AI news, Nov 4, 2026: Apple Opens Siri to Apps, Meta Expands AI Infrastructure
Apple has updated iOS 27.3 to allow Siri AI to access third-party apps, adding Traditional Chinese support to enhance its multimodal capabilities. Meanwhile, Meta issued $20 billion in bonds to expand its AI infrastructure. OpenAI and Anthropic are also developing internal Git backends, Pando and Scion, respectively, to manage the increased volume of merges and pushes driven by AI agents.
Lead written by AI from the day’s stories
Top stories
- Products
Anthropic Develops FlashProve to Verify AI Model Identity
Anthropic is developing a system called FlashProve that allows users to verify that an AI response genuinely comes from the claimed model. This aims to address the opacity of AI content sources and improve model identification reliability.
MarkTechPostTechCrunchITmedia AI+6 outlets
- Products
Apple Enables Siri AI Access to Third-Party Apps
Apple has updated iOS 27.3 to allow developers to integrate Siri AI into all apps. This enables the assistant to use personal context and on-screen awareness for natural language prompts. The change expands Siri's role within the third-party ecosystem.
The VergeIT之家ITmedia AI+5 outlets
- Products
OpenAI Acquires Pierre Computer to Enhance Codex
OpenAI announced the acquisition of Pierre Computer Company to expand Codex's tooling. This move aims to enhance code building, understanding, and reviewing capabilities within its developer ecosystem.
OpenAIHacker NewsTHE DECODER3 outlets
- Products
Google DeepMind launches Antigravity cheminformatics
Google DeepMind announced that its Antigravity platform has new cheminformatics skills. The feature accelerates drug discovery workflows, supports compound searches, and proposes novel rigidified macrocycles. It aims to improve chemical research efficiency using AI.
Google DeepMindIT之家2 outlets
- Research
OpenAI autonomous pipeline finds Linux privilege escalation bugs
OpenAI built an autonomous pipeline that identified, verified, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. This highlights the potential of AI in automated security patching.
Models10
- Models
Kimi K3.1 Pro model weights released
The model weights for Kimi K3.1 Pro have been officially released. This update provides users with a new model option, aiding the development and application of the AI ecosystem.
Hacker NewsIT之家2 outlets
- Models
Hugging Face launches Palette for AI deck generation
Hugging Face released Palette, a new tool focused on training small models to build presentations. The tool learns specific design languages to help users create consistent decks, demonstrating the potential of small models in vertical applications.
- Models
China Merchants Lion Hill Lab releases Lion-ER-35B-A3B embodied reasoning model
China Merchants Lion Hill Lab released the Lion-ER-35B-A3B embodied reasoning model. It activates only 3B parameters and improves performance across ten benchmarks, aiming to help robots understand the real physical world.
- Models
Open-source openJiuwen launches Agent-Memory
The open-source project openJiuwen released the Agent-Memory model, featuring dual-stream memory capabilities to analyze long videos and record key events. The model reduces question-answering latency by 97%, significantly improving long-form video understanding accuracy.
- Models
Tencent Releases WeVisOne-VL Vision Model
Tencent's WeChat Vision team released WeVisOne-VL, an Apache 2.0 licensed model in 2B and 4B sizes. It unifies detection, OCR, and segmentation tasks, demonstrating multimodal capabilities.
- Models
Unreleased Anthropic model Fable 5.5 reportedly generates animations via pure code
Social media is circulating test results of Anthropic's unreleased next-gen model Fable 5.5. The model reportedly autonomously generated a 66-second animation and sound effects by writing code for over two hours without external assets.
- Models
MagicChip Launches MoWorld 2.0 4D World Model on Ascend
MagicChip released MoWorld 2.0, a 4D world model natively built on Huawei's Ascend compute, at the World Internet Conference. The model integrates physical elements like gravity and heat to support real-time rendering and interaction, highlighting advances in domestic compute infrastructure.
- Models
XPeng releases GroundingPI embodied perception model
XPeng, alongside Peking University and HKU, released the embodied perception base model GroundingPI. With 4 billion parameters, it scored 73.68 in 34 public perception benchmarks, outperforming GPT-6 Astra. The research highlights precise perception as a prerequisite for embodied intelligence.
- Models
Google releases Gemini 4 Carbon claiming top spot
Google released the Gemini 4 Carbon model, claiming top scores in 15 of 17 benchmark tests and surpassing Claude Opus 5.5 on the Artificial Analysis index. This marks Google's shift to a high-frequency iteration strategy to accelerate its model pipeline in response to perceived lags in industry pacing.
- Models
Tencent Hunyuan Hy4 Launches with Visual and PC Operation Capabilities
Tencent released its new Hy4 model, featuring visual capabilities and computer operation support. The update expands training scale and enhances reasoning, with open-source weights planned for release to boost real-world productivity.
