3 ms·

Not commenting on open AI specifically, but I feel like actually acquiring the data is borderline impossible if you only follow legal channels. Where can I ins

Not commenting on open AI specifically, but I feel like actually acquiring the data is borderline impossible if you only follow legal channels. Where can I insert my money to buy every penguin random house book? How do I pay every deviant art artist whatever amount they're owed?

MBCook 2y ago
You could go to Penguin directly. Or perhaps the Authors Guild would help you set something up across publishers. All of deviant art? Probably not. But do you need to train on all of that? You could certainly run ads telling people you’d be willing to pay a small amount to train on their art and let them choose. Would it be legal to train against the national archives? Options exist. No you won’t get as much stuff as you do by taking whatever you want, but people are being compensated for their work or at least being given the choice to opt in.
- rockemsockem 2y ago
  Those sound like really large obstacles. Also, projecting this thinking forward, if we have humanoid robots walking around assisting humans, should they not be able to learn from their environment without paying every time they look at a piece of art or read something?
- tavavex 2y ago
  Ignoring the actual legality of using training data in machine learning, let's look at these "options" in a purely objective-driven way. If you do that, you'll quickly realize that what you describe would strengthen megacorporations to an even greater extent. If training on public data is banned, then the entities who have complete ownership of all their data would be granted a de facto monopoly. Stock image services, media conglomerates, music labels, industry giants would start in the winning position. When one side just gets their way for free and the other has to beg for scraps, do you really think this is a fair proposition? The reason why open-source AI exists at all is that we've always been allowing use of public data - it was okay when Google did it, it was okay when the Internet Archive did it, it was even okay when text translation services used that same data to train their models - or really, that applies to basically anything ML-driven before generative AI. There's, like, a sea of reasons to criticize OpenAI for - but arguing for extending IP laws even further and calling out opponents for "literal theft" is one of the weaker options that caught on with many people.
qbxk 2y ago
you're not wrong, and i think the technology exists to accomplish something that did this, and has for a few years. but the end result would be a system that funnels money from large numbers of people, to different large numbers of people. and a lot of overlap and cross payments too. maybe there's a good business in transaction fees on that? but it seems like big numbers going to other people while a much smaller number goes to the party owning the system is not an appealing business in our world. ipso facto why does it not exist while spotify buys podcasts and kills them

MBCook 2y ago

You could go to Penguin directly. Or perhaps the Authors Guild would help you set something up across publishers. All of deviant art? Probably not. But do you need to train on all of that? You could certainly run ads telling people you’d be willing to pay a small amount to train on their art and let them choose. Would it be legal to train against the national archives? Options exist. No you won’t get as much stuff as you do by taking whatever you want, but people are being compensated for their work or at least being given the choice to opt in.

rockemsockem 2y ago
Those sound like really large obstacles. Also, projecting this thinking forward, if we have humanoid robots walking around assisting humans, should they not be able to learn from their environment without paying every time they look at a piece of art or read something?
tavavex 2y ago
Ignoring the actual legality of using training data in machine learning, let's look at these "options" in a purely objective-driven way. If you do that, you'll quickly realize that what you describe would strengthen megacorporations to an even greater extent. If training on public data is banned, then the entities who have complete ownership of all their data would be granted a de facto monopoly. Stock image services, media conglomerates, music labels, industry giants would start in the winning position. When one side just gets their way for free and the other has to beg for scraps, do you really think this is a fair proposition? The reason why open-source AI exists at all is that we've always been allowing use of public data - it was okay when Google did it, it was okay when the Internet Archive did it, it was even okay when text translation services used that same data to train their models - or really, that applies to basically anything ML-driven before generative AI. There's, like, a sea of reasons to criticize OpenAI for - but arguing for extending IP laws even further and calling out opponents for "literal theft" is one of the weaker options that caught on with many people.

qbxk 2y ago

you're not wrong, and i think the technology exists to accomplish something that did this, and has for a few years. but the end result would be a system that funnels money from large numbers of people, to different large numbers of people. and a lot of overlap and cross payments too. maybe there's a good business in transaction fees on that? but it seems like big numbers going to other people while a much smaller number goes to the party owning the system is not an appealing business in our world. ipso facto why does it not exist while spotify buys podcasts and kills them