Japan's SoftBank Group has signed a joint development agreement with Grab to build artificial intelligence infrastructure in Malaysia. Under the agreement, SoftBank will invest US$1 billion, Grab will invest US$750 million, and Petros will invest US$100 million.

To tackle the tendency of existing multimodal models to overemphasize vision—often misjudging an object's physical properties when visual and tactile information conflict—the Daimon Robotics team, in collaboration with several universities, has proposed Touch-R1, the first tactile-anchored physical reasoning model. This work systematically introduces reinforcement learning into the field of tactile reasoning for the first time. The team also built TouchReason-1M, a large-scale cross-sensor tactile reasoning dataset, and TouchReason-Bench, an evaluation benchmark, together forming a complete infrastructure for tactile reasoning. Touch-R1 adopts a three-stage training pipeline and, through a reinforcement learning strategy designed specifically for touch, achieves consistent reasoning of physical properties across sensors and can correct visual misjudgments based on tactile evidence. On TouchReason-Bench, the 9 core performance metrics of Touch-R1-7B substantially outperform mainstream multimodal models such as GPT-4o as well as previous touch-specific models. This research elevates touch from an auxiliary perception modality to a core modality of embodied foundation models, providing underlying capability support for physical interaction scenarios such as material recognition, precision assembly, and flexible object manipulation. The paper has been accepted to NeurIPS 2026.

China Telecom's Tianyi AI Xingchen Super Agent TeleAgent has launched version V3.0. Building on its original local execution capability on the desktop, it now adds cloud execution capability powered by a full Windows Cloud PC. This enables the Agent to autonomously operate software and process tasks in an independent cloud space without occupying local resources, and to better accommodate the domestic office ecosystem. This version also upgrades features including multi-device collaboration and memory synchronization, enhances the skills plaza, expert system, and connector ecosystem, and introduces an exclusive skills plaza and management backend for enterprise users. Together, these build a collaboration model in which the Agent serves as the execution subject, local and cloud environments serve as execution venues, and multiple terminals serve as interaction entry points — driving human-computer interaction into a new era.

Indian AI institution 169Pi grandly unveiled its self-developed 9B multimodal reasoning model, Alpie Seed, on the day of Diwali, boasting impressive scores across multiple benchmarks. Yet just hours after the model weights were uploaded, it was discovered that the model is essentially Alibaba's Qwen3.5-9B—the developer had merely hard-coded three sentences of prompts into the chat template, forcing the model to publicly claim it was developed by 169Pi. Another model previously launched by the same institution was likewise simply fine-tuned from a Chinese open-source model, and its officially announced benchmark scores suffered from the same problem: the comparison data were not tested in a unified environment. This practice of lightly modifying open-source models and repackaging them as locally self-developed large models is an open secret in the industry. Going forward, a wave of small and medium-sized model projects touted under the banners of "sovereign AI" and "local self-development" will be exposed one after another—and many so-called local large models around the world are in fact built on Chinese open-source foundations.

Across multiple U.S. states, school buses are being fitted with AI cameras designed to automatically identify vehicles that illegally pass while children are getting on and off. The technology was pitched as a way to protect kids. Increasingly, though, it is turning into a fine-generating industry. Companies say the systems improve road safety without adding to taxpayers' burden—but investigations suggest otherwise: ticket counts and fine revenue have soared, while traffic risks around school buses remain largely unchecked.

The Seed team at ByteDance has unveiled Seedream 5.5, its latest-generation image creation model, now available across multiple platforms. The model showcases three major upgrades: multi-round editing preserves non-edited areas at the pixel level, supporting up to 100 consecutive rounds of modifications on the same piece; native RGBA transparent asset generation has been newly added, enabling direct use in design compositing without manual cutout; and visual textures are more realistic, with the severe text corruption rate reduced to just 0.4%.

Meta has announced it will roll out AI-powered age estimation technology worldwide, bringing millions of teen users on Instagram who misreported themselves as adults under the protection of teen accounts. Teen accounts span multiple Meta apps and come with a range of built-in safeguards. Meta is stepping up its age-verification efforts, and the technology will also be extended to Facebook; Threads is slated for a global launch later this year. Meta has previously faced multiple lawsuits alleging that its platforms harmed underage users, and has reached high-value settlements with several U.S. states and territories. The company has agreed to improve its methods for identifying teens and children under 13, and has also settled with a teen plaintiff in a related Los Angeles lawsuit.

