1. Google Launched Three AI Search Features in One Week, All Skipping the Text Box On Monday, Google released Gemini 3.1 Flash Live, a model tuned for low-latency voice interaction, across its products.

2. Apple Opens Siri to Rival Chatbots; Google Moves to Poach Their Users Apple's iOS 27 will let users plug third-party AI chatbots into Siri, Bloomberg's Mark Gurman reports.

3. Wikipedia Bans AI-Written Articles After Editors Find Fabricated Citations English Wikipedia's volunteer editors have spent two decades building the internet's largest encyclopedia entry by entry. Last week, they voted to keep it that way.

In Brief

• Senators Warren and Hawley Demand Mandatory Energy Reporting for Data Centers The bipartisan pair sent a letter to the Energy Information Administration requesting "comprehensive, annual energy-use disclosures" from data centers, made publicly available. The push would establish mandatory reporting requirements amid growing concern over AI infrastructure power consumption. The Verge

• Meta Prepares Two New Ray-Ban AI Glasses Models FCC filings reveal Meta and EssilorLuxottica are readying the next generation of Ray-Ban AI glasses. The filings point to two distinct models, building on the current smart glasses line that added multimodal AI features last year. The Verge

• Google DeepMind Publishes Research on AI Manipulation Risks DeepMind released a study examining how AI systems could harmfully manipulate users across domains including finance and health. The research led to new internal safety measures targeting persuasion and deception vectors. DeepMind

• Google Research Introduces TurboQuant for Extreme Model Compression TurboQuant is a new quantization method aimed at compressing AI models far below standard precision levels while preserving output quality. The technique targets deployment scenarios where memory and compute budgets are tight. Google Research

• Webtoon Adds AI Translation and Localization to Creator Platform Webtoon's Canvas platform for user-uploaded comics will roll out AI-powered localization tools to help creators reach international audiences. The update is part of a broader overhaul designed to increase creator revenue and global distribution. The Verge

• Apple Music's AI Playlist Playground Struggles with Genre Accuracy Apple's new AI-powered playlist feature frequently mismatches genres, returning vocal tracks for instrumental requests and mixing unrelated styles. Early testing shows the tool fails on specific subgenre prompts that streaming competitors handle better. The Verge

• CUA-Suite Releases Largest Open Dataset for Computer-Use Agents Researchers published CUA-Suite, a collection of human-annotated video demonstrations for training desktop automation agents. The dataset addresses a key bottleneck: prior open datasets topped out at roughly 20 hours of video, far too little for training general-purpose computer-use agents. Hugging Face

• UniGRPO Proposes Unified Reinforcement Learning for Text-and-Image Generation A new framework applies group-relative policy optimization to models that interleave text reasoning with image generation via flow matching. The method trains the model to reason about a prompt before generating the image, improving prompt adherence. Hugging Face

• EVA Uses Reinforcement Learning to Make Video Understanding Models More Efficient EVA replaces manually designed video-agent workflows with a learned policy that selects which frames to process, cutting redundant computation. The approach outperforms uniform sampling and prior agent-based methods on standard video QA benchmarks. Hugging Face

Read the full edition →