AI firms pulping rare books to train models, Dutch seller finds

The Reddit post, which went viral, details how a Dutch bookseller was approached with a bulk order for a specific title that is essentially impossible to source in large quantities. The bookseller initially flagged it as phishing or spam, but the reality is far more troubling: AI training companies are acquiring physical copies of books, cutting off the spines, and running them through high-speed scanners to digitize content that isn't available electronically. The physical copies are then discarded, often pulped. The issue is that many of these books are rare, out-of-print, or culturally significant works that exist only in limited physical runs. Libraries and collectors hold the only remaining copies, and once destroyed, the knowledge is gone forever—even if the text is digitized, the artifact itself is lost. This practice highlights a growing tension between the insatiable demand for training data and the preservation of physical media. While companies argue they are extracting knowledge for the public good, critics point out that the destruction is unnecessary and reckless. For authors and publishers, this raises urgent questions about consent and compensation. For the rest of us, it signals that the AI gold rush is now consuming the physical world, not just the open web.

The post, which quickly climbed the Hacker News front page, broke down the token cost of a simple web fetch. While the visible text of a Wikipedia article might be 2,000 words, the underlying HTML, CSS, navigation menus, and JavaScript tags balloon the token count exponentially. For a typical AI agent using a tool to browse the web, the model sees the entire raw document unless specifically instructed to parse it. This means a single lookup can cost fractions of a cent, but at scale—thousands of agents making millions of calls—the cost becomes a significant line item. The implication is that developers need to rethink how agents access data. Solutions include using APIs that return clean text (like the Wikipedia REST API), implementing pre-processing pipelines that strip HTML before sending to the model, or using smaller, specialized models for extraction tasks. The broader lesson is that the AI ecosystem is still in its 'dial-up' phase, where efficiency matters as much as capability. As agents become more autonomous, the cost of 'just browsing' will become a critical factor in whether applications are economically viable.

The release signals OpenAI's push into vertical-specific models rather than just one massive general model. GPT-transcribe is optimized for batch processing, ideal for converting recorded meetings, podcasts, or video content into text. GPT-live-transcribe is built for real-time streaming, aiming to provide low-latency captions or live agent assistance. Both models leverage the deep language understanding of GPT-5.6, which means they can handle accents, code-switching, and industry jargon better than older dedicated speech models. The strategic move is clear: OpenAI wants to own the entire pipeline from raw audio to structured insight. For developers, this means simpler integration—no more gluing together a speech-to-text service with an LLM for post-processing. The models can now output timestamped, punctuated, and even summarized text directly. Early benchmarks suggest they outperform Whisper on noisy audio, but the real test is pricing and latency in production environments. If the cost is competitive, this could disrupt established players like Deepgram and AssemblyAI.

The AMD and Cerebras collaboration is a significant technical move. Cerebras's WSE-3 chip is enormous, allowing for extreme memory bandwidth and compute density, which is ideal for running large models at scale. By integrating with AMD's ROCm software stack, they are offering an alternative to the CUDA lock-in that Nvidia enjoys. This could lower costs for enterprises running massive inference workloads. On the safety front, the initiative involving Nvidia, SpaceX, and Microsoft is more vague. It aims to establish best practices for advanced AI development, but lacks concrete governance or binding commitments. Given that SpaceX is involved, the focus may be on autonomous systems and edge deployment rather than just data center safety. The industry is clearly bifurcating: one path is racing to build faster hardware, the other is trying to build guardrails. This week's news suggests the race is still winning.