parsed every line and absorbed nothing.

02Between passes.

Then I tried to describe the paper to a friend who hadn't read it. I got through the setup. I got through what the authors claimed. I could not, for the life of me, explain why the result was supposed to be surprising. Which means I hadn't read it. I'd parsed it.

// heuristic If you cannot say what the paper would have predicted if it were wrong, you have not read the paper. You have watched the paper happen.

03Second pass.

The second read I went in with one question — what would have been the boring version of this result, the one that wouldn't have been worth publishing? — and almost everything in the paper rearranged itself around that question. The architecture choice the authors spent half a page defending? It existed entirely to rule out the boring version. The ablation that had felt decorative? It was the whole point.

I want to be careful not to make this sound like a method. It wasn't a method. It was just asking, slowly, what would I have expected, if I'd never seen this? and then noticing where the paper had quietly anticipated me.

04What I kept.

  • The point of the first read is to find out which paper you're going to read on the second one.
  • If a paper has nothing to argue against, it usually has nothing to argue.
  • The interesting part is almost never in the abstract. The abstract is the version that goes on a slide.
  • I take better notes after closing the PDF. Whatever I remember without looking is the part that landed.

I think I'm going to start reading every paper I actually care about twice, on purpose, with a day in between. The first pass is to find out what the paper looks like. The second pass is for what it says.

end of file · 05.002
// filed under
paperslearningaimethodnote
all writing