HarfBuzz 13.0.0 released with new experimental features: `hb-vector` for vector output of glyph outlines to SVG and `hb-raster` for rasterizing glyphs to A8 / BGRA32 images, among other features.

[liquor_american@universeodon.com]

@behdad Well that's an own-goal isn't it

[gkrnours@mastodon.gamedev.place]

@behdad

[jerojasro@col.social]

@behdad I'm curious: how do you validate the output from the LLMs?

do you review it as you would a regular PR?

are there automated tests, and who/how are they written?

[mcc@mastodon.social]

@jerojasro @behdad What are your plans if it is someday discovered your LLM(s) replicated a non-trivial section of code from the training data, and the replicated code in question turns out to have a stronger-than-MIT license encumbrance (for example it could be GPL)?

(LLMs have so far lived a legally charmed life, but this may not be the case in all jurisdictions forever.)

[lhp@mastodon.social]

@behdad I am not interested in any features that aren't possible without LLMs. Eww and yikes.

[chx@chx.contact]

@behdad by including LLM output, you have made one of the best open source libraries unusable from at least ethical, legal and software security angles. More angles are possible, these are the ones that come to mind immediately.

[jackemled@furry.engineer]

@behdad Sad to see you go. Are there any recommended drop in replacements?

[linear@nya.social]

@behdad@typo.social cool, looks like i'll be permanently locking harfbuzz's version in my distro to the previous version, and looking for AI-free forks

[lumi@snug.moe]

@behdad yikes, that's disgusting

[bnlandor@mastodon.social]

@behdad @dysfun Is that a rhetorical question?

[aburka@hachyderm.io]

@dysfun @behdad with what?

[julianh@mas.to]

@behdad @dysfun have you enclosed yourself in a rock with monkeys on typewriters, tech bros, and Elon Musk for the last ten years?

Asking this question is the equivalent to shitting on sometimes lawn and asking the neighbors to elaborate when they ask you to stop shitting on their lawn.

[julianh@mas.to]

@behdad @dysfun in case you need a list:
Energy use.
Forceful integration of ai by large companies who often steal copywrited/licensed material and scrape data in bad faith (such as ignoring robots.txt).
Major vulnerabilities and issues in software supposedly written by AI (see: Windows 11).
The lack of legislation making it unclear if AI generated material is compatible with code licenses.
I could go on.

[isurandil@mastodon.online]

@behdad "Claude & Codex"? It's not yet April 1, is it?

[barubary@infosec.exchange]

@behdad Thanks for the warning.

[iorsh@kishkush.net]

@mcc @dysfun @behdad
That's an interesting approach. I admit I'm a bit cynical here, as I myself just started to use it quite extensively after a long period of scepticism. But at least forking is somehow practical.

[a1ba@suya.place]

@behdad
maybe that's good that I don't need harfbuzz rendering with languages I know.

[jlink@det.social]

@mcc @iorsh @dysfun @behdad Maybe we should agree on a common naming scheme and strategy for forking non-ai branches of projects. This way it would be easy to check if there’s a related non-ai project. In addition there should be a „nonai“ TLD so that domain-based package names are obvious as well.

[valpackett@social.treehouse.systems]

@behdad maybe you should've renamed it to SlopBuzz first

[david_chisnall@infosec.exchange]

@dysfun @behdad

Does HarfBuzz parse things like the hinting tables in fonts? That’s traditionally been a source of a lot of security vulnerabilities and I wouldn’t be surprised if 90% of the things in the training set have vulnerabilities.