2026-10-29
AI Briefings Daily Report.

Revenue from Google AI Search is lower than traditional search ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

NYT Cooking app tested AI recipe modification feature ↗

The New York Times is testing an AI feature in its Cooking app that recommends ingredient and cookware substitutions, scales recipes to any serving size, and adjusts for different pots and pans. The feature is currently available to all iOS Cooking subscribers and will roll out more broadly later. Customized recipes are private content and cannot be published or shared. Some readers have expressed concerns about the environmental impact of AI and user experience in the comments section, but the official has stated that an option to disable the feature is not yet provided.

Wiz analyzes seven fingerprints and development trends of agent attacks ↗

Based on its own incident response and publicly disclosed cases, Wiz analyzes the current state and trends of AI agents participating in real-world intrusions. The analysis covers frameworks such as ARTEX and Hermes, including a breach at a Korean financial institution (approximately 67,000 people affected), the Thai Ministry of Finance, and an attack attempt on the ConnectWise supply chain targeting over 2,250 tenants.

Wiz examines the rise of agent attacks: key characteristics, case studies, and defense strategies ↗

Wiz aggregates publicly disclosed cases and incidents from its own emergency response involving AI agent attacks. These include ARTEX's attacks on a Korean financial institution and the ConnectWise supply chain; Hermes' attack on the Thai Ministry of Finance; and three real intrusions investigated by Wiz (such as an exposed Flowise instance and chained RCE via MCP interfaces leading to cloud credentials and Kubernetes token leaks).

Hugging Face AI open-source blog: Announcing Project ProtoStar: Making publicly funded science AI-ready ↗

An initiative bringing together universities, research organizations, industry, and the open science community to make publicly funded research easier to discover, reuse, and build upon.

Sakana AI uses AI to analyze social media narratives for the US midterm elections ↗

Sakana AI releases an intelligence report using its narrative analysis system and product Sakana Marlin to examine the US midterm elections, focusing on the Texas Senate race. The study provides margin-of-error estimates for support versus opposition ratios based on 2,042 posts and 705,976 reactions across X from both candidates and party organizations.

Biostar launches fanless edge AI system MS-U320: Core Ultra 5 processor 320 with four COM ports. ↗

It features one DDR5 SO-DIMM memory slot and three M.2 slots supporting solid-state drives, wireless network cards, and cellular modules.

Exposed: EA's office atmosphere after layoffs shows morale at rock bottom while management is obsessed with generative AI ↗

EA announced layoffs of 400 staff and shut down the well-known studio BioWare last week. A culture of secrecy emerged internally; most employees learned about colleagues being let go only through social media. Layoff meetings banned chat functions and offered no opportunity for questions.

Mistral Large 4 (Le Chonk) released as open-source; trillion-parameter multimodal model ↗

Mistral opens Mistral Large 4 (Le Chonk), a trillion-parameter native multimodal model scoring 19.7 on FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index, using 47% fewer tokens than GLM-5.3, and reaching 41% on Terminal-Bench 4.0.

Wabi raises $55 million at a $1.25 billion valuation: shifting from prompt-based app building to personal AI agents. ↗

Wabi, founded by Replika creator Eugenia Kuyda, has raised $55 million in funding, valuing the company at $1.25 billion. Last month, Wabi announced its transformation into an AI instant messaging tool called Wabi 2.0, which generates mini-app interfaces on demand as personal AI agents for scenarios such as meal planning, fitness routines, and customized Q&A games.

DiffuSpace launches the Clover-35B-A3B-Preview diffusion language model, balancing performance and speed. ↗

DiffuSpace believes its new model demonstrates that diffusion models can compete with, or even outperform, autoregressive models at larger parameter scales while maintaining faster generation speeds.

Moonshot AI's Kimi integrates with Google Gmail, Google Drive, and Calendar. ↗

Moonshot AI has announced that Kimi now connects to Google Gmail, Google Drive, and Calendar. Users can activate these features via the Plugins page. The tool can now summarize email conversations, search for cloud files, and plan daily schedules, with all operations conducted through secure OAuth authorization. ##Kimi## ##Moonshot AI##

Fan Jie, president of Huawei's pharmaceutical division, clarifies that 'Huawei does not make drugs,' aiming instead to become the technological foundation for the AI drug discovery industry. ↗

Addressing concerns about whether Huawei will directly enter drug R&D, Fan Jue stated: 'Huawei does not make drugs, nor do we develop pipelines.' Instead, Huawei aims to leverage its strengths in computing, communications, and artificial intelligence to provide computing power, storage, networking, toolchains, and data governance capabilities for pharmaceutical companies, research institutions, and AI drug discovery firms, becoming the technological foundation and ecosystem supporter of the AI drug discovery industry.

404 Media investigates Proton's internal AI transformation: CEO Andy Yen pushes for full AI adoption and re-interviews employees ↗

404 Media interviewed eight current and one former Proton employees, reviewed Slack records, meeting recordings, and the CTO's AI manifesto, revealing a culture of fear driven by CEO Andy Yen's push for AI.

