Comment Re:The last one was pretty lame... (Score 2) 49
Comment Re:one thing i've learned (Score 2) 86
Comment Re:Could have been worse (Score 1) 135
Comment Legal requirement (Score 1) 103
It's certainly a legal requirement to disclose LLM usage in the UK, because if you just use a human-sounding pen name, you're making a false implied representation that you're selling a human-written book with intent to gain (profit) from it, which is fraud by false representation. The police may not want to prosecute anyone for doing anything which corporations do because they exist to serve the rich, but the law as written seems very clear.
It may also be a legal requirement to have permission from the authors of books still under copyright before you use a machine to produce a derivative of their work, but that's completely uncertain at the moment of course.
Comment Re: What we made ourselves .... (Score 1) 66
Comment Re:What we made ourselves .... (Score 1) 66
Comment Re:Bejesus - Beer and Circuses (Score 1) 53
Comment Statistics (Score 1) 53
Comment Re:I keep hearing this but (Score 1) 191
There's already been several courts that have ruled that you can't look at individual inputs to an LLM's training data, and say "well, if 0.000001% of the training data was X, then the output of the LLM is X"
How does this work if you, say, take a C file from the Linux kernel which has been modified by hundreds of different people, include it in your proprietary program, and when asked for the source under the GPL, claim "no individual contributor's copyright actually covers a meaningful proportion of this file, so who am I violating the copyright of exactly"?
If this doesn't work because it;'s all GPL, then what if no individual person has provided a significant proportion of the training data, but 20% is GPLv2 only and 20% more is GPLv3 only?
Of course, when it comes to suing people in the courts, you need a plaintiff, but copyright also exists within criminal law in many places, and for criminal copyright infringement when pirates get arrested, maybe it's "what gives you the right to distribute copies of something that's only 0.0001% your own work?
Comment Re:Uhh... Cool? (Score 1) 191
First off, you can always just use a LLM with an open training dataset (OLMo, Pythia, etc) to prove that the data isn't in it, and have that handle the "clean" side of the equation.
Note that since free software projects with compatible licenses can share code freely, simply not finding that source amongst an enumeration of the sources is not going to be enough to prove none of the code was there. If all of the training data was explicitly provided with authorisation from the writers to feed it to a LLM and do whatever you want with the results, that would be better.
Secondly, you are under the mistaken impression that a LLM that has been exposed (which, BTW, does not mean "memorized", or even any rough understandings of the workings - ask the LLM (ideally an abliterated one to avoid refusals) to reproduce the original code verbatim, it'll almost certainly fail badly) automatically fails in such a suit. It does not. There is no such thing as "homeopathic copyright contamination". Copyvio requires substantial similarity.
The grey areas around copyright violation can change a lot depending on which legal system you're thinking of. Since you haven't specified which legal system you're thinking of, it's hard to know what you're claiming. But another formulation might be, say, "copyright violation is averted if the new input dominates the copied input", which would be game over for LLMs. All of this territory is a big fat "we don't know yet".
Comment Re:Uhh... Cool? (Score 1) 191
Comment Re:Uhh... Cool? (Score 1) 191
Comment Re:Begone. Give us the UNBIASED facts. (Score 1) 73
Comment Re:Will AI ever make it better/more-secure? (Score 1) 26
If you use LLMs to find problems, but have humans fix them and generally keep your code 100% human-written, then LLMs will help find vulnerabilities but are unlikely to insert any. Plus you can still be confident of having the right to distribute your code since you haven't just made it a derivative of a ton of GLPv2-only code and also a ton of GPLv3-only code...
Still got to reconcile yourself to participating in the "make hot countries which have contributed little to the problem unliveable and then call the people who lived there 'illegal' when they try to escape" project and such, of course.