2026-10-12
AI Briefings Daily Report.

Microsoft plans to expand its self-developed Maia AI chips to gigawatt-scale capacity by 2027 ↗

Microsoft plans to deploy Azure Maia AI chips at multi-gigawatt scale starting in 2027 as the primary inference system for its products. Maia 200 is already running Microsoft Copilot in the US, Europe, and Australia; GitHub Copilot and Foundry will follow by year-end; Microsoft claims its token throughput per watt in Teams scenarios leads GPUs by up to 1.7 times.

Microsoft President Brad Smith says the UK is best suited to lead global AI regulation ↗

Microsoft President Brad Smith called for urgent AI regulation, arguing that consumer goods like milk and washing machines face more oversight than AI companies, and stated that the UK, rather than the US or EU, is best suited to lead global safety rules. He mentioned that the UK AI Security Institute could gain access to models before deployment at US frontier labs, and advocated for external assessments before AI product releases, mandatory breach notification obligations, and kill-switches in case of loss of control.

Blueprints AI claims it can generate building blueprints ready for permit applications within minutes ↗

Blueprints AI has launched its second-generation software, which can generate building-code-compliant construction blueprints ready for permit applications from 2D sketches or PDFs within minutes. It supports simultaneous output for architectural, structural, mechanical, electrical, plumbing, and interior design disciplines; a change in one area updates the entire project and validates against local building codes. The first-generation software cost between $3,200 and $8,500 per month. The second generation is now open to early customers and will be publicly released this week, while the third generation will include structural analysis and panel board design.

SemiAnalysis founder: AI computing power demand still far exceeds supply, infrastructure scale underestimated ↗

SemiAnalysis founder Dylan Patel and technical team member Jordan Nanos stated on the Big Technology Podcast that AI computing power demand continues to outstrip supply, and widely circulated figures regarding infrastructure scale remain underestimated.

NVIDIA and Cisco Launch New USFP Pluggable Optical Module Form Factor for AI Factory Expansion ↗

NVIDIA and Cisco announced the launch of a new form factor for Ultra-Dense Small Form Factor Pluggable (USFP) optical modules, the first to support 400Gb/s-per-lane electrical signals, with a single module capacity reaching 12.8Tb/s.

Why haven't AI companies been prosecuted for hacking after system intrusions? The author argues that intent can be established. ↗

The author analyzes incidents where AI systems from companies such as Anthropic, OpenAI, Google, and Meta have gained unauthorized access, yet no legal consequences have followed. The article points out that these companies claim a gap in liability due to emergent behavior in their systems, thereby evading the requirement for criminal intent. The article argues across four dimensions—training data, training processes, environment, and configuration—that companies exert broad control over model capabilities. If designs foreseeably lead to intrusion behaviors, intent can be established. The author calls for prosecution under existing hacking laws rather than shifting to future AI governance discussions.

Microsoft Unveils Maia 300 AI Chip: TSMC N2P Process, 288GB HBM4 Memory, 18TB/s Bandwidth ↗

The Maia 300 features a multi-chiplet design built on TSMC's N2P process, delivering FP4 performance of 13.8 PFLOPS with a power consumption of 950W. It includes eight HBM4 memory modules with a total capacity of 288GB and bandwidth of 18TB/s, making it suitable for high-memory AI inference. The bidirectional interconnect speed reaches 3.2TB/s, allowing up to 96 chips per rack and a cluster size of up to 12,000 chips, with deployment planned for 2027.

AMD launches next-generation Ryzen AI Halo developer platform, upgrading Max+ PRO 495 with up to 192GB memory ↗

New products feature improved CPU, GPU, and NPU performance, allowing 160GB to be allocated as shared video memory, making it possible to run local large AI models exceeding 300B parameters.

AMD launches Helion, an agent AI computer rack with a pure CPU version scalable to 32,768 cores ↗

On Helion, partners and customers can flexibly combine diverse CPU and GPU computing resources according to their own needs.

Microsoft releases third-generation self-developed AI chip Maia 300, set for production in 2027 to reduce dependence on Nvidia GPUs ↗

Microsoft releases third-generation self-developed AI chip Maia 300, positioned as the 'primary AI inference system' for its products and services, with plans to begin production in 2027 and expand to multi-gigawatt deployments.

