3 ms·
> It’s trained on 15T tokens. So how many did you provide that were genuinely novel? Are we suggesting that we should ditch creators' rights and instead value
by logifail 2y ago
> It’s trained on 15T tokens. So how many did you provide that were genuinely novel?
Are we suggesting that we should ditch creators' rights and instead value intellectual property along the lines of "I should be able to copy all your stuff as long as long as I copy lots of other stuff too, and give it all away for free or almost free...?"
- ripped_britches 2y agoNo not at all, just that their damages are going to be measurably fairly low
- llm_trw 2y agoThe question if training an llm is fair use is one that will have to be answered by the courts.
- yencabulator 2y agoThe point (often) is to stop the practice, not to ask money for it. Like, you get fined for speeding, but if you keep speeding you'll get you're license revoked, and if you keep driving after that you get jail time. The payment required is punitive, but the point is to stop you.
- llm_trw 2y agoAgain, that an opinion that will need to be tested in the courts. You will have to explain why Hunter Thompson copying every word of every Hemingway novel isn't copyright infringement, but a computer doing the same is.
- yencabulator 2y agoThree things: 1. Humans are not machines; arguments saying because a human can learn LLMs must be allowed to copy is not interesting. 2. Did Thompson publish the work? It sounds like you're referring to an activity Thompson did in private, to improve his skills as an author. Meanwhile, lawsuits are alleging that LLM services reproduce copyrighted materials. 3. What can be fair use at small scale is no longer fair use at large scale.
- llm_trw 2y agoIf you think that pretraining is copying then your opinions are irrelevant.
- yencabulator 2y agoIf you think courts follow your personal opinion and technical definition, you're silly. The reality is we don't yet know how the courts will decide.
- logifail 2y ago> Hunter Thompson copying every word of every Hemingway novel isn't copyright infringement I make a high-res photo of a banknote. I can print that out at home. The bad stuff starts at the step after that...
- marssaxman 2y agoThat sounds like a pretty good deal to me - but I've always believed that the entire concept of "intellectual property" does overall more harm than good.
- logifail 2y ago> I've always believed that the entire concept of "intellectual property" does overall more harm than good It's fairly broken, but on balance it seems the creators are the ones getting screwed. I did years of research in a scientific lab which resulted in <drum roll> 2 (yes two ... count them) peer-reviewed papers. My colleagues and I did the work, wrote up the damned papers, yet to get them published we had to sign over copyright to what I'd now suggest is essentially a rent-seeking scientific publishing mafia. All a long time ago, but I never had (and still don't have) the ability to either legally download or legally redistribute my own work...
- jncfhnb 2y agoYes. As long as you don’t reproduce other people’s stuff specifically.
- latexr 2y agoWhich they do. That’s what the New York Times lawsuit is about. And in Meta’s case, they went specifically out of their way to remove the copyright notices to hide their actions.
- jncfhnb 2y ago“They” might be doing that. But this is not intrinsic to LLM usage
- latexr 2y agoThe conversation is specifically about OpenAI and Meta.
- jncfhnb 2y agoI disagree. This thread seems to be about LLMs on a fundamental level.
- latexr 2y agoThe submission is about Meta, and the comment that started the thread specifically mentioned OpenAI and Meta and no other LLMs or providers.
- jncfhnb 2y agoAnd the thread is about how LLM training interacts with copyright. Not whether OpenAI or Meta coincidentally blatantly copied other works.
- scotty79 2y agoOf course. The alternative is that creators dictate the price for any of the infinite number of zero cost copies which is and always has been ridiculous.
- tomrod 2y agoI think the statistical arguments cover this.