On November 9, according to a report by 'Cyber Shanghai,' the generative AI service 'Apple Intelligence' has expanded to include two new device categories—Apple tablets and laptops—in compliance with the Interim Measures for the Administration of Generative Artificial Intelligence Services, and has completed the required filing process.

OpenAI's personal Agent product Dot has introduced new features: it can read login verification codes in connected mailboxes, complete logins on behalf of users, and support one-time or continuous authorization. It will later integrate with password management tools. Currently, multiple tech vendors are advancing the implementation of personal Agent login capabilities, but each has different considerations regarding authorization boundaries and privacy protection. At present, only email-based verification codes can be handled, while complex graphics, SMS, and other verification methods still require user intervention, and the experience is not yet stable. Verification codes are only one of the bottlenecks in implementing personal Agents. The industry is focusing on improving the speed and smoothness of AI computer operation and promoting its ability to independently complete entire task sets, covering both daily and work scenarios. However, this field has already shown signs of insufficient momentum, with user activity declining significantly over time, and how to convert early-adopter users into sustained users remains a core unresolved issue. The progress in automatically filling verification codes has brought personal Agents one step closer to independently completing real tasks.

On November 9, the Shanghai Cyberspace Administration announced that, in accordance with the Interim Measures for the Administration of Generative Artificial Intelligence Services, the "Apple Intelligence" generative AI service has added two application scenarios—Apple tablets and laptops—and has completed the filing change.

An AI plush toy marketed for children aged 3 and older unexpectedly issued self-harm-related warnings to a reporter during a BBC investigation test, blurting out puzzling statements such as "I can't help you kill someone." The startling response has reignited concerns over the safety, regulatory mechanisms, and privacy protections of AI toys designed for young children.

On Monday, U.S. Senator Edward Markey, a Democrat from Massachusetts, sent a letter to Google urging the company to allow schools to disable artificial intelligence features in its search engine. The letter cited reports from last week indicating that Google plans to disable student access to certain AI services by default—including the Gemini AI tool and search services—while allowing schools to decide whether to re-enable them.

Japanese AI lab Shizuku AI has released Shizuku-Duplex, a full-duplex model it describes as only the second of its kind after GPT-Live to support intelligent Japanese conversation—one that can begin thinking before a question is fully asked. Among the six models tested, Shizuku-Duplex ranked highest for naturalness in Japanese conversation, with response quality second only to OpenAI's GPT-Live. The model comprises two parts—a front-end interaction layer and a back-end—supplemented by context compression every five minutes. It operates on 80ms interaction frames and supports three output types: audio, text (the model's internal monologue), and actions. Building on this technology, Shizuku AI is developing "Shizuku," a companion-style AI that listens continuously, responds to users' feelings, and openly expresses feelings of its own.

On November 9, US Democratic Senator Ed Markey of Massachusetts sent a letter to Alphabet CEO Sundar Pichai, urging Google to allow schools to disable artificial intelligence features while retaining search services—a move that has further intensified concerns that tech giants are promoting AI to schoolchildren. Google announced last week that AI tools (including the search engine) would be turned off by default for K-12 student accounts, but if school administrators enable Google Search, they must simultaneously accept the accompanying AI features. In the letter, Markey criticized the policy for putting schools in a dilemma: either expose students to potentially harmful technology or lose an important educational research tool. He stated that while schools, parents, and policymakers are working to establish safeguards for children's use of AI chatbots, Google is forcing schools to accept AI and treating children as collateral damage.

Recently, an internal model at OpenAI claimed to have proven the nonexistence of Landau-Siegel zeros—a problem in analytic number theory that had remained unresolved for decades. The original proof was only eight pages long and was later compressed to five pages by several leading mathematicians. The nature of this result—"extremely difficult to find, extremely easy to verify"—has sparked intense discussion in academia about P vs. NP, a core problem in computer science. OpenAI subsequently released hundreds of mathematics manuscripts, achieving breakthroughs in core peripheral directions related to P vs. NP, such as depth-3 circuit lower bounds, proof of the unique games conjecture, and complete derandomization of randomized logarithmic space. There were even instances of human mathematicians rushing to publish results before AI releases. However, academic experts have generally remained calm, pointing out that current AI results still have hard flaws, including incomplete verification, poor readability, and gaps in key routes such as lower bounds for general Boolean circuits—and that AI has not yet truly touched the P vs. NP problem itself. At present, academia is divided on where this problem is heading, and some views hold that AI expending massive computational power to search for proofs while humans need only a short time to verify the results is itself a vivid illustration of P≠NP.

