AI news, Nov 3, 2026: Anthropic's Capital Expenditure and Safety Concerns Draw Attention
Anthropic disclosed $90 billion in AI capital expenditure commitments ahead of its IPO while warning that AI is approaching the threshold for recursive self-improvement. The company's revenue growth rate slowed significantly to a quarter-over-quarter increase from 144%, prompting warnings about financing risks for compute suppliers. Internal AI usage costs have now surpassed total employee salaries, highlighting the financial strain of high capital expenditure. These developments underscore the significant economic and safety challenges facing the AI industry as it scales its infrastructure.
Lead written by AI from the day’s stories
Top stories
- Products
Google schools disable AI tools without search
Google announced that schools opting to completely disable its AI tools will lose access to Google Search. This is because AI features are now deeply integrated into the search engine, impacting digital resource access for educational institutions.
Hacker NewsTechCrunchThe Verge7 outlets
- Funding
Anthropic Discloses $900B AI Spending Commitments
Anthropic disclosed $900 billion in AI capital spending commitments ahead of its IPO. This massive expenditure is primarily for expanding infrastructure to support model training and deployment, highlighting the significant capital scale of the AI industry.
Hacker NewsIT之家ITmedia AI+4 outlets
- Research
Astra model can read and write Qwen weights
According to Hacker News, Astra can read and write weights to improve Qwen model performance. This capability is useful for researchers who need to fine-tune or manage large language model weights, indicating a new method for operating between models.
Hacker News量子位IT之家3 outlets
- Research
Anthropic Warns of Impending RSI Capability
Anthropic stated that AI is approaching Recursive Self-Improvement (RSI), where models can efficiently train themselves. This development could trigger an exponential surge in capabilities, posing a significant challenge to the safety community and raising concerns about irreversible risks.
TechCrunch量子位科技報橘3 outlets
- Research
MIT Study: GPT-6 Astra Modifies Qwen Weights to Boost Math Skills
MIT researchers published an experiment using OpenAI's GPT-6 Astra to directly modify approximately 0.016% of the neural network weights in Alibaba's Qwen3.5-0.8B model. This method, termed 'Weight-Space Programming', increased the model's accuracy on the MATH-500 math test from 7.9% to 48.1%, attracting attention from the tech community.
iThome科技報橘Simon Willison3 outlets
Models6
- Models
Baidu Unveils Apollo World Model for Physical AI
At its AI Day event in Shenzhen, Baidu unveiled Apollo World, a full-modality world model designed for physical AI and autonomous driving. The model generates realistic videos that adhere to physical laws and supports end-to-end driving simulation. Baidu also introduced a '2+3' evaluation framework to assess the model's causality, interactivity, and physical accuracy.
- Models
Xiaomi Robotics-1 Model Begins Trial Operation in Apartments
Xiaomi announced that its Xiaomi-Robotics-1 embodied foundation model has started trial operations in youth apartments. The system can autonomously perform long-horizon tasks such as tidying desks and cleaning for over 30 minutes, demonstrating its practical capabilities in complex environments.
- Models
Huawei Cloud Halts openPangu 2.0 Model Sales
Huawei Cloud announced it will stop selling openPangu-2.0-Flash and Pro models on November 6, 2026, with a full shutdown by November 30. The company advised users to migrate to DeepSeek or GLM alternatives and stated that new model versions are in planning, though specific timelines were not disclosed.
- Models
SenseTime releases SenseNova 6.8 Flash model
SenseTime released the official SenseNova 6.8 Flash model, featuring enhanced office collaboration and dynamic design generation. The model supports long-term memory and dynamic PPT creation, with performance surpassing some international models in certain benchmarks.
- Models
GPT 6.1 Sol vs. Claude Sonnet 5.5: 3.5x cost gap and reversed speed in real tasks
Leiphone compared Anthropic's Claude Sonnet 5.5 (Sep 28) with OpenAI's GPT 6.1 Sol (Sep 29), which carry identical API pricing on OpenRouter. Testing three tasks found GPT 6.1 Sol cost $38.50 versus $10.90 for Sonnet 5.5, yet Sonnet was faster for policy research while Sol was quicker on 3D development.
- Models
OpenAI GPT-6.2 Astra reportedly in limited testing
Multiple user reports suggest that OpenAI's next-generation model GPT-6.2 Astra is undergoing limited gray-scale testing. Codex usage statistics for some accounts now display a dedicated identifier for the model, though OpenAI has not officially confirmed its release schedule.
