> on the same level like Anthropic scraped copyright protected material for their training.
I see no problem with distillation, on the other hand the complete dismissal of copyright by AI labs is pretty bad, I don’t think we should put them at the same level
> on the other hand the complete dismissal of copyright by AI labs
Courts keep ruling over and over that an LLM trained on copyrighted works qualifies as a transformative work and is therefore fair use. They don't have to dismiss copyright law, this has always been allowed.
The only thing they get in trouble for is pirating the works to get their hands on them.
> Courts keep ruling over and over that an LLM trained on copyrighted works qualifies as a transformative work and is therefore fair use. They don't have to dismiss copyright law, this has always been allowed.
*USA only.
the UK has fair dealing, which is more restrictive
> "Keep ruling over and over" is way too strong. There have maybe been two rulings, nothing nationally binding, and most of the litigation is still ongoing.
> This ambiguity has resulted in extensive litigation on the limits of Fair Use to AI development. Currently, we only have 3 first instance decisions out of the 53 cases being tried. It will likely take a decade before we understand how Fair Use applies to any one step in AI training, let alone all.
> In the three lower court decisions so far, one held Fair Use did not apply (Thomson v Ross), one held Fair Use could apply (Kadrey v Meta) with the court suggesting more evidence was needed on the fourth factor ‘harm to the market’, and the third case held Fair Use may apply to some AI. As Fair Use is dependent on the specific facts at issue, none of these cases help educate the market or the public as to the limits of Fair Use in AI contexts.
Not to mention the cases where the AI labs would have lost in court so bailed and settled for billions. Just this week, Anthropic agreed to pay $1.5B in a settlement to avoid losing a pretty cut and dry case.
To be clear that was one of the few resolved cases where the judge agreed training was fair use. But the piracy was enough of a distraction that I don't consider that a particularly useful precedent. I am much more interested in the NYT case, which quite clearly shows GPT was trained on NYT articles and can spit them out verbatim (and has since been validated by academic research; all the commercial models are capable of mass plagiarism).
> Anthropic agreed to pay $1.5B in a settlement to avoid losing a pretty cut and dry case.
It takes two parties to agree to a settlement. That the other party agreed to a settlement instead of taking it to court implies this was not the slam dunk you may think it was.
Settling just says that they expected the internal costs or risks to be more than 1.5 billion cashflow.
In the $65B in Series H funding at $965B post-money valuation they said their run-rate revenue crossed $47B annualised.
With those numbers, there can be sound financial reasons for wanting to just get rid of the lawsuit.
Also if it ends up that other competitors also need to pay $1.5 billion, then maybe that does or doesn't have a competitive advantage.
Anthropic's business and legal strategies are not public. I would expect there to be multiple legs/reasons for settlement even for a decision below 1%. Trying to create a single narrative is what us spectators do.
I'm telling you what the law plainly says. I personally think that the exceptions for fair use made sense before the transformer was invented, and they still make sense now.
The two main problems with copyright also haven't changed: copyright lasts too long and is too expensive to defend.
I don't think anyone is really dismissing it, just pointing out the audacity of complaining about distillation after stealing so much themselves is comical.
The amount of original, copyrightable and trademarkable IP actually created by the AI labs themselves is dwarfed by their staggeringly vast infringement activities.
I generally agree, in the same sense that it's "fair" for the US and China to spy on each other. It's not a moral outrage, but it is something that the targets can and should try to prevent.
Except it’s not just a dump of the internet, which Moonshot also did themselves (and probably used even more pirated content as laws in China are different without any recourse for the entire world). I don’t know why this is so unclear to folks.
Chinese IP law is actually quite solid. You do have to register your trademarks and copyrights properly in China, and then lawsuits have to be filed appropriately according to Chinese law. Is that a problem?
I'm by no means taking the side of the AI companies, but it's possible that Anthropic "added value" to the data they harvested. Stealing that does seem kind of uncool.
Regardless, it was always inevitable—will continue to happen.
I wonder how the "added value" argument can apply to Anthropic and not to the Kimi team. If Claude's value is that you don't have to pay a team of slow expensive subject-matter experts, and Kimi's value is that you don't have to pay Claude, it just seems like the same thing.
Valuation is hard to perform when it's deep inside a black box. Ther "API" may be easier to evaluate. The problem with this angle is that Moonshot is actually producing _better_ value from Anthropic's blackbox.
Technically, providing better value from your competitor's private holdings could be theft (of trade secrets), but might it also be fair use? "Schrodinger's IP" be damned.
I don't think the 1.5B settlement has resolved this. The 2 cases need to be merged!
The issue seems to be the US only likes competition when it is winning.
Markets in Asia are meant for cheap labor and resources, they're not meant to actually compete. /s
It turns out U.S. law doesn’t have jurisdiction across the entire world, nor does Anthropic and OAI’s rather blatant attempt to buy government influence.
For example, if OpenAI / Anthropic were actually open, other US labs could be building near-frontier open weights models by distilling off OpenAI / Anthropic. But because US companies don't want to be sued, US labs who obey terms of service, will be at a disadvantage to Chinese peers.
Maybe US labs need to just not care and distill from OpenAI / Anthropic anyways?
It's about the claim of whether these companies could develop a similarly powerful model without larger companies building their own first, which is an important point, and it's likely not the case.
It's also about the larger companies explaining why they can't be as efficient, of course they can't, they're not just ripping the outputs of another model that someone else invested billions to train.
Why would that matter? OpenAI or whatever frontier lab couldn't have built their frontier models without the entirety of humanity unknowingly developing their training set for 5000 years.
It would be one thing if Moonshot was breaking into OpenAI servers and stealing trade secrets, but the only thing they are doing is looking at the output of the program, which is exactly the service that OpenAI offers. So, at best, this is a ToS violation. Sucks for the frontier labs I suppose, but live by the sword - die by the sword.
Yeah agreed, from one standpoint I couldn't care less that they did a "distillation attack", but I am interested in knowing if China is able to develop open weight frontier models without the prior existence of a huge model to distill from.
> they're not just ripping the outputs of another model that someone else invested billions to train.
If they payed for inference, doesn't they own the output? So if I pay for a model to generate code, isn't that code mine to do with it whatever I want? Just curious.
Not arguing for the morality of it, but if we're going by the law because that's what you're using in your comment ("don't I own" which only matters wrt the law), then you explicitly accepted a Terms of Use which excludes distillation as a use case.
Now of course they themselves trained on the whole Internet for free, etc.
I see no problem with distillation, on the other hand the complete dismissal of copyright by AI labs is pretty bad, I don’t think we should put them at the same level