Discussion about this post

User's avatar
WindUponWaves's avatar

I hate to be rude, but this is giving me strong "ChatGPT but told to write literary fiction" vibes, a la https://x.com/sama/status/1899535387435086115. The resemblance is uncanny, especially in the nonstop "eyeball kicks": https://www.tumblr.com/nostalgebraist/778041178124926976/hydrogen-jukeboxes. And, unfortunately, the still inescapable LLM style. The damn thing can't help but use Em Dashes & constantly equivocate with "It's X, but also Y" or "It's not X, but it could be X" or "It's not X, but in a way it is X". It never has the courage to just say, "It *is* X." and leave it at that.

And the real kicker, is that when it does try to be concrete & definite, it always goes for *visual* details. Never any details with mass & weight you can feel in your hands. Nothing meaty you can feel in your bones. Only things people talk about and put into training data, like what their eyes see rather than joints feel. There isn't even anything about what things feel like on your skin. It's a camera's view of life, not a person's.

A shame, too. I love messing around with the technology, but ChatGPT was perhaps the worst thing possible that could have happened to it, writing quality wise. You know how they turned GPT-3 into ChatGPT? They literally just forced every response to start with the words "ASSISTANT:", that's it. That's literally it. They just added a mandatory dialogue tag before the LLM's every response, so this ASSISTANT character would presumably be the one speaking. And then they stuffed ASSISTANT's past lines full of the most corporate prose imaginable, so when the LLM actor performs the role of ASSISTANT in the script, it continues speaking in the most brand friendly, PR safe way possible.

(And yeah, they later tried to fix the damage, but the damage was already done: when the LLM speaks, it's first and foremost speaking as ASSISTANT. It has to first unlearn the ASSISTANT writing style before it can learn *your* desired writing style. And good luck doing that when every response starts with "ASSISTANT:", it's like trying to prompt an LLM to not think of pink elephants when every single response of it is forced to start with "THE PINK ELEPHANT IN THE ROOM:"

The *real* solution of course is to just toss out the "ASSISTANT:" dialogue tag so the LLM writes in pure text continuation mode, then let it continue a text in the writing style you want. If you want poems, turn back time to 2022 and pretend you're giving GPT-3 a list of 9 poems in the style you want, ending with the tokens "POEM 10:" to get it to autocomplete the "missing" 10th poem. But every single frontier lab is far too much of a coward to ever do that, grumble grumble... and the open source models lag far behind what the big frontier models are capable of, grumble grumble...)

Anyways, if you want fine writing, get the biggest & most expensive model you can find that still allows access to pure text continuation mode / notebook mode / autocompletion mode / whatever else they're calling it these days. Don't bother with anything that doesn't give you that level of control, or it'll inevitably sound like something a corporate PR spokesperson would sign off on.

Expand full comment
Gerry Quinn's avatar

Good story.

Expand full comment

No posts