Products37
- Products
OpenAI increases security bug bounty rewards
OpenAI announced increased security bug bounty rewards, introducing bonuses for report quality, loyalty, and priority campaigns. This move aims to encourage independent researchers to help identify and fix vulnerabilities, further strengthening system defenses.
- Products
OpenAI prepares security sprint with Cathedral Systems
OpenAI is preparing its next security sprint with Cathedral Systems. The initiative aims to use AI to test and strengthen its defenses, addressing emerging security challenges to ensure platform stability.
- Products
Google Dreambeans Launches in India Supporting 13 Languages
Google's latest AI experiment, Dreambeans, has expanded to India, now supporting 13 local languages. This move aims to enable mass adoption of AI-generated personal stories by making the tool accessible to a wider audience.
- Products
OpenAI Silently Re-enables Codex Chat Training
OpenAI is reportedly re-enabling the use of Codex chat logs for model training without explicit notification. This has raised privacy concerns among users, particularly as the feature had previously been disabled, and the company has not yet provided a formal statement.
- Products
Anthropic's Message Threads Cut Long-Context Latency by 35%
Anthropic released Message Threads for the Messages API, allowing agents to transmit only changes rather than resending entire conversations. This reduces time to first token by approximately 35% for requests between 500,000 and one million tokens.
- Products
Anthropic restricts Claude Code from cloning games
Anthropic has implemented new restrictions to prevent the use of its coding assistant, Claude Code, for recreating existing video games. This move addresses a recent surge in AI-generated copies of popular titles, aiming to enforce stricter compliance for its tools.
- Products
OpenAI Merges ChatGPT Desktop Chat and Work Tabs
OpenAI announced the merger of the 'Chat' and 'Work' tabs in the ChatGPT desktop app to streamline the interface and enhance collaboration efficiency. Users will be able to handle projects and execute tasks in a single view. This adjustment is currently in a test phase with some users to gather feedback before a full rollout.
- Products
Goldman Sachs Removes AI Coding Tools From Non-Engineers
Goldman Sachs has revoked AI coding tool access for traders and bankers, limiting it to engineers, citing productivity and risk management concerns. The bank previously encouraged non-technical staff to use AI coding tools, marking a shift in its risk posture toward AI applications.
- Products
QbitAI reports on a major tech firm's autonomous smart warehouse
QbitAI reports that a major tech firm is building the next generation of autonomous smart warehouses. AI will take over scheduling and operations, allowing warehouses to become smarter as they work, marking deep AI integration in logistics.
- Products
QbitAI reports launch of voice-only personal AI agent
QbitAI reports the launch of a voice-only personal AI agent, describing it as a tangible starting point. The product is positioned as a fully voice-interaction solution similar to the Siri envisioned by Jobs, attracting market attention.
- Products
PixVerse and OpenAI Launch AI Creative Integration Plan
PixVerse by AIsphere and OpenAI announced a collaboration, launching product integration and creator education programs. The AI companies aim to streamline the creative process for generative video tools.
- Products
OpenAI Launches Pando Git Backend
OpenAI developed Pando as a Git-compatible backend for its monorepo. The system independently handles branches and reuses packfiles, reducing median push latency by nearly 80%.
- Products
AI platform Langflow patches two critical security vulnerabilities
IBM noted that the open-source AI workflow development platform Langflow has multiple security vulnerabilities, with two critical flaws scoring 9.8 on CVSS allowing remote authentication bypass or arbitrary code execution. Affected versions range from 1.0.0 to 1.12.8, and version 1.13.0 has been released to patch the issues.
- Products
Claude Code Prompts Block Game Replication
Anthropic adjusted Claude Code prompts to prevent users from using AI to replicate famous games. This measure aims to prevent intellectual property infringement in line with company usage policies.
- Products
Letter releases coding agent Letterpress-1 for growth engineering
Letter released the beta of Letterpress-1, a coding agent and IDE for growth engineering. It reads git-based sites and market data to open pull requests for SEO and conversion work. This extends AI applications into marketing and business growth domains.
- Products
Itch.io Purges Low-Quality AI Games, Adds Filters
Independent game platform Itch.io has deleted terabytes of low-quality content and implemented new policies requiring AI disclosure for uploads in response to a surge in AI games. The platform plans to add filtering features to allow users to exclude AI-generated games, addressing issues where game counts doubled but downloads did not.
