A group of authors filed a lawsuit against Meta, alleging the unlawful use of copyrighted material in developing its Llama 1 and Llama 2 large language models....
There actually isn’t a downside to de-duplicating data sets, overfitting is simply a flaw. Generative models aren’t supposed to “memorize” stuff - if you really want a copy of an existing picture there are far easier and more reliable ways to accomplish that than giant GPU server farms. These models don’t derive any benefit from drilling on the same subset of data over and over. It makes them less creative.
I want to normalize the notion that copyright isn’t an all-powerful fundamental law of physics like so many people seem to assume these days, and if I can get big companies like Meta to throw their resources behind me in that argument then all the better.
Humans learn a lot through repetition, no reason to believe that LLMs wouldn’t benefit from reinforcement of higher quality information. Especially because seeing the same information in different contexts helps mapping the links between the different contexts and helps dispel incorrect assumptions. But like I said, the only viable method they have for this kind of emphasis at scale is incidental replication of more popular works in its samples. And when something is duplicated too much it overfits instead.
They need to fundamentally change big parts of how learning happens and how the algorithm learns to fix this conflict. In particular it will need a lot more “introspective” training stages to refine what it has learned, and pretty much nobody does anything even slightly similar on large models because they don’t know how, and it would be insanely expensive anyway.
There actually isn’t a downside to de-duplicating data sets, overfitting is simply a flaw. Generative models aren’t supposed to “memorize” stuff - if you really want a copy of an existing picture there are far easier and more reliable ways to accomplish that than giant GPU server farms. These models don’t derive any benefit from drilling on the same subset of data over and over. It makes them less creative.
I want to normalize the notion that copyright isn’t an all-powerful fundamental law of physics like so many people seem to assume these days, and if I can get big companies like Meta to throw their resources behind me in that argument then all the better.
Humans learn a lot through repetition, no reason to believe that LLMs wouldn’t benefit from reinforcement of higher quality information. Especially because seeing the same information in different contexts helps mapping the links between the different contexts and helps dispel incorrect assumptions. But like I said, the only viable method they have for this kind of emphasis at scale is incidental replication of more popular works in its samples. And when something is duplicated too much it overfits instead.
They need to fundamentally change big parts of how learning happens and how the algorithm learns to fix this conflict. In particular it will need a lot more “introspective” training stages to refine what it has learned, and pretty much nobody does anything even slightly similar on large models because they don’t know how, and it would be insanely expensive anyway.