AI News Roundup – AI labs are consuming and destroying rare books, DeepMind AI models improving hurricane forecasting, Nvidia to purchase Hugging Face as it plots AI future, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• Major AI companies have quietly been buying up, scanning, and destroying copies of rare books to support AI model training, according to The Wall Street Journal. Speaking with several booksellers, the WSJ found that major AI companies, often using straw purchasers with innocuous names such as the “Red Sparrow Project,” were buying up rare books to obtain new content to train their AI models. One bookseller, suspicious of the bulk purchases, placed trackers inside the shipments, determining that once the books were scanned, they were often destroyed (as removing the spine of a book allows for much more efficient scanning). Outside of internet-scraped content, books remain a major part of the corpus of works used as training data for AI models. Anthropic notoriously used pirated digital copies of books to train its models, a key focus of a copyright infringement lawsuit brought against the company by a group of authors in 2024 (which the company later settled for $1.5 billion). Some booksellers were happy to sell their “dead stock,” books which had been on their shelves for years, but others have raised concerns over the destruction of some rare books as cultural artifacts. An Anthropic spokesman told the WSJ that “none of our data acquisition programs buy and destroy rare or antiquarian books,” and Elon Musk said in a post on X that he instructed SpaceXAI to “preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning.” However, it is likely that destructive scanning will continue as a practice so long as the insatiable demand for AI training material continues.

• The New York Times reports on a new paper from Google’s DeepMind AI unit analyzing the performance of its AI-powered hurricane forecasting models. The paper, published earlier this month in Nature, analyzed two years of hurricane forecasts and found that their newest AI weather model, WeatherNext Cyclones, greatly outperformed other predictive models. The researchers demonstrated that their AI system outpaced traditional baseline models, including the European Center for Medium-Range Weather Forecasts’ gold-standard Ensemble Prediction System, by a day or more across vital parameters such as hurricane path, intensity, and structure. For instance, DeepMind’s model reportedly achieved an accuracy lead time of over 30 hours compared to its top public rival when predicting five-day storm tracks. The real-world value of this technology was demonstrated during October 2025’s Hurricane Melissa, where the model successfully anticipated the storm’s erratic trajectory across the Caribbean and its rapid jump from a Category 1 to a Category 5 hurricane over the course of two days. Forecasters at the U.S. National Hurricane Center praised the model’s guidance during the crisis, saying that developments in forecasting give authorities critical extra time for life-saving evacuations and other storm preparations. DeepMind’s weather-related work has been years in the making: this AI Roundup covered an earlier DeepMind weather forecasting model in 2024.

• Nvidia has agreed to purchase AI model platform Hugging Face for nearly $13 billion, according to CNBC. Citing a report by The Information, the $12.9 billion deal began after Hugging Face received acquisition interest from another suitor, with a source confirming to CNBC that the acquisition was part of recent, ongoing talks. If finalized, the deal would place one of the world’s most popular platforms for hosting and sharing open-source AI models under Nvidia’s control, expanding the chipmaker’s footprint further into software and the broader AI ecosystem. Bloomberg reports that the acquisition, which commands a valuation of roughly 86 times Hugging Face’s approximately $150 million annualized revenue, represents a dramatic shift for Hugging Face, as the company previously rejected a $500 million investment offer from Nvidia to safeguard its neutrality in the AI space. The acquisition could offer Nvidia strong strategic leverage, as controlling the central hub where millions of developers download open-weight AI models could help Nvidia counter closed-source rivals like OpenAI, which is actively building competing in-house AI processors. By promoting open-source models, Nvidia could steer coders toward its own computing hardware, which remains the industry leader and for which demand remains incredibly high, driving the company’s valuations to record highs in recent weeks.

• The Economist reports on a new study that explores the effects of AI usage on student academic performance. In the study, researchers at Stockholm University and the University of Hong Kong tracked nearly 27,000 Chinese students in grades seven through 12. 80% of students reported using AI models such as DeepSeek and Doubao, while the other 20% served as a control group for the experiment. After six months, the AI-using students saw their homework scores rise by 18% on average across all subjects, while completing those assignments 30% faster. However, on exams in those same subjects, the AI-using students scored 20% lower than their classmates who did not use AI. The results are striking: traditionally, homework scores generally correlated with test scores, leading some to question whether this phenomenon should lead to changes in AI adoption in schools. The drop in scores was primarily among AI-using students who spent less time on their homework, while those who spent more time had similar scores to students who did not use AI. The researchers speculated that students who relied on AI tools for quick answers rather than tutoring or learning assistance were more likely to do poorly on exams. The researchers also recommended further research on the topic as well as more in-person, closed-book exams and attention paid to study effort to incentivize students to use AI tools in a more productive manner.

• A U.S. court has ruled in favor of Anthropic in its ongoing dispute with the U.S. Defense Department over the company’s designation as a supply chain risk, according to Bloomberg. As this AI Roundup covered earlier this year, the dispute grew over Anthropic’s refusal to remove safety restrictions on the AI products it provided to the Pentagon, stating that they should not be used for mass surveillance of American citizens or for fully autonomous weapons systems. Anthropic was eventually designated as a supply-chain risk, which generally banned Anthropic from federal government contracting work, at the direction of President Trump and Defense Secretary Pete Hegseth. In response, Anthropic sued the federal government, sparking the current lawsuit. In an opinion by Judge Rita F. Lin, Anthropic won its motion for summary judgment, successfully arguing that the Pentagon, in its actions, violated the company’s First Amendment freedom of speech rights. The court also ruled that the Pentagon did not follow proper procedures in designating Anthropic as a supply chain risk, and that the company regardless did not meet that definition. While the order does not require the Pentagon to adopt Anthropic technology (as the department is “winding down” its use of Anthropic products by Sept. 30, according to a government lawyer), the victory is still a major milestone for Anthropic as it approaches an expected initial public offering later this year. U.S. government spokesmen did not respond to Bloomberg’s request for comment, though the Pentagon is likely to appeal the decision. The case is Anthropic v. U.S. Department of War, 26-cv-01996 in the U.S. District Court for the Northern District of California.