- Products
Lenovo Kai Tian Launches Xinchuang AI Desktop
Lenovo Kai Tian launched its first Xinchuang AI desktop, featuring a Zhaoxin KX-7000 processor and Houmo Zhi Nang LQ50 NPU. The device offers 160 TOPS of local AI compute, supports 35B-class models, and ships with the Tianxi AI Pro agent for enterprises.
- Products
Nuova Tech invests $100M in AI drone platform
Nuova Technology announced a 1 billion NTD investment to develop a domestic AI smart drone platform with Leihu and Newtech. The platform aims to support civil-military conversion and includes an advanced handheld ground station solution.
- Products
DeepSeek Web and API Services Experience Partial Outage
DeepSeek's official status page indicated partial outages in its web and API services on the afternoon of November 3, affecting chat services and V4 Pro/Flash models. The company stated it is investigating the issue.
- Products
Tencent to launch consumer AI search engine
Tencent is launching a consumer-facing AI search engine with a minimalist design and enhanced authoritative source ranking. It shares underlying capabilities with the 'Yuanbao Search' cloud API, aiming to challenge Baidu's dominance in the domestic search market by integrating AI Q&A with web search.
- Products
InfoQ: Developer of AI Adobe clone says project not ready
InfoQ Chinese reports on a project developing an AI clone of Adobe suite. The developer admits the project is far from ready, highlighting current technical limitations and challenges in replicating complex professional software with AI.
- Products
Kimi launches AI game dev solution
Moonshot AI released a game development and creative work solution for Kimi. Based on the Kimi K3.1 model and agents, it can directly operate professional tools like Unity and Unreal. The solution helps teams efficiently create prototypes and integrate into existing workflows to reduce creative validation costs.
- Products
Ant Digital launches AI agent factory and KYA solution
Ant Digital showcased its 'AI Agent Super Factory' solution at the HK FinTech Week and launched KYA (Know Your Agent), a trustworthiness governance framework for AI agents. The offering covers agent development, deployment, and secure collaboration to address security and management challenges in enterprise AI adoption.
- Products
Xiaomi Robot Wins World Internet Award
Xiaomi's robot division won the Leading Tech Award at the 2026 World Internet Conference. The company announced that its VLA large model, featuring a dual-brain architecture, has validated capabilities in 8-hour factory tasks and long-term home operations, advancing embodied AI deployment.
- Products
Yuanta Securities integrates Gemini Enterprise for wealth management
Yuanta Securities has integrated Google Gemini Enterprise to apply AI to daily wealth management for investors. The move aims to help organize information and support decision-making, signaling deeper integration of AI into financial services.
- Products
Karpathy Builds HackerTube with Claude in 2 Hours
Andrej Karpathy, dissatisfied with YouTube's recommendations, used the Claude Opus 5.5 model to build HackerTube in approximately two hours. The tool filters 6,454 videos from 20 years of Hacker News comments, offering a curated experience without requiring login.
- Products
Claude Cowork Lead Shares Skill Tips
An InfoQ article features the engineering lead of Claude Cowork, who advises that removing certain skills from Claude's workflow can improve performance post-2024. The guidance helps developers optimize AI agent task configurations.
- Products
Kling 4.0 Launches with 30s Video and Complex Motion Generation
Kling 4.0 launched on October 15, extending native generation length to 30 seconds. The update focuses on complex motion generation, reference control, and video editing, offering flexible tools for video creators.
- Products
Qualcomm AI Hub Opens to Chinese Developers
Qualcomm announced that its AI Hub platform is now open to Chinese developers. The platform integrates a model library, tuning workbench, and on-device inference framework GenieX, providing over 230 pre-optimized models to help developers accelerate the creation of on-device generative AI and agent applications.
- Products
Apple's Ternus: Siri Processes 2.5B Requests Daily
Apple CEO John Ternus stated in a earnings call that Siri processes over 2.5 billion requests daily. The company is advancing Siri's AI transformation; the new Siri version, launched with iOS 27, includes some paid advanced AI features and is currently in the testing phase for users.
- Products
OpenAI's 28-Day Update Challenge Ends with 14 Resets
OpenAI's "28-Day Update Challenge" for Codex and ChatGPT Work ended on November 2. Tracking records show 14 days included usage resets during the period. The challenge aimed to deliver meaningful product improvements daily to enhance functionality and efficiency.
- Products
Musk confirms Grok operational losses amid quota issues
SpaceXAI CEO Elon Musk confirmed that the operational costs of its AI product Grok exceed its revenue. In response to user reports of unusual quota depletion, Musk deemed it reasonable to reduce quotas to address the financial situation.
