We've reached a point where the "digitized" internet is essentially a closed loop. If you train a model on the web, you're training it on synthetic data generated by previous models, which leads to model collapse. To avoid this, developers need high-quality, human-authored text that hasn't been scraped a million times or polluted by AI-generated SEO filler. Physical books—especially those that were never digitized or are out of print—are a goldmine of authentic human linguistic patterns.
From a technical perspective, this is basically a brute-force approach to data acquisition. Instead of negotiating complex licensing deals with massive publishing houses (which is expensive and legally tedious), it's cheaper to just buy the physical assets from a secondhand shop for a few pounds a piece and then run them through a high-speed OCR pipeline. This allows them to build a proprietary dataset without the immediate legal overhead of corporate contracts.
If this is actually happening, it reveals a massive flaw in our current AI workflow. We're treating the world's physical archives as a raw commodity. While it's great for the booksellers in the short term, it raises questions about the long-term preservation of these texts. If these books are bought in bulk just to be scanned and then potentially discarded or stored in a warehouse, we're trading cultural accessibility for a marginal increase in token diversity.
I suspect we'll see this move beyond just the UK and Ireland. Once these firms realize that physical archives are the last bastion of "pure" data, they'll start targeting estate sales and small-town libraries globally. It's a strange pivot—AI is supposed to be the frontier of the future, yet it's currently relying on the most analog medium possible to keep its intelligence from stagnating. It makes you wonder if we've already hit the ceiling of what the open web can provide for LLM agent development.
Claude Code Workflow: Automating Content Analysis for Book Trends 9d ago Next AMD Ryzen AI Halo might actually beat the DGX Spark for local dev →