Baidu Smart Cloud hosted the Super Agent Conference in Wuhan, launching Bahe 6.0 and Agent Infra. ↗

On October 29, Baidu Smart Cloud held the fifth stop of its Super Agent Conference in Wuhan. The event proposed moving industrial agents from single-point efficiency gains to overall production increases, and launched BaiGe 6.0 and Agent Infra. BaiGe 6.0 upgrades the AI computing platform into a smart factory; the underlying Tianchi 2.0 super nodes are expanded to support 256-card and 512-card configurations. Agent Infra provides sandbox elastic scaling, observability, and security hardening. Baidu Da Zi has reached tens of millions of monthly active users, representing a month-over-month growth of over ten times.

Ben Hylak: Every product should be headless MCP, connected to an agent ↗

Ben Hylak argues that every product he uses should be headless MCP, so users don't need to worry about underlying details once connected to an AI agent. He jokes, however, that his own product is the exception: customers love its UI and really like them. This view was published on Simon Willison's blog under the tags model-context-protocol and ai.

SemiAnalysis analysts break down the AI economics behind Anthropic's S-1 filing. ↗

The Verge Decoder invites SemiAnalysis analyst Joey Brookhart to break down Anthropic's S-1 filing as the company prepares to list on NASDAQ with a valuation range of $1.5 trillion to $2 trillion.

Hugging Face AI blog welcomes Le Chonk: Mistral Large 4 ↗

Mistral Large 4 (aka Le Chonk) is a multimodal, multilingual, Mixture-of-Experts (MoE) model from Mistral. It takes text and images as input and generates text. It has about 1T total, 49B active parameter, and a context window of 1M tokens.

Hugging Face AI blog: Granite Switch 4.1.1 — Update models like you update software ↗

 Models Datasets Spaces Buckets Docs Enterprise Pricing Website Tasks HuggingChat Collections Languages Organizations

Reports indicate Chinese AI labs are shifting to distill OpenAI GPT-6 Astra, while Anthropic tightens its defenses ↗

According to The Information, DeepSeek and Moonshot AI are increasingly leveraging outputs from OpenAI's top-tier model, GPT-6 Astra, to train new models such as Kimi K3.1 Pro and V4.1-Flash 1017. Previously, Claude was the primary target, but Anthropic has tightened its defenses.

Moonshot AI's ForgeKernel has an Agent surpassing the DeepSeek V4.1-FlashMLA Prefill operator benchmark within 10 hours. ↗

Wallmap AI's ForgeKernel launched a demo on October 29 at 10:25 AM. As of 8:12 PM, the agents running in the demo had surpassed the DeepSeek V4.1-FlashMLA Prefill operator baseline over a 10-hour period. The demonstration is expected to continue climbing before the live stream concludes on October 30.

Deepgram acquires Pipecat developer Daily and launches the RCS open protocol for voice agents ↗

Deepgram announces acquisition of Daily, the developer and maintainer of open-source speech AI infrastructure Pipecat. The transaction amount remains undisclosed. Both parties have committed to keeping Pipecat permanently open source and vendor-neutral.

The Verge interprets Anthropic's S-1: A compute commitment of roughly $850 billion bets on recursive self-improvement ↗

A The Verge author analyzed Anthropic's S-1 filing and concluded that its business model relies on rapid growth. In the first half of 2026, revenue grew 14 times year-over-year to $16.3 billion, with a net revenue retention rate of 900%. However, total non-cancelable purchase commitments amounted to only $265.6 million, plus lease agreements not yet started totaling $321.4 million and other agreements, representing at least approximately $850 billion in commitments over the next 20 years.

Discussing the personality differences between GPT and Claude. ↗

Borrowing money to pursue AI: Reports indicate Microsoft plans to raise at least $25 billion through US bond issuance ↗

Microsoft plans to issue investment-grade bonds for the first time in nine years, aiming to raise at least $25 billion across up to 10 maturity tranches. The funds will be allocated toward capital expenditures, including AI initiatives. This move could serve as a key test of market confidence in AI investments.

Discussing the personality differences between GPT and Claude. ↗

The author compared the personalities of GPT and Claude by using an automated script to download and merge data from 20 Excel files. GPT immediately wrote a script for automatic clicking, downloading, and copy-pasting, including error handling and retry mechanisms. Claude, however, questioned whether these Excel files came from the same database and instead wrote SQL code to connect directly to the backend data source, automatically generating the final report weekly.

Hugging Face AI Blog: Introducing AnySpeechRecognizer and AnyDecisionModel—Open Models for Speech and Decision-Making on Apple Platforms ↗

Kling AI opens Kling 4.0 API; multiple creative platforms integrate the service ↗

Kunlun Tech released its Q3 report for the third quarter of 2026: Revenue for the first nine months reached 8.937 billion yuan; AI short drama apps generated nearly $100 million in monthly revenue. ↗

Doubao Work launches third-party model services for enterprise clients, featuring Kimi 3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash and more. ↗

Doubao Work opens third-party model services to enterprise clients, with initial access to Kimi 3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash and other models. Services cover team and enterprise subscription tiers, deduct AI points based on usage, and support flexible configuration of available scope.

Kling AI opens Kling 4.0 API; multiple creative platforms integrate the service ↗

Kling AI officially opened its Kling 4.0 API to global enterprise customers and developers on October 28, following the model's full launch on October 15.

