alt Hacker NewsThe thing I find most hypocritical though is that they are probably never share their libraries with anyone. After scanning, downloading, stealing, overloading websites and everything in between, to acquire enough data for their stupid machine, they're not going to share their data? I get that most of it can't be shared, but a lot can. There's no reason why you need to destroy multiple copies of a book from 1880, when it's free to share.
At the same time I can understand keeping track of when each books enters public domain might also be an absolute nightmare, and I wouldn't blame the AI companies for not wanting to deal with that. For the stuff they absolutely know is clear, they should provide dumps for everyone to download.
Replies
Since they seem to leapfrog each others’ models every few months, the training data is one of the few ways they can build competitive advantage, and that explains why they don’t share, even if we don’t have to like this.
This is solved by law, which is solved by 'we the people' and I bet many AI companies would be fine with something like the equivalent to patent law with bankruptcy escrow to the library of congress, where they must release the scans in 10 years for books that the vast majority will not give a flying shit about. By then the advantage is long gone in data moat.