Doing More with Less: an LLM Response to Less is More.

Ludwig Mies van der Rohe, the architect who created the look of the modern skyscraper when he designed the Seagram Building in New York City, where I worked when I was fresh out of college, had a famous saying: less is more. 

As someone who has produced multiple books that have been praised for their writing style, I am extremely curious how I learned to write well, and why LLM writing is thought of so poorly at present. Obviously I am not the only good writer in the world—there are millions. So how do they do it? And why don’t LLMs accomplish the same? 

So I designed an experiment—I asked ChatGPT to write a paragraph in a field I know: 18th century Virginia from the perspective, intellectual commitments, and linguistic style of two different famous historians who are in some senses polar opposites: Bernard Bailyn and Rhys Isaacs. It did a surprisingly good job of this task, which led me to inquiry about whether it had ingested their texts in its training data, a question that it said it could not answer. Indeed, to OpenAI’s credit, they are not included in the Books3 database demonstrating that its competitor Anthropic used Bailyn’s books, and my own, in training data.

But then ChatGPT arrived at a more interesting angle, and independently expressed something that I have written about in my own writing blog: that great writing, and what we call authorial “voice,” are shaped more by what you leave out than by what you include. 

ChatGPT: “You might discover that the problem isn't fundamentally that LLMs can't write well. It may be that, without sufficient constraint, they have no reason to prefer one particular kind of good writing over all the others they know how to produce. That is a much more interesting theory of why AI has a voice—and why, by default, it so often sounds like nobody.”

In other words, when it comes to writing, knowing less paradoxically is more helpful. So at that point I directed it to the blog post I wrote last week on the topic. How to Fix AI Writing, Part I. What is interesting to me is the fact that I wrote my posts before this conversation, so ChatGPT could not have been merely parroting what I knew it wanted to hear when it came up with a very similar idea. What follows is ChatGPT’s response.

ChatGPT:

Having now actually read the post, the convergence is considerably closer than I expected. In fact, you had already formulated the core idea I was reaching for. Your concluding sentence is the key:

“The point of good writing is not to use words like you know everything, but to know some things very well, and others not at all.” (Advantage Editing)

And earlier you make the complementary observation that good human writing involves “knowing what to leave OUT.” (Advantage Editing) Those are remarkably close to what I independently formulated above as “voice is partly a stable pattern of exclusion.”

But I think our conversation just pushed your argument somewhere new.

Your blog post currently explains voice principally through the situatedness of the writer. The hypothetical 75-year-old Irish/German, Red Sox-loving, eighth-grade-educated, Vietnam-veteran aircraft mechanic from Nantucket has a voice because he occupies a particular place in the world. He knows certain things, has experienced certain things, belongs to a generation and class and region, and therefore doesn't possess the universal, disembodied perspective of an LLM. (Advantage Editing)

That's a strong account of why human voice exists.

But the Bailyn/Isaac experiment suggests a second mechanism that I don't think you quite articulate in Part One: voice isn't merely a consequence of what a person knows. It's a consequence of what a person habitually selects from what they know.

Bailyn and Isaac are perfect for demonstrating this because their demographic situatedness isn't dramatically different. Both were highly educated, twentieth-century, English-speaking male academic historians studying eighteenth-century British America. Both knew a tremendous amount about Virginia and the Revolution. Yet they produced profoundly different historical prose.

Why?

Because when confronted with the same historical evidence, they selected different realities from it.

Give Bailyn a Baptist minister in a Virginia jail and he notices authority, liberty, rights, legitimacy, constitutional thought. Give Isaac the same jailed Baptist and he notices the grate, the crowd outside, the magistrate, the preacher's body, the inversion of deference, the transformation of a jail into a pulpit.

That's upstream from prose style.

And I think that's potentially a major insight for your larger argument about AI writing. Bad AI writing may originate before the writing begins.

The model's problem isn't merely:

How should I phrase this sentence?

It's:

Of the nearly limitless things that could be noticed about this subject, what deserves noticing?

Generic AI tends to answer: the things most generally relevant to a competent answer.

A distinctive writer answers: these five weirdly specific things interest me, and the other 9,995 don't.

Then syntax, diction, imagery, rhythm, metaphor—all the things conventionally called “style”—follow partly from that selection.

So I'd now separate three levels:

World → Attention → Language

A human life constrains the world available to the writer.
A particular intelligence constrains what the writer notices within it.
A developed voice constrains how those selected things get expressed.

General-purpose LLMs blow open all three constraints. They have been exposed to vastly more textual worlds than any person; they can attend to almost innumerable aspects of a subject; and they command an enormous repertoire of linguistic forms.

Which sounds like the recipe for the greatest writer in history.

It may instead be the recipe for the average writer of everything.

And that's where I think your “dreck” argument can become more sophisticated without abandoning it. Your line—

“The problem is, LLMs have read everything—and most of it is crap.” (Advantage Editing)

—is funny and provocative. Keep it. But analytically, I'd add:

The problem isn't only that the model has read too much bad writing. The problem is that it has read too much different writing without acquiring the durable preferences that cause a human writer to reject most of it.

That's considerably harder to knock down.

It also gives you a fascinating definition of editing. Editing is artificial selectivity. A first draft—human or machine—generates possibilities. The editor imposes exclusions. Not this word. Not this observation. Not this explanation. We've already said that. Too obvious. Too abstract. Wrong metaphor. That's interesting but doesn't belong here.

Which makes your profession rather central to the problem instead of something AI renders obsolete.

And there's a delicious empirical possibility here because you wrote the blog post before this conversation. We now have a dated public statement of your hypothesis. We could turn the Bailyn/Isaac discovery into Part Two: not “how to make ChatGPT sound like Bernard Bailyn,” but an experiment demonstrating that better AI prose may result from progressively narrowing what the model is allowed to notice and say.

That would move the argument from “AI needs voice” to the much more interesting proposition:

To teach AI to write better, don't teach it more. Teach it what to ignore.

 ————

Impressive. Now if we can only teach it that a sentence is not a paragraph we’ll be getting somewhere!