Comment by remus
1 month ago
While I agree on a moral level, I think there is a distinction to be made. Training a SOTA model takes a huge amount of resources and expertise so the people doing the training are adding a lot of value along the way. I think this is much less true for distillation (which is kind of the whole point).
ed: to clarify, I totally agree that a huge chunk of the value in LLMs is coming from the source material. My point was just that training an LLM takes more resources and expertise than distilling from an existing LLM so I don't think the equivalence between training and distilling is entirely justified.
I like this comment because its argument only makes sense if you assume that the entire world's output of books and art did not require a huge amount of resources and expertise to make, nor did it add any value.
It's the most CS-major take ever!
If turning other peoples copyrighted work into a model is transformative enough to be protected then so is distilling that model into a different, better, model.
The models were built using copyrighted works, so why can't models be built using other models?
21 replies →
there is a major god complex here.
MBAs and non technical managers = inept Catbert-type charlatans.
Software engineers, devs, etc = geniuses capable of mastering any domain, innate ability to be right on any topic.
I don't think that's what it's saying at all. It's saying that there's a level of creativity in model creation that isn't present in distillation.
Maybe, but it's not like their AI is likely to repeat it back verbatim so it's unlikely to be a copyright violation. It seems like at most, they would be breaking Anthropic's terms of service?
Or maybe they're going through an intermediary "transfer station" that's breaking terms of service:
https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
5 replies →
>a level of creativity in model creation that isn't present in distillation.
the same argument - a level of creativity in the world knowledge creation that ins't present in the model training on that knowledge.
Or in other words - model creation and training is just a distilling of the world knowledge.
2 replies →
Yes, this is what I was getting at.
5 replies →
No? They outright say the opposite!
Like look, I'm not a native speaker, sure. But I think when someone says "value add", that means there was value there (which you claim they're rhetorically erasing), and then that was added to. Under no interpretation of this phrase do I get an erasure of prior value.
So certainly, as long as words mean anything, no, they absolutely did not say or suggest what you claim they did, and what you extract a thus unreasonable amount of obnoxious schadenfreude from, while throwing in a cheap insult for funsies at the end.
It's the second time I feel compelled to reach for this just today: https://i.kym-cdn.com/photos/images/original/002/659/979/108...
This is a misrepresentation though.
The LLM output, is not the same as the input - there is value add.
Of course works used as raw inputs to LLMs required work and are reasonably subject to IP concerns - but they are different.
It's possible that the LLM makers 'owe' the content creators that created the content they used to make their products - it's an interesting but separate question.
We could very well end up where content IP is protected, LLM output is not and visa versa with reasonable legal founding, doubtful but plausible.
Lossly storing IP in LLM itself, and using IP for training (so it’s lossly stored in LLM), without licensing these works or otherwise following license agreements (eg GPL) is infringement. Using then this product for commercial activity is a smoking gun.
1 reply →
> but they are different.
How, and why?
> We could very well end up where content IP is protected, LLM output is not and visa versa with reasonable legal founding, doubtful but plausible.
That is the current state of legal rulings - LLM output is public domain, not copyrightable.
2 replies →
> raining a SOTA model takes a huge amount of resources and expertise
Writing books, building Wikipedia, and answering questions on online forums takes a lot of resources and expertise that scraping didn't. So at the very least, we're already one rung down the "maybe you should've asked" ladder.
I suspect that, in aggregate, all of the informational output of humanity prior to 2020 has taken more resources to produce than the last few years of LLM research.
I don't know man. This reads like "yeah we stole your grain, but making bread is hard."
It sure is, but it doesn't matter. Whatever position that generates more economic activity is declared legal using some nonsense retconned logic "because we said so".
Why is it less true for distillation? Everyone technically has access to Fable but Moonshot came up with the model. How can you objectively claim one is adding value while the other is not?
If that is the whole point you need to clarify why this is the case on an objective level.
I would say building a comparable model using any means necessary (just like what Anthropic and OAI did) at a lower cost is actually more valuable to soceity and Monshoot is arguably generating more value with less.
Probably not as much effort as writing books and creating art the models were trained on.
The value of LLM's come from replacing what generated its training data.
If the distilled model is cheaper, then it's just LLM's getting LLM'ed.
Still, AFAIK Kimi's architecture (just like that of other LLMs from Chinese labs) is different from those of OpenAI and Anthropic's model in a nontrivial way. So the expertise is still there, and I guess resource use too (although Chinese labs tend to optimize this, thanks to the restrictions they have on GPU use).
EDIT: just wanted to add that resource optimization is usually where the contribution of Chinese labs is, so you shouldn't reaad the above parenthesis as a negative comment.
As an author, that's a genuinely disheartening thing to read.
It took me a year to write a book. It took OpenAI and Anthropic a fraction of a second to ingest it. Do you understand now why I give zero shits if it takes Anthropic a billion to train a model, and Moonshot 10k in API cost to distill it?
> training an LLM takes more resources and expertise than distilling from an existing LLM
This is not automatically true. Training and distillation use the same underlying infra and method and there is no intrinsic differences in between.
They add value on top of other people’s work, often against licensing, and then commercialize this product, ie profiting from making a product out of other people’s IP.
>Training a SOTA model takes a huge amount of resources and expertise so the people doing the training are adding a lot of value along the way.
producing the entire body of human knowledge that Silicon Valley companies absorbed like the Borg did not just take more resources but also a fair amount of blood and sweat, certainly more than the LLM so on that front that comparison also seems entirely justified.
I'm sure it takes a lot of time and resources to plan and pull off an epic heist but it is unusual to see people like Thomas Crown being accused of creating value, as they're usually accused of committing theft.
Yeah, there's a difference. One party spends a bunch of resources doing something illegal and extremely immoral. The other party spends little money doing something legal and morally neutral.
You can argue that reverse engineering anything is as hard if not harder than engineering something. I can’t imagine distillation is any different.
Distillation is objectively easier than training a model from scratch, that's why all these Chinese labs are doing it.
Training a model is objectively easier than generating the sum total of human creative output prior to 2020. That's why the big labs are doing it. What's the difference here?
1 reply →