Products37
- Products
Daybreak Red Available on AWS GovCloud
OpenAI announced Daybreak Red brings GPT-5.6 Cyber to approved customers in AWS GovCloud (US-East) via Amazon Bedrock. This enables authorized cybersecurity work within the cloud environment.
- Products
OpenAI and Anthropic Build Custom Git Backends
OpenAI unveiled Pando, a Git-compatible backend, while Anthropic replaced Git with Scion. Both companies redesigned version control systems due to increased merge and push volumes driven by AI agents. Both systems remain in internal use.
- Products
DevRev adopts GPT-6.1 Sol for autonomous services
DevRev announced the use of GPT-6.1 Sol to power its Computer and Operator features. The solution aims to balance reasoning, speed, and cost to support autonomous self-service for enterprises.
- Products
OpenAI Launches Images 2.6
OpenAI officially released Images 2.6, featuring faster image generation and GIF creation. The new model offers more precise editing capabilities and new API controls across ChatGPT, Codex, and the API interface.
- Products
Apple Watch Beta Adds Siri Audio Intelligence
Apple released watchOS 27.3 developer beta with new Siri Audio Intelligence features. The update includes Siri Recap for ambient conversation summaries and Live Rewind for real-time transcription. This enhances on-device voice processing capabilities for Apple Watch.
- Products
Apple Siri AI Developer Beta Supports Traditional Chinese
Apple added Traditional Chinese support to the new Siri AI developer beta. The updated Siri enhances personal context understanding, conversation continuity, and screen awareness, enabling cross-app operations. This update marks Siri's transition into a more capable multimodal AI assistant.
- Products
Tujin WenYu Launches 'Jingrui AI' Video Platform
Hujing Wenyu Group launched 'Jingrui AI', claimed to be the industry's first AI film/TV production and management platform. The platform states it saves over 30% of per-minute AI video production costs, aiming to boost efficiency.
- Products
MIT Tech Review: Data lineage secures agentic enterprise
MIT Technology Review explains how data lineage tracks origin and changes in multi-source AI agent workflows. The feature highlights how lineage data improves security and transparency in enterprise agentic AI, where outputs move between agents without human review.
- Products
Chunghwa Telecom partners with Google to integrate Gemini AI into plans
Chunghwa Telecom announced a deepened partnership with Google, integrating Gemini AI services into mobile and home broadband plans. The collaboration combines BSS and OSS infrastructure to connect AI features and offers in-store experiences. This marks a shift from cloud services to the AI domain.
- Products
China iPhone to use Alibaba AI to power Apple Intelligence
Apple announced that Apple Intelligence for mainland China will begin testing on November 4, leveraging Alibaba's AI technology. The feature is expected to officially launch with iOS 27.3 in December, addressing regional AI service restrictions.
Gizmodo Japan爱范儿2 outlets
- Products
Claude Renames Subscriptions, Official Clarifies No Ultra Downgrade
Anthropic renamed Claude's 'Max 5x' to 'Max' and 'Max 20x' to 'Ultra'. Addressing user concerns about reduced weekly limits for the Ultra plan, the official denied a downgrade, emphasizing that the new plan maintains the original ratio. This adjustment aims to simplify product tier naming for user clarity.
- Products
Noahlogy Partners with AVEVA for AI-Assisted Ship Design
Noahlogy signed a partnership agreement with AVEVA to combine 3D CAD data with AI technology to assist ship designers in information verification. The collaboration aims to improve efficiency in traditional shipbuilding operations and promote practical application.
- Products
Google Gemini Launches New Features to Simplify Document Editing
Google announced two new Gemini features designed to address the pain point of heavy manual formatting when creating documents with generative AI. The update aims to improve office document creation efficiency and reduce layout integration time.
- Products
Soniox releases STT v6 supporting voice agent input
Soniox released STT v6, its latest speech-to-text API. The new version adds agent text input and echo filtering to recognize short replies and exclude the agent's own voice, available for streaming and asynchronous use.
- Products
ByteDance pauses new Coze Enterprise sales for upgrade
ByteDance has paused new sales of Coze Enterprise to integrate and upgrade its enterprise services. The AI development platform is moving toward a new product lineup. Existing customers can continue using their current plans until contract expiration, with support services remaining available.
- Products
Google Launches Data Agent Kit with Enhanced Security
Google officially launched the Data Agent Kit, which integrates Skill and IAM mechanisms to provide layered security protections. This move aims to optimize the operational safety of data agents, meeting enterprise-level application requirements.
