One thing I wondered is whether high prestige writing is encoded into the models, but it doesn't seem far fetched that there's various linkages inside the data to say "this kind of thing should be weighted highly."

https://youtu.be/1d4JOKOpzqU?si=xXDqGEXiawLtWo5e&t=569

I don't have evidence that that's true, but it's what I assume and I'm surprised it's not even mentioned as a possibility.

When I studied author profiling, I built models that could identify specific authors just by how often they used very boring words like 'of' and 'and' with enough text, so I'm assuming that OpenAI plays around with some variables like that which would much harder to humans to spot, but probably uses several layers of watermarking to make it harder to strip, which results in some 'obvious' ones too.

Regular computers for human use only support ASCII in US or ISO-5589-1 in EU still to this day, and Unicode reliant East Asian users turn off Unicode input modes before typing English words, leaving the Asian part mostly in pure Unicode and alphanumeric part pure ASCII. So Unicode-ASCII mixed text is just odd by itself. This in turn makes use of em dashes odd.

Same with emojis. LLMs generate Unicode-mapped tokens directly, so they can vocalize any characters within full Unicode ranges. Humans with keyboards(physical or touchscreen) can mostly only produce what's on them.

Yeah, this is me. I've always liked good type and typography. 5 or 6 years ago I added em-dash to my keyboard configs to make typing it in convenient - mostly because I just think it just looks nicer. But lately I don't use it much because... AI.

However, in recent weeks someone accused an HN post of mine as being from a bot, despite the fact I used a plain old hyphen and not an em-dash. There was nothing in the post which seemed AI-like except possibly that hyphen. At the time, I realized that person probably just couldn't tell a hyphen from a real em-dash. So maybe that means I have to not use any dash at all.

Never spent too much time thinking about em-dashes. Writers I like probably use them all the time—again, never really thought about it.

There are many other language model artifacts that are genuinely shite and are worth criticizing. Though, come to think of it, they have been getting stamped out with each iteration in model. Used to spend a lot of time trying to get models to refrain from words like "crucial".

What I do find strange is how the latest SOTA models appear to write with contractions by default, which began sometime in the past year. Anthropic models, in particular.

Also, I wrote a small extension that automatically replaces ChatGPT responses with em dashes with alternative punctuation marks: https://github.com/nckclrk/rm-em-dashes

Actually, they wouldn't have to go and buy these old books: The texts are already available copyright free, due to legislation stating that copyright expires 70 years after the author's death (any book published in the USA before 1923 is also reproducible without adherence to copyright laws), making the full texts of old books much easier to find on the internet!

The truth is ... most people don't read much. So it's not too surprising they think it looks weird if all they read is posts on the internet, where the average writer has never even learned how to make one on the keyboard.

Delve on the other hand, that shit looks weird. That is waaay over-represented.

And so AI models are prone to using them because they require less computation than rewriting a sentence.

Instead of

"The time has come," the Walrus said,

"To talk of many things:"

... it would be spelled as

— The time has come, — the Walrus said,

— To talk of many things:

I wonder how much of russian language content was in training model.