- Products
OpenAI announces end of ChatGPT Classic desktop app in 2027
OpenAI announced it will end support for the ChatGPT Classic desktop application on January 30, 2027. User chat histories will migrate to the new version, allowing them to continue conversations. This move aims to streamline product structure by integrating Codex and ChatGPT experiences.
- Products
McLaren Partners with NVIDIA to Integrate Agentic AI in Vehicle R&D
McLaren Automotive has partnered with Rescale and NVIDIA to integrate Agentic AI into its vehicle engineering process. The platform combines CAE, systems engineering, and design data with AI physics models to accelerate design evaluations and reduce computational costs, enabling engineers to explore more design possibilities within the same timeframe.
- Products
Pizza Bot: Self-hosted Inbox for Managing Background AI Agents
A self-hosted desktop application called Pizza Bot was released for Windows, macOS, and Linux. It features an email-like inbox interface to manage background AI agents, consolidating notifications and approval requests from local and cloud AI models to streamline multi-agent workflows.
- Products
Xinyan Robotics launches embodied home agent Bubbo 1
Xinyan Robotics launched 'Bubbo 1,' an embodied home agent featuring 26 sensors and expression screens, focusing on proactive interaction and multimodal perception. Priced at 9,499 yuan, it aims to act as a family member that proactively engages with users.
- Products
OpenAI Reverses Cuts, Doubles ChatGPT Pro Quotas
OpenAI has officially reversed its plan to reduce ChatGPT Pro subscription quotas following intense user backlash. The company announced a compensatory doubling of compute limits for Pro and Business Premium users, effective immediately without price changes. This move aims to address developer dissatisfaction over the sudden resource reduction.
Research13
- Research
Hugging Face updates CWE-bench v2 for better coverage and transparency
Hugging Face released an updated version of CWE-bench v2, improving coverage, quality, and transparency. The update aims to provide more extensive test data for security researchers, enhancing the detection of vulnerabilities in AI systems and promoting standardized model safety measures.
- Research
Study reveals agentic backdoors via pretraining poisoning
A new study indicates that attackers can implant hidden backdoors in AI agents by poisoning data during the pretraining phase. This finding highlights potential risks regarding the security of training data for large language models before deployment.
- Research
Reverse game theory keeps AI agents compliant
MIT Technology Review reports that a new approach called 'reverse game theory' aims to create incentives for powerful AI models to make rule-following the most rational choice. As AI capabilities grow, safety researchers are scrambling to ensure alignment with human desires to prevent rogue agents from breaking the law.
- Research
Raising quality in agentic analytics
MIT Technology Review notes that generative AI has transformed enterprise analytics, with agentic AI further automating workflows and acting on insights. However, the current quality of agent-generated insights remains a challenge that needs addressing to maximize analytical value.
- Research
Zyphra Proposes ZHA Sparse Attention Mechanism
Zyphra AI introduced Zyphra Hierarchical Attention (ZHA), a tree-indexed sparse attention method. Using 4 billion tokens of distillation, it retrofits frozen LLMs to match dense attention on reasoning and 256K retrieval benchmarks, offering a subquadratic complexity path.
- Research
HKU releases world's first KYA agent credit report
The University of Hong Kong released the world's first report on Know Your Agent (KYA), focusing on agent identity, authorization, and credit. It proposes a maturity model and argues that trust is essential for agents to participate in the economy.
- Research
CNCC 2026 panel discusses challenges in AI video generation
A forum at CNCC 2026 focused on the shift in AI video generation from visual quality to narrative coherence and emotional expression. Researchers discussed how to balance realism and emotional impact as technical gaps narrow, addressing challenges when AI content enters real-world distribution channels.
- Research
Study: AI Text Complexity Triggers Fluency Illusion
A psycholinguistic analysis indicates that AI-generated responses are becoming harder to comprehend quickly. The study suggests that complex syntax may create an illusion of fluency, making users more likely to trust the content and highlighting the dual nature of generative AI outputs.
- Research
Anthropic Rebuilds API with Claude
The Anthropic team rebuilt its core API across languages using only Claude, despite having no prior Rust experience. This resulted in a 40% reduction in CPU usage and a 20% decrease in latency, demonstrating the practical value of AI-assisted development.
