The Trade-Off

Most writers online, I think, give up relatively soon. This is not because we ran out of things to do—there’s always another typo we could fix… But when we choose to spend our limited time and energy hunting for typos or fixing formatting, there’s a definite sense in which we decide it’s no longer ‘worthwhile’ trying to improve old writing after some point, and we’re better off doing new writing.

But this is a little odd: if we’re not ‘done’ with old writing, why should we start something new? And if we have a large backcatalogue, we can hardly plead poverty of material; how can I argue with a straight face that my readers lack for things to read and so it is imperative I go write something new rather than polish an old popular essay?

Marginal returns. The economics or optimal foraging answer would be the heuristic ‘you should switch between whichever has the highest marginal return at any given moment’—but marginal return of what? Even assuming we have some ‘utility’ definition we’re satisfied with, how, exactly, does writing turn into that? If I write an essay or I polish an old one, what does that mean for a reader?

A toy model that seems plausible to me is that what is going on is that each reader has a certain finite size or appetite or window of time, which a writer tries to fill up as best he can. (I’ll use time as my unit here instead of number of words or essays, because I think that increasingly, it is time and attention which is the bottleneck.) One reader has a window of 5 minutes of ‘Gwern reading time’ a year, while another has 50 minutes, and a big fan has 500… The absolute size of the backcatalogue doesn’t matter for each reader—only the best k minutes of your writings.

So, when you write, if a new piece doesn’t make it into the top-k and is only the k+1^th best reading-minute, then a reader doesn’t care. They’re never going to see it and it doesn’t exist for them.

This helps account for why writing is such a “hit-driven business”: after a while, no matter how prolific you are, it makes zero difference to the reader, because you filled up the top-k in their window and nothing budges after that.

Self-Cannibalization

To move the needle, you need to beat your best and push something out of their window. And frustratingly, once you do that, the reader is only slightly happier with you!

Net, not total. This is because you are cannibalizing yourself. If you wrote something which was a +10 in some Quality Unit, and then you write something which was +11, you feel proud of yourself for adding a +11 and thus writing a total +21 of text—that was hard work—but as far as the reader is concerned, it’s just a net +1 because your +11 pushed out the +10.

Perversely, the more you write, the harder it gets to improve your window. If a reader has a window of 10 essays and you’ve written 100 essays to date, then your next essay has to beat >90 of your prior essays to make it into the window at all!

Simply put, I think you should polish in inverse proportion to how much of your writing a reader will see; the more diverse your writing, the better off you are writing more rather than better.

Order Statistics

Rank entry. We can treat this statistically using order statistics; for example, if we imagine some normally-distributed quality statistic, our expected gain is only:

set.seed(2026-05-11)

windowGain <- function(backlog_n = 100, window_k = 10, simulations = 1e7) {
    mean(replicate(simulations, {
        old_quality <- rnorm(backlog_n)
        new_quality <- rnorm(1)
        cutoff <- sort(old_quality, decreasing = TRUE)[window_k]

        max(0, new_quality - cutoff)
    }))
}

windowGain()
# [1] 0.0474129394

~0.05σ better—which is almost invisible percentile-wise.

Equal odds. The simplification of treating pieces as independent samples is crude, but it resembles the equal-odds model summarized by Jung et al 2015: hit-rate is approximately stable, so quantity matters because it buys more rolls of the dice. For writing, this assumption is partly defended by offsetting forces: positive factors like experience are offset by negative ones like using up ideas.

Lumpy returns. More realistic distributions do not change the basic rank fact. With n old pieces and a k-piece window, a new independent piece enters the window with probability k⁄(n+1), distribution-free. A power law or other heavy-tailed quality distribution changes the payoff conditional on entry, not the entry probability: the process becomes more hit-driven, with long stretches of zero marginal value punctuated by rare large replacements.

Polish the Window

Guaranteed gain. This also adds a little insight to the value of polish: while it’s true fixing typos in your best writings doesn’t make them much better, it does make them a little better… And importantly, it’s making the writings in the window better. When you write a new one, you usually fail to replace something in the window, rendering it moot, and if you do, it’s probably a modest improvement; but when you edit the writings in the window, you are getting a small but guaranteed improvement.

Tiny windows. This sheds some light on why you might escalate your quality standards going from a blog post to, say, a New York Times article. (Jasmine Sun mentioned in a recent Q&A at Inkhaven 2 that her NYT piece on peptides might have 10 million readers total, vastly exceeding her lifetime total of Substack readers—so of course she spent months sweating every detail!)

Decay. Meanwhile, a breaking-events periodical newsletter like Zvi Mowshowitz’s will often not bother to fix typos or broken URLs, because the temporal decay guarantees even a good issue will soon drop out of the window, and it needs to rush to write the next one.

For the former, you are talking to a small steady audience with a large window; there’s little point in polishing it. But for the latter, you may get only one, and so the window is small—you’d better make it as good as you can!

Heterogeneity

So, to the extent that your readers are homogenous and have a small window and a universally shared ranking, the better off you are, after a while, trying to polish your writing winners, because you aren’t going to easily beat them but you can still make them a bit better.

But to the extent that they have large windows or heterogeneous different rankings (eg readers prefer completely different articles), you’re better off just writing more in the hopes of getting a new better one into their windows. The marginal return on every new essay may be quite low (because its topic will appeal to only a small fraction of readers) but it is also true that the return is even lower on polish (because those too are niche but the effort is constant). In the extreme case, every reader has a window size of 1 and prefers a different essay… So you are best off spending 100% of your time writing on a new topic every time, because each one will find a different audience.

In full generality, we could advise a writer to polish in proportion to:

And to write more in inverse proportion to all 4.

Eclectic Generalists

My case. For my writing, I think it falls into the latter group: I do not write for prestige or popular periodicals, but on many topics for different readers (eg darknets vs poetry vs AI vs genetics…). This also describes writers I am sometimes grouped with, like Scott Alexander or Tyler Cowen’s Marginal Revolution. We all share, if nothing else, a generalist eclectic approach. Under this model, Scott Alexander, who has a recognizable ‘beat’, should probably invest more in polish, while I should emphasize more diverse new writing rather than over-polishing, and Tyler Cowen is probably fine in his omnivorous niche of writing about anything with no polish.

Self-indulgence. So I should be writing a lot more on more topics, instead of falling into the artisanal/hobbyist/indie-creator trap of spending as much time as I do on polish and infrastructure. What can I say? I don’t write to be popular, and the polish is for me a personal esthetic and an ethos, not a calculated strategy. (I try not to pretend otherwise, and I tell any aspiring writer who asks that they definitely should not try to imitate Gwern.net but focus on writing. They have my permission to fiddle around with typography after they have, let’s say, a million words—at which point they may have some writing worth polishing.)

But other writers might find this a helpful frame: do your readers have a lot of appetite each, or a little? And are they homogenous or heterogeneous? This can help guide you in deciding how much to invest in copyediting, illustrations, marketing, etc.