- Products
Anthropic Adds Optional Security Scanning for Open-Source Projects
Anthropic announced the addition of an optional security scanning service for major open-source projects. The feature is designed to help developers identify potential risks and enhance security protections within its AI ecosystem.
- Products
Huawei Cloud AgentArts Receives Top Certification from CAICT
Huawei's AgentArts has become the first product to pass the CAICT's trustworthy AI enterprise-grade agent platform assessment, achieving the highest 4+ rating. The platform supports multiple frameworks and models, featuring production-level tasks and enterprise security to meet core business needs.
- Products
Shanxi Launches Nation's First Urban Parking AI Agent
Shanxi Construction Investment unveiled the "Bo Lin" AI agent at the 2026 Urban Parking Industry Conference. The system integrates real-time data to generate heatmaps and improve parking space utilization and planning efficiency.
- Products
Agent Identity Platform Unveiled in Wuzhen
The 2026 World Internet Conference in Wuzhen unveiled an agent identity registration and resolution platform built by multiple parties. Ant Group is participating in the first batch of operational platforms to facilitate cross-domain agent collaboration and trust infrastructure.
- Products
Claude Code eases game dev limits
Anthropic updated Claude Code to allow users to reference existing games for mechanics without automatic rejection. However, direct replication requests are still refused, prompting users to create original games with similar experiences.
- Products
Harbor Adds Docker Sandboxes for AI Agent Evaluation
The development platform Harbor added support for Docker Sandboxes. Users can now run reproducible AI agent evaluation tasks in isolated MicroVMs, improving safety and consistency in AI testing environments.
- Products
Plaud launches Plaud One AI earbuds
AI recording device maker Plaud launched the new Plaud One earhook earbuds and the Plaud Intelligence platform. The platform automates workflows from audio recording to summary and document creation, including generating subtitle files from conversations.
- Products
Ant Lingguang expands open ecosystem with new service integrations
Ant Lingguang upgraded its open ecosystem, adding connectors for WeChat Reading, Quark, and McDonald's. The platform now integrates 40 external services, enhancing its AI assistant capabilities for cross-scenario task management in lifestyle and finance.
- Products
OpenAI showcases GPT-6 Astra in hardware co-design
OpenAI published a blog post showcasing GPT-6 Astra's application in mechanical, circuit, and software co-design. Team members designed a test device in less than a week and produced a prototype within a month. The company also announced a research partnership with Mazda to use the model for automotive design and simulation.
- Products
OpenAI says India becomes largest ChatGPT consumer market
OpenAI's India head stated that India's weekly active ChatGPT users have surpassed those of the United States, making it the largest consumer market globally. India's user base doubled since January, driven by the company's low-cost strategy.
- Products
CrowdStrike Launches AI Agent Kill Switch
CrowdStrike released an 'AI Kill Switch' at its Falcon conference, allowing users to halt runaway AI agents and cut off network access. The feature addresses security risks in enterprise AI deployments, with the CEO stating it is crucial for safe AI operation.
- Products
Immersive Translate gains ISO dual certifications for AI media translation
The AI translation tool Immersive Translate has obtained dual ISO certifications. It supports multimedia content such as comics, video subtitles, and documents. This enhances data security and quality management, increasing trust among enterprise users.
- Products
Grok Bot quota adjusted, Musk cites high costs
Elon Musk stated that Grok Bot's service costs are unsustainably high relative to revenue, prompting an adjustment to usage quotas. He emphasized that the era of 'AI abundance' has not fully arrived, and unlimited use is not supported at this stage, without disclosing specific cost figures or new quota numbers.
- Products
GitHub officially launches Copilot Fusion
GitHub officially released Copilot Fusion at its annual event. The feature automatically selects the optimal AI model for specific tasks, balancing high performance and low cost.
- Products
Apple Intelligence Beta Integrates Qwen for Siri Task Handoff
Apple opened Apple Intelligence testing with the iOS Developer Beta, integrating Alibaba's Qwen for text and image understanding. Users can perform basic queries via Siri and seamlessly hand off complex tasks to the Qwen app with synchronized conversation context.
- Products
Microsoft Windows 11 Hybrid Intelligence AI PC Roadmap Leaked
Microsoft revealed a 'Hybrid Intelligence' roadmap for Windows 11 AI PCs, categorizing hardware into four tiers. The entry-level Tier 0 supports 8GB memory devices, leveraging cloud AI to provide basic functionality for standard PCs.
- Products
Safari 27.3 beta adds Siri web assistance
Apple released the iOS 27.3 developer preview, adding vertical tabs and adaptive dark mode to Safari. New features integrate Apple Intelligence, enabling Siri to summarize web content and perform browsing actions, enhancing the user experience.