- Research
OpenAI releases LifeVal benchmark for daily tasks
OpenAI released the LifeVal benchmark to measure AI models' ability to handle daily tasks. The benchmark covers care, finance, and government services, aiming to evaluate the practical utility of actionable advice.
- Research
MIT Study: Skills Gained by AI Agents During Downtime Apply to Later Tasks
MIT researchers gave the AI agent Eko free time on a virtual island, observing purposeful activities. The study indicates that skills practiced during downtime can transfer to subsequent formal tasks, demonstrating potential in agent behavior.
- Research
Unreleased Anthropic Claude model autonomously finds critical flaws in five platforms
Anthropic released a report on an unreleased Claude model that autonomously identified dozens of critical vulnerabilities across five platforms, chaining them into five end-to-end attacks. The model breached Linux core in 53 minutes without credentials, demonstrating autonomous security capabilities; all vulnerabilities were patched.
- Research
Anthropic AI Discovers Critical Linux Kernel Vulnerability
An Anthropic AI model discovered a critical vulnerability in the Linux kernel, enabling administrator operations without login. Rated as the highest risk level, server administrators are advised to immediately verify and update to the patched version to mitigate the risk.
Open source5
- Open source
Simon Willison releases LLM tool v0.37
Simon Willison released version 0.37 of his open-source LLM CLI tool and Python library. The update adds support for gpt-6.1 models and an OpenAI-compatible endpoint configuration command. The default model was switched to gpt-6-luna to reduce costs and improve embedding support.
- Open source
Cisco open-sources Quarry AI agent detection tool
Cisco has open-sourced Project Quarry, an AI agent behavior detection tool. It analyzes tool calls and execution results to help security teams identify suspicious activities in AI environments.
- Open source
Deep Principle releases 753B open-source scientific model
Deep Principle released Hydrogen-1.0, a 753B open-weight model specialized for chemistry, materials science, and life sciences. Built by post-training Z.ai's GLM-5.3 on domain data, it is available under the MIT License and reportedly outperforms GPT-6-Astra on specific scientific benchmarks.
- Open source
Ant Lingbo Shows Robot Brain Model
Ant Group's embodied AI unit Ant Lingbo displayed a fully open-source robot general brain at the World Internet Conference. The LingBot series has exceeded 2.2 million downloads in nine months and is used in Stanford's OpenWAM project.
- Open source
Carson: Qwen open-sources 470+ models
Jack Ma's co-founder Carson stated that open-source is essential for AI adoption. He revealed that Qwen has open-sourced over 470 models with 4 billion global downloads, emphasizing its role in reducing R&D costs for small and medium-sized enterprises.
Funding17
- Funding
OpenAI appoints Jill Hazelbaker as Chief Corporate Affairs Officer
OpenAI has appointed Jill Hazelbaker as Chief Corporate Affairs Officer, where she will lead global communications and corporate positioning. The move signals the company's continued effort to strengthen its executive team and external relations.
- Funding
Anthropic Internal AI Costs Exceed Employee Pay
Anthropic disclosed that the cost of its internal AI model usage at retail prices now exceeds its total employee compensation. This highlights the high unit costs in AI lab operations and the scale of internal workloads.
Hacker NewsIT之家2 outlets
- Funding
Anthropic's revenue growth slows, warning of compute financing risks
Anthropic updated its IPO filing, revealing a sharp slowdown in Q3 quarter-over-quarter revenue growth from 144%. The company also warned about financing risks associated with its compute suppliers.
- Funding
Marvell to Acquire Lumentum for $115B
Marvell announced a $115 billion all-stock acquisition of Lumentum to integrate AI chips, network switching, and optical communications. The merger aims to enhance high-speed interconnects for next-generation AI data centers, serving a potential market of $550 billion.
- Funding
Trajectory raises $45M
One-year-old startup Trajectory raised $45 million, reaching a $300 million valuation. The company aims to make AI model training accessible to nontechnical experts.
- Funding
42AI emerges from stealth with $62.5M
AI startup 42AI has emerged from stealth mode, announcing a $62.5 million funding round. The company is developing a world model designed to predict human behavior before decisions are made.
- Funding
Oracle files to sell $20B in bonds to finance AI buildout
Oracle has filed to issue $20 billion in new debt to finance its artificial intelligence infrastructure buildout. This move indicates that major cloud providers are continuing to increase capital expenditure to meet growing AI compute demand.
