US accuses China of “industrial-scale” AI theft. China says it’s “slander.”
- Thread starter JournalBot
- Start date
I have this big pile of stuff I stole and someone's been rummaging through it all. That hardly seems fair...
Upvote
1
(1
/
0)
Upvote
0
(0
/
0)
Of course they aren't. They are doing it to catch up at 1/10 the cost by distilling the models and to undermine these billion dollar labs spending hundreds of millions of dollars a year to train the models, and probably to try and destabilize our economy which is more and more dependent on the growth of AI revenue.
In this case we (the regular people) are getting open weight models that we can run on our own hardware without any internet connection... And the Chinese labs are still publishing their research unlike basically all of the big US labs now.
I'm just hoping we get a local models that can do agentic coding before the big labs make it too expensive to use their services — I literally don't care who releases it.
Upvote
3
(3
/
0)
China: "Mister Trump, we read your book, art of the deal. Now we have many great succes. Now please stop slandering us or we put tariff. Please comply or we block gulf of America." Many thank yous. "
Upvote
-1
(0
/
-1)
There's no IP law that protects model weights when they're connected to exposed external endpoints. They can't be a trade secret, because if the output is capable of exposing the weights then it's inherently not a trade secret, under trade secret law.
And if any form of copyright could conceivably apply to "AI" output, then it's because of "AI" input, and all these companies are screwed because they trained their LLMs on unlicensed content.
And while copying a model directly would possibly violate copyright (if the model itself weren't a mass of derivative work of all the copyright violations it was trained using) as a static collection of information in a specific format (much like copyrights of telephone directories, databases, etc), distilling model weights from running against the working model doesn't: they're functional in that context, and what's being distilled isn't an actual specific copy of the model's underlying data and format, it's a derivation of the functioning the model performs.
So there's no "stealing". At best there's some "unlicensed use". Kind of like "unlicensed use" of all that material they crawl. Boohoo. This whole LLM crap is both unsustainable and turning into a race to the bottom where if you don't use it you're going to get stomped by it, even though the primary thing it produces is same-same mediocrity; please break the process of these commercial thieves burning electricity and all of the chips in existence on modeling stolen content faster so we can get past it. "Stealing" from them to do it is just the icing on the cake.
And if any form of copyright could conceivably apply to "AI" output, then it's because of "AI" input, and all these companies are screwed because they trained their LLMs on unlicensed content.
And while copying a model directly would possibly violate copyright (if the model itself weren't a mass of derivative work of all the copyright violations it was trained using) as a static collection of information in a specific format (much like copyrights of telephone directories, databases, etc), distilling model weights from running against the working model doesn't: they're functional in that context, and what's being distilled isn't an actual specific copy of the model's underlying data and format, it's a derivation of the functioning the model performs.
So there's no "stealing". At best there's some "unlicensed use". Kind of like "unlicensed use" of all that material they crawl. Boohoo. This whole LLM crap is both unsustainable and turning into a race to the bottom where if you don't use it you're going to get stomped by it, even though the primary thing it produces is same-same mediocrity; please break the process of these commercial thieves burning electricity and all of the chips in existence on modeling stolen content faster so we can get past it. "Stealing" from them to do it is just the icing on the cake.
Last edited:
Upvote
3
(3
/
0)
Not sure it was stolen... infringed perhaps. It was handed to them by American companies looking for cheaper-made alternatives so they could make more profits. Chinese manufacturers just kept the IP/specs and started spinning off knockoffs for less. Now they have the infra/tooling/capabilities to outpace the world AND build their own designs.
Upvote
0
(0
/
0)