David Gerard1d ago
https://bsky.app/profile/scalzi.com/post/3mj6q3pdblc2y
Show threadGustavo22h ago

@davidgerard We should consider one thing: if one calls something AI, it is mostly like that's the tool the arson group uses, useful things often aren't called AI.

Even if they share the same mathematical principles! Take good old machine translation: great to make people understand each other, yet uses the same transformer model as LLMs. No one calls it AI (unless there is a shitty LLM inside).

How about speech recognition? It is awesome for accessibility yet often has the same ethical issues as generative AI: many models are trained on stolen work from humans who never subtitled things wanting their work to be used by big AI companies, many volunteers even.

Show threadTim

@qgustavor @davidgerard

I don't want to be 'that guy' and I'm not.

But there is one thing I use that is AI based, and isn't ripping anyone off - noise/wind reduction.

Sure when it's music extraction, that can mean training on unpaid work/stems. But in that case, the wind isn't going to be ripped off for royalties.

Not sure whether Whisper was trained on legal or dodgy sources, but AI transcription has really helped with accessibility, so it's a hard one there...it's shitty if people have had their work stolen though.

But it is niche, and a lot of AI IS exploitative, but there are a few examples of tools where I don't see a down side.

Unlike GenAI, which is genuinely a massive problem and needs to die in a fire.

Apr 12 at 2:27pmWeb
Show threadGustavo19h ago
@radioclash @davidgerard
I use Whisper daily and I'm sure it was: I set my computer to transcribe every audio message I get, mostly due to accessibility issues (like when people prefer watching with subtitles, it's easier to understand), but when it transcribes silence or noise it returns things like credits (like "subtitles by example.org" or "transcribed and translated by Jonh Doe"). It mostly happens on my aunt's audios, since she whispers a lot, the model confuses her audios with movie credits.
Show threadTim19h ago

@qgustavor @davidgerard yeah that is suspicious when it hallucinates, makes me wonder what it was trained on...likely a whole load of amateur subtitles or even commercial ones.

Unlikely OpenAI got permission. This is when it gets tricky over net good.

Cos it helps a load of people - but those who worked on the subtitles should get credit and some of those AI billions?

Show threadTim19h ago

@qgustavor @davidgerard I stopped using Whisper and now use whatever Capcut or Resolve uses.

Not sure if it's the same.

Show threadGustavo17h ago
@radioclash @davidgerard I don't use it for subtitling. I tried using it once, to be fair, I don't even remember if the results were good or not. I guess they weren't, at least for the kind of content I was working with.
Show threadTim16h ago
@qgustavor @davidgerard it's not bad, better than the Capcut internal one. Davinci Resolve's subtitling is accurate and easier to deal with.