404 Media reported that Amazon has purchased rare and out-of-print books, removed their spines and scanned the pages at a facility in Las Vegas identified as VGT3. The outlet said it placed a tracking device in a rare book that ultimately arrived at the site, which is marked by a logo of a dinosaur holding a book.
How the deliveries and scans were documented
The report describes a chain of custody traced via the tracking device and asserts the book reached the VGT3 facility. Amazon told 404 Media it purchases books through commercial channels to improve products and services customers use. The account says the company acquires books that are difficult to find online to add to training data.
The piece also mentions that, in Anthropic’s case, some books were illegally pirated, presenting an additional example of how companies have sought offline sources for text. The report did not provide a response from Anthropic within its coverage.
Why rare books are sought for LLM training
The report explains that companies building large language models (LLMs) require very large volumes of text and that rare, out-of-print works can supply material not available on the public internet. It notes that texts published before 2022 are unlikely to have been written by an LLM, increasing their perceived value for training.
404 Media’s coverage also raised a risk described as “model collapse,” which the report characterizes as a potential drop in output quality when LLMs ingest substantial amounts of AI-generated text. The report presents this as a rationale for seeking human-written sources that predate the widespread use of generative models.
The article published by 404 Media on August 17, 2026, includes the tracking-device evidence and the description of the VGT3 facility used to support its claims.
Original source: TechCrunch AI