These are English-language models. They’re trained on an English-language corpus. Ideally, the entire history of published works. They’re in public, for public consumption. That’s what “published” means. Nobody broke into an author’s house to steal their secrets. The machine that writes is skimming the whole library to find out how to write gooder.
If they used a literal library - shelves full of dead trees - would people still insist the authors were never paid? Are we supposed to pretend a robot scanning a book a thousand times is fundamentally worse than a hundred people borrowing it? I guarantee you each human got more out of it. These stupid robots have to plow through an entire shelf just to figure out what “fantasy” means.
The more works we shovel into these things, the less each individual work counts. The model does not get bigger. Each work’s unique impact shrinks as statistical patterns get generalized.
Of course books were used - and why not?
These are English-language models. They’re trained on an English-language corpus. Ideally, the entire history of published works. They’re in public, for public consumption. That’s what “published” means. Nobody broke into an author’s house to steal their secrets. The machine that writes is skimming the whole library to find out how to write gooder.
If they used a literal library - shelves full of dead trees - would people still insist the authors were never paid? Are we supposed to pretend a robot scanning a book a thousand times is fundamentally worse than a hundred people borrowing it? I guarantee you each human got more out of it. These stupid robots have to plow through an entire shelf just to figure out what “fantasy” means.
The more works we shovel into these things, the less each individual work counts. The model does not get bigger. Each work’s unique impact shrinks as statistical patterns get generalized.