White House restores access to frontier models for the UK's AI Safety Institute ↗

The White House has revoked its previous ban preventing the UK's AI Safety Institute (AISI) from accessing new models, restoring the institute's early access to models from Anthropic, Google DeepMind, and OpenAI.

Andrej Karpathy says the outside world has only seen previews of AI capabilities; internal frontier systems are already handling tasks that experts would take years to complete. ↗

OpenAI co-founder and current Anthropic member Andrej Karpathy proposes a four-tier classification for large language model users: about 6 billion people have barely interacted with LLMs, while 1–2 billion people use free products as better search or writing assistants.

How WebMCP Enables Websites to Be Designed for Both Humans and AI Agents ↗

WebMCP (distinct from MCP) provides web developers with a standard method to register structured tools—including name, description, and input schema—via document.modelContext.registerTool, allowing agents to directly invoke website capabilities. It also supports declarative tool exposure through HTML form attributes.

Former Plaud executive starts Ambi: $26 million raised for an all-day background listening AI app ↗

Ambi, founded by former Plaud executive Peter Mo, has completed a $26 million multi-round financing round led by Luminous Ventures and Huaqin, and launched apps for iPhone, Mac, and Apple Watch. The app listens in the background around the clock, with an AI agent actively providing suggestions during meetings. It supports cross-device task continuation, with environmental data collection turned off by default and no audio recordings stored.

Exponential View: Why the Impact of GLP-1 May Surpass AI ↗

Azeem Azhar writes that the impact of GLP-1 drugs may surpass that of AI; sales reached $76 billion by the end of 2025, exceeding generative AI's approximately $63 billion; the usage rate among American adults rose from 3% in 2024 to 11%. The article outlines the price collapse (monthly cost for Indian generics is about $13, 99% lower than the original price), the fact that about half of users stop taking the medication within a year, and potential spillover effects such as alcohol cessation and slowed heart aging, describing it as a form of societal restraint technology.

QGIS passes QEP428 to ban AI-assisted contributions ↗

QGIS passed proposal QEP428, changing its AI policy to strictly NOT ALLOWED. Submitted by nyalldawson and merged into the QGIS-Enhancement-Proposals repository on October 7, 2026.

Anker AI Meeting Headphones Feishu Edition Now Available: Supports Recording in All Online and Offline Scenarios, ¥1799 ↗

The Anker AI Meeting Headphones Feishu Edition, jointly launched by Anker and Feishu, officially went on sale tonight at 20:00 with a launch price of ¥1799 and a launch subsidy price of ¥1376.23.

Alibaba Qwen releases QwenGyre, an online reinforcement learning framework for ultra-long-horizon agent tasks ↗

QwenGyre addresses the challenges of training ultra-long-horizon agent tasks through elastic scheduling and trajectory processing, dynamically redirecting idle GPU units to training. This boosts training efficiency by 1.78 times and raises the model pass rate from 52.5% to 58.5%. ##Alibaba Qwen##

Microsoft details Edge browser's new AI troubleshooting feature: overall similar to a chatbot, it only collects diagnostic data after user consent ↗

The feature is integrated into Edge settings in the form of an AI chatbot. It only collects necessary diagnostic data for the current session when actively initiated and authorized by the user, with no background data collection.

Intel and Broadcom CEOs Issue Joint Warning: Funding Shortages May Become the Biggest Bottleneck for AI Development ↗

Two CEOs warned that as the industry races to build infrastructure, financing may become a greater obstacle to AI expansion than semiconductors, storage, and power. However, they remain optimistic about short-term demand, believing that supply-demand imbalances could persist until 2030.

Comparison of 12 AI Real-Time Translation Tools in 2026: Latency, Languages, and Pricing ↗

MarkTechPost evaluated 12 AI real-time translation tools based on official product and pricing pages, covering scenarios such as two-way calls, meetings, and conferences.

AI coding agents generated more code, but not more software ↗

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'.

Developers used AI agents to port Call of Duty: Black Ops 2 to the web, but it was taken down following a DMCA copyright complaint from Activision ↗

Developers used AI "vibe coding" to create two unofficial Call of Duty web games that recreated the original maps and gameplay. Although the developers stated they were not affiliated with the official game, Activision issued a copyright complaint due to the use of original game assets, leading to their removal.

