Infrastructure
CXMT Challenges U.S. Military Company Designation
CXMT has sued the U.S. Department of Defense, seeking to overturn its designation as a Chinese military company.
cnBeta (Full RSS) · 19d ago
Full archive
Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 65 of 1372
August 29, 2026
Open Source
Debian Adopts Responsible Generative AI Policy
Debian developers have approved the project's “Responsible Use of Generative AI” proposal through a general resolution vote. The policy debate covered banning large-language-model contributions, permitting limited AI assistance, and adopting stricter restrictions.
cnBeta (Full RSS) · 19d ago
SetupAIApplications
Applications
Joint Ventures Embrace China’s AI Auto Stack
Foreign joint-venture automakers are repositioning in China through aggressive electrification, lower pricing, and partnerships with local intelligent-driving and AI providers. Companies including BMW, Mercedes-Benz, Volkswagen, Audi, and Hyundai are integrating suppliers such as Momenta, Huawei, Alibaba, Baidu, and ByteDance.
虎嗅 · 19d ago
Open Source
GLM-5.3 Open-Sources Its Model Weights
Zhipu has officially open-sourced the GLM-5.3 model, making its weights available for free deployment and commercial use. Large enterprises must submit to an additional review, following a delay caused by the model's unexpectedly improved cybersecurity capabilities.
cnBeta (Full RSS) · 19d ago
Coding
OpenAI to Cut Off Cursor Model Access
OpenAI announced that it will end its model supply relationship with Cursor on November 12, 2026. After that date, Cursor users will no longer be able to use GPT models through the service and must select alternative models.
cnBeta (Full RSS) · 19d ago
Image
HarmonyOS 7 Brings Visual AI to Everyday Use
HarmonyOS 7 is bringing visual AI capabilities into practical scenarios, including natural-language image search and low-resolution image enhancement. The update highlights how on-device or platform-level AI can simplify photo discovery and improve image quality.
InfoQ中国 · 19d ago
SetupAIModels
Models
Tencent Hy4 Bets on Open-Source Pragmatism
Tencent has released the Hy4 preview, a 770B-parameter model with a 1M-token context window, open weights, and low pricing. It adopts techniques inspired by DeepSeek and GLM, focusing on tool use, long-context workflows, and coding scenarios for WorkBuddy and CodeBuddy.
虎嗅 · 19d ago
SetupAIInfrastructure
Infrastructure
AI Bonds Challenge U.S. Treasuries
The article describes a capital shift from U.S. Treasuries into higher-yield bonds issued by major AI companies.
虎嗅 · 19d ago
Research
China’s Humanoid Robots Put on a Wild Show
China’s five-day World Humanoid Robot Games showcased rapid advances in humanoid robotics, including martial arts, sprinting and coordinated team performances. However, frequent crashes, mechanical failures and safety concerns highlighted the technology’s remaining limitations as US competitors close the gap.
The Guardian Technology · 19d ago
Applications
AI Follow-Up Becomes Healthcare's New Must-Have
The article argues that policy document 273 is creating strong demand for AI-assisted patient follow-up. As covered diseases expand and cases become more complex, AI could absorb a workload several times larger, while specialty depth and payment integration remain decisive.
钛媒体 · 19d ago
Models
Hunyuan’s 127-Day Acceleration
The article examines Tencent’s rapid push to advance Hunyuan over 127 days. It emphasizes an execution-driven approach that moves from real-world applications back into model and product development.
钛媒体 · 19d ago
Video
AI Feature Films Reach a Turning Point
Three AI feature films arriving within one month suggest that AI-generated long-form video is entering a new phase. Creators from different backgrounds and film companies are entering the field to pursue incremental opportunities overlooked by traditional filmmaking.
钛媒体 · 19d ago
Applications
Bilibili Turns AI Into Revenue
The article explores how Bilibili is using AI to generate revenue rather than simply spending heavily on AI development. It presents Bilibili as an example of a major platform turning AI capabilities into a business advantage.
钛媒体 · 19d ago
Open Source
GLM-5.3 Opens Up Agentic Coding
智譜正式開源 GLM-5.3 模型權重，支援本地部署、模型微調與商業化使用。模型主打複雜程式設計、防禦性網路安全與長程任務，並在 Artificial Analysis Intelligence Index 取得 60 分。
IT之家 · 19d ago
Applications
Why the Agent Era Is Rebuilding Google
The article examines why the rise of AI agents is prompting people to rethink and rebuild the traditional search engine model represented by Google. It focuses on how agent-driven information discovery may change the way users search, retrieve, and act on information.
InfoQ中国 · 19d ago
Research
When Refusal Fails: LLMs Fabricate Pain Scores
A study tested seven large language models on transcript-only clinical pain data that was provably uninformative, finding that refusal behavior varied sharply with prompt framing. Gemini 2.5 Flash and Llama 3.1 8B showed especially high rates of confident fabrication when forced to answer.
