Post

Using LLMs to write but not wanting to engage with LLM-generated content is wild to me. The hypocrisy. Also, why would I read something you didn't bother to write? The value of writing is more than the output it produces. Writing generates understanding and ideas. Why take shortcuts?

dan@danabra.mov

on one hand i use llm’s quite a bit, and sometimes to edit writing on the other hand if i see an llm-ism in something im reading i’m just gonna disengage and scan the rest or close the tab i’m not sure that’s internally consistent

i usually ask llm to *shorten* what i wrote and then sometimes i use the edits. i think it’s a bit different from having it create thoughts but i also feel strange about this continuum!

An LLM shortening what you write is still changing the meaning of the ideas you're trying to convey, even if it appears minimal. LLM-editing influences what you write, thus shaping your future ideas. Writing and editing are not just how we teach and share, but also how we learn.

Editing is hard, yes, but that doesn't mean it doesn't add a ton of value for your readers and for yourself. The fact that it's hard _is_ the point: shaping ideas into words is difficult and important.

0 reposts2 likes

I've always found your writing insightful: you have a great "voice", and you're easy to tell apart from other "known documentation/tutorial writers". That's amazing, I think you should cherish it. Don't cheapen your work.

LLM-editing is a big red "make this text more average" button. I don't want to read the "average sentence on the internet", I want to read what you write. The process *is* the work. Sorry for the rant, have a great day <3

i don’t think that’s actually always a good model for what it does. i think of it more as an effect pedal in a semantic space. true most outputs are boring just like most things you can do with an effect pedal are boring and overdone. doesn’t mean it’s useless or always averages out. needs taste

I see what you mean by "an effect pedal in a semantic space", but LLMs have no understanding of the semantics, and I don't think stolen data is a good "synthetizer" iykwim. I wouldn't write a piece of text and then "dip it" in the toxic corners of the internet just to see what output it produces.

i mean i feel like i’m able to judge whether what i want to express is more closely conveyed by variant A or variant B and i do have some experience editing myself too. sometimes it helps to process text through a compressor and see what comes out

i don’t see this as fundamentally different from sound engineer accidentally putting too much overdrive on a part and then deciding keeping it crunched is maybe better. or it gives an idea. i’m sure there’s something about that in oblique strategies. llm is just another tool to fuck with the output