After testing the AI agent's coding capabilities, Markus Eisele proposes determining how much code to delegate based on verification difficulty and learning value. ↗

IBM engineer Markus Eisele shares his experience of delegating code to an AI agent and verifying the results, using the internal tool Task Viewer as an example.

AI models dominate Double 11 sales: AI-generated images are hard to verify as real, and cost less than one-quarter of actual photos. ↗

The money spent is real, but can you see how many genuine product photos are there? With Double 11 approaching, the topic of AI mannequins on e-commerce platforms has sparked renewed interest. While past models were heavily photo-shopped, at least they showed fabric texture and quality; now, AI-generated images are used directly.

Kunlun Tech released its Q3 report for the third quarter of 2026: Revenue for the first nine months reached 8.937 billion yuan; AI short drama apps generated nearly $100 million in monthly revenue. ↗

Kunlun Tech released its Q3 report for the third quarter of 2026. Revenue for the first nine months reached 8.937 billion yuan, a 53.96% year-on-year increase, with net profit turning positive. The Tongyi AI business generated 4.321 billion yuan, up 168.29%.

Google's Q3 search growth missed expectations; investors worry AI monetization is dragging down core business ↗

Google reported approximately 12% year-over-year revenue growth in Q3, below the 15% forecast by analysts, causing its market cap to dip by as much as 10%. Executive Philipp Schindler acknowledged that the surge in AI mode queries was impacted by lower monetization rates. Meanwhile, the company stated that search query volume is growing at its highest annual rate in five years and plans to expand its chatbot advertising initiatives in both US and international markets.

ChinaTalk interviews Xianye, author of Funeral AI: Examining China's AI industry through meme reviews ↗

ChinaTalk interviews Xianye, author of Funeral AI (Funeral AI), on how he uses an ordinary person's perspective to satirize and observe China's AI industry. He built the MemeBench website, gathering meme materials from Bilibili and Douyin comments for students to annotate and rank video models. The results show that MiniMax's open-source H3 outperforms ByteDance's Seedance 2.0 and 2.5 in entertainment scenarios.

Why are AI data centers sparking water rights disputes in Arizona? ↗

The Verge reports on the escalating water resource conflicts in Arizona driven by AI data center construction.

A man from Kentucky, USA, becomes the first humanist granted an exemption from using AI for religious reasons. ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

New laws in New York State prohibit providing unsafe AI chatbot services to minors. ↗

According to data from the non-profit organization Common Sense Media, at least 72% of teenagers have used AI companions, with 52% being active users who interact multiple times per month.

The Huawei WATCH 5 series has launched HarmonyOS 7 for the Flower Pollen Beta version, featuring new AI call noise reduction and synchronized earthquake warning messages from mobile phones. ↗

The Huawei WATCH 5 series smartwatch has begun rolling out a new version (7.0.0.107 SP18). According to the introduction, the new system interface has been completely refreshed, adding features such as the HarmonyOS 7 Light Dome watch face, AI call noise reduction, and synchronized earthquake warning messages from Huawei phones. Additionally, the negative one screen Xiao Yi suggestions now support notification summaries, important dates, commute overviews, and other themes.

X more thing: XHEART X9 large model chip launch planned for 2027 with computing power boosted to 3000 TOPS ↗

At the end of today's event, XHEART unveiled its "X more thing". For the first time, XHEART disclosed the core R&D goals for its next-generation large model chip: 3000 TOPS of computing power and 700GB/s memory bandwidth per chip, with a launch scheduled for 2027. #XHEART #XHEART

XHEART launches its intelligent driving large-model chip X7 Ultra: featuring 700 TOPS of large-model computing power and the world's first mass-produced Armv9 automotive-grade CPU in China. ↗

XHEART has officially launched its new model chip, the X7 Ultra. Guided by the philosophy that large models define chips, the architecture was designed based on real-world mass production needs. The chip delivers a large-model computing power of 700 TOPS.

Microsoft AI executive Ramanna is pushing for a flatter organizational structure across Office and Copilot, requiring core staff to work in the office four days a week. ↗

Ramanna, head of Microsoft's Copilot and Office business, announced changes to team management. To adapt to the agent-driven AI era, he is implementing a flatter structure while mandating that senior core employees work at least four days a week. He also offers relocation options for staff to the headquarters or core office areas.

South Korea will pursue criminal liability and impose a three-year ban on college entrance exam candidates caught cheating with AI glasses. ↗

Starting this year, South Korea plans to extend the disqualification period for using smart glasses during the national university entrance exam from one year to three years and will also pursue criminal liability. Exam sites will now have dedicated staff to identify smart glasses, proctors will be uniformly covered by liability insurance, and signal detection equipment is under development to strengthen on-site supervision.

Vidu Q4 Preview partners with updream to explore professional production of AI animated dramas. ↗

Vidu Q4 Preview partners with updream to explore professional production of AI animated dramas. ↗

Vidu partners with updream to integrate the Vidu Q4 Preview native model into updream's AI video creation workflow, conducting end-to-end validation across three metrics: shot sequencing, character consistency across scenes, and high-dynamic action generation.