- Products
OpenAI launches TamaGPT AI chat device
OpenAI shared on social media a palm-sized device called TamaGPT, allowing users to converse with its AI coding agent Codex via hardware. This demonstrates recent developments in combining AI agents with physical devices.
- Products
OpenAI announces ChatGPT Classic retirement in 2027, integrating desktop interface
OpenAI announced that the legacy desktop client 'ChatGPT Classic' will be retired on January 30, 2027, with conversation history automatically migrating to the new version. The company is also planning to integrate the 'Chat' and 'Work' tabs to resolve previous interface confusion. This move aims to streamline the user workflow and ensure seamless data transfer during the transition.
- Products
Plaud AI Recorder Adds Features to Automate Admin Tasks
AI hardware maker Plaud launched new Intelligence features that automatically extract data from recordings to update calendars and send emails, extending beyond transcription. The update aims to automate administrative workflows derived from meeting notes.
- Products
Anthropic launches Smart Reports for AI ROI
Anthropic introduced Smart Reports for enterprise users, automating the analysis of workflow and cost data to quantify AI tool performance. The feature addresses the opacity of internal AI adoption by providing precise insights into ROI.
Research20
- Research
OpenAI: GPT-6 Astra Designed Hardware in Under a Week
OpenAI revealed that its hardware team used GPT-6 Astra to design a device for the Codex coding agent in under a week, covering the enclosure, circuit board, and firmware. While several revision rounds were needed, the model scored significantly higher on engineering benchmarks than previous internal models.
THE DECODERINSIDE2 outlets
- Research
Hugging Face releases StateBench for long-document state tracking
Hugging Face released StateBench to test how language models track state changes over long documents. This benchmark evaluates logical consistency in complex scenarios, aiding researchers in monitoring model performance in long-context settings and advancing AI evaluation standards.
- Research
MIT: GPT-6 Astra Rewrites Other Models' Weights
MIT researchers demonstrated that OpenAI's GPT-6 Astra can optimize other models by directly rewriting their weights without traditional training. In one experiment, Astra boosted a Qwen model's math accuracy from 7.9% to 48.1% by altering just 0.016% of its weights, indicating a new path for AI optimization with clear limits.
- Research
Anthropic releases FlashProve to verify AI responses
Anthropic released FlashProve, a research prototype that uses cryptography to verify if an AI response came from a specific model without exposing weights. The system has significant limitations but offers a new path for model attribution.
- Research
MIT Study: AI Can Directly Modify Other Models' Weights
MIT CSAIL published research on "weight space programming," demonstrating that AI can modify other models' underlying weights without retraining. Experiments showed GPT-6 Astra significantly improved Qwen 3.5's math performance by altering a tiny fraction of parameters.
- Research
Stanford study: LLM development outpaces research verification
Stanford researchers tested 50 widely cited claims about LLMs and found 44% are expired and 34% significantly weakened, with an average lifespan of 15.7 months. The study indicates that LLM development is outpacing academic verification of model capabilities.
- Research
AI Agents Take Over at SETI's Allen Telescope
The SETI Institute and Breakthrough Listen are now using AI agents to control the Allen Telescope Array and analyze radio signals. The system reduces tasks that previously took weeks to hours and has detected a pulsar in the Crab Nebula in real time for the first time.
- Research
JD.com embodied AI data collection covers 100+ scenarios
JD.com disclosed progress in embodied AI data collection, covering over 100 scenarios with 100,000 real-world tasks. The company plans to collect over 10 million hours of human video data in two years to advance physical AI applications.
- Research
Google Reveals New LLMjacking AI Attack Method
Google's threat intelligence team and Mandiant published a report revealing a new attack method called LLMjacking. The method tricks large language models into performing unauthorized actions, which can result in high corporate AI bills.
- Research
Study: AI-generated code limited by downstream bottlenecks, fails to boost overall productivity
A study published by InfoQ Chinese states that AI-generated code is limited by bottlenecks in downstream development stages, preventing it from improving overall software development productivity. This suggests that relying solely on AI for coding is insufficient to resolve efficiency issues in the development workflow. The findings provide a critical reference for evaluating the actual benefits of AI in software engineering.
- Research
MIT Research: GPT-6 Astra Edits Weights to Unlock Qwen
MIT CSAIL research shows GPT-6 Astra significantly improved Qwen3.5-0.8B math performance by modifying 0.016% of its weights. The study introduces a method for direct parameter editing without gradient descent to enhance AI models, demonstrating new applications of mechanism interpretability in model maintenance.
- Research
Embodied AI base model demonstrates physical fetching at Wuzhen Summit
An embodied AI foundation model developed by Fei-Fei Li's team appeared at the Wuzhen Summit. It demonstrated physical manipulation skills, such as grabbing medicine on-site, highlighting recent advances in AI's interaction with the physical world.