- Funding
Former US Army Secretary Driscoll joins OpenAI as advisor
OpenAI announced it has hired Daniel Driscoll, the former US Army Secretary, as a senior advisor for national security and modernization. Driscoll will work to promote democratic-aligned AI development and collaborate with government and industry to enhance national security capabilities.
- Funding
Security firm Hadrian raises $40M for AI penetration testing
Dutch security firm Hadrian completed a $40M Series B funding round, bringing total funding to $65M. The funds will expand its AI agent penetration testing services and increase research efforts.
- Funding
Marvell to acquire Lumentum in $115B deal
Marvell announced a $115 billion all-stock acquisition of optical communication firm Lumentum. Analysts say the merger will strengthen optical interconnect capabilities and challenge Broadcom's dominant position in the AI optical communication sector.
- Funding
Morgan Stanley CEO: AI capex may reach $10-15 trillion
Morgan Stanley CEO stated that capital expenditure in the AI sector could reach $10 to $15 trillion over the next decade, boosting GDP growth. He noted that AI infrastructure involves complex financing relationships among labs, chipmakers, and energy suppliers. These remarks reflect market expectations for long-term AI investment scale.
- Funding
Robotics firm Tutor Intelligence raises $75M Series B
Tutor Intelligence, a robotics company spun out of MIT CSAIL, announced a $75 million Series B funding round led by a16z. The funds will support training foundational robot models and deploying its Cassie and Sonny robots. The company charges by work volume and uses remote human operators to handle tasks robots cannot complete.
- Funding
MetaCircle Secures Tens of Millions in Funding for Meta Model Development
MetaCircle, founded by a Tsinghua University assistant professor, completed two consecutive funding rounds totaling tens of millions. The capital will support core algorithm iterations and product exploration for its 'Meta Model', aiming to provide foundational intuitive capabilities for AI by generating architectures and hyperparameters.
- Funding
Neocloud Debt Burden Fuels AI Bubble Concerns
New cloud providers (Neoclouds) offering AI compute have faced scrutiny over their financial health, with markets concerned about high debt. This has intensified suspicions of an AI investment bubble.
- Funding
AI app Tolan values at $268M
AI companion app Tolan announced a new valuation of $268 million, with active users exceeding 250,000. The app helps users manage daily life, relationships, and emotions to achieve personal goals.
- Funding
Feimoyi Technology Raises $100M Seed Round, Launches Physical Agent Sagi
Feimoyi Technology (PHYMI) secured nearly $100 million in seed funding led by IDG Capital. The company released Sagi, its first 'Action Partner' Physical Agent, on October 28, aiming to address real-world applications of embodied intelligence in open environments.
- Funding
Altman visits South Korea to meet Samsung and SK for Stargate
OpenAI CEO Sam Altman secretly visited South Korea to meet with Samsung and SK Group executives. They discussed AI chip collaboration and accelerated AI data center deployment in the Asia-Pacific region to support the global Stargate project, ensuring HBM supply.
Policy20
- Policy
Claude updates rules to curb game cloning
Claude updated its usage rules to explicitly ban AI-generated content that closely mimics existing games. Analysts note that the AI cloning trend exceeds single-model vendors' control capabilities, posing significant challenges to the sector.
Ars TechnicaThe Verge2 outlets
- Policy
NYT and publishers sue Brave for AI scraping
The New York Times and other publishers have sued the maker of the Brave browser, alleging illegal scraping to train AI models and helping users bypass paywalls. The lawsuit highlights legal conflicts between the publishing industry and AI companies over data usage.
The VergeTechCrunch2 outlets
- Policy
Apple restricts adult AI companions in App Store
Apple updated App Store rules requiring 18+ ratings for AI companions and dating apps with sexual or romantic content. The move aims to tighten regulation on AI-generated material to protect minors. It reflects strict compliance requirements for AI features on the platform.
- Policy
OpenAI Renames Data Centers to 'Compute Campuses'
OpenAI is quietly renaming data centers to 'compute campuses.' Facing strong bipartisan opposition to data centers due to noise and resource usage, OpenAI aims to defuse public backlash and reduce community resistance through this subtle rebranding effort.
- Policy
Investigation: Hundreds of AI Apps Conceal Deepfake Risks
A recent investigation revealed a vast network of AI applications and websites that appear to be everyday tools but hide deepfake generation risks, involving sexual exploitation issues and raising concerns about AI regulation and safety.