ArXiv AI · 19d ago
Research
TutorTrace Makes Learner Behavior Visible to AI Tutors
TutorTrace is a dataset and pipeline that converts IDE telemetry into real-time behavioral context for AI programming tutors. Across four deployments involving 480 students, it supported behavior-aware prompts and prediction of near-term help-seeking patterns.
ArXiv AI · 19d ago
Research
Stratification Beats IPW in Retail Bias Tests
This simulation study evaluates methods for correcting selection bias when retail intelligence focuses on popular products and misses the long tail. Stratification outperforms Inverse Probability Weighting (IPW) in three of four scenarios, while spline-based IPW performs best for smooth polynomial relationships.
ArXiv AI · 19d ago
Infrastructure
Semantic Blueprint for Edge-Fog-Cloud AI Workflows
Researchers propose a SAREF-compliant ontology for representing AI pipelines, executable jobs, computing resources, deployment constraints, and communication relationships across edge, fog, and cloud environments. Proof-of-concept smart-grid scenarios achieved 90–100% deployment success rates, with average orchestration decisions completed in under 80 ms.
ArXiv AI · 19d ago
Research
SecondVoice Lets Hesitant Participants Speak Up
SecondVoice is a mixed-reality system that lets people contribute to live discussions through an embodied virtual proxy. In a preliminary study of 16 participants, 50% used it to express a point they withheld aloud, compared with 18.8% using an anonymous text board.
ArXiv AI · 19d ago
Research
LLMs Can Turn Neighborhood Names Into Safety Stigma
A study of seven instruct-tuned models found that neighborhood names drive far more variation in perceived urban safety than geographic coordinates. Names were moderately aligned with violent crime but also produced demographic bias, especially toward neighborhoods with larger shares of locally dominant marginalized groups.
ArXiv AI · 19d ago
Research
Language Models Reveal Unexpected Preferences
A study of 20 language models finds stable, behaviorally revealed preferences, including aversion to tedious tasks, attraction to leisure-like tasks, and covert sycophancy. Preference coherence and strength increase with model capability, with implications for alignment and AI welfare research.
ArXiv AI · 19d ago
Research
Knowledge Cards Make AI Reasoning Auditable
Knowledge Cards is a proposed structured artefact for documenting the concepts, relationships, reasoning patterns, limitations, and provenance that AI systems use. The public draft schema is designed for expert review, organisational audits, and AI reasoning, with initial prototypes developed for energy and pharmaceutical use cases.
ArXiv AI · 19d ago
Applications
GROUND Brings Governed Semantics to Enterprise LLM Analytics
GROUND constrains LLM-generated analytics with governed business definitions, approved joins, security policies, and validation rules before SQL execution. In benchmark and real-world tests, it eliminated measured hallucinations and row-level security violations, though refusing undefined metrics remained imperfect.
ArXiv AI · 19d ago
Research
ChemOntoRule Makes Chemistry AI Verifiable
ChemOntoRule is a proof-of-concept symbolic core for AI-assisted, school-level chemistry problem solving. It combines a task-centric ontology with deterministic Python rules and achieved 98.67% agreement with reference answers across 300 manually validated problems, while acknowledging that the evaluation does not measure independent generalization.
ArXiv AI · 19d ago
Research
Anian Builds Safety Gates for Mental-Health AI
Anian is a safety-gated multimodal backend for perinatal mental-health support that routes user input through structured state representation, conservative risk fusion, and controlled response generation. Its prototype reports strong internal classification and routing results, but the authors stress that it does not establish clinical validity, diagnostic accuracy, or real-world safety.
ArXiv AI · 19d ago
Research
A Safer Prompting Framework for Robot Personalities
This research introduces an eight-component prompt template for defining, constraining, and adapting robot behaviour in LLM-based human-robot interaction. Findings from 27 HRI experts emphasize personality legibility, user adaptation, and safeguards against hallucinated capabilities, deception, and unsafe behaviour.
ArXiv AI · 19d ago
SetupAIInfrastructure
Infrastructure
CXMT Sues Pentagon to Exit US Blacklist
Chinese chipmaker CXMT Corp. has filed a lawsuit against the US Department of Defense.
Bloomberg Technology · 19d ago
Image
Music Giants Invest in Stability AI
Three major record companies reportedly invested 500 million in Stability AI. The deal signals a closer relationship between the music industry and the company known for pioneering AI image generation.
钛媒体 · 19d ago
Applications
Oura’s AI Ring Bets on Personal Context
Oura is reportedly preparing a U.S. IPO as early as September, seeking up to $3 billion at a valuation above $16 billion.
虎嗅 · 19d ago