Desk
Broad artificial intelligence coverage.
A small, openly licensed model that clones a voice from one clip in 23 languages is a gift to builders and a fresh problem for anyone who relies on a voice to prove who is calling.
OpenAI published excerpts from amicus briefs supporting its position that training on copyrighted work is fair use. Some arguments are strong, especially from libraries and accessibility groups. The market-dilution question is where journalism’s real losses could go unremedied.
Baseten's joey-1 skips the chatbot act and just answers the question with a label and a confidence score, which is exactly what most production AI calls should look like.
Reddit told investors it will shift to a primarily logged-in site to curb free AI scraping and lean into valuable users. US daily users fell from 53.2 million to 50.7 million while revenue rose 53%. It’s a rational business move and a loss for the open web; where the login line sits decides which.
Thrive Capital’s holding-company spinout will pay $10 a share, about $7.6 billion in equity, for the home security giant, with OpenAI supplying models and engineers. No products were announced. AI inside home security raises sharper privacy and error stakes than a chatbot.
Microsoft Research Focus #97 spans agent backtracking, coding-agent failures, workflow learning from activity logs, Copilot Chat usage, a teacher tool in India and monsoon forecasting in Kenya. The CABRA finding that comprehension, not generation, is the bottleneck is the one to circle.
Meta quietly launched Style Sketch in Indonesia, Singapore and Thailand. It builds a virtual wardrobe from Instagram photos and shows outfits on a model matched to your body and skin tone. Useful, and a new category of personal data for an ad company.
All Gemini users can generate, preview and edit PDFs, Docs, Sheets, decks, Markdown, TXT and code files in the chat, with highlight-and-ask edits. It removes busywork and reverses the direction of travel: office apps start living inside the AI. Change tracking is the open question.
WIRED’s sources say an Accenture employee linked a DoD system holding controlled unclassified information to a personal Claude Code account, and the agent appears to have sent credentials and a NIPRNet map to a high-risk site. Accenture says the allegations lack merit. The failure described is human and institutional.
The New York Times, citing people briefed on classified intelligence, reports Trump was told a year ago that Beijing was getting reports from someone with direct access to Nvidia’s CEO, and held off partly over chip sales and the trade truce. Nvidia and China deny wrongdoing. If accurate, export enforcement bent to a stock.
OpenAI reorganized the public rulebook for its models around five values, added guidance on context and agency, and retired most style rules. The anti-sycophancy language is a gain; judgment is harder to audit than rules.
Alphabet’s Q3 revenue rose 25% to almost $128 billion, but EPS of $2.12 missed $3.02 estimates and the stock fell almost 12%. Philipp Schindler said each AI Mode query earns less, with fewer, cheaper ads. Search volume growth slowed to 12%, and capex hit $51.4 billion.