ai training

Bilingual

Company Offering Printed Books to Train AI Stops After 404 Media Coverage
2 months ago
- ISBNdb deleted a portion of its website that offered to source printed books for AI training after 404 Media's reporting.
- The company walked back its claims, describing the offering as 'a test of market interest' and stating no such service was ever implemented.
- On July 30, ISBNdb added a note to its homepage and news page denying it ever purchased, scanned, or sold books for AI training.
- ISBNdb emphasized it does not train AI models and has never done so, reiterating its role as a provider of book data, not physical books.
AI Firms Are Buying Up Old Books, Then Scanning and Destroying Them
2 months ago
- AI companies like Anthropic are buying millions of old books to scan and then destroy them for training large language models.
- Anthropic's Project Panama used hydraulic cutting machines to remove book spines and slice pages for industrial scanning before discarding.
- Internal documents reveal Anthropic wanted to keep the project secret, acknowledging negative public reception.
- A judge approved a $1.5 billion class action payment for pirated works, but physical book scanning was ruled fair use under the first-sale doctrine.
- Booksellers have mixed feelings: benefiting financially from selling unsold inventory but disliking the destruction of uncommon books.
The Vanishing Page: AI Firms Scan Then Destroy Rare Book Editions
2 months ago
- AI companies are bulk-buying rare and out-of-print books, cutting the spines, scanning the pages, and disposing of the originals.
- Pre-2022 books are preferred as they are considered free of AI-generated content, with orders of up to one million volumes facilitated anonymously by services like ISBNdb.
- A federal judge ruled that digitizing purchased books and destroying physical copies qualifies as fair use, citing the replacement of print with digital without redistribution.
- Anthropic’s Project Panama aimed to destructively scan all books in the world, using hydraulic cutters and industrial scanners, with physical materials recycled or disposed of.
- Booksellers in multiple countries reported unusual bulk orders for niche titles, not matching typical collector patterns, often from anonymous buyers.
- Destructive scanning is faster and cheaper than non-destructive methods, and the elimination of physical copies ensures only one digital version exists, aiding fair-use arguments.
- AI tools like Character.ai and others can rewrite or alter public domain works, raising concerns about digital modification of original texts and the role of physical books as static references.
- Used-book dealers worry about the loss of rare editions with few surviving copies, while some compare the practice to themes in George Orwell’s '1984' where records are destroyed and rewritten.
- No official government or library guidance has been issued, but readers are increasingly buying physical copies to preserve personal records outside digital systems.
AI companies are shredding rare books
2 months ago
- AI companies are buying rare books, scanning them with high-speed machines that cut off spines, and then shredding the originals.
- A service called ISBNdb facilitates bulk orders of up to a million books and keeps buyers anonymous.
- Pre-2022 books are considered premium because they contain no AI-generated text.
- A federal judge ruled the practice as fair use, reasoning that eliminating the original means only one copy exists at a time.
- Anthropic hired the former head of Google Books partnerships to obtain 'all the books in the world.'
- Rare books with few surviving copies—those that survived wars, fires, and centuries—are being destroyed for AI training data.
- ISBNdb acknowledges the negative headline potential but built its business around discreet operations, including NDAs.
- The author argues this practice is worse than previous AI data scraping because the destruction of physical originals is irreversible.
- The legal ruling is expected to accelerate the destruction of rare books for AI training.
Why Would Meta Download So Much Porn?
2 months ago
- Meta is accused of downloading large volumes of porn, nonconsensual celebrity nudes, 3-D-printable handgun blueprints, and hacked passwords.
- Strike 3, a porn producer, is suing Meta for allegedly downloading nearly 3,000 of its copyrighted videos.
- Evidence suggests Meta downloaded content like 'GirlsDoPorn' archives, deepfake celebrity porn, and password lists.
- Meta claims the downloads were for personal consumption, not AI training, but the pattern and volume suggest otherwise.
- Judge Eumi Lee found Meta's arguments unconvincing, citing algorithmically coordinated downloads across multiple networks.
- Meta used BitTorrent to acquire content, which also involved sharing (seeding) the files with others.
- Internal Meta documents show employees were uncomfortable with downloading copyrighted books via BitTorrent for AI training.
- The case highlights tech companies' heavy reliance on mass data collection, often involving piracy.
If AI companies avoid AI slop, shouldn't we?
2 months ago
- AI companies prefer human-generated content for training to avoid 'model collapse' caused by AI-generated data.
- ISBNdb provides pre-2022 books for AI training, as they contain no AI-generated content.
- Reddit data is valued for its human-generated content and voting system that judges quality.
- Model collapse leads to declining accuracy, generic output, and validated hallucinations when AI trains on AI data.
- Human-in-the-loop AI use can also cause homogenization and model collapse, as AI-influenced content gets fed back into training.
- Studies show AI-generated ideas reduce diversity and creativity, even when humans are involved.
Samsung Health app threatens data deletion if users opt out AI training
2 months ago
- Samsung is requiring users to consent to the use of their private health data for AI training in the Samsung Health app.