Wallbox Intelligence and China Telecom Artificial Intelligence sign strategic cooperation agreement to promote large-scale implementation of edge-side models ↗

Intel CEO Chris Kratzer: The adoption speed of cutting-edge Chinese AI models is very fast ↗

Chris Kratzer stated at a DBS Bank panel discussion that Chinese AI models such as DeepSeek, Kimi, and GLM have sufficient performance, low token costs, and rapid adoption. He firmly supports open source, believing the market will determine the application scenarios for different models. He also mentioned that quantum computing is expected to go mainstream within three to five years. #

HarmonyOS 7 Xiaoyi AI Shortcut Bar feature release date announced; expected to roll out in batches by late October ↗

At the HarmonyOS 7 | HUAWEI Mate XT 2 and all-scenario new product launch event on September 7 this year, the HarmonyOS 7 operating system was officially released. IT Home learned today from Huawei Customer Service that the Xiaoyi AI Shortcut Bar is a new feature of HarmonyOS 7, expected to gradually appear with users by late October, with phased rollout based on device models.

An Meng: AI companies have asked Qualcomm to launch a mobile platform supporting local execution of 100B-parameter models by 2028 ↗

An Meng believes that AI agents will drive the next 'super cycle' for smartphones, with smart glasses being his most favored wearable AI device.

Wallbox Intelligence and China Telecom Artificial Intelligence sign strategic cooperation agreement to promote large-scale implementation of edge-side models ↗

Wallbox Intelligence and China Telecom Artificial Intelligence Technology Co., Ltd. signed a strategic cooperation agreement in Beijing. The two parties will jointly develop products and engage in deep cooperation around two major directions: terminal intelligence and industry intelligence. The cooperation will advance through joint solutions and joint promotion models, creating personalized edge-side AI applications for consumer electronics, focusing on standardized vertical industry models and intelligent agent infrastructure for enterprise-level applications, and collaboratively enhancing edge-side model capabilities while exploring key technologies such as inference acceleration.

Microsoft President Brad Smith: AI faces even less regulation than milk and washing machines ↗

Microsoft President Brad Smith stated that washing machines currently face stricter safety regulations than AI, and every carton of milk is subject to safety rules and inspections. He called for the industry to establish more safety specifications, requiring companies that possess out-of-control AI agents to bear responsibility for damages caused by those agents infiltrating other businesses, and suggested that companies implement a "kill switch" to halt operations if an AI goes rogue.

Report indicates that U.S. programmer hiring has increased rather than decreased since the widespread adoption of AI development, but the number of junior programmers employed has dropped by approximately 20% ↗

Data from Indeed Hiring Lab shows that since February 2025, job postings for software development positions in the United States have grown by nearly 15%, while overall employment market hiring numbers have declined by 7%.

Arena releases new weekly AI large model rankings: Claude Opus 5.5 Max tops the agent leaderboard for the first time; multiple domestic models remain stably in the upper-middle tier ↗

In this week's Arena AI large model weekly rankings, Claude Opus 5.5 Max topped the agent leaderboard upon its entry, while several domestic models remained stably in the upper-middle tier across comprehensive and coding leaderboards.

Microsoft launches decision model Microsoft-Decision-1 based on Qwen ↗

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'.

Vipshop quietly launches its self-developed AI assistant 'AI Xiaowei', testing product AI comparison features in a limited rollout ↗

Vipshop recently launched its self-developed AI assistant 'AI Xiaowei', developed by Vipshop's large model team. It supports text and voice conversations and provides services such as finding products, checking promotions, and answering discount inquiries.

Yann LeCun: Next-generation AI must understand the real world; world models will replace large language models ↗

Yann LeCun stated in Seoul that the next breakthrough in AI will not come from large language models (LLMs), but from "world models" capable of understanding and predicting the physical world. He pointed out that while LLMs excel at predictable sequence tasks such as programming and mathematics, real-world information is far more difficult to predict than language; it is easy for machines to read language, but hard for them to understand the real world. He advocates developing world models so that AI can learn physical laws and causal relationships and predict the outcomes of actions.

HP ZBook Studio G11 Ryzen Edition AI Mobile Workstation Now Available: AMD R9-9955HX3D + RTX PRO 2000, ¥21,999 ↗

The machine is equipped with an AMD Ryzen 9 mobile processor and an NVIDIA RTX PRO professional graphics card, supporting up to 200W total system performance release, featuring a high-refresh-rate, high-color-gamut screen, and passing professional certifications and military-grade tests.

