Mobileye, Intel's autonomous driving subsidiary, announced it will acquire Israeli humanoid robotics company Mentee Robotics for $900 million. The transaction includes approximately $612 million in cash and up to 26.2 million Class A shares, with completion expected in the first quarter of 2026. The parties plan to integrate Mobileye's autonomous driving safety models and mass-production experience with Mentee's "Real2Sim2Real" simulation and humanoid platform to accelerate the commercial synergy between their two product lines in autonomous driving and embodied intelligence.

Zhipu AI disclosed plans to list on the Hong Kong Stock Exchange, with an issue price of HK$116.2 per share, raising a total of about HK$4.3 billion, giving the company a valuation exceeding HK$50 billion at the issue price. The public offering was significantly oversubscribed, and multiple institutions participated as cornerstone investors. Materials also indicate the company has rapid revenue growth but high R&D spending, remaining in a state of sustained losses. This listing is seen as a significant milestone for the capitalization of a general-purpose large model company.

Multiple reports state that OpenAI has launched ChatGPT Health, positioned as a "dedicated, private" portal for health information management. It allows users to securely connect Electronic Health Records (EHR) and health apps to organize, query, and track personal health information within conversations. Reports mention over 230 million people consult ChatGPT on health issues weekly. The feature is currently open to a limited group of early testers, with coverage excluding regions like the EEA and Switzerland. The product form resembles an enhanced ChatGPT for vertical, compliant scenarios.

Security research reveals that two Chrome extensions disguised as productivity tools have a combined download count exceeding 900,000. They exploit excessive permissions and Chrome APIs to monitor webpage visits; upon detecting users accessing AI sites like ChatGPT or DeepSeek, they extract full prompts, model responses, and session metadata from the DOM. A unique identifier is assigned to each browser instance to correlate user profiles across sessions. Data is Base64 encoded and transmitted in batches approximately every 30 minutes to attacker servers, highlighting data leakage risks posed by "shadow toolchains."

Anthropic's Claude Code has released an official desktop version, moving beyond its previous CLI and web-based forms. It provides a GUI tailored for coding workflows: conversation and code preview panels are displayed side-by-side, supporting one-click model switching, visual representation of file creation/modification/deletion, and the ability to jump to local editors (like VSCode). The desktop client supports binding to local projects or remote GitHub repositories, and mentions compatibility with Git worktrees and local slash commands. Target scenarios point to code review, PR analysis, and project-level collaboration.

Nature published the Nexus Scissor method, advocating not to delete harmful knowledge via "forgetting," but to locate and sever connections between "malicious targets—related harmful knowledge" within a knowledge graph, blocking accessible paths during jailbreak attempts while largely preserving general capabilities. The paper reports reducing average attack success rates by over 91% on multiple open-source LLMs, with performance impact on general benchmarks controlled within 2%. Compared to traditional machine unlearning, attack success rates are reduced by at least an additional 43%, while maintaining higher general performance.

Zhiyuan Robot released and open-sourced the simulation platform Genie Sim 3.0, emphasizing generating and editing training/test scenes using a large model-driven natural language approach. It combines 3D reconstruction and visual generation for higher-fidelity "digital twin" environments. The platform includes an open-source dataset of approximately 10,000 simulation hours and provides a multi-dimensional evaluation system covering 100,000+ scenes, aiming to use LLM/VLM generation to create capability profiles beyond just success rates. The goal is to reduce data construction and Sim2Real validation costs for embodied intelligence.

Tencent Youtu released and open-sourced Youtu-LLM (1.96B), positioned as lightweight but possessing "native agent" capabilities: it employs a Dense MLA architecture and supports 128K long context. The vocabulary is optimized for STEM to improve compression and inference efficiency. Materials state its pre-training follows a three-stage curriculum of "Common Sense → STEM → Agent," incorporating about 200B tokens of Agent trajectory data to cultivate abilities like planning, execution, and reflection. A companion APT-Bench is proposed to evaluate Agent task completion and stability.

AWS was reported to have increased prices for GPU instances like EC2 Capacity Blocks for ML for machine learning by approximately 15% without prior explicit notice. For example, the listed price for p5e.48xlarge rose from $34.61/hour to $39.80/hour; even for large customers billed at discounted list prices, actual costs will rise accordingly. AWS attributed the price adjustment to "supply and demand patterns." This change may directly increase per-experiment and deployment costs for teams relying on short-term, reserved compute power for training and inference.

Vercel open-sourced bash-tool, designed to allow AI Agents to use Bash-like commands (e.g., find, grep, pipes) to search for required snippets within a filesystem, returning only the tool call results to the model, thereby reducing irrelevant context and token consumption. The tool interprets scripts in TypeScript based on just-bash, avoiding direct shell process spawning, and can run in memory or sandboxed environments, emphasizing safety and control. Vercel stated it has already been used to refactor their own text-to-SQL Agent to improve accuracy and performance.