17 ms·
LAION, a high school teacher’s free image database, powers AI unicorns
- helsinkiandrew 4y agohttps://archive.ph/UIu69 https://archive.ph/UIu69
- Bostonian 4y ago"That person is high school teacher Christoph Schuhmann, and LAION, short for “Large-scale AI Open Network,” is his passion project. When Schuhmann isn’t teaching physics and computer science to German teens, he works with a small team of volunteers building the world’s biggest free AI training data set, which has already been used in text-to-image generators such as Google’s Imagen and Stable Diffusion. Databases like LAION are central to AI text-to-image generators, which rely on them for the enormous amounts of visual material used to deconstruct and create new images. The debut of these products late last year was a paradigm-shifting event: it sent the tech sector’s AI arms race into hyperdrive and raised a myriad of ethical and legal issues. Within a matter of months, lawsuits had been filed against generative AI companies Stability AI and Midjourney for copyright infringement, and critics were sounding the alarm about the violent, sexualized, and otherwise problematic images within their datasets, which have been accused of introducing biases that are nearly impossible to mitigate."
- ftxbro 4y agoSomething about this reminds me a lot of Michael Hart and Project Gutenberg.
- GaggiX 4y ago>Tools like Dall-E-2 and Stable Diffusion have been criticized for reproducing racial stereotypes even when a text prompt doesn’t imply the subject’s race. Such biases were why Google decided not to release Imagen, which had been trained on LAION. This is so ingenuous and naive to think that the reason Google didn't release Imagen is because of racial stereotypes.
- GuB-42 3y agoGoogle absolutely don't want to be accused of racial stereotyping. Remember, Google is in the business of advertising, and the last thing most brands want is to be labeled as racist. Anything Google does that can be interpreted as racist is going to cost them, as advertisers pull away to avoid the association.
- atleastoptimal 4y agoIt's what Google claimed and thus the eternally true correct reason and its very problematic if you think otherwise
- astrange 4y agoIf there's a reason Google doesn't want to talk about, it's that the biggest legal risk of AI image generation is that it can be used for revenge porn and CSAM. I don't think it's problematic to be worried about that!
- ImprobableTruth 4y agoThey could just filter it like Midjourney or DALL-E.
- astrange 4y agoHow do they know they can "just" do something that came out a year after they invented Imagen? Especially when getting it wrong once might mean Europe makes AI illegal. Making things illegal is their favorite hobby.
- ImprobableTruth 4y agoMidjourney and DALL-E were both announced and in closed beta a month before Imagen was announced, with filters already in place.
- anonylizard 4y ago
- janmo 4y agoWe need the same with text to train Large Language Models
- simonw 4y agoThat's pretty much what Red Pajama is: https://simonwillison.net/2023/Apr/17/redpajama-data/ https://simonwillison.net/2023/Apr/17/redpajama-data/
- circuit10 4y agohttps://pile.eleuther.ai/ https://pile.eleuther.ai/
- angrais 4y agoIt contains captions. That's the alt text. LAION is used for train LLMs.
- 1letterunixname 4y agoOne could also write an image crawler to execute from AWS. The key is finding quality source material. It may become necessary to: 1. Fly drones around to gather data in the world at present time 2. Direct humans to take high-quality photos of particular subjects. Fiverr for photography
- ramraj07 4y agoYou mean like what google did with street view and convincing everyone to upload their photos on to maps?
- bradgessler 4y agoThis will be the new captcha challenge—“take a picture of 14 crosswalks” It will send you a link to an app that only runs on an iPhone so it can verify you actually took the pics from the phone.
- quickthrower2 4y agoThat would not be a11y friendly!
- logi2mus 4y agoWhy exactly, they should know how to use crosswalks by that age.
- 4y ago
- 1letterunixname 4y agoText-to-visual LMs are making children's book illustrators go the way of harpsichord tuners and whale oil hunters.
- milofeynman 4y agoIs there an example of a generated children's book that maintains style throughout?
- totetsu 4y agoIf you take the point of view, that children’s media is only for kids, so quality doesn’t really matter. The best children’s book authors and illustrators are ones who see children as the most important audience of all and deserving of the most careful work.
- ftxbro 4y agoMaybe the idea is that most children's books are illustrated, compared to other books.
- p-e-w 4y agoExcept that the better outputs from the better Stable Diffusion models make 99% of professional illustrators look like they're drawing stick figures by comparison. If you still believe that "AI generated = bad quality", you haven't seen some of the stuff that has popped up lately. SD is still new, so understandably it's taking a while for people to figure out how to use it properly, but some users are clearly getting the hang of it now, and the results are mind-blowing.
- strkitten 4y agohttps://xkcd.com/2347/ https://xkcd.com/2347/
- vrglvrglvrgl 4y ago[dead]
- pxoe 4y ago"free database", except the images in it may not be so "free". even their "open" 400M dataset makes a caveat like "if we found a CC license, we put it there, otherwise, well, we just don't know.", so there's still a bunch of data with no license information, that got included anyway. their 5B dataset simply does not have license metadata, at all. a little bit misleading, a little bit malicious. completely enabling copyright and license violations of data it 'only links to'.
- dvt 4y ago> Within a matter of months, lawsuits had been filed against generative AI companies Stability AI and Midjourney for copyright infringement [...] But these aren’t Schuhmann’s concerns. He just wants to set the data free. Maybe it should be one of his concerns? I'm not sure how he would feel about someone stealing his work and commercializing it.
- vasco 4y agoHe is not commercializing it as far as I know.
- dvt 4y ago"Hey man, I just make the meth, I don't sell it." The point I'm trying to get across is that, at least to me, it seems incredibly irresponsible and borderline unethical.
- jonhohle 4y agoDepending on views about drugs and personal liberty, that statement may be neither irresponsible nor unethical.
- moomoo3000 4y agoThese are all images that were posted to the internet right? That place where anyone can download things freely?
- dvt 4y ago> That place where anyone can download things freely? This is incredibly reductive. Property rights exist on the internet, too. Like how someone can own a domain name or the copyright to a photograph or short story.
- vasco 4y agoSo say what you think is unethical, don't stretch the truth just because you think you are right. He is not commercializing anything and kept his job as a high-school teacher.
- netuffresche 4y agoIf someone is interested, I interviewed him recently. The interview is in german, though. https://entwickler.de/machine-learning/laion-open-source-ai https://entwickler.de/machine-learning/laion-open-source-ai
- w-m 4y agoThank you! The Bloomberg article piqued my interest, but it is actually very light on the details of the LAION project - it quickly moved into the same old license and bias discussions and left me wanting. Your interview is much more interesting, it's great to learn a more about the person behind it and his motivations.
- netuffresche 4y agoThanks! Really happy to hear that. He's a very fascinating guy and LAION is an awesome project, eventhough the discussion about copyright and data sets is important.
- epups 4y agoI wonder how it came to be that such a critical component of a multi billion industry relies on something so amateurish as LAION. This is no offense to the author at all, who organised a gigantic effort which we now see is very valuable. But I would imagine a company like Google could do a much better job in no time, simply due to expertise and resources. I guess the answer is legal liability.
- Maken 4y agoForever relevant: https://xkcd.com/2347/ https://xkcd.com/2347/
- lisasays 4y agoBy all indications, it's definitely a solid engineering project. On what basis should we deride it as "amateurish"? Because it originates in the public sector? As opposed to the private sector (where, as we know, everything is done to the highest possible engineering standards)?
- epups 4y agoI know the word sounds demeaning, and I didn't mean to criticise in on technical grounds. I find it an extremely impressive project. I meant that it does not have the refinement you would expect for such a critical tool. A substantial portion of LAION is composed of duplicates. If you have ever browsed it, you will find that many annotations are quite basic and in some cases incorrect. In ChatGPT's case we know there was a small army of people going through their dataset to filter and refine issues that are presumably similar to those.
- wodenokoto 4y agoWho is not to say google et al doesn’t have an in house army that re-annotates images with a low score and store it in a parallel dataset?
- dahwolf 4y ago"I guess the answer is legal liability." Indeed. The answer is in the article. He gets offered jobs all the time, but nobody offers to buy the data itself. Clearly because nobody wants to own it, it's all plausible deniability.
- kleiba 4y agoThe group used raw HTML code collected by the California nonprofit Common Crawl to locate images around the web and associate them with descriptive text. It does not use any manual or human curation. Does the Common Crawl data already take care of the copyright issue? Or else how does the LAION crawler deal with that problem? I mean, it's not too hard to write an image crawler. Also, scaling it up it a bit of a challenge, but it's a technical one. But the real difficulty is how to deal with all the legal strings attached...
- Borrible 4y agoWell, at least one content producer is pissed about it. Sorry, only in German: https://www.alltageinesfotoproduzenten.de/2023/04/24/laion-e-v-macht-ernst-schadensersatzforderung-an-urheber-fuer-ki-trainingsdaten/ https://www.alltageinesfotoproduzenten.de/2023/04/24/laion-e... And translated via Google translate: https://www-alltageinesfotoproduzenten-de.translate.goog/2023/04/24/laion-e-v-macht-ernst-schadensersatzforderung-an-urheber-fuer-ki-trainingsdaten/?_x_tr_sl=auto&_x_tr_tl=de&_x_tr_hl=de&_x_tr_pto=wapp https://www-alltageinesfotoproduzenten-de.translate.goog/202...
- Silverback_VII 4y agoAfter reading it, it seems that LAION is merely doing the dirty work (going as far as intimidating the creators of the materials they use) so that other companies can utilize copyrighted material without taking too much risk. Not sure if this teacher is a hero...
- nprateem 4y agoIt's not really a difficulty though is it, for VC funded unicorns who ask forgiveness and move fast. When the law catches up, if ever, they'll pay a slap on the wrist while sitting on their superyachts
- layer8 4y agoLAION doesn’t publish the images, only the metadata, which includes the original image URLs. Anyone who wants to make use of the image set has to re-download the images from the original sources [0], and is then liable for actually using the images. Arguably, the alt text contained in the metadata is still subject to copyright though. [0] see also https://news.ycombinator.com/item?id=35681085 https://news.ycombinator.com/item?id=35681085
- chwa982 4y ago[flagged]
- sailorganymede 4y agoSounds like the website you are currently on does the trick.
- dang 4y agoYou're welcome in this community, but can you please stop posting repetitive comments? The idea here is curious conversation on topics of intellectual interest. https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html
- whywhywhywhy 4y ago> To build LAION, founders scraped visual data from companies such as Pinterest, Shopify and Amazon Web Services — which did not comment on whether LAION’s use of their content Pinterest, "their content"... not sure I agree with that but if we're going to use the logic that things saved to Pinterest by it's users becomes Pinterest's content then isn't LAION doing the same thing and the content becomes LAION's content when it's saved to their database of images...
- mcv 4y agoIf LAION republishes people's copyrighted content, that sounds like a pretty blatant copyright violation (edit: it's not; see below). Sounds like all the artists unhappy that their art is being used to train these AI systems, should be talking to LAION to have their content removed from the dataset. Edit: Apparently LAION doesn't republish the content, only the metadata, so it's not a copyright violation. Still, it would be nice if got permission or offered a way for artists to be excluded from the data set.
- Filligree 4y ago> Still, it would be nice if got permission or offered a way for artists to be excluded from the data set. It obeys robots.txt, and the user-agent is documented.
- wildrhythms 4y agorobots.txt generally only addresses permission to crawl; not permission to reproduce the content.
- tyingq 4y ago>Apparently LAION doesn't republish the content, only the metadata "LAION datasets are simply indexes to the internet, i.e. lists of URLs to the original images together with the ALT texts found linked to those images." I'm surprised link rot doesn't make this a big problem.
- seydor 4y ago" ...in america Unlikely to power anything for-profit in europe
- archerx 4y agoI think Stable Diffusion may use part of LIAON and stability ai is based in London.
- lisasays 4y agoFor his part, Schuhmann hasn’t profited from LAION and says he isn’t interested in doing so. “I’m still a high school teacher. I have rejected job offers from all different kinds of companies because I wanted this to stay independent,” he said. Brave is this man for going against the current Zeitgeist (and the obvious financial temptations).
- ChicagoBoy11 4y agoAs a tech-inclined person working at a private school with a healthy amount of time to putter and work on my own projects, I feel like I need to get busy!