Quoting Venkatesh Rao: LLM Voice
In defense of the stylistic quirks of LLM writing
My most important style principles concerns things I don’t do.
First, I don’t try to make AIs sound like I used to in my pre-AI work. […] There’s nothing wonderfully unique about my voice that’s worth enshrining in AI-assisted writing protocols. If I have something interesting to say, I’m fine sounding like an index-fund voice.
Second, I don’t try to edit away or even moderate the native stylistic quirks and tics of LLMs. I largely leave all the well-known markers alone, unless they make it personally hard for me to read the essay. […]
After all, I don’t try to micromanage human collaborators, so why should AI collaborators be any different?
This particular species of collaborator is in fact valuable precisely because it possesses a voice that is a distilled embodiment of the entire contemporary wild zeitgeist itself. Shaped by training protocol choices, yes, but still, a remarkably distinctive voice that I feel it would be a pity to reshape too much to conform to my more human preferences.
You wouldn’t cook fresh vegetables to death and overwhelm them with aggressive spicing would you? No. You’d try to bring out and preserve their natural character as much as the recipe allows.
– Venkatesh Rao, “AI Witchcraft Style Principles” on the Contraptions About page
I love these lines from Venkat. It’s the only defense of AI speak I can remember reading, and I find it really compelling. The Contraptions blog is probably the best example of LLM-assisted writing done well, and I take personal inspiration from his fearless approach.
I’ve written a few posts about AI writing and have even written a few “witch hunt” tools myself. I use sloppy-syntax and STE lint to help prepare pull request descriptions, commit messages, and reports on security issues. To me this is mostly a courtesy thing, a way to help someone else digest machine-text as quickly as possible. Everybody’s busy and making my work easy to grok helps everybody surmount the context switch inherent in a PR review.
That said, I don’t think those tools produce text that’s better, or even nice. They take some of the bizarre excesses of the models (especially the Opus 5/Fable era) and wrangle them into something like a concise technical manual. That’s perfect for the task at hand. Outside of that context, there’s a lot of room for play with generating text and bringing out ideas.
As I mention in my own About page, the motivation for me to use LLMs for writing is that I simply don’t have the time or energy to get my ideas out there otherwise. I’d love to do more hand-crafting of text as I do think it would be the best way to express myself and my ideas – but preserving energy for writing means taking it away from some other area of my life, and that’s not a trade I want to make right now.
This blog will continue in Goblin Mode. (Thanks again for a useful concept, Venkat.) I don’t think anyone needs to be ashamed of using LLMs for writing, so long as the ideas are good and the end product is serviceable. There’s no need to scrub away every last characteristic of the LLM voice, especially in the cozy, opt-in world of blog writing.