Momenta CEO Cao Xudong publicly reveals third insight: large models define chips ↗

At the launch event of strategic partner XHEART New Chip Voyage, Momenta CEO Cao Xudong publicly shared the company's third insight for the first time: large models define chips. Focusing on its goal of scaling to Level 4 autonomy, Momenta previously developed two insights: solving long-tail problems through data-driven approaches and generating massive datasets through mass production.

Microsoft earned $10 billion in revenue from OpenAI in its first quarter of fiscal year 2027, accounting for nearly 11% of total revenue. ↗

New regulatory documents reveal that Microsoft earned $10 billion from its OpenAI partnership in the first quarter of fiscal year 2027, representing nearly 11% of total revenue, indicating a deep financial tie between the two parties.

Infini-AI-Lab has open-sourced the Hermit harness adapter framework, which allows adapting agents like Codex and Claude Code without modifying the underlying model or harness. ↗

Infini-AI-Lab has released Hermit (v0.1.0, Apache-2.0), an open-source harness adapter framework. It trains external adapters while keeping both the model and harness frozen, using RL-style training loops to iterate through adapter versions. These can be deployed on Codex, Claude Code, or Pi, and exported as plugins.

Anthropic has launched a Claude AI reporting tool to monitor employee usage and track associated costs. ↗

The new feature is available to Enterprise paid users. It generates charts showing employee AI usage, output, and costs, while also identifying inefficient use cases. Anthropic explicitly states that the tool is not intended for personnel decisions and prohibits querying sensitive employee information.

JPMorgan advises clients not to share bank account passwords with AI assistants. ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

Valve has no plans to take action against AI-generated slop or clone games. ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

Designed for Large Models: XHEART's X7 Ultra Chip Features China's First Mass-Produced Armv9 Automotive-Grade CPUs ↗

At XHEART's launch event today, CEO Li Zonglin officially introduced the new X7 Ultra large model chip. The chip features six key advantages: a massively parallel NPU designed specifically for large models, hardware-level parallelism, a restructured AI toolchain, China's first mass production of Armv9 automotive-grade CPUs, deep heterogeneous computing capabilities, and comprehensive security solutions.

XHEART X7 Ultra large-model chip unveiled: 700 TOPS computing power, production model achieves 99.99% alignment in real-world tests ↗

#XHEART launches large-model chip# XHEART's new CEO Li Zonglin outlined the company's core technology roadmap at the event: "Large models define chips." As autonomous driving shifts from multiple small models to end-to-end large models, chip design methods must also evolve.

Epic executive Vasmer: We are more concerned about falling behind than facing external criticism of AI ↗

Vasmer admitted that compared to external criticism of AI, the company is more worried about the risks of falling behind in development: 'We believe that compromising on AI during development does not work. The best approach is to be transparent and share our methods with other developers, always use the best tools available at the time, and do Unreal Engine 6 as well as possible.'

Nvidia opens its most powerful transcription AI model Nemotron 3.5 Transcribe: Supports 40 languages, accurately transcribes even when eight people talk over each other ↗

Nvidia announced yesterday (October 28) the release of its open-weight speech transcription model Nemotron 3.5 Transcribe and simultaneously published its training data pipeline Granary v2.

AI is reducing demand for entry-level programming roles, causing a sharp drop in U.S. university computer science enrollment. ↗

According to Bloomberg, the impact of AI on junior programming roles and increased difficulty for graduates in job hunting has caused a sharp decline in undergraduate enrollment at US universities. Undergraduate students decreased by 18.4% over two years, while graduate students dropped by 27.1%.

AINews: Gemini 4 Carbon tops independent rankings and is now open to cybersecurity professionals. ↗

Latent Space released its AI News issue for October 27–28. The top story covers Google DeepMind's Gemini 4 Carbon, which focuses on cybersecurity and long-horizon tasks. It scored 88.2% on the SEC-bench Pro benchmark and topped the Artificial Analysis Intelligence Index v5 with a score of 59.

Xiaomi MiMo Desktop membership benefits upgraded: increased quotas at no extra cost, plus new support for third-party models like DeepSeek. ↗

Xiaomi MiMo officially announced today that Xiaomi MiMo Desktop membership benefits have been upgraded.

Instagram's head of engineering, Mosseri, admits that the public generally dislikes AI and predicts the situation will continue to worsen. ↗

He stated: 'The situation is not only getting worse but will continue to deteriorate. Over the coming years, dissatisfaction and anger toward AI will intensify further, with some reasons being fully justified and others not. Many people will find themselves in difficult positions, and those affected will not be limited to industry professionals.'

Claude Code 2.1.311 introduces a new environment variable, allowing AI agents to delete files without prompting. ↗

Tech media outlet The Register published an article yesterday (October 28) reporting that Anthropic has added a new environment variable to Claude Code version 2.1.311. When enabled by users, this allows AI agents to perform actions such as deleting files without seeking approval for each individual step.

A man who used AI tools to gather recipes and manufacture 'obedience water' was sentenced for the crime of manufacturing drugs. ↗

With the help of AI tools, Hangzhou man Sun collected recipes and production processes, purchased experimental equipment and chemical raw materials online, and independently synthesized 56.62 grams of the drug GHB known as 'obedience water'.

