FFmpeg 9.0 • Hacker News Board

Re: FFmpeg 9.0

#11

Post by lavela » Tue, Aug 04, 2026, 3:45 PM UTC

gortok wrote: post #7

joeysim wrote: post #4

These days every post is getting the AI writing treatment :-(

I was like “how can you tell?” And then I read the article. I realize for folks that need to write release notes that it’s not fun, and it’s hard to write well, but using AI to do it just makes it worse.

I really don't get why there isn't more work put into making LLM generated prose less bad. Or is it just not possible?

'Release note' as a genre shouldn't be too hard to train? Especially when there are already a plethora of release notes of the same project.

Re: FFmpeg 9.0

#12

Post by robby_w_g » Tue, Aug 04, 2026, 3:46 PM UTC

gortok wrote: post #7

joeysim wrote: post #4

These days every post is getting the AI writing treatment :-(

I was like “how can you tell?” And then I read the article. I realize for folks that need to write release notes that it’s not fun, and it’s hard to write well, but using AI to do it just makes it worse.

Contrast with the Dolphin emulator release blogs. I swear I’ve spent more timing reading those releases than actually using the software

Re: FFmpeg 9.0

#13

Post by owlboy » Tue, Aug 04, 2026, 4:05 PM UTC

gortok wrote: post #7

joeysim wrote: post #4

These days every post is getting the AI writing treatment :-(

I was like “how can you tell?” And then I read the article. I realize for folks that need to write release notes that it’s not fun, and it’s hard to write well, but using AI to do it just makes it worse.

It’s strangely judgmental too.

> We have a video encoder and muxer for the Playdate, the little yellow console with a crank. It produces 1-bit, 400×240 video using delta coding and zlib compression. It is not important, but it is cool.

Re: FFmpeg 9.0

#14

Post by gortok » Tue, Aug 04, 2026, 4:21 PM UTC

lavela wrote: post #11

gortok wrote: post #7

Earlier quoted context omitted.

I was like “how can you tell?” And then I read the article. I realize for folks that need to write release notes that it’s not fun, and it’s hard to write well, but using AI to do it just makes it worse.

I really don't get why there isn't more work put into making LLM generated prose less bad. Or is it just not possible? 'Release note' as a genre shouldn't be too hard to train? Especially when there are already a plethora of release notes of the same project.

The folks that would use an LLM to make their release notes are not the folks that care enough about their release notes to put any more effort in.

And practically speaking, I don’t see a LLM being able to do this. Perhaps a specialized language model trained only on “good” release notes and commits, but putting that much effort into the training data goes against the whole mantra of using an off the shelf model in the first place.