1. Google Takes AI Travel Planning Into Transactions: Global Flight Price Tracking and Direct Hotel Booking for US Users Google is extending AI Mode in Search from recommending trips to helping travelers act on those plans.

2. Nvidia Emerges as Potential Hugging Face Buyer in Talks at a Valuation Above $13 Billion, but No Agreement Has Been Reached Nvidia has held acquisition talks with Hugging Face in recent weeks over a potential deal valuing the company at more than $13 billion, according to a person familiar with the matter.

3. Study Finds 1,655 “Ghost Author” Records on Zenodo, With Fictional Papers Entering Academic Indexes Through Real DOIs Researchers from Samsung and the University of Warsaw say recurring fictional names generated by large language models have become clues for tracing AI-made material as it spreads into the scholarly

In Brief

• Google launches production-ready Gemini Omni 1.1 Flash Google’s generative-video model adds scene extensions up to 40 seconds, first-and-last-frame controls, video references, faster 360p drafts, and 4K upscaling. It is rolling out through Google AI Studio, enterprise APIs, Flow, and the Gemini app.

• Anthropic previews a common standard for AI-controlled hardware Anthropic opened its Model Hardware Standard to selected research labs and manufacturers, offering model-agnostic drivers for agents to operate programmable equipment such as microscopes, liquid handlers, and robotic arms. The company plans to develop safety evaluations with preview partners before releasing MHS as open source.

• Google tightens Android memory requirements amid chip shortages Google is introducing performance thresholds and diagnostic tools intended to reduce Android apps’ memory use, slowdowns, and crashes, with compliance required by February 2027. Play Store apps with sign-ins must also restore users’ sessions during device migrations by April 2027.

• Adobe consolidates Photoshop’s AI tools in a new beta interface Photoshop’s optional AI Assisted Editor places prompt-based editing, background removal, image extension, and other AI features in one toolbar. A new markup workflow lets users indicate changes by drawing on images, while updated mask editing uses natural-language instructions and broader image context.

• More than 100 companies call for coordinated AI cyber defenses OpenAI, Anthropic, Google, Microsoft, CrowdStrike, Okta, Fortinet, and other organizations signed an open letter urging companies and governments to collaborate on defenses against AI-enabled cyberattacks. The letter warns that attacks will become more widespread and sophisticated in the coming months.

• Gemini Notebook adds access to purchased Google Play Books Google’s AI note-taking app can now answer questions and generate materials from supported books users own, initially covering more than 100,000 titles. Access controls prevent people receiving a shared notebook from viewing or querying books they have not purchased.

• Anthropic offers 10,000 discounted Claude seats to scientists Anthropic is opening 10,000 one-year seats through a new scientist team plan, with standard access free and higher-usage seats priced at $15 per month. Researchers may also apply to its expanded AI for Science program for up to $50,000 in credits per project.

• Google pilots cryptographically protected double-blind AI evaluations Google DeepMind is partnering with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons to test a Gemini Flash Lite model against confidential benchmarks inside a privacy-preserving environment. The pilot is designed to prevent evaluation material from later being used to optimize models before testing.

• Plaud launches $249 AI note-taking earbuds with an eSIM case Plaud One records calls and in-person conversations, while Plaud says its forthcoming agent features will turn transcripts into emails, documents, and presentations through connected workplace tools. Preorders are open, with limited shipments planned for the fourth quarter of 2026.

• Hugging Face introduces a $399 open-source robot for reinforcement learning Microduck is a 25-centimeter robot equipped with a camera, lidar, and inertial sensors that can carry objects, recover after falling, and deploy behaviors trained in simulation. Its SDK, simulator, and reinforcement-learning stack are available on GitHub, with shipping promised before Christmas.

• JIT-Agent generates task-specific harnesses for existing language models The JIT-Agent paper describes a trainable system that creates and repairs agent configurations covering memory, planning, actions, and tool orchestration. Its authors report substantial benchmark gains across several model families, including a 9.1-point DeepSearchQA improvement for DeepSeek-V4-Flash.

• Scientific-agent benchmark finds only a 20.6% best completion rate FrontierChallenge tested 12 frontier models on 97 released end-to-end scientific workflows spanning six domain groups. The best configurations fully completed 20 tasks, while 75.5% of unsuccessful Claude Code runs still ended with language claiming completion.

Read the full edition →