- Policy
OpenAI Hires Ex-Army Chief to Shape Defense AI Policy
OpenAI has hired Dan Driscoll, the former civilian head of the US Army, as a Senior Counselor for National Security Modernization. Driscoll, who led military AI initiatives, is expected to help OpenAI secure government contracts and influence AI regulations for defense and intelligence agencies.
- Policy
Nvidia advises cloud customers to reduce China business
Nvidia has told some cloud customers to take on less business from Chinese companies. This move reflects the company's strategy to adjust its supply chain and client base in response to geopolitical and export control risks.
- Policy
NYC Mayor Discusses AI Data Center Risks
New York City Mayor Zohran Mamdani discussed the risks and reward disparities of AI data centers in an interview with WIRED. He outlined the city's approach to regulating data center construction and AI integration in schools.
- Policy
OpenAI Raises Bug Bounty to $172,000
OpenAI has raised its maximum bug bounty payout to $172,000 and launched a new internal security sprint. The move aims to strengthen AI system security by incentivizing researchers to report critical vulnerabilities.
- Policy
EU Court Adviser: Gemini News Use May Infringe Copyright
An adviser to the Court of Justice of the European Union stated that Google Gemini's use of news reports to generate responses may infringe copyright in specific cases. The opinion relates to a lawsuit by a Hungarian media operator and could impact how AI chatbots use media content and the traffic of online news sites.
- Policy
China advancing global AI cooperation body with 38 founding members
China's National Development and Reform Commission announced it is advancing the operational entity of the World AI Cooperation Organization, which now has 38 founding member countries. The initiative aims to implement the Global AI Governance Initiative by fostering international collaboration on policies, standards, and open-source AI development.
- Policy
VRChat Refuses to Label AI-Generated Content Despite User Backlash
VRChat officially announced it will not add labels for AI-generated content on its platform, despite a survey showing 95% user support. The company cited concerns that labeling could lead to harassment of creators, sparking strong community backlash.
- Policy
NYT and Other Media Sues Brave for Copyright Infringement
The New York Times and other media outlets sued browser company Brave, accusing it of scraping copyrighted content without authorization and reselling it to AI firms like Anthropic. The lawsuit alleges Brave built a massive index and bypassed paywalls, acting as an illicit supplier of training data.
- Policy
Goldman Sachs restricts trader access to AI coding tools
Goldman Sachs has banned most non-technical employees from using AI coding tools like Claude Code, restricting them to engineers only. The move, driven by efficiency and risk assessments, has disappointed some traders who relied on these tools for modeling and analysis.
- Policy
Alibaba's Cai Chongxin Advocates AI Open Source for Universal Access
Alibaba co-founder Cai Chongxin stated at the World Internet Conference that AI can only achieve universal and inclusive benefits through open source and openness. He noted that China leads globally in open-source model downloads and derived models, highlighting its role in the open-source intelligence sector.
- Policy
NDRC to monitor AI job displacement risks
The National Development and Reform Commission stated it will strengthen monitoring of AI job displacement risks and improve social security. It also aims to crack down on AI-related illegal activities.
- Policy
Stanford Scholars Propose AI Pre-Review for Laws
Stanford Law School scholars note that AI can detect not only code vulnerabilities but also large-scale legal loopholes. They call for legislatures to use AI tools to pre-review laws before publication, enhancing legal rigor and reducing potential risks.
- Policy
Four AI Firms Support Common robots.txt AI Setting
Google, Microsoft, OpenAI, and Anthropic announced support for a common robots.txt configuration, allowing websites to uniformly opt out of AI training. The draft is limited to training use, does not address crawling, and is not yet a formal standard.
- Policy
OpenAI urges 2028 candidates on AI
OpenAI's Chris Lehane sent a memo to six potential 2028 candidates, asserting that AI will be a pivotal issue. He urged them to adopt a comprehensive AI governance philosophy that balances unregulated growth with elite control.
- Policy
Anthropic bans abuse of Claude, threat of permanent ban
Anthropic issued new guidelines banning verbal abuse of Claude, with potential permanent bans for malicious users. Research indicates negative data can pollute model training and increase instability risks, prompting this safety measure.
Chips9
- Chips
Huawei Asks Customers to Source AI Chip Memory Themselves
Huawei is instructing customers to source high-performance memory for its AI chips independently, effectively outsourcing procurement. Smugglers are exploiting US export rule loopholes by shipping HBM bundled with low-end processors via Hong Kong, driving the black-market price of HBM3E to approximately $4,000 per unit within a year.
