The issue here is not whether Anthropic used Common Crawl, Alibaba also does that.
The issue is that by distilling Claude, Alibaba reuses the IP anthropic used to train the model that's more akin to historical Chinese reverse engineering methods and disrespect of IP
If using Common Crawl or Anna's Archive in your training data is legal, then surely the same is true for using conversations with Claude. I don't see a reasonable framework where training AI on copyrighted data is ok if and only if that data is not generated by AI
(granted, only meta got caught using Anna's Archive, but it seems safe to assume it's common practice. And even if it wasn't, the websites in Common Crawl are still covered by copyright)
The practical implementation of IP? Sure, that's debatable. But the concept of IP is rooted in favoring progress. The thought process being, that if one's intellectual work can be copied and reused and modified and what not without issues, why should anyone invent things anymore? Just wait for the next person to do it and then copy their work, that's way less effort than inventing things yourself. IP aims to protect progress by making sure inventors have actual incentive to invent stuff. They way it's implemented is fundamentalst flawed, I agree, but the concept itself? I'm not so clear on that
> fair bunch of scientific and technological breakthroughs
Outside of military technologies they had massively fallen behind the west by the 80s. Without the western tech they licensed or copied they were permanently stuck in the 50s. Even their crappy cars were licensed copies of cheap European cars from the 60s.
When it comes to consumer electronics, vehicles and a bunch of other things they were comically behind. So it’s really not a good example..
> monopolizes technologies and ossifies development.
As bad as it might be empirical evidence shows that historically a superior system has never existed (it might be feasible but everything that was tried underperformed).
What absolute bollocks. Human ingenuity and innovation is only limited by the greed of elites, not due to something as damaging as "IP."
Good grief. All one has to do is look at how humanity has consistently progressed due iterating on what has existed is how we progress, not whether some corporation that wants to rat fuck us all for a few pts in share value.
> progressed due iterating on what has existed is how we progress
Progress was extremely slow until the 1800s. Coincidentally corporation and modern capitalism in general developed around the same time. Of course I’m not necessarily saying it was the main or direct course since it isn’t exactly possible to create an experiment comparing it to other systems (of course that was tried an failed completely in the USSR, Maoist China and similar places)
It's more complicated than that because Google has been legally displaying other people copyrighted material for years.
In any case there's still a difference between publicly available copyrighted data and whether you can use it for model training, and the innovation around model training, RLHF, etc which you presumably have some interest as a country to allow companies to invest in with some legal protections (like the diff between patent law vs copyright law)
No, I am saying that there is a good chance that for the good of humanity, society decides that for miracle AGI we collectively forfeit copyright in LLM training yet IP protections for model development is still kept.
There are many cases in the early 2000s were copyright protections were relaxed for tech advancements
LLM output is not copyrightable, though? So effectively if you pay for it you can do whatever you want from it. That seems perfectly fair and reasonable.
Historically most evidence seems to point to the contrary.
Amongst other things after the printing press was created it was impossible for anyone who was an author to survive from their work unless they were independently wealthy or had rich patrons.
> Alibaba reuses the IP anthropic used to train the model that's more akin to historical Chinese reverse engineering methods and disrespect of IP
Why is this any worse than Anthropic's disrepect of IP? You've apparently drawn a distinction between the two here, but I'm failing to see what it actually is.
Copyright law and IP law is not the same although everyone seem to conflate the two.
Search engines for example historically ignored copyright law by copying excerpts or serving other site images, it doesn't mean someone copying Google's code has some moral frepass
> Excerpts are often considered fair use, but it depends on country.
That had happened progressively, thumbnails for example were ruled as fair use later on, DMCA safe harbor was a huge gift for tech companies because otherwise it would curtail the ability to create platforms (relaxing copyright protections in exchange of innovation)
> Nobody copied Anthropic's code. They used it's output to train another model. At most they violated some terms of service
Distilling a model is a method that can push the entire market to low margins and prevent companies from making money off such research. It also copies the Anthropic special parts (RLHF and other specific methods) rather than the "copy of the entire web" part
This is similar to what happened with Chinese reverse engineering of American manufacturing or PC clones killing IBM PCs.
Is it in the interest of the USA, probably no, that's why I assume this will be backed by law eventually
> Search engines for example historically ignored copyright law by copying excerpts or serving other site images, it doesn't mean someone copying Google's code has some moral frepass
Not sure that's the best example as they lost that battle and had to pay, eventually it's been codified in law in most countries.
I wish people would stop using Anthropics incorrect use of the term distill. They don’t share logits so you can’t distill. You can generate training data, which doesn’t sound nearly so scary.
Distillation is a specific named technique from a pretty famous paper "Distilling the Knowledge in a Neural Network," by Hinton and others.
What makes it different from just training on any data is that you train the student model on "soft targets" that includes the full output distribution (logits) from the teacher model. Regular training uses one-hot targets and penalizes anything else; distillation will partly reward a student for getting in the distribution. This teaches the student to think like the teacher, not just imitate it.
Generating training data is not distillation, not in the technical sense, and I dislike Anthropic undermining the nomenclature to set a narrative.
I think the popular use of distill (one-hot rather than logits) is not due to Anthropic but due to the DeepSeek-R1 tech report: https://arxiv.org/abs/2501.12948