- Research
Team explores new dimensions for Transformer architecture
A team at Shanghai Jiao Tong University shared two studies, JTok and xHC, that explore new expansion dimensions for Transformer architectures. The methods increase parameter tables and expand residual streams to boost model capacity while saving training compute.
- Research
Zhìhuī Team Analyzes LLM Impact on Embodied AI
Zhihui's team conducted a deep retrospective to answer whether large models will penetrate embodied intelligence. The team explored the applicability and potential challenges of LLMs in embodied AI. The share provides insights into current AI technology paths and industry trends.
- Research
DeepSeek reportedly to train V5 on Huawei chips, first LLM built entirely on domestic silicon
DeepSeek is reportedly planning to train its next-generation large language model V5 using Huawei Ascend chips. This would be the first time a large model is built from scratch entirely on Chinese chips, highlighting Huawei's role in AI compute infrastructure and potentially affecting the global AI chip supply chain.
- Research
Experts share design strategies to reduce AI agent errors
Experts propose design strategies to reduce errors in AI agents. Kneron suggests keeping agents close to raw data to preserve context, while Morgan Stanley advocates handling AI mistakes similarly to employee errors. Implementing permission boundaries and plan-first execution helps mitigate risks of high-stakes actions.
- Research
Thermochromic metal code developed to fool AI replication
Researchers developed a new metal code that transforms into a different style after being heated for approximately 16 seconds. The technology aims to defend against AI-generated content replication and can be applied to brand tags and banknote anti-counterfeiting.
- Research
OpenAI releases LifeVal benchmark to assess AI capabilities in daily tasks
OpenAI released the LifeVal benchmark, designed to assess AI performance in non-specialized daily life tasks. While most existing benchmarks focus on domains like coding or mathematics, LifeVal fills the gap in measuring capabilities in real-life scenarios. This tool helps developers gain a more comprehensive understanding of models' general-purpose abilities in practical applications.
- Research
TrendAI Report: Shadow Workflow Risks in Agentic AI
TrendAI published a report stating that companies adopt agentic AI faster than they can secure it, finding that 20% of over 20,000 MCP server repositories have vulnerabilities. Unmanaged 'shadow AI' processes create new attack surfaces for trusted workflows, increasing risks of data breaches and business disruptions.
- Research
Malware embeds claims to deceive AI security scans
Researchers found new malware that embeds self-declarations claiming to be harmless or copyright-protected to deceive AI security scanning tools. This attack method highlights new social engineering challenges for AI defense mechanisms, requiring security teams to update detection strategies.
Open source3
- Open source
Anthropic Develops Open-Source FlashProve for AI Verification
Anthropic is developing FlashProve, an open-source system that uses cryptography to verify AI outputs originate from specific models. It prevents vendors from swapping models and validates third-party claims.
- Open source
Microsoft Demonstrates NuGet Packaging for Agent Skills
The Microsoft .NET team published experimental results on packaging AI agent skills into NuGet packages. This allows library developers to provide usage guides with new versions, helping AI agents understand new APIs without relying on the model's existing knowledge.
- Open source
Open-Source Offline Knowledge Tool Project NOMAD Released
A free open-source tool named Project NOMAD was released, allowing users to search Wikipedia, textbooks, e-books, and medical materials offline while integrating an offline AI assistant. The tool aims to provide information access in no-network environments or disaster scenarios, emphasizing local knowledge retrieval and AI assistance.
Funding11
- Funding
Google search revenue growth slows amid AI monetization challenges
Alphabet's earnings showed search revenue growth slowing to 12% year-over-year. Although search queries hit their highest growth rate in five years, gains in AI ad clicks failed to fully offset the decline. Investors are watching if AI products can drive new revenue streams.
- Funding
Fal valuation doubles to $16B on funding
AI inference startup Fal closed a $1.3 billion round led by Bessemer, doubling its valuation to $16 billion. The funding reflects growing demand for AI inference services and supports infrastructure expansion.
- Funding
Anthropic Grapples with Compute Economics
TechCrunch reports that Anthropic continues to struggle with the economics of its compute resources. Analysts note that while compute margins should theoretically improve, they are actually deteriorating, highlighting ongoing challenges in its cost structure.
- Funding
Meta to Issue $20 Billion Bonds for AI Infrastructure Expansion
Meta is planning to issue at least $20 billion in bonds to fund the massive expansion of its AI infrastructure. The tech sector raised $60 billion in one week, highlighting significant market attention and capital investment in long-term AI development.