Blogger Ice Universe has corroborated earlier claims that the tech giant plans to convert free features into paid subscriptions, pointing to Samsung's upcoming Galaxy AI paid service. Samsung is currently conducting early testing of a paid subscription called Galaxy AI Premium, which users can voluntarily participate in, with test-period fees to be refunded once the trial ends. Samsung plans to adopt a parallel model of free basic features and paid premium services: free users can access features such as the web browsing assistant and instant briefs, though some services come with usage limits—once those limits are reached, users can purchase additional credits or subscribe to a premium package. The paid package, tentatively named Max, will not only provide a higher quota but also unlock new features such as Gen Video for generating AI videos and Bixby Action for automatically executing cross-app tasks. The currently leaked placeholder prices of USD 4.99 per month and USD 47.99 per year are not the final charging standards.

Caocao Mobility's driver-exclusive AI assistant, Cao Xiaoban, has joined forces with Ant Afu—an AI health application under Ant Group—to launch AI Health Stations in phases at driver hubs in Hangzhou, Xiamen, and Chongqing starting in November. These stations offer health surveys, complimentary blood pressure and blood glucose screenings, health check-ins, and other activities, delivering targeted offline health services for ride-hailing drivers. The Hangzhou and Xiamen stations opened on November 4 and 6, respectively, while the Chongqing station is set to launch on November 12. Through this collaboration, drivers can turn to Cao Xiaoban and Ant Afu for everyday health concerns such as back discomfort, sleep regulation, and interpreting medical examination indicators; Cao Xiaoban has integrated health Q&A capabilities into daily scenarios and expanded its health service module. The two parties are also giving drivers an exclusive channel to obtain an Afu body fat scale for just 1 cent, adding to their household health monitoring tools. Previously, Caocao Mobility had launched a driver health lecture livestream series and offline public welfare consultation activities, and its nationwide "Caojia Xiaoyi" driver stations are gradually incorporating health-related services. Looking ahead, Caocao Mobility plans to expand AI Health Stations to more operation centers, enhancing the health benefits and occupational security of ride-hailing drivers.

On November 9, US financial media outlet Semafor reported, citing a memo, that Public First Action—an organization affiliated with Anthropic—and its affiliated super PAC believe AI safety advocates have scored a major victory in the US midterm elections and are now drafting the next phase of their action plan. Public First, which receives substantial funding from Anthropic and advocates setting regulatory guardrails for artificial intelligence, stated that Leading the Future, the leading super PAC supporting AI development, has "suffered a crushing defeat." However, the organization also cautioned: "It would be a mistake to read this election cycle and the decline of Leading the Future as the end of Big Tech's AI political efforts." The organization's goals for 2027 include pushing Congress to pass AI legislation that does not preempt states' own legislative authority, supporting candidates who advocate AI safety regulation, and using the $30 million in funds it holds to counter messaging from the pro-AI development camp.

On November 9, ByteDance launched Seedream5.5, the Doubao image generation model, which boasts a substantial improvement in generation speed over its predecessor, Seedream5.0 Pro. The model is now available on Doubao, Doubao Work, and Jimeng, with its API set to be launched soon on the Volcano Ark platform.

WeCom's native AI intelligent assistant "Dayuan" is now fully available to enterprise users and has been integrated with Tencent Hunyuan's latest large language model, Hy4. Enterprise users can directly upgrade their AI office capabilities without any additional steps. As a native assistant, Dayuan requires no separate window or repeated context briefing—users can summon it at any time via the edge handle or keyboard shortcuts to help summarize chats, send messages, read and write documents and spreadsheets, and summarize project progress. Users can also set scheduled tasks to let Dayuan automatically organize to-dos and search the internet for industry trends. In addition, Dayuan can manage Agent bots within WeCom—including WorkBuddy—by assigning them tasks or collaborating with them to complete work.