Applications
Xiaodu Sets September 8 AI Hardware Launch
Baidu's smart-device business Xiaodu will unveil new AI hardware on September 8 at 2 p.m. Beijing time.
TechNode · 10d ago
Full archive
Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 19 of 1373
September 7, 2026
Infrastructure
China’s Automotive Chip Replacement
The article reviews how the 2021 global chip shortage accelerated domestic substitution in China’s automotive semiconductor supply chain. It covers the barriers to automotive qualification, early mass-production successes, and growing local presence across power semiconductors, cockpit SoCs, MCUs, memory, and safety chips.
虎嗅 · 10d ago
Open Source
Qwen Open-Sources an Autonomous Driving Model
Alibaba’s Qwen team has released Qwen-Drive-1.0-4B, an open-source model that combines driving-scene understanding with vehicle-motion planning. Developed with Huazhong University of Science and Technology, it builds on Qwen3.5-4B and adds dedicated 3D perception and trajectory-generation components.
TechNode · 10d ago
Infrastructure
DeepSeek Plans Massive Huawei AI Chip Cluster
DeepSeek reportedly plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center under construction in Inner Mongolia. The proposed cluster would rank among the largest known deployments of Huawei AI chips, although its timing depends on production capacity.
TechNode · 10d ago
Models
Meta’s AI Model Bets Face the Monetization Test
The article examines whether Meta’s improved AI model can translate technical progress into meaningful business returns. It also questions how investors and practitioners should evaluate the value of AI spending at application-layer companies.
钛媒体 · 10d ago
Applications
Hangzhou’s Token Cards Bring AI Support to Founders
Hangzhou is introducing Token cards for entrepreneurs as AI tools become increasingly central to office work. The initiative reflects an effort to adapt city talent services to the needs of the AI era.
钛媒体 · 10d ago
SetupAIApplications
Applications
Qwen Office Launches Multi-User Workspace
Alibaba’s AI agent product Qwen Office has launched a new multi-user workspace. The feature is described as the first of its kind in the industry, targeting collaborative office workflows.
量子位 · 10d ago
Business
Microsoft Moves AI Governance Into Runtime
Microsoft is shifting AI governance from policy-level guidance toward enforcement during runtime execution. The move suggests that governance controls will increasingly operate directly within deployed AI systems rather than remain solely as organizational rules.
InfoQ中国 · 10d ago
Applications
China’s New AI Customer Service Rules Take Effect
A new national standard for AI customer service has taken effect, requiring companies to take responsibility for the systems they deploy. The change aims to address misleading, unhelpful, and unaccountable AI customer-service experiences.
钛媒体 · 10d ago
Applications
Tesla FSD Faces Europe’s Regulatory Maze
Tesla FSD Supervised has resumed supervised L2 testing on French public roads after France initially rejected the system. The European approval process remains fragmented, while Tesla’s safety data shows fewer reported collisions than human driving but lacks detailed incident and causality disclosures.
虎嗅 · 10d ago
SetupAIBusiness
Business
中科類腦完成數億元B+輪融資
中科類腦宣布完成數億元人民幣的 B+ 輪戰略融資，並由產業龍頭領投。此輪融資將進一步支持公司在類腦智能相關技術與業務上的發展，但文章未披露投資方及資金用途細節。
量子位 · 10d ago
SetupAIResearch
Research
Why 4-Bit Memory Breaks Recurrent Models
This arXiv study shows that deterministic 4-bit storage of recurrent states can massively degrade a fixed GRU's accuracy, increasing errors by about 70x and 300x on two fluorescence-lifetime parameters. Error feedback, residual memory, and direction memory recover accuracy without retraining, while experiments with an LSTM confirm that state-storage interfaces are critical to low-precision inference.
ArXiv AI · 10d ago
SetupAIResearch
Research
ResLearn-XR Targets Bursty XR Traffic and QoE Risk
ResLearn-XR is a residual learning framework for forecasting extended reality network traffic and estimating Quality-of-Experience risk. Its authors also introduce the DDA feature-construction module and an XR Traffic-QoE dataset, reporting substantial gains over single-stage baselines.
ArXiv AI · 10d ago
SetupAIModels
Models
Removing Hidden Influences from LLM Explanations
This research introduces a model-agnostic, test-time method for improving LLM explanation faithfulness. It removes input concepts not cited in the explanation, then re-queries the model to reduce incomplete or hidden reasoning influences.
ArXiv AI · 10d ago
SetupAIResearch
Research
PerfReasoning Exposes LLM Hardware Reasoning Gaps
PerfReasoning is a benchmark for evaluating how well LLMs reason about hardware performance and generate analytical performance-model code. Leading closed-source models exceed 90% on reasoning questions, but most models perform poorly when constructing reliable models.
ArXiv AI · 10d ago
SetupAIResearch
Research
Machine Learning Improves Power Grid Contingency Screening
This study applies KNN, Random Forest, and SVM to classify power-system contingencies as safe, moderate, or severe. On IEEE-14 and IEEE-30 bus systems, Random Forest achieved the strongest F1 scores, while PCA generally contributed more than SMOTE to overall performance.
ArXiