The 40-year-old teacher and trained actor helped find LAION two years ago.
The word “LAION” is scribbled in pencil on a mailbox in front of a suburban house in the northern German city of Hamburg. This is the only indication that the person behind the massive data-gathering operation at the heart of the artificial intelligence boom that has captured the world’s attention belongs to this house.
That person is high school teacher Christoph Schumann, and his passion project is LAION, which stands for “Large-scale AI Open Network”. When Schuhmann isn’t teaching physics and computers to his teenagers in Germany, he works with his team of small volunteers to build the world’s largest free AI training dataset. I’m here. This dataset is already used by text-to-image generators such as his Imagen and Stable at Google. diffusion.
Databases like LAION are at the heart of AI’s text-to-image generators, which convert vast amounts of visual material used to deconstruct and create new images. The debut of these products late last year was a paradigm shift. It has overheated his AI arms race in the tech sector and caused a myriad of ethical and legal issues. Within months, copyright infringement lawsuits were filed against generative AI companies Stability AI and Midjourney, with critics alleging violent, sexual, and other problematic images in the dataset. I was ringing the alarm. Almost impossible to mitigate.
But these are not Schumann’s concerns. he just wants to release the data.
big language
The 40-year-old teacher-trained actor helped find LAION two years ago after hanging out on a Discord server for AI enthusiasts. His DALL-E, the first iteration of OpenAI’s His DALL-E, is a Deep His learning model that generates digital images from verbal prompts. For example, I create an image of a pink chicken sitting on a couch for such a request. Inspired and concerned to encourage big tech companies to make more data their own.
“We quickly realized that if this concentrated in one, two, or three companies, it would have a very negative impact on society,” Schumann said.
In response, he and others on the server decided to create an open-source dataset to help train image-to-text diffusion models. It’s a multi-month process similar to teaching someone a foreign language with millions of flash cards. The group used his raw HTML code collected by Common Crawl, a California nonprofit, to find images on the web and associate them with descriptions. We do not use manual or human curation.
Within weeks, Schumann and his colleagues had 3 million image-text pairs. Three months later they released his dataset of 400 million pairs. That number now exceeds 5 billion, making LAION the largest free dataset for images and captions.
As LAION’s reputation grew, the team worked unpaid and received a one-time donation from machine learning company Hugging Face in 2021. One day, a former hedge fund manager appeared in a Discord chat.
Emad Mostaque offered to bear the cost of computing power. No conditions attached. He wanted to start an open source generative AI business and wanted to utilize LAION to train his product. The team initially scoffed at the suggestion and thought he was a weirdo.
“We were very skeptical at first, but in about four weeks we had access to GPUs in the cloud that typically cost between $9,000 and $10,000.”
When Mostaque launched Stability AI in 2022, he used LAION’s dataset for his flagship AI image generator, Stable Diffusion, and hired two of the organization’s researchers. A year later, the company is now seeking a $4 billion valuation, largely thanks to data provided by his LAION. Schuhmann said he has not and does not intend to profit from LAION. “I’m still a high school teacher. I refused job offers from all kinds of companies because I wanted to remain independent,” he said.
new oil?
Many of the images and links in databases like LAION have been sitting in plain sight on the web for decades in some cases. The bigger and more diverse the dataset, and the higher the quality of the images in it, the sharper and more accurate the images generated by AI, so the AI boom will reveal its true value. it took time.
That recognition raises a number of legal and ethical questions about whether publicly available material can be used to feed databases and, if the answer is yes, whether creators should be compensated. was filed.
To build LAION, the founders collected visual data from companies such as Pinterest, Shopify and Amazon Web Services (who did not comment on whether LAION’s use of their content violated their terms of service). bottom. Also YouTube thumbnails, images from portfolio platforms such as DeviantArt and photos from his websites for governments including EyeEm, the US Department of Defense, and content from news sites such as The Daily Mail and The Sun.
Ask Schuhmann, and he says anything freely available online is fair game. However, there are currently no regulations on AI in the European Union. The AI law, whose language is expected to be finalized early this summer, does not specify whether copyrighted material can be included in big data sets. Rather, as lawmakers are debating whether to include a clause requiring the companies behind AI generators to disclose the material contained in the datasets used to train their products, the Gives authors the option to take action.
