Writers: Generative AI models were built on our stolen works, are deeply unethical, and risk devaluing our entire profession.

[xenonchromatic@mastodon.social]

@jamesthomson "risk devaluing our entire profession..." To the people developing these models, this isn't a bug, its intended behavior.

[eggfreckles@mastodon.mit.edu]

@jamesthomson DragThing now by ChatGPT.

[jamesthomson@mastodon.social]

@xenonchromatic Indeed.

[davidhmccoy@mastodon.world]

@jamesthomson

I know a manager who suggested someone use AI to write a simple description of a piece of code.

Nuts.

[jamesthomson@mastodon.social]

@amyinorbit I mean, aside from the personal, ethical, societal, financial, and environmental issues, it's just great.

[colincornaby@mastodon.social]

@jamesthomson I think the problem is developers don't really consider any unattributed use of open source as stealing - just a mild grey area. (They should consider it stealing.)

[jaredwhite@indieweb.social]

@the_other_jon @jamesthomson John Siracusa seems to offer the most balanced perspective there, but yeah it's pretty grim. I mostly don't listen any more. And don't get me started on MacStories omfg ‍

[dancingtreefrog@mastodon.social]

@jamesthomson Also developers: Vibe-coding generative AI models were built on our stolen source code, we're all being laid off, please hire us.

[estranged@indieweb.social]

@jamesthomson
I would be interested in training a model based on my own code. I spend a decent amount of time looking through my own code to find something I know I’ve done before.

[jaredwhite@indieweb.social]

@colincornaby @jamesthomson Open source != public domain, and free software != free (it's free speech, not free beer), but apparently many developers are clueless re: all those nuances. ️

Perhaps if all LLM-generated code was legally automatically placed in public domain, we'd see a bit of a light bulb moment.

[drwave@mastodon.social]

@jamesthomson I’ve been struggling with this cognitive dissonance for years.

[jamesthomson@mastodon.social]

@estranged My understanding is that there isn't enough data in a small set like that to actually train usefully, as a standalone thing. So it's always going to be based on other models, with your training on top.