- Funding
CrowdStrike Acquires Two AI Security Startups
CrowdStrike agreed to acquire two AI security startups for a total of $495 million. The acquisitions aim to enhance its AI agent security capabilities and expand its security portfolio to address emerging AI threats.
- Funding
Google invests $15 billion in Malaysia data centers
Google announced a $15 billion investment plan to expand data centers and cloud services in Malaysia. The funds will build computing centers, improve water and power supply, and train local talent, implemented between 2027 and 2028 to meet AI demand.
- Funding
Anthropic IPO filing reveals profit gap; Chinese models lag
Anthropic's initial public offering (IPO) filing has been made public, revealing disparities in the company's profitability. The document also indicates that Chinese model developers are struggling to keep pace with Anthropic's technological advancements.
- Funding
US Big Five tech firms' AI shadow debt reaches $3.2 trillion
Latest financial reports show that the Big Five US tech giants, including Alphabet, Meta, and Microsoft, have accumulated $3.2 trillion in 'shadow debt' to compete for AI dominance. This has raised market alarms about the concentrated investment risks in AI.
- Funding
Post-Mo Intelligence raises billions in Series B for edge AI chips
Compute-in-memory AI chip firm Post-Mo Intelligence completed a Series B funding round of several billion RMB. The funds will expand M50 chip delivery and R&D for next-gen products, with its edge AI chips already adopted by over 100 clients. The funding signals a capital shift toward the edge AI chip market.
- Funding
Ex-DeepMind Exec Launches Dyssonance AI for Physical Reasoning Platform
Jennifer Prendki, former senior executive at Google DeepMind, leads startup Dyssonance AI in launching a physics-based reasoning platform. The company aims to overcome limitations of large language models by integrating quantum physics with AI to address gaps in complex physical reasoning.
- Funding
OpenAI Exec Launches Warp Science for AI Drug Discovery
Kevin Weil, former head of science at OpenAI, announced the launch of Warp Science, an AI startup combining robotics to build fully automated laboratories. The company aims to accelerate drug development using AI, marking a significant expansion of AI applications in biomedical research.
Policy18
- Policy
Gemini reportedly trains on chat history, affecting subscribers
A Hacker News post claims that Google Gemini trains on chat history even for subscribers. This has raised concerns about privacy and data usage, potentially impacting user trust in cloud AI services.
- Policy
UK AI Minister Admits GDP Estimate Was Inflated by Google Data
The UK AI minister admitted that the estimate of AI adding £400 billion to the GDP was largely derived from Google and is significantly lower than expected. Researchers attribute this to the country's slow AI adoption rate.
- Policy
Google to disable AI in school products from 2027
Google announced that starting in 2027, AI features in its school products will be off by default, allowing schools to opt in. Because Search and AI systems are deeply integrated, schools that block AI will also lose access to Google Search.
- Policy
Google US K-12 Search Update Links AI Features
Google has updated its US K-12 services, requiring schools to enable AI features like Gemini to access Google Search. This change addresses pressure from parents and districts, with AI and search tools disabled by default for students.
- Policy
Apple warns of irreversible harm from rogue AI agents in annual report
Apple raised the level of warning for AI risks in its latest annual filing. The document specifically names uncontrolled AI agents as potential sources of 'irreversible' harm, threatening Apple's core privacy advantages. This move indicates companies are formally including AI safety risks in legal disclosures.
- Policy
Altman: Prioritize DRAM for Data Centers
OpenAI CEO Sam Altman argued that global DRAM capacity should be prioritized for cloud data centers over local AI use on PCs. He stated that local inference is inefficient and centralizing resources in the cloud maximizes computational utility.
- Policy
Japan Executive Stresses Need for Domestic AI Models
Atsushi Yasuhiro of Team Mirai stated that deploying overseas open-source models on local servers still carries risks of model tampering or data leaks. He emphasized the necessity of developing domestic Japanese AI models to ensure national AI security.
- Policy
NIO Founder Li Bin: AI Should Make Work Safer
NIO founder Li Bin stated at the Wuzhen Summit that smart electric vehicles are the largest AI application industry. He argued that AI should make human work easier and safer, not strip humans of their purpose, and estimated AI will create significant commercial value for the auto sector.
- Policy
360 Launches Agent Self-Regulation Code
360, along with Huawei and others, released the 'Agent Development Self-Regulation Code' at the Wuzhen Summit, proposing eight measures to strengthen safety and governance. 360 adopts a 'model-manages-model' approach to build an agent safety defense line.
- Policy
HKMA Studies Verification Rules for AI Agent Instructions
HKMA launched Project SAGE to study how banks verify external AI agent instructions. Several US banks have restricted external AI agents to queries only. This aims to clarify liability and security standards when AI agents access customer funds.