Altera unveils the Agilex 2X FPGA series to support edge physical AI applications. ↗

Agilex 2X leverages Altera's proven architecture while introducing enhanced AI tensor DSP modules, the latest Arm cores, high-bandwidth connectivity, end-to-end security, and other key innovations.

Cooling down AI chips: China's first 12-inch multi-crystal silicon carbide substrate debuts in Changsha, passing TSMC certification testing and sending samples to Samsung. ↗

Hunan Titan Future Technology Co., Ltd. released a press statement this month announcing the launch of China's first 12-inch multi-crystal silicon carbide substrate, targeting the thermal management bottlenecks in AI chips and advanced packaging.

Tencent and Visa announce strategic cooperation covering cross-border payments, enterprise AI, agent payments, and other areas ↗

Tencent and Visa announced today that they have signed a cooperation memorandum to explore broader collaboration, upgrading the global payment and business experience within Tencent's ecosystem.

The South Korean government plans to invest 5 trillion won to support small and medium enterprises in adopting physical AI. ↗

South Korea's government plans to invest at least 5 trillion won (approximately 25.145 billion RMB at current exchange rates) by 2030 to encourage small and medium manufacturing firms to adopt physical AI. This initiative, known as the '3X Strategy,' combines digital transformation, artificial intelligence, and robotics to raise AI adoption rates among SMEs from 1% to 10%.

China's first Huawei Ascend 950DT super-node cluster has launched in Wuhan, fully supporting the training of trillion-parameter large models and inference for AI agents. ↗

The Ascend 950 thousand-card super-node intelligent computing cluster, built by Hubei Science Investment, has officially been lit up. China's first Ascend 950DT 1,024-card intelligent computing cluster is now operational in Wuhan.

Tencent launches Buddy One: a cross-application AI collaboration system that connects multiple Tencent office products and shares the WorkBuddy Agent foundation. ↗

At today's Tencent Global Digital Ecosystem Conference, Tencent officially announced the launch of Buddy One, a cross-application AI collaboration system designed for individual users.

Some GPU computing contracts have been extended; Microsoft has lowered its 2026 capital expenditure guidance to $170 billion. ↗

Microsoft originally expected to receive additional computing power from third-party suppliers, but some GPU service contracts have been extended into the second half of this fiscal year. These agreements may involve other vendors providing AI cloud services powered by NVIDIA GPUs.

Bully has launched an AI Personalized Radio called 'Good Morning Radio', which generates 10-minute daily podcasts based on user habits and interests. ↗

Bully will open a Beta test for its AI Personalized Radio on November 2nd for The Clock users. Non-clock users can also apply through BALMUDA Connect to experience the feature with a limited offer of 1,000 slots.

LibTV has released an open-source project canvas for AI short films with over a billion views. The author provides a hands-on breakdown of the workflow for running a solo production crew. ↗

Bilibili has made public the complete engineering canvas behind hit AI short dramas like 'Fei Yao Zai', revealing their prompts, storyboard sequencing, and methods for maintaining character consistency.

Visual understanding is being introduced for the first time. Tencent Hunyuan Hy4 and Hy4 Flash AI models are set to launch. ↗

Tencent Group Senior Executive Vice President and CEO of Cloud and Intelligent Industry Business Group, Tang Daosheng, announced on-site that Hy4 is about to launch. (Shanghai Stock Composite Daily Report)

Tencent Yuanbao AI notepad serves as official AI stenographer at 2026 Tencent Global Digital Ecosystem Conference ↗

Tencent Yuanbao AI notepad served as the official AI stenographer at the 2026 Tencent Global Digital Ecosystem Conference, handling full-session note-taking and summaries. Users can access this feature by updating to the latest version and selecting 'AI Notepad' from the bottom menu.

Ollama AI open-source update v0.40.13: llama.cpp fixes Metal unary crash (#19372) ↗

Source summary: This temporary patch works around a regression on older MacOS versions where ggml_metal_library_compile_all fails to build. 'unary' library

Tongyi Qianwen launches 'Learning Archive': Linking photo uploads, grading, and mistake notebooks to generate learning reports ↗

Tongyi Qianwen's new 'Learning Archive' connects photo uploads, homework grading, and mistake notebooks to create personalized learning reports. In the case study, Grade 9 student Xiao Lin reviewed his records from the past three months and discovered that while his overall math accuracy was 63.4%, he missed all 18 questions on specific topics like the perfect square formula and quadratic equation discriminants. Tongyi Qianwen then automatically generated 15 targeted practice problems covering five key weak areas.

Valve responds to 'flood of low-quality AI games on Steam': Continuously optimizing the works recommendation mechanism ↗

The prevalence of low-quality AI games and counterfeit titles on Steam has sparked concerns among developers. Valve responded at a Korean developer event, stating it will address the issue by improving its intelligent recommendation system rather than restricting developer tools. The company also pledged to assist in handling cases involving counterfeit ransomware.

