Worth repeating: Claims of "AI" democratizing anything (coding, creative endeavors, etc) are always BS. The thing locking out people from doing that is not having the time/resources.
You want to democratize coding/art/creativity? Give people paid time off to do it and access to infrastructure. Easy.
@tante this is an intellectually dishonest argument.
Everything is solvable by "more time and more access to infrastructure" when your limits approach infinity, similar to how, if you gave an infinite amount of monkey an infinite amount of typewriters, one of them will eventually type up with the works Shakespeare.
Take creating an alternative to Android for example, it's a Herculean task not because of the OS itself, but because of a
> wide range of nonfree software blobs that commonly occur in even the most progressive "free software" operating systems.
LLMs are great at pattern recognition by design, including identifying function signatures, recognising common driver patterns, reading Ghidra output, etc.
This takes a problem that was so far considered in the realm of needing state backing to a problem that can be tackled by a relatively small team of dedicated researchers.
So yeah, my bad if I don't give a damn about moralising "AI use" when the alternative is rewriting and drying up the moat of American tech companies that enable surveillance and brutalising of the Global South
@budududuroiu @tante Whatever gets you through the night.
The LLMs are abusing the generosity of the Free and Open Software movements to destroy them by commoditizing the work of the Community.
No matter how noble the goal, these beginnings are unclean.
@DataGhost @tante where is the abuse of generosity if you train and distribute a GPL-licensed LLM? Chinese labs proved it's possible to do on a small budget, efficiently.
Why is me using GPL snippets in my code and further distributing my code as GPL ok, but using GPL snippets in training an LLM and further distributing that LLM as GPL wrong?
As I said above, it seems it's an intellectually dishonest argument
@budududuroiu @tante If you use GPL snippets and comply with the GPL by publishing the work you incorporated it into, you’ve honored the agreement that gave you those snippets in the first place. Good.
An LLM scraping them all, mashing them together, and regurgitating it out of context does not honor the gift of the code. It’s also a violation of the GPL, which uses copyright to enforce the continued freedom of that code.
GPL licensing the model doesn’t help.
> It’s also a violation of the GPL
Under what? GPL gives you the freedom to run, study, share, or modify the software as long as you publish the derivative work as GPL too. Releasing a truly open source model with training data, training scripts, hyperparameters, and everything needed to reproduce the model would not violate GPL.
@budududuroiu @tante The output of that open model will incorporate the GPL code it was trained on, generate derivative works, and strip all credit and licensing.
Apparently this is an unsettled area of copyright law, so please set aside “violation of the GPL”.
> generate derivative works, and strip all credit and licensing.
Isn't this already an issue we solve by expecting developers to not strip licensing and credit?
Also, GPL-ness isn't static, when GPL didn't prevent projects from offering managed solutions for GPL code as SaaS, people came up with AGPL.
@budududuroiu @tante Do we expect LLMs to not strip licensing and credit?
It seems fundamental to their operation.
tante
Bogdan Buduroiu
Data Ghost