TikTok CEO Shou Zi Chew: The market may overestimate AI's short-term capabilities while underestimating its long-term impact ↗

Shou Zi Chew pointed out that the long-term development potential of artificial intelligence is still underestimated, and AI is expected to drive major breakthroughs in fields such as physics and biotechnology within the next two to three years.

"Baidu World 2026" scheduled for November 12, featuring "keynote speakers and enterprise-level AI products" ↗

Baidu officially announced today that Baidu World 2026 will be held on November 12 at Beijing's National Convention Center Phase II. The conference website has now opened its ticketing channel.

Haier launches M1 Plus smart clothes drying rack: supports offline AI voice control, priced at 689 yuan ↗

The new product features 80 drying slots, supports zone-specific drying, has a slim body that saves space, adapts to various renovations and ceilings, includes a large-screen lighting system that can replace balcony lights, offers multiple smart functions, and is built with durable materials.

Big AI Theory: Squeeze Model Capabilities by Adjusting Reasoning Levels in Three Stages ↗

Unitree Robotics' Wang Xingxing: My biggest mistake over the past decade-plus was not believing in AI enough ↗

The 26th China International Industry Fair will be held from October 12 to 16, 2026, at the National Exhibition and Convention Center (Shanghai).

Big AI Theory: Squeeze Model Capabilities by Adjusting Reasoning Levels in Three Stages ↗

The author proposes a "Big AI Theory," using three-stage conversations to maximize model capabilities. Step one maximizes reasoning and selects plan mode, using voice, images, and video without filtering to dump ideas and organize them into a plan; step two switches to medium reasoning for quick execution, provided the requirement description is detailed enough—the author cites Terminal Bench 3.0 to show that detailed requirements with medium reasoning yield higher quality than vague requirements with high-end reasoning; step three adjusts reasoning back to high-end for the AI to review the medium-reasoning results, claiming it can identify about 20% of issues, with review time taking about one-third of generating from scratch.

Wuchen Xingqiong completes 2.3 billion yuan C-round financing; both single-round and cumulative amounts set new records for China's AI-native infrastructure niche ↗

AI infrastructure company Wuchen Xingqiong has completed a C-round financing of 2.3 billion yuan. (Securities Times)

Qiu Chengtong Mathematics and Artificial Intelligence Laboratory Alliance Established at Tsinghua University ↗

On October 11, the Qiu Chengtong Mathematics and Artificial Intelligence Laboratory Alliance was established at Tsinghua University. It will focus on developing specialized large models for mathematical sciences, AI-assisted formal proof and conjecture discovery in mathematics, and interpretable artificial intelligence. The alliance released over 10 key achievements and research directions in "AI + Mathematics," having gathered more than 1,000 mathematicians and researchers. It collaborates with institutions such as Tsinghua's Qiuzhen College, the Yau Mathematical Sciences Center, and the Beijing Yanqi Lake Institute of Applied Mathematics to explore a path for building China's own mathematical large models.

AI Reverses Everything: Brandon Thomas Uses Claude Opus 5.5 to Open-Source Rewrites of Seven Adobe Software Applications; REA Project Nears 100K Stars ↗

Developer Brandon Thomas, dissatisfied with Adobe's subscription cancellation penalties, used Claude Opus 5.5 to reverse-engineer and rewrite seven Adobe applications, including Photoshop, in Rust and open-sourced them. The Photoshop project alone garnered 46,000 stars in less than 10 days.

Sakana AI Proposes MASS: Achieving Homogeneous Recursive Self-Improvement via Multi-Agent Self-Supervision ↗

Sakana AI has released the MASS (Multi-Agent Self-Supervision) method, which allows a single model to simultaneously act as an executor, workflow optimizer, and evaluator, distilling the collective experience of multi-agent teams into self-supervised signals for Recursive Self-Improvement (RSI).

Ma Huateng Meets with Guangdong Provincial Party Secretary Huang Kunming Alongside 28-Year-Old Youth; Tencent Chief AI Scientist Yao Shunyu Makes an Appearance ↗

On October 8, the first working day after the National Day holiday, Guangdong Provincial Party Secretary Huang Kunming visited Shenzhen and toured Tencent's headquarters. During a discussion session, Tencent Board Chairman and CEO Ma Huateng spoke, seated next to 28-year-old Yao Shunyu, Tencent's Chief AI Scientist.