MEP Dragos Tudrash told Bloomberg that the basic idea behind the provision is simple.
Such restrictions are not a problem for Stability AI, but can be a problem for other text-to-image generators. Examples of how technology companies lock public data. It also upends the current state of data collection.
“It’s become a tradition in this space to think that we don’t need consent, we don’t need people to know it, or they don’t even need to know it. just,” said Abeba Birhane, a senior fellow in Trustworthy AI at the Mozilla Foundation, who studies LAION.
LAION has never been directly sued, but has been named in two lawsuits. One is Stability and Midjourney accusing an artist of using copyrighted images to train their models, and the other is Getty Images accusing Stability that he scraped 12 million images. claimed to have been Used by LAION to train Stable Diffusion.
Since LAION is open source, it’s impossible to know how many other companies are using this dataset. Google acknowledges using LAION to help train Imagen and Parti AI text-to-image models. Schumann believes other big companies are doing the same quietly, simply not disclosing.
the worst web
Sitting in the living room while his son played Minecraft, Schumann likened LAION to a “little research vessel” on top of a “big information technology tsunami,” taking samples of what was below to explore the world. published on
“This is just a small part of what is publicly available on the Internet,” he said of the LAION database. “If you have a $10,000 budget from a donor, even we can do it, so it’s really easy to get.”
But what’s available to the public isn’t always what the public wants. Nor is it legally permissible to view. In addition to his SFW photos of cats and firetrucks, LAION’s dataset includes pornography, violence, child nudity, racist memes, hate his symbols, copyrighted art, and private Contains millions of images of his work collected from his website for companies. Schuhmann said he was unaware that his set of LAION data contained child nudity, but admitted that he had not examined the data in detail. He said he would remove links to that content immediately.
Schuhman consulted a lawyer and ran an automated tool to filter out illegal content before he started building the database, but there’s more to learning from it than sanitizing LAION’s holdings. Not interested. “We could have excluded violence from the data we published, but we chose not to because it would accelerate the development of violence detection software,” he said. LAION provides a deletion form to request removal of the photos, but the dataset has already been downloaded thousands of times.
Offensive content taken from LAION appears to have been merged into Stable Diffusion. Stable Diffusion makes it easy to generate fake IS beheadings and Holocaust imagery, despite the recent enhancements to its filters. Some experts believe such material could create bias in the AI generators themselves.
Such bias is why Google decided not to release LAION-trained Imagen.
When asked for comment, Stability AI said it trained Stable Diffusion on a curated subset of LAION’s database. In an email, the company added that it aimed to “give the model a much more diverse and broad dataset than the original SD” and tried to remove “adult content using LAION’s NSFW filter.” rice field.
Even proponents of open source-based AI warn about the implications of training AI on uncurated datasets. Generative AI tools based on tainted data reflect that bias, according to Yacine Jernite, who leads the Machine Learning and Society team at Hugging Face. “The model is a very direct reflection of what it was trained to do.”
Jernite added that just putting guardrails in place after a product goes live is not enough, as users will always find ways to circumvent security measures. “This is what happens when you use a model trained to emulate what people typically do on the internet and say, ‘Okay, but don’t do that. People will find a way to make it happen,” they said.
Gil Elbaz, founder of the data nonprofit Common Crawl, wondered “if there is a straight line that can be drawn from the training set to what was generated”, and turned the process to museums for inspiration. It was likened to an artist who was prevented from producing what he did. A replica of the work. Instead, he said, “it’s important for society to decide which use cases are legal and which are not.”
Don’t just leave it up to society. As European regulators create legislation to navigate the use of artificial intelligence, the data currently being mined due to the current AI boom has been generated for years in a legal gray area. grappling with the fact that “Without years of accumulated data, AI at this level of complexity would not have been possible,” said Tudorache, who is a member of the European Parliament.
But for Schumann, it’s not the dataset that should be monitored. In his eyes, the worst-case scenario for AI is one in which Big Tech can lock out developers by aligning its tools with regulatory frameworks. “If we try to slow things down and over-regulate, there is a great danger that ultimately only a few large companies will be able to meet all the formal requirements,” he warned.
(Except for the headline, this article is unedited by NDTV staff and published from a syndicated feed.)
