7 ms·
For all the leaks on: Secret projects, novelty training algorithms not being published anymore so as to preserve market share, custom hardware, Q* learning, int
by belter 3y ago
For all the leaks on: Secret projects, novelty training algorithms not being published anymore so as to preserve market share, custom hardware, Q* learning, internal politics at companies at the forefront of state of the art LLMs...A thunderous silence is the lack of leaks, on the exact datasets used to train the main commercial LLMs.
It is clear OpenAI or Google did not use only Common Crawl. With so many press conferences why did no research journalist ask yet from OpenAI or Google to confirm or deny if they use or used LibGen?
Did OpenAI really bought an ebook of every publication from Cambridge Press, Oxford Press, Manning, APress, and so on? Did any of investors due diligence, include researching the legality of the content used for training?
- ethbr1 3y agoWould be fascinated to hear from someone inside on a throwaway, but my nearest experience is that corporate lawyers aren't stupid. If there's legally-murky secret data sauce, it's firewalled from being easily seen in its entirety by anyone not golden-handcuffed to the company. They may be able to train against it. They may be able to peek at portions of it. But no one is downloading-all.
- alfiedotwtf 3y agoWhy isn't robots.txt enough to enforce copyright etc? If NYT didn't set robots.txt properly, is their content free-for-all? Yes I know the first answer you would jump to is "of course not, copyright is the default", but it's almost 2024 and we have had robots.txt as industry de jure to stop crawling.
- cj 3y agorobots.txt is not meant to be a mechanism of communicating the licensing of content on the page being crawled nor is it meant to communicate how the crawled content is allowed to be used by the crawler. Edit: same applies to humans. Just because a healthcare company puts up a S3 bucket with patient health data with “robots: *” doesn’t give you a right to view or use the crawled patient data. In fact, redistributing it may land you in significant legal trouble. Something being crawlable doesn’t provide elevated rights compared to something not crawlable.
- cj 3y agoFurthering the S3 health data thought exercise: If OpenAI got their hands on an S3 bucket from Aetna (or any major insurer) with full and complete health records on every American, due to Aetna lacking security or leaking a S3 bucket, should OpenAI or any other LLM provider be allowed to use the data in its training even if they strip out patient names before feeding it into training? The difference between this question or NYT articles is that this question asks about content we know should not be available publicly online (even though it is or was at some point in the past). I guess this really gets at “do we care about how the training data was obtained or pre-processed, or do we only care about the output (a model’s weights and numbers, etc)
- stuartjohnson12 3y agoHIPAA is about more than just names. Just information such as a patient's ZIP code and full medical history is often enough to de-anonymise someone. HIPAA breaches are considered much more severe than intellectual property infringements. I think the main reason that patients are considered to have ownership of even anonymised versions of their data (in terms of controlling how it is used) is that attempted anonymisation can fail, and there is always a risk of being deanonymised. If somehow it could be proven without doubt that deanonymising that data wasn't possible (which cannot be done), then the harm probably wouldn't be very big aside from just general data ownership concerns which are already being discussed.
- Sai_ 3y ago
- cogman10 3y agoWe all remember when Aaron Swartz got hit with a wire tapping and intent to distribute federal crime for downloading JSTR stuff right? It's really disgusting, IMO, that corporations that go above and beyond that sort of behavior are seeing NO federal investigations for this sort of behavior. Yet a private citizen does it and it's threats of life in prison. This isn't new, but it speaks to a major hole in our legal system and the administration of it. The Feds are more than willing to steamroll an individual but will think twice over investigating a large corporation engaged in the same behavior.
- SideQuark 3y agoCircumventing computer security to copy items en masse to distribute wholesale without transformation is a far cry from reading data on public facing web pages.
- cogman10 3y agoHe didn't circumvent computer security. He had had a right to use the MIT network and pull the JSTR information. He certainly did it in a shady way (computer in a closet) but it's every bit as arguable that he did it that way because he didn't want someone stealing or unplugging his laptop while it was downloading the data. He also did not distribute the information wholesale. What he planned on doing with the information was never proven. OpenAI IS distributing information they got wholesale from the internet without license to that information. Heck, they are selling the information they distribute.