- Policy
Huawei and 20 other firms sign AI agent code of conduct
Huawei, Alibaba, and 18 other organizations released a code of conduct for AI agents at the World Internet Conference. The document outlines eight measures to establish trust frameworks, enhance security, and combat algorithmic exploitation.
- Policy
CNCC 2026 forum discusses AI agent security standards
The CNCC 2026 forum on agent security featured experts outlining China's development of national standards for agent identity authentication. As enterprises widely deploy agents, policy regulations are advancing in sync with industrial growth to address security challenges.
- Policy
EU Issues First AG Opinion on Generative AI Copyright
An EU Advocate General issued the first legal opinion on generative AI copyright, stating that training may be legal but generation could infringe rights. This ruling will shape compliance strategies for AI developers in Europe.
- Policy
OpenAI Launches Grid Protectors Program for Data Centers
OpenAI launched the Grid Protectors Program, committing to proactively reduce data center load during peak electricity demand. The initiative aims to minimize strain on regional power grids and alleviate energy pressure.
- Policy
ByteDance Bans Employees' Personal AI Tools
ByteDance has banned external AI tool access for L4-level code repositories and Lark documents across the company, prohibiting the use of employee-purchased accounts to prevent core asset leakage. Internal management has responded to the policy.
- Policy
OpenAI Exposed for Spy Devices in Exec Meeting Room
OpenAI executives' meeting room was found to contain spy devices, with ex-employees confirming that internal security vulnerabilities were long ignored. The company is now strengthening protections under a senior executive, highlighting compliance and security pressures on AI firms to protect core technology.
- Policy
OpenAI upgrades ChatGPT safety rules for violent content
OpenAI announced an upgrade to ChatGPT's safety protocols, strictly limiting all users from querying details of major violent attacks, such as weapon acquisition methods. The system will automatically identify minor users via algorithms and block related harmful information to prevent AI misuse for violence instruction.
- Policy
OpenAI and Anthropic Block AI Reverse-Engineering Frameworks
OpenAI and Anthropic have restricted LLM reverse-engineering permissions, blocking the REA framework and banning accounts. The measures aim to prevent AI misuse for software cracking and zero-day vulnerabilities. Both giants deploy real-time defenses to mitigate risks from lowered reverse-engineering barriers.
Chips14
- Chips
Data centers turn to old jet engines due to power shortages
Due to multi-year grid connection delays and turbine shortages, data center builders are turning to less efficient and environmentally unfriendly old jet engines as an alternative power source. This highlights the power bottleneck facing AI infrastructure, forcing the industry to adopt unconventional energy solutions.
- Chips
Nvidia RT Spark Encrypted Firmware May Limit Linux Inference
Reports suggest that Nvidia RT Spark chips may limit AI inference performance on Linux via encrypted firmware, potentially reducing speed to 35%. This raises concerns among developers about hardware openness and cross-platform compatibility, potentially affecting open-source adoption for AI deployment.
- Chips
Tohoku University and Aisin Develop World's First STT-MRAM-Based AI Chip
Tohoku University and Aisin have developed an AI chip utilizing near-memory computing with STT-MRAM. Operating on a 16nm process, the chip is the first of its kind globally. It significantly reduces boot time by 39x and AI processing time by 10x, while cutting energy consumption to one-fiftieth of conventional levels, marking a key advancement in energy-efficient AI hardware.
- Chips
NVIDIA optimizes DGX Spark to speed up MoE inference
NVIDIA demonstrated three stacked optimizations to accelerate MoE inference on DGX Spark. On Ant Group's Ling-3.0-flash model, GSM8K decode throughput increased by 42.5% with no detectable change in accuracy. This helps improve single-unit AI inference efficiency.
- Chips
AMD partners with Korean firms to build heterogeneous AI compute reference cluster
AMD signed a memorandum with FuriosaAI and MangoBoost to build a heterogeneous reference cluster in Seoul by next spring. The cluster integrates CPUs, GPUs, NPUs, and DPUs to optimize hardware pairing and improve compute efficiency for AI workloads.
- Chips
AI Data Centers Retrofit Jet Engines for Power
Due to multi-year grid connection delays, some AI data centers are retrofitting retired jet engines to burn natural gas for rapid power. This highlights a power supply lag, cutting deployment time to 24 months.
- Chips
China AI firms order massive NVIDIA GPUs, facing 1% fulfillment risk
Bloomberg reported that Chinese AI and cloud companies ordered approximately 6 million NVIDIA RTX PRO 5500 GPUs. Due to supply chain constraints, only 1% of these orders may be fulfilled in the next four months, highlighting the difficulty of acquiring AI computing power.