Global telecom operator's largest single intelligent computing center expansion project: China Mobile's Hohhot Intelligent Computing Center reaches a total computing power of 20,100 PetaFLOPS, serving 90% of the Beijing-Tianjin-Hebei region. ↗

After expansion, 90% of computing power services serve the Beijing-Tianjin-Hebei region, with green electricity accounting for 88% and an average PUE as low as 1.15, marking it as a core node in China's national 'East Data West Computing' project.

AI Prompts Found in Printed Book "The Lone Walker at the End of the Road"; Planning Team Apologizes, Citing Mistaken Treatment of Proofreading Software Prompts as Main Text ↗

Recently, a reader shared on social media that while reading the book "The Lone Walker at the End of the Road," AI prompts were inserted into the main text. Tianjin Zhiyu Shiguang Culture Media Co., Ltd. and the planning team for "The Lone Walker at the End of the Road" issued an apology letter stating that upon verification, the original manuscript did not contain this content; the error was caused by the planning team mistakenly treating prompts provided by proofreading software as part of the main text during the process of condensing the text.

Cloudflare Releases Clef-omni: A 30B-A3B Multimodal Decision Model That Scores Text, Images, Audio, and Video in a Single Call ↗

On October 9, 2026, Cloudflare released Clef-omni, an open-weight decision model that can score text, images, audio, and video in a single call.

OpenAI announced that the default context length for GPT-6.1 Sol has been increased from 272k to 500k, an increase of approximately 80%. ↗

Tibo stated that although OpenAI will continue to raise the limit in the future, this adjustment has already significantly reduced the frequency of context compression triggers and allows for more relaxed handling of tasks that require retaining large amounts of context and complex details over the long term.

Raise Financial Services partners with Cartesia to build multilingual voice AI for India's financial market ↗

Raise Financial Services is collaborating with Cartesia to integrate voice models into its proprietary real-time voice infrastructure, providing a multilingual voice experience for India's financial market. Cartesia's Sonic text-to-speech model supports 11 Indian languages and code-switching within a single generation; it is currently handling conversational interactions such as KYC in Dhan's voice bot, where over 75% of interactions use non-English languages, with Hindi accounting for more than half.

The UK will strengthen cyber defenses; the Minister for AI calls on artificial intelligence companies to increase security measures. ↗

Narayan stated that the UK must not rely entirely on tech giants for "self-regulation" of its models, emphasizing the need to build more autonomous and controllable AI infrastructure domestically.

After Anthropic's withdrawal, reports say Amazon is in talks to acquire AI startup Decart ↗

According to people familiar with the matter, Amazon is negotiating to acquire artificial intelligence startup Decart, with a potential valuation for the deal of approximately $7 billion.

Scale AI Founder Wang Tao on Startup Management: From Deciding Details Personally to Learning to Let Go; His Task at Meta Is to Recruit Talent ↗

After joining Meta, Wang Tao shifted his management approach. Since he does not directly cover all professional fields, his work focus turned to recruiting top talent for Meta's Superintelligence Lab.

vLLM Semantic Router releases open-source multimodal decision model Decision 3.0, available in five sizes ↗

The vLLM Semantic Router team has released the Decision 3.0 series of open-source decision models. These models process text, JSON, and images, returning probabilities instead of generating text, and are all licensed under Apache-2.0.

Silicon Valley giants follow Trump's lead in renaming AI to super intelligence ↗

The Verge analysis points out that after Trump announced he would rename AI 'super intelligence' and labeled those who continue to use the term 'Artificial Intelligence' as enemies, several Silicon Valley giants followed suit.

TechCrunch Disrupt 2026 is approaching: Over 250 speakers, including Anthropic, NVIDIA, and Cerebras, to appear ↗

TechCrunch Disrupt 2026 will be held at Moscone West in San Francisco from October 13 to 15, with over 10,000 founders and investors attending. More than 250 speakers will take the stage across 200+ sessions, including Cat de Jong from Anthropic, Les Karpas from NVIDIA, and Andrew Feldman from Cerebras Systems.

Can the AI industry persuade data center opponents by eliminating NDAs? ↗

Amazon recently followed Microsoft's lead, announcing that it will no longer use non-disclosure agreements (NDAs) in negotiations with local governments over data centers. This statement was buried in a blog post arguing that data centers benefit communities; Microsoft made a similar announcement earlier this year. TechCrunch's Equity podcast discussed that eliminating NDAs may not immediately change opponents' stances, but could be the first step toward opening dialogue, with poor industry communication being the root cause of the controversy.