- SideQuark 3y ago> right to use the MIT That right ended when he used it to break the law. It was also for use on MIT computers, not for remote access (which is why he decided to install the laptop, also knowing this was against his "right to use"). The "right to use" also included a warning that misuse could result in state and federal prosecutions. It was not some free for all. > and pull the JSTR information No, he did not have the right to pull en masse. The JSTOR access explicitly disallowed that. So he most certainly did not have the "right" to do that, even if he were sitting at MIT in an office not breaking into systems. > did it in a shady way The word you're looking for is "illegal." Breaking and entering is not simply shady - it's illegal and against the law. B&E with intent to commit a felony (which is what he was doing) is an even more serious crime, and one of the charges. > he did it that way because he didn't want someone stealing or unplugging his laptop Ah, the old "ends justifies break the law" argument. Now, to be precise, MIT and JSTOR went to great lengths to stop the outflow of copying, which both saw. Schwartz returned multiple times to devise workarounds, continuing to break laws and circumvent yet more security measures. This was not some simply plug and forget laptop. He continually and persistently engaged in hacking to get around the protections both MIT and JSTOR were putting in place to stop him. He added a second computer, he used MAC spoofing, among other things. His actions started to affect all users of JSTOR at MIT. The rate of outflow caused JSTOR to suffer performance, so JSTOR disabled all of MIT access. Go read the indictment and evidence. > OpenAI IS distributing information they got wholesale No, that ludicrous. How many complete JSTOR papers can I pull from ChatGPT? Zero? How many complete novels? None? Short stories? Also none? Can I ask for any of a category of items and get any of them? Nope. I cannot. It's extremely hard to even get a complete decent sized paragraph from any work, and almost certainly not one you pre-select at will (most of those anyone produces are found by running massive search runs, then post selecting any matches). Go ahead and demonstrate some wholesale distribution - pick an author and reproduce a few works, for example. I'll wait. How many could I get from what Schwartz downloaded? Millions? Not just even as text - I could have gotten the complete author formatted layout, diagrams, everything, in perfect photo ready copy. You're being dishonest in claiming these are the same. One can feel sad for Schwartz outcome, realize he was breaking the law, and realizing the current OpenAI copyright situation is likely unlike any previous copyright situation all at the same time. No need to equate such different things.
- lumost 3y agoChatGPTs birth as a research preview may have been an attempt to avoid these issues. It would have been unlikely to trigger legal anger for a free product which few use. When usage exploded, the natural inclination would be to hope for the best. Google may simply have been obliged to follow suit. Personally, I’m looking forward to pirate LLMs trained on academic content.
- graphe 3y agoIs there already a dataset? Before llama Facebook had one too I forgot what it was called.
- devindotcom 3y agofor what it's worth, i asked altman directly and he denied using libgen or books2, but also deferred to murati and her team on specifics. but the Q&A wasn't recorded and they haven't answered my follow-ups.
- belter 3y agoReally? Because the GPT-3 paper talks about "...two internet-based books corpora (Books1 and Books2)..." (see pages 8 and 9) - https://arxiv.org/pdf/2005.14165.pdf https://arxiv.org/pdf/2005.14165.pdf Unclear what that corpora might be, or if its the same books2 you are referring to.
- simonw 3y agoMy guess is that this poster meant books3, not books2. books1 and books2 are OpenAI corpuses that have never (to my knowledge) had their content revealed. books3 is public, developed outside of OpenAI and we know exactly what's in it.
- devindotcom 3y agosorry, books3 is indeed what I meant.
- jprete 3y agoWhy would he know the answer in the first place?
- Jensson 3y agoThe legal liabilities of the training data they use in their flagship product seems to be a thing the CEO should know.
- hhsectech 3y agoI'm not for or against anything at this point until someone gets their balls out and clearly defines what copyright infringement means in this context. If you give a bunch of books to a kid all by the same author and then pay that kid to write a book in a similar style and then I go on to sell that book...have I somehow infringed copyright? The kids book at best is likely to be a very convincing facsimile of the original authors work...but not the authors work. It seems to me that the only solution for artists is to charge for access to their work in a secure environment then lobotomise people on the way out. The endgame seems to be "you can view and enjoy our work, but if you want to learn or be inspired by it, thats not on"
- flextheruler 3y agoI think you’re skipping over the problem. In your example you owned the work you gave to the person to create derivatives of. In a more accurate example you would be stealing those books and then giving them to someone else to create derivatives.
- slyall 3y agoHow about if I borrowed them from the library and gave them to the kid to read? How about if I got the kid to read the books on a public website where the author made the books available for free?
- graphe 3y agoIronically these artists cant claim to be wholly original as they were certainly inspired. Artists that play live already "lobotomize" people on their way out since it's not easy to recreate an experience and a video isn't the same if it's a good show. Artists that make easily reproducible art will circulate as these always have along with AI in a sea of other jpgs.
- twoodfin 3y agoThere are two problems with the “kid” analogy: a) In many closely comparable scenarios, yes, it’s copyright infringement. When Francis Ford Coppola made The Godfather film, he couldn’t just be “inspired” by Puzo’s book. If the story or characters or dialog are similar enough, he has to pay Puzo, even if the work he created was quite different and not a literal “copy”. b) Training an LLM isn’t like giving someone a book. Among other things, it involves making a derivative copy into GPU memory. This copy is not a transitory copy in service of a fair use, nor likely a fair use in itself, nor licensed by the rights-holder.