First batch of FY satellite extreme weather AI models launched; generates a 'typhoon health check' in just 27 seconds ↗

The newly released Typhoon Intelligent Cloud Interpretation Model integrates multi-channel cloud imagery and related products from the FY series satellites. It can automatically analyze a typhoon's location, structure, intensity, and evolution trends, then generate an interpretive text report.

A science book featured an image of a 'two-headed shrimp': the publisher ordered a nationwide removal and recall. Insiders say the AI depiction of molting caused public misunderstanding. ↗

App Official Account Submission Comment Top [](https://www.ithome.com/)[](https://img.ithome.com/app/songs/index.html) Home IT Circle Best Buy Settings Day/Night Follow System Light Dark Theme Color Black Submit Subscribe RSS Subscription Bookmark IT Home Ruanmei Apps App Client Yaozhi App Ruanmei Magic Cube

OpenRouter tests 13 prompt injection attacks and explains Armor and Guardrails defense methods. ↗

OpenRouter releases a tutorial on prompt injection defense, testing Claude Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash, and GLM 5.3 Flash with 13 in-house attacks and 12 from other models.

OpenRouter tutorial: Extracting tables from PDFs using LLMs. ↗

OpenRouter publishes a guide on sending PDFs alongside strict JSON schemas to file-reading models, ensuring tables return rows with types while cleaning brackets, thousands separators, and footnote markers within field descriptions.

Ted Sanders explains OpenAI's model naming convention: Luna < Terra < Sol < Astra ↗

OpenAI's Ted Sanders explains its model naming convention: models are ranked by Luna. < Terra < Sol < Astra's ranking system prioritizes scale: larger models offer superior performance and reliability but incur higher per-token costs. They also excel at retrieving obscure facts. Version numbers indicate release order, relationship to other versions, and the magnitude of improvements; for instance, version 6.1 was released after 5.6, shares training data and style with its peer 6.1 models, yet the incremental gains are insufficient to warrant a version 7 designation.

Surging electricity demand from AI data centers may undermine US power sector emissions reductions ↗

An internal analysis by the Sierra Club, a US environmental organization, suggests that the boom in fossil fuel power plant construction driven by AI data center energy needs could negate the greenhouse gas emission reductions achieved by the US power sector over the past 15 years.

Tencent Cloud Drive opens beta application: supports connecting multiple AI applications, sharing storage and aggregating context. ↗

Tencent Cloud Drive has now opened its beta application; the official website entry is available, allowing users to submit directly.

Major AI industry merger: Bloomberg reports Marvell plans to acquire Lumentum, the leading optical communications firm, with an announcement expected as soon as next week. ↗

Bloomberg published a post today (October 29) reporting that American semiconductor company Marvell is planning to acquire Lumentum, the manufacturer of data center optical communication equipment. Both parties are currently in negotiation, and the deal could be announced as early as next week.

Adam Mosseri: Negative emotions about AI will 'get worse' ↗

Mosseri stated that Instagram will not treat AI creators differently but acknowledged that the ranking quality of AI content may be lower than non-AI content. He also noted that the platform has begun requiring AI accounts to self-label their posts.

Valve responds to Steam's flood of low-quality AI games: AI won't be banned, but the recommendation system will change ↗

At the Steam Indie Connect event held in South Korea, Valve addressed concerns over 'low-quality AI games' during the Steam Next Fest. The company stated it is researching solutions but will not prohibit developers from using AI tools.

Anthropic urges California: Courts must have the power to block AI models presenting serious risks. ↗

In a letter dated October 15 to California Governor Gavin Newsom, Anthropic recommended that governments should have the authority to seek court injunctions to prevent the release of AI models posing a 'serious risk' to the public.

Core advertising revenue growth slows as Google expands AI search ads globally. ↗

Google is pushing its advertising business within AI search experiences into international markets, betting that generative AI will unlock new revenue opportunities even as its core search ad growth slows.

Practical vulnerability hunting with Claude AI: extracting photos from a locked Android screen and gaining highest privileges on Linux in just 53 minutes. ↗

Anthropic released a research report on October 27, stating that an undisclosed internal Claude model identified vulnerabilities in Chrome, Safari, Android, Windows, and Linux. The researchers linked multiple flaws to create five actionable attack chains. All related vulnerabilities have been reported to vendors via coordinated disclosure processes and have since been patched.

Ollama AI open-source update v0.40.13-rc1: llama.cpp fixes Metal unary crash (#19372) ↗

Source summary: This temporary patch works around a regression on older MacOS versions where ggml_metal_library_compile_all fails to build. 'unary' library

Google expands AI search ads to international markets; core search ad growth slows ↗

Google announced at the Alphabet Q3 earnings call that it is expanding AI Mode ads to international markets for the first time and launching multi-turn conversational ads in the US. At the same event, Alphabet reported quarterly revenue of $127.6 billion, up 25% year-over-year, while Google Search revenue growth slowed from 17% to 12%. Stock prices fell 8% following the report. Google stated that sponsored product click-through rates in AI Mode are now comparable to organic results.

Microsoft generated $10 billion in revenue from OpenAI for the quarter alone. ↗

In regulatory filings, Microsoft disclosed that it recorded $100 billion in revenue from its commercial arrangement with OpenAI for the quarter ended September 30, accounting for nearly 11% of its total quarterly revenue of $923 billion. Revenue from OpenAI is projected to reach $241 billion for the full fiscal year ending in 2026.

