5 ms·
I would speculate this is true of all the leading commercial LLM models. Don’t have enough training data? Just steal some!
by ungreased0675 2y ago
I would speculate this is true of all the leading commercial LLM models. Don’t have enough training data? Just steal some!
- Havoc 2y agoOn true for all - you’d need to split it by era I think During the early Llama 1 days The Pile dataset was in heavy use by many. Bit later people figured out that a subset of it - Books 3 - was especially problematic. I’m guessing all the big houses threw that piece out in later models since it’s extra radioactive
- gooosle 2y agoCopy some*
- dehrmann 2y agoCourts have yet to decide on which it is, and it might depend on how well the model can transform vs. recite.
- BSDobelix 2y agoThat's exactly the difference, one does not steal in the digital world. If i could download/copy a car i would do it ;)
- pera 2y agoIt depends who you are: - if you are an individual then it's called "pirating copyrighted work" - if you are a multi-billion dollar corporation then it's called "use of uncleared material for training"
- throw5959 2y agoIntent matters. I'm very happy we don't live in a world where it doesn't matter.
- taskforcegemini 2y agoone is personal use, the other is making money off of it?
- throw5959 2y agoMeta open sourced the result.
- bodiekane 2y agoIf we're going to use absurd hyperbole like "steal", I think we should just keep going further. Zuckerberg murdered some old library books to train a model. Zuckerberg genocided training data! Heck, everyone who read your comment here stole it. I'm so sorry for your loss.
- aprilthird2021 2y agoIt's not absurd hyperbole. If I took the text of, say the NYT Bestseller this week, stored the text along with various Projects Gutenberg books, then created a program to randomly deliver you a chapter from any such book. That would probably get me a lawsuit. This is just that with lots of levels of indirection.