Elliot

i may start doing bedtime reading livestreams on tiktok. i really enjoyed doing that tonight.

entirely fitting

Who’s a Better Writer: A.I. or Humans? Take Our Quiz. A.I. chatbots contain the sum of all human knowledge. That can make them pretty good writers.

Anyway, this thread was inspired by this NYTimes quiz that asks you whose writing is better.

As humans (1) is inescapable for us, and basically impossible for a generically trained chatbot to achieve short of becoming a "person". (2) is something that agents could probably get to with sufficiently large context windows and enough coaching, but it would be obscenely expensive.

(2) a short term memory sufficiently deep to be able to hold an entire piece of text as well as all the evidence and influences, the imagined reader's reaction to it, and the emotional weight of each piece of it, in mind together.

I suspect that the limitations of AI with respect to writing good prose have a lot to do with its lack of both (1) a contingently and particularly formed longterm memory that lends a specific perspective from which it can speak about or imagine things, as well as its lack of

In any case, watching a book I'd thought up during a thirty minute session rambling into the voice memo app on my phone materialize in this way was very entertaining. I will never publish it, I wouldn't trust it on the whole, but it was a fun exercise.

The other one is "Ockham’s conclusion was not reckless." This sort of thing also shows up a lot. I can't explain why it's so offensive, but it is. Perhaps because, like "the escape was real", it's completely uninformative.

Mostly it's pretty decent. It kind of botches the explanation of why the poverty dispute was so important, but it gets all the basics right, the sentences are pretty good, etc. Two weird things stick out: First, "The escape was real". This is a formulation Claude really loves for some reason.

But you can still see the weird habits of the AI slip through periodically. Take a look at this (completely AI-generated) page from a section on William of Ockham.

The result was much better! Sentence structure increased in complexity, bland parallelisms weren't littering every paragraph, and the text did actually flow decently through each chapter.

After this I spent a while talking through the intellectual perspective it should inhabit as putative author of this book, and what sort of relationship the author should have with the reader, what sort of pedagogical philosophy it should work from, etc. Then I had it re-write the book from scratch.

So I gave it a long booklist, to represent the authors it should have its voice influenced by. I talked through which authors should have the most influence and in which ways. I told it to model its literary style on Waugh and Mann, and so on.

Finally I took a totally different approach. It occurred to me that one of the things that could give the text cohesion was to give it an intellectual background and to model its own writing style on the idea of a person formed by a specific set of philosophical and literary texts.

The revision guide was not much more successful than the style guide. It got the agent to remove specific offending patterns that were identified in the edits, but the narrowness of the edits was frustrating, and it clearly was not generalizing effectively from my input.

Then I tried a different strategy: I picked a random section of the book and did line edits on it myself, and told the agent to diff my edits against the original version and talk through the intent of each change and why the original was unsatisfactory. It used this to draft a revision guide.

Not only did it not substantially improve the paragraph-level style of the writing, it did nothing to address the myopia of the text or its weird predilection for lists of parallelisms and self-aggrandizing descriptions of its own arguments.

My first approach to try and get it to write well was to describe what sort of style I wanted, by referring to constructions, rhetorical devices, the way paragraphs connect to each other, etc. I gave it positive examples and things it should avoid. This didn't turn out well.

I took this (please don't hate me) as a challenge. Some things simply cannot be done reliably by current models (e.g. code review). Is writing in a human voice one of them? I wanted to find out.

But when it came to draft the thing, it was crap. The citations and ideas were good; individual paragraphs were well written; but the text had no cohesion, no narrative arc, no fellow-feeling. It had a weirdly defensive posture. It insisted on its own correctness a lot. Lots of meta-narration.

It was an entertaining exercise, mainly because I found it easier to get at what I wanted to say by having to push the AI in the right direction than I would have when drafting a detailed outline on my own. Claude was functioning as "midwife to my thoughts" as Nietzsche puts it.

A month or so ago I had a ton of Claude credits to burn and tired to have Claude write a book for me. I gave it an outline, core themes, and a research guide, and engaged in a bunch of back and forth with it, trying to capture the perspective and intent of the book and the job of each chapter.

the ending was majestic. just amazing.

i just read the last dozen pages of The Name of the Rose live on TikTok. It was kind of fun.

*interdimensional

i did. it’s got very little to do with the movie. really weird dimensional time travel thing. howl is a grad student at oxford

My pet theory is that bsky should primarily be a journaling app with the option of sharing select entries to a public feed.

what if he nukes tehran?

alas

i will be kind to you.