Brando Miranda — September 2026 · ~4 min read

TL;DR. Someone asked me on Instagram what my “technique” is for doing good research. The honest answer is I don’t have one — I just work consistently, exercise, keep good relationships, and work with people I enjoy on things I actually care about. The best anecdote I have for that is Mirage: my friend Rylan and I argued about emergent abilities in LLMs for months without ever agreeing we were writing a paper. I found out it was a paper when I saw my name on the draft.

A few people asked me the same question in the comments of a post about the Mirage paper — “Are Emergent Abilities of Large Language Models a Mirage?” — and it’s a fair question: what’s your technique for doing good research? I tried to give an honest answer instead of a generic one, and the honest answer is boring: work consistently, exercise daily, keep good relationships, work with people who are better than you so you can learn from them, and work on things your heart actually cares about. There’s no trick in there.

But there’s an anecdote that explains it better than the list does.

Rylan and I didn’t set out to write Mirage. We didn’t even know we were writing a paper. It started the way most of our conversations start — he brought something up, I told him he was wrong, and we argued about it. This time the thing he brought up was emergence: the idea, all over the literature at the time, that large language models suddenly acquire new abilities at some model size, as if a switch flips. I thought that was suspicious. He thought I was missing something. We went back and forth on it basically every day for months — on the couch, over text, whenever one of us thought of a new angle.

Somewhere in there the arguing turned into writing things down so we could stop repeating ourselves, and the writing turned into a wall covered in notes, plots, and half-finished claims we were trying to pin down. Neither of us was tracking that as “we are now writing a paper.” We were just trying to win the argument.

I found out it was a paper when I saw a draft with my name on it. I remember being genuinely surprised. And the funny part is that the disagreement resolved in the opposite direction from how it started — I was arguing that he was wrong, and it turned out I was the one who was wrong. Emergence, in a lot of the cases people were pointing to, wasn’t a property of the model quietly waking up. It was an artifact of the metric you chose to score it with — pick a discontinuous metric and you get a dramatic jump, pick a smooth one and the same underlying model traces a smooth, predictable curve. (I’ve written up the technical version of that argument elsewhere — the bit flip, the figure, the six-sentence structure — for anyone who wants that instead of the personal story.)

A friend, César, asked me about this over DM once he’d skimmed the arXiv link: how do you not know you’re writing a paper? I sent him a voice memo, and the gist of what I told him was this — when you’re younger in your career, you’re more naive about it, in a good way. You’re doing your PhD, you’re studying, you’re talking about research in general, and you don’t even know yet what counts as “research” and what doesn’t. So you’re just in an app, your best friend brings something up, you tell him no, he tells you yes, and you argue about it every day. And after a few months it turned into a paper, and honestly I didn’t even know I was making one — I was just having fun.

That’s the whole technique, if there is one. My best papers and my best memories both come from the same place: doing work with people I genuinely enjoy working with, on things that actually bother us, without either of us managing the process toward a deliverable. The paper was a side effect of a good argument between friends. I’d take that trade every time.

If you’d like to cite this post:

@misc{miranda2026miragestory,
  author = {Miranda, Brando},
  title  = {We Didn't Know We Were Writing a Paper: How Mirage Happened},
  year   = {2026},
  month  = {September},
  howpublished = {\url{https://cs.stanford.edu/people/brando9/2026/09/26/how-the-mirage-paper-happened.html}},
  note   = {Blog post}
}