Working on a blog post with a group of colleagues I discovered that writers fall into two camps. Those of us who learned to write on the internet wanted to write only what felt exciting, what our startup to say that readers haven’t heard before. Then my coworkers who learned to write for their PhD thesis made us delete anything that could possibly be nitpicked by a hostile reviewer, which was about two thirds. I was confused: Why would a hostile reader even bother to read some startup’s technical post? It didn’t matter: the hostile reader was in the academic writers.
In academia, “this point is interesting but how do you prove it” is a threat of vastly more work to be done, if not outright rejection. On the internet, that’s a 99th percentile outcome. 99% of people who you’d want to read your essay probably don’t know you exist, or don’t give a fuck, or wouldn’t be bothered to engage even if they read it. The academic reader is hostile but captive — they can fight you but they have to read you. The online reader is free. Free to love the essay, dunk on it, be transformed by it, moved to tears. And also free to close the tab at any moment.
To write for the free reader, they must also be in you — a simulation. You must imagine them at every turn asking, “why am I reading this?” At every turn, you should give them motivation and a promise.
Motivation is the structuring principle of an essay: every part contains the reason for its existence and its place in the sequence. It’s what separates an essay from a disordered braindump. Motivation is a feudal affair: the sentence serves the purpose of the paragraph, which serves the section, which serves the essay, which serves the goal of writing the damn thing in the first place. The purpose of a section could be: to explain or expand on what came before, to introduce and give a reason for what comes next, to fulfill an expectation and create a new one.
These expectations are the promises that a writer must make and keep. Every piece of the writing incurs a debt that is repaid when the reader is satisfied with the answer to why have I read this? Promise too little and the reader wanders off bored. Leave too many debts unpaid too long and the reader closes the tab in exasperation.
Why are you reading this? I imagine you’re hoping for some new ideas about writing and AI. I started with “two types of writers” to introduce the concept of a free reader by way of contrast and to raise the question of how you would write for such a reader. The section on motivation and promises gave an answer but left two debts unpaid: the advice was too general and needed an example, and also you expect this all to somehow tie to AI in the end. This paragraph is a demonstration both of “writing in promises” and of reader stimulation, the idea we started with and the one that brings us back to AI.
Last time we discussed that LLMs write to get a grade, not with the goal of moving the reader. The RL grader, whether human or machine, is not a free reader: they are captive and static. LLMs themselves are the perfect captive reader: they never get bored or confused, never miss a reference or have an emotional reaction, never ask themselves why am I reading this?, never close the tab. They can infer something about the free reader’s soul from examples of things written for such a reader, though these are a small fraction of the training corpus. They aren’t able yet to simulate such a reader, to imagine that reader saying ok Jacob, I get the point now, I hope this essay wraps up soon.
But I can.
PUTANUMONIT 