- The consent toggle allows Samsung to use personal metrics like sleep, medications, medical records, and cycle tracking details to improve the app's algorithms.
- Withdrawing consent results in loss of data sync and deletion of health data, unless retention is required by law.
- Humans, including employees and contractors, may review some of the collected data.
- The app recently added generative AI features, such as Vitals for sickness warnings and a Heart Health Score.
- New features include a Cardio Load metric and Fitness Index based on metrics like VO2 max.
Samsung is pushing users to train AI with their personal health data or lose it
2 months ago
- Samsung Health users must consent to sharing health data for AI training to maintain syncing features; opting out results in data deletion and loss of syncing capabilities.
- The AI will use anonymized health data, including activity, health records, and menstrual cycles, to provide personalized insights such as workout plans and sleep analysis.
- These AI-driven features will debut with the Galaxy Watch 9 series and expand to older wearables, aiming to improve accuracy and relevance over time.
- The consent requirement raises privacy concerns, as users face a trade-off between functionality and data sharing, with no clear option to keep syncing without consent.
AI developers are buying up old books in Germany in large numbers
2 months ago
- German booksellers report a surge in demand, with orders sometimes matching annual revenue, facilitated by Zoom Books.
- High demand for scientific/academic literature, regional history, linguistics, economics, and law books, especially with historical linguistic features for AI training.
- Interest increased as AI developers exhaust public online texts; Anthropic's Panama project involved cutting book covers to digitize, destroying physical copies.
- Booksellers fear this practice spreading to Europe, comparing it to a gold rush and raising copyright and cultural heritage concerns.
- Legal controversy exists; U.S. companies may use 'fair use' to copy copyrighted works without consent, a key argument in AI training lawsuits.
- Focus on copyright-protected books from the 1970s onward, risking cultural loss, as earlier German works are already digitized.
Publishers Are Preparing to Opt Out of Google Search
2 months ago
- Publishers historically relied heavily on Google Search for traffic and revenue, but some are now considering blocking Google's crawler to protect their content from being used for AI training without compensation.
- Cloudflare announced default settings to block 'multi-purpose crawlers' on ad-supported pages, primarily targeting Google, forcing publishers to choose between allowing AI scraping or losing search visibility.
- Google offers Google Extended to allow opt-out from AI training without affecting search rankings, but publishers are skeptical about potential penalties and prefer technical solutions over trust in Google's discretion.
- USA Today Inc. is preparing to delist from Google within six to twelve months, leveraging licensing agreements with other tech companies as an alternative revenue source, while Beehiiv has enabled creators to block Google's crawler.
- The decline in search traffic has paradoxically empowered publishers to consider leaving Google, as they diversify audience sources like newsletters and social media, reducing dependence on search.
- Blocking Google could degrade search quality and harm consumers, but publishers view it as a bargaining tactic; the industry is at a symbolic milestone where threats to opt-out are becoming credible.
- Other media executives express uncertainty, noting a lose-lose situation: allowing free scraping is self-defeating, while opting out offers unclear benefits and risks future compensation if Google eventually pays for content.
Learning to Replicate Expert Judgment in Financial Tasks
3 months ago
- Outperforming the market is challenging due to the need for unique insight from investor judgment, which is difficult to articulate or teach directly.
- LLMs struggle with simple financial tasks like filtering and processing documents, even though these are routine for investors.
- The post explores automating information triage using LLMs, showing that with expert annotations, proprietary models can achieve expert-level judgment.
- Frontier models (e.g., Gemini, Claude, GPT) underperform on six filtering tasks, with accuracy around 50-80%, below the 80% threshold for trust.
- Improved prompting boosted accuracy to the mid-70s, but fine-tuning with high-quality human-labeled data was necessary for further gains.
- A custom training dataset was built using expert verification to correct non-expert labels, enhancing data quality.
- The training recipe used Qwen3-235B as a base model, with techniques like interleaved batching, CISPO loss with asymmetric clipping, and on-policy distillation.
- The final proprietary model achieved 84.7% accuracy, making 29.8% fewer mistakes than frontier models, with a 13.8x reduction in inference costs.
- The conclusion highlights that custom models tuned to organizational needs outperform frontier models in accuracy and cost, enabling differentiated intelligence.

ai training

Company Offering Printed Books to Train AI Stops After 404 Media Coverage

AI Firms Are Buying Up Old Books, Then Scanning and Destroying Them

The Vanishing Page: AI Firms Scan Then Destroy Rare Book Editions

AI companies are shredding rare books

Why Would Meta Download So Much Porn?

If AI companies avoid AI slop, shouldn't we?

Samsung Health app threatens data deletion if users opt out AI training

Samsung is pushing users to train AI with their personal health data or lose it

AI developers are buying up old books in Germany in large numbers

Publishers Are Preparing to Opt Out of Google Search

Learning to Replicate Expert Judgment in Financial Tasks