reading Star Trek: Lower Decks and it has a lot of jokes about personal logs and when and how exactly everyone records their personal logs (since they're voice-based and like what, people are just telling other ensigns "one sec I just need to talk into my log here—") and this made me think of when I first watched the original Twin Peaks and got a tape recorder so I could record my thoughts like Agent Cooper talking to Diane.
but my projects in this direction never went very far. like I have so many *thoughts* all the time—you may have noticed it from this feed—and I like talking but when I try to record them in voice it somehow feels high friction, and listening/organising/backuping/etc. the logs after the fact is even more high friction, so even though that *feels* like something fun and that I want to do, when I try I actually don't
I heard something about "AI" being demonstrably good for audio transcription and then I thought, yeah, after this whole bubble pops I guess that's the only useful technology that will be left—machine learning models for a first-pass at speech-to-text, translation, and other tough language problems. They're language models, after all.
Then I saw as a reply the criticism that ChatGPT transcription actually has a problem where it keeps fabricating content whole cloth that has nothing to do with the original audio and that's hard to spot and end up being more trouble than worth for the transcription corrector to fix, and I was like, because of course it does. It's reverse centaurs all the way down