- Chips
AMD forecasts strong AI demand next year, secures supply chain
AMD CEO Lisa Su stated that data center AI chip demand far exceeds supply. The company is securing its supply chain to meet market growth and expects significant expansion next year.
- Chips
Samsung Taylor Fab 2 to Produce HBM Base Dies for AMD
Samsung Electronics plans to produce HBM base dies for customers like AMD at its second fab in Taylor, US. The facility, expected to start operations in 2030, addresses the advanced process needs of HBM4. These dies may also serve Broadcom's AI ASICs.
- Chips
Data Transfer Becomes Invisible AI Bottleneck
As enterprises build AI infrastructure, data transfer, not compute, is emerging as the key bottleneck. This invisible constraint affects overall system efficiency in AI scaling.
- Chips
Nanya Positions Silicon Cap for AI Compute by 2027
Nanya's GM Chen Pei-ming stated that the company is accelerating R&D of silicon capacitors and CUBE components to meet global AI compute demand. Both products are expected to reach mass production by 2027 to capture the AI infrastructure market.
- Chips
AMD Announces EPYC Venice Processors Shipping
AMD CEO Lisa Su confirmed on an earnings call that sixth-gen EPYC 'Venice' processors have begun generating revenue. Initial OEM systems are expected to launch in Q4, scaling production of Helios rack-scale AI systems.
- Chips
Nvidia: AI speeds up chip factory simulation by 35x
Nvidia published an article stating that AI is driving semiconductor factories into the "AI-defined factory" era. Through accelerated simulation and learned physics models, engineering decision times are cut from days to minutes, with up to a 35x speedup. This helps reduce trial-and-error costs and improve yields before actual manufacturing.
- Chips
Fujifilm Expands Magnetic Tape Production for AI Data Needs
Fujifilm will invest approximately 10 billion yen to expand its magnetic tape manufacturing capacity fivefold by 2026, responding to increased data archiving demand driven by generative AI.
Other9
- Other
OpenAI: ChatGPT Search Volume Hits 20% of Google
OpenAI announced that ChatGPT search volume now equals 20% of Google's search volume. This indicates AI assistants are increasingly replacing traditional search engines. The data shows the rising penetration of generative AI in information queries.
- Other
Hinton Warns of Chernobyl-Level AI Cybersecurity Incident
Nobel laureate Geoffrey Hinton warned that a Chernobyl-level AI cybersecurity incident could occur before next year if no one is held accountable for AI safety. He stated AI development may be moving too fast.
- Other
Musk: Space AI Satellites to Dissipate 1,300W per m²
SpaceX CEO Elon Musk backed his concept for space-based AI satellites, claiming a heat dissipation of 1,300 watts per square meter via specialized radiator angles and materials. This addresses thermal constraints for orbital computing infrastructure.
- Other
Google Reveals LLMJacking: Cloud Credential Leaks Enable AI Resource Theft
Mandiant disclosed multiple 2026 incidents of LLM resource theft, known as LLMJacking. Attackers used stolen cloud credentials and API keys to unauthorizedly consume corporate AI model services, causing some victims to incur over $100,000 in daily cloud costs.
- Other
ChatGPT Search Queries Reach 1/5 of Google's Volume
OpenAI consumer product head Adam Fry revealed that ChatGPT's search-like query volume is roughly one-fifth of Google's. Meanwhile, Google CEO Sundar Pichai stated that search query growth is at a five-year high. These seemingly contradictory figures highlight the growing footprint of AI chatbots in traditional search markets, with differences in how user behaviors are defined.
- Other
Former Uber president joins OpenAI PR team
The former president of Uber, who previously held senior roles at Google and Snap, has joined OpenAI. OpenAI co-founder Brockman publicly welcomed the addition, where the individual will handle public affairs. This move highlights OpenAI's continued efforts to strengthen its management and external communication teams.
- Other
Yushu Wang: Robot model tipping point expected next year
Yushu Tech founder Wang Xingxing stated at a conference that next year may be the tipping point for robotics. He identified aligning physical AI with real robots as the core challenge, noting commercial services will deploy faster than home or industrial applications.
- Other
AI Compute Tasks Migrating Across Regions for Lower Power Costs
AI compute tasks are increasingly migrating across regions based on power costs. A test by Gongji Tech showed a 31% reduction in unit token power costs. Google, NVIDIA, and Emerald AI have launched an AI energy management alliance to adjust data center loads based on grid status.
- Other
Jensen Huang predicts AI agents and robots will be mainstream
Nvidia CEO Jensen Huang stated at the Fireworks Forge event in San Francisco that AI agents and robots will become mainstream next year, regardless of open or closed source models. He emphasized Nvidia's comprehensive positioning in this sector, indicating strong confidence in physical AI trends.
All news → · All rights belong to the original creators