You’re absolutely right!
I continue to be one of those people using AI, but in a way that I hope isn’t too insufferable; or at least, I try to be sensitive to the suffering it may cause to those around me. I also feel like I am using it in a certain way: an emerging sub-sect of practice for a certain clientele. It may be age-related: I am someone who believes in hypertext, in composable tools, and malleable interfaces, in people writing and sharing their code, of owning the means of computation, and a free culture of sharing. I am also very Unixy, which is not unrelated. Why yes, I do own a mechanical keyboard, why do you ask?
My etiquette is tied up with all this, I think: I resonated with Alex Martsinovich’s It’s rude to show AI output to people, whose guidelines I find myself following. I don’t, as he warns, just say “Hey I asked ChatGPT this and it said”, or (even worse) paste the slop directly into the chat. I do think, like Fedora may, that one should be transparent about AI usage. And my added twist is that I really want to make my use of LLMs transparent and reproducible, to the extent that LLMs are.
To give an example, a friend of mine pointed me to the writer Samantha Hancox-Li, and I was curious to understand her journal, Liberal Currents, better. I found a Youtube video explaining it, but I didn’t really need to spend the time watching the whole video. So I grabbed the subtitles, fed that into a LLM, and asked it for a detailed summary. That answered my question, but I wanted to check in to make sure the model wasn’t on crack, so I threw a tiny bit of it into my chat with my friend, along with a “does this sound right?”. I sat between my friend and the LLM, I’d signalled that it may be unreliable, I’d edited it so that it was relevant and interesting.
But also, I wanted to show the process, including my own potential mistakes and biases that might have led to anything wrong. So I stuck this at the end of my mesage to show my working:
|
1 2 3 |
yt-dlp --write-auto-sub --skip-download -o "/tmp/subs.%(ext)s" "https://www.youtube.com/watch?v=FfZWlFjmmo0" files-to-prompt /tmp/subs.en.vtt | llm "Can you turn this transcript into a (detailed) description of the conversation" |
This lets you regenerate the text that I used (of course, it won’t be exactly the same, because llms rarely repeat themselves, but as time goes on, it’ll probably be a better summary). It also shows what I am depending on, including my own prompt and the programs I used.
This seems to be both polite in the way that Martsinovich would like us to be, lives up to my own personal set of ethics. It’s also a bit of an affectation and I’m not sure how long or consistently I’ll do it — but hey, sometimes these nerdish quirks become as fleeting and ephemeral as writing your geekcode or hand-crafting micro-formats, and sometimes they become smileys and markdown. I hope dearly that llm’s skill at coding (and explaining coding) will mean we can throw around such executable fragments until we all become a little literate in programs as well as words.