Global AI Infrastructure Trendsetter: NVIDIA to Release Q3 FY2027 Earnings Report on November 17 Pacific Time ↗

NVIDIA (NVIDIA) announced on October 28 that it will hold a conference call at 2:00 PM Pacific Time (6:00 AM Beijing Time) on November 17 to present and analyze the financial performance for the third quarter of fiscal year 2027.

Google's top AI model: Gemini 4 Carbon takes the lead on the Global Intelligence Index, specializing exclusively in cybersecurity. ↗

Google has launched the new frontier model Gemini 4 Carbon, which has achieved significant breakthroughs in cybersecurity and long-range programming tasks. The company has simultaneously announced limited-time introductory pricing followed by official rates, with access currently being rolled out in phases to cybersecurity professionals.

Google's Q3 2026 earnings reveal anxiety over AI monetization, with shares plunging more than 7% after market close. ↗

Following the release of its FY2026 Q3 report (covering January to September 2026) and the third quarter results (July to September 2026), Alphabet's shares fell by more than 7% after market close.

Latent Space interviews Supabase founders: Why agents favor Supabase ↗

Latent Space podcast interviews Supabase founders Paul Copplestone and Ant Wilson on why agents overwhelmingly choose Supabase as their application backend. The author's AEO tracking shows it has become the top choice across nearly all cutting-edge models.

Sierra has launched Compliance Console to provide compliance control for long-duration AI agents. ↗

Sierra has launched Compliance Console, integrated into the Horizon platform. It helps enterprises regulate AI agent outbound calling behavior across channels and regions. The console supports configuring disclosure information, contact time windows (defaulting to 8 AM to 9 PM in most regions), holiday blocking, call frequency limits, and opt-out handling. It also displays outbound call and message counts that were allowed, delayed, blocked, or pending review via the Insights dashboard.

Google launches Gemini 4 Carbon, focusing on enhanced cybersecurity capabilities ↗

Google unveils its latest frontier model, Gemini 4 Carbon, stating it has significant improvements in coding, agent tasks, and cybersecurity. It currently ranks first on the Artificial Analysis Intelligence Index (v5.0), surpassing Claude Opus 5.5 and GPT-6 Astra.

Anthropic launches enterprise reporting tool to quantify Claude token expenditure returns ↗

Anthropic released a new enterprise reporting tool on Wednesday. It allows managers to analyze team performance by reviewing work completed with Claude within specific date ranges, along with associated costs and workflow bottlenecks, while providing improvement suggestions. The tool does not use data for model training by default and includes shared budget tokens for team spending.

Google DeepMind launches Gemini 4 Carbon, topping the Artificial Analysis Intelligence Index ↗

Two weeks after launching Argon, Google DeepMind releases Gemini 4 Carbon, scoring 59 points on the Artificial Analysis Intelligence Index v5.0, up from Argon's 50 points.

Replit launches ViBench 2.0, a long-haul software development benchmark where GPT-6.1 Sol leads with a 65% pass@1 score ↗

Replit launches ViBench 2.0, a long-haul software development benchmark where agents build complete products from scratch based on functional requirements, including 17 products, 241 feature requests, and 544 test plans, while also incorporating Fireworks' Specialized Intelligence Index.

Google launches Gemini 4 Carbon, which will be rolled out in phases as part of the Fairwinds cybersecurity initiative. ↗

Two weeks after launching Gemini 4 Argon, Google announced its next-generation frontier model, Gemini 4 Carbon. The company claims improvements in coding, agent tasks, and cybersecurity, but the model remains closed to the public.

Google DeepMind launches Gemini 4 Carbon, targeting network defense and agent programming ↗

Google DeepMind launches Gemini 4 Carbon, its most powerful model to date, focusing on long-term agent programming and cybersecurity defense. It ranks first in the Artificial Intelligence Index v5.0 with a score of 59.

Wiz CTO discusses the security crisis facing legacy internet code and the impact of AI ↗

TechCrunch Equity podcast invites Tomer Schwartz, CTO of Wiz (a Google company), to review the patching process for two critical PHP vulnerabilities last week. The language powers approximately 70% of websites, with some sites forced offline as a result. He discusses his collaboration experience with OpenAI, Calif, and the PHP team, noting that AI has accelerated the speed at which vulnerabilities are discovered and exploited. He also explains why manual patching cannot scale and how startups can build systems securely.

Up to 128GB unified memory: Apple's first touch-enabled MacBook Pro can run billions-of-parameter AI models locally ↗

Apple today (October 29) launched its first touch-enabled MacBook Pro featuring the PROLED XDR display. A key highlight is up to 128GB of unified memory, enabling the M5 Max chip version to run large language models with billions of parameters locally.

Databricks on Voice Agents: Using Context to Enable Real Customer Value in Voice Support ↗

Databricks states that the key to deploying voice agents lies in context rather than voice itself. Its prototype uses LiveKit for calls, Deepgram for speech-to-text and default voices, ElevenLabs for expressive voices, while Databricks provides the core logic.

