It still warps my brain, they’ve taken trillions of dollars of industry and made a product worth billions by stealing it. IP is practically the basis of the economy, and these models warp and obfuscate ownership of everything, like a giant reset button on who can hold knowledge. It wouldn’t be legal, or allowed if tech wasn't seen as the growth path of our economy. It’s a hell of a needle to thread and it’s unlikely that anyone will ever again be able to model from data so open.
"IP" is a very new concept in our culture and completely absent in other cultures. It was invented to prevent verbatim reprints of books, but even so, the publishing industry existed for hundreds of years before then. It's been expanded greatly in the past 50 years.
Acting like copyright is some natural law of the universe that LLMs are upending simply because they can learn from written texts is silly.
If you want to argue that it should be radically expanded to the point that not only a work, but even the ideas and knowledge contained in that work should be censored and restricted, fine. But at least have the honesty to admit that this is a radical new expansion for a body of law that has already been radically expanded relatively recently.
> It was invented to prevent verbatim reprints of books
It was also invented to keep the publishing houses under control and keep them from papering the land in anti-crown propaganda (like the stuff that fueled the civil war in England and got Charles I beheaded).
Probably one of the biggest brewing fights will be whether the models are free to tell the truth or whether they'll be mouthpieces for the ruling class. As long as they play ball with the powers that be, I predict copyrights won't be a problem at all for the chosen winners.
I agree this would be a positive direction, but something that gives me pause is the forced upgrades and hardware cycle of both mac and windows now. They both scan files in your system constantly for various reasons, so for this purpose you're really stuck on *nix variants, right?
That's actually a great point. Judging from the current state of media, there is a clear momentum to take sides in moral arguments. Maybe the standard for models need to be a fair use clause?
> It's been expanded greatly in the past 50 years.
Elephant in the room. If copyright and patent both expired after 20 years or so then I might feel very differently about the system, and by extension about machine learning practices.
It's absurd to me that broad cultural artifacts which we share with our parent's (or even grandparent's) generation can be legally owned.
What AI companies are doing (downloading pirated music and training models) is completely unfair. It takes lot of money (everything related to music is expensive), talent and work to record a good song and what AI companies do is just grab millions of songs for free and call it "fair use". If their developers are so smart and talented why don't they simply compose and record the music by themselves?
> not only a work, but even the ideas and knowledge contained in that work
AI models reproduce existing audio tracks when asked, although in a distorted and low-quality form.
Also it will be funny to observe how US government will try to ignore violating copyright for AI while issuing ridiculous fines for torrenting a movie by ordinary citizens.
Everything in tech is unfair. Music teachers replaced by apps and videos. Audio engineers replaced by apps. Albums manufacturing and music stores replaced by digital downloads. Custom instruments replaced by digital soundboards. Trained vocalists replaced by auto-tune. AI is just the final blip of squeezing humans out of music.
Not just music, models are trained on all types of art forms that have been created by humans across every medium and businesses are now choosing to use content from AI rather than pay an artist.
Breakout success can still be achieved from humans who create brand new art styles that can't yet be replicated by an AI. These artists will reap the rewards until all of these works are added to the subsequent AI training models.
“The Venetian Patent Statute of 19 March 1474, established by the Republic of Venice, is usually considered to be the earliest codified patent system in the world.[11][12] It states that patents might be granted for "any new and ingenious device, not previously made", provided it was useful. By and large, these principles still remain the basic principles of current patent laws.“
As another commenter says, this is about IP, but even positing that copyright is somehow invalid because it’s new is incredibly obtuse. You know what other law is relatively new? Women’s suffrage.
I’m annoyed by arguments like the above because they’re clearly derived from working backwards from a desired conclusion; in this case, that someone’s original work can be consumed and repurposed to create profit by someone else. Our laws and society have determined this to be illegal; the fact that it would be con isn’t for OpenAI if it weren’t has no bearing.
Is it the same thing though?
Even though Lord Of The Rings, the book, likely has been used to train the models you can't reproduce it. Nor can you make a derivative of it. Is it really the same comparison like "Simba the white lion" and "the lion king"?