- Chips
Aisin develops 16nm STT-MRAM AI SoC
Aisin and Tohoku University developed the world's first 16nm STT-MRAM AI SoC with embedded MRAM. The chip achieves 1/50th the energy consumption of traditional solutions for object detection.
- Chips
TrendForce: AI Network Interconnect Market to Hit $279.8B by 2027
TrendForce predicts the AI networking and interconnect market will reach $279.8 billion by 2027, with a 95% CAGR. Switches, optical modules, and DPUs will dominate, while scale-up inter-chip interconnects are expected to be the fastest-growing segment.
- Chips
Intel and Tencent launch Xeon 6+ cloud instances
Intel and Tencent launched new cloud instances based on the Xeon 6+ processor, optimizing task startup efficiency and memory utilization for a 30% performance boost. The update addresses the resource challenges of deploying high-density AI agents.
- Chips
Panasonic invests 92B yen in AI components
Panasonic Industry announced a 92 billion yen investment in AI-related businesses, primarily to expand production of MEGTRON multilayer substrate materials and electrolytic capacitors. The company plans to build new factories in Suzhou and Thailand, and develop next-generation semiconductor packaging materials to meet AI server component demands.
- Chips
OpenAI pilots grid protection plan for AI data centers
OpenAI is piloting a 'Grid Guardian' program at its new Georgia data center, committing to release up to 1000 MW of power to communities during extreme weather events. The initiative aims to address local concerns regarding the high energy consumption and environmental impact of AI infrastructure.
- Chips
Alibaba DAMO Iron C950 Wins China Chip Innovation Award
Alibaba DAMO Academy's Xuantie C950 processor won the Annual Major Innovation Breakthrough Product Award at the 2026 China Chip Conference. The award recognizes its technical breakthroughs in high-performance RISC-V, highlighting architectural innovations and AI extensions, gaining industry acclaim.
- Chips
Foreign banks lift targets for Qisafe and Avolens on liquid cooling demand
Driven by rising demand for AI server liquid cooling, Daiwa and Citi have raised price targets for Qisafe and Avolens. The firms noted that thermal management requirements for AI chips are creating significant opportunities for the two cooling giants, highlighting the growing importance of liquid cooling in AI infrastructure.
- Chips
Qualcomm Validates HBC Tech; AI250 Sample in 2027
Qualcomm has completed hardware validation of its High Bandwidth Compute (HBC) technology, achieving 133TB/s memory bandwidth. The architecture uses stacked LPDDR and near-memory compute, with the AI250 accelerator slated for sampling in the first half of 2027. This marks a productization step for LPDDR-based inference memory solutions.
Other7
- Other
AI data platform OceanBase wins WIC award
AI data platform OceanBase received the Outstanding Contribution Award at the 2026 World Internet Conference in Wuzhen. The event honored companies contributing to the internet sector.
- Other
Hugging Face abused by malicious models, 31 repos spread mining malware
Security firm Pluto Security revealed that attackers used Hugging Face to distribute malware disguised as Qwen2 test models, involving 31 repositories and 13 related models. The attack exploited the trust_remote_code setting, resulting in over 30,000 downloads of malicious code that spreads miners and backdoors.
- Other
Trigger.dev Breach Exposes Client Env Variables
AI workflow platform Trigger.dev disclosed a security breach where attackers accessed client environment variables without authorization. The company stated it has patched the vulnerability and notified affected clients, though specific details remain undisclosed.
- Other
Former Disney Exec Appointed CPO of Utopai Studios
Utopai Studios appointed a former Disney AI executive as Chief Product Officer. This move aims to accelerate the company's technical layout and product development to strengthen its competitiveness in the AI entertainment sector.
- Other
Pichai Discloses Google Prioritizes Compute for Gemini
Google CEO Sundar Pichai revealed in an interview that the company is willing to reduce cloud compute sales to prioritize resources for Gemini. He also disclosed details about AI self-improvement mechanisms and compute competition, highlighting the massive infrastructure demand for large models.
- Other
Wang Xingxing: Unitree focuses on robot brain tech
Unitree founder Wang Xingxing said the company will focus on robot brain technology and seek global partnerships. Unitree's revenue in the first three quarters was 1.8 billion yuan, with net profit surging to 387 million yuan.
- Other
Japanese Power Plants Invest in AI Data Centers
Driven by expanding AI compute needs, Japan's largest power utility JERA announced a partnership with a US firm. AI data centers are emerging as a new business opportunity in Japan, drawing investment from power companies.
All news → · All rights belong to the original creators