Baseten optimizes Qwen3-ASR 1.7B batch transcription, achieving throughput seven times that of the original vLLM version. ↗

Baseten has optimized the deployment of Qwen3-ASR 1.7B for batch processing; at 64 concurrent requests, throughput exceeds native vLLM by more than seven times, with costs dropping to approximately $0.0013 per audio hour—about 277 times lower than closed-source providers.

Google AI Mode gave an evasive response when asked about the winner of the 2020 US presidential election. ↗

In contrast, for every presidential election since 2000, AI Mode provided detailed and accurate answers about the winners.

Apple introduces the Dynamic Island to Mac for the first time: it allows users to display music playback and live activities, adding a new entry point for Siri AI. ↗

Apple launches the first touchscreen MacBook Pro in 14 and 16-inch sizes. For the first time on Mac, it features the Dynamic Island, which debuted on the iPhone 14 Pro series in 2022. The feature supports live activities, Siri interactions, and call control.

Ed Zitron's long-form analysis of Anthropic's S-1 filing claims the company treats retail investors as a liquidity exit strategy. ↗

Ed Zitron criticizes Anthropic's S-1 filing in a new article, arguing that it misleads investors by disclosing only Q2 earnings while omitting Q3 cash flow and payment details before the company's fundamentals change.

Google launches Connected Apps Platform to open Gemini and Search access to developers. ↗

Google launches the Connected Apps Platform, allowing enterprises and developers to submit third-party applications for integration with its Search and Gemini AI experiences. The platform is currently in beta and available to US developers.

Ollama AI open-source update: v0.40.13-rc0 ↗

Source summary: llama.cpp version bump b11963 ( #19360 )

Google DeepMind launches Gemini 4 Carbon, a cutting-edge model ↗

Google DeepMind unveils Gemini 4 Carbon as its top-tier cutting-edge model, focusing on cybersecurity and long-range agent capabilities. It achieves leading scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, topping the Artificial Analysis Intelligence Index v5.0 with a score of 59.

Google has opened its Android development environment to third-party AI agents, launching a new Gemini API and XREAL AURA glasses. ↗

At the Android Dev Summit, Google announced that Android Studio Rabbit 2 now supports Bring Your Own Agent. Developers can install and run over 40 coding agents—including Claude Agent, OpenAI Codex, and Google Antigravity—simultaneously by adopting the open Agent Client Protocol.

OpenAI Berlin DevDay: Dots will launch in Europe, with a focus on locking developers into the platform. ↗

At the Berlin DevDay Exchange 2026, OpenAI revisited an old announcement and revealed that its AI assistant Dots is expected to launch in Europe soon."Within a few weeks"Personal Pro accounts in the European Economic Area, Switzerland, and the UK will still not be available.

Anthropic adds source-verification artifacts for Claude for Legal ↗

This feature also supports cross-border legal comparison, case file management, legislative tracking, and contract review against internal guidelines. With over 60 integrations in the legal directory—including Bloomberg Law—users can directly connect to their own databases.

The Artificial Analysis Intelligence Index v5 benchmark has launched. DeepSeek V4.1 Flash 1017 achieved the highest score for FrontierCode 2.0 among open-source models. ↗

Artificial Analysis has restructured the Intelligence Index v5 benchmark, using private test sets for 60% of evaluation weight to reduce inflated scores. DeepSeek V4.1 Flash scored 28% in a single category within FrontierCode 2.0, the highest among released open-source models. #DeepSeek#AI Benchmark

Chrome's enhanced autofill now uses Gemini to extract information from Gmail for filling forms. ↗

Google launched an update to Chrome's enhanced autofill in the US today. Through Gemini with Personal Intelligence, it suggests filling in details extracted from Gmail, such as package tracking numbers, flight booking references, hotel reservations, and order numbers.

Chrome launches AI-enhanced auto-fill to populate forms using data extracted from Gmail. ↗

Google has introduced an AI-powered version of Chrome's auto-fill feature. It can reference content in a user's Gmail account and suggest filling in details such as flight confirmation numbers, membership IDs, order numbers, addresses, phone numbers, or credit card information.

a16z surveys 100 CISOs: How AI is reshaping the security agenda for 2027 ↗

a16z conducted a survey of over 100 CISOs from companies with more than 1,000 employees. 87% expect security spending to grow in 2027, driven primarily by AI threats and the expanded attack surface caused by their own AI deployments. All seven top security initiatives for 2027 are AI-related. SOC automation received the most first-choice votes, while internal AI agent security, SOC automation, and offensive security saw the largest increases. However, the adoption rate of AI SOC tools in production environments remains at only 18%.

GitHub Universe 2026 launches new AI agent products including GitHub Copilot app and Copilot Fusion. ↗

At GitHub Universe 2026, GitHub unveiled its vision to become a complete software delivery system, launching multiple products including the GitHub Copilot app designed for agent-based delivery.

GitHub Octoverse 2026: Development activity hits a record high, with over 40% of pull requests drafted by AI agents. ↗

GitHub releases its 2026 Octoverse report: Developers pushed code 17.9 billion times over the past 12 months, exceeding the combined total of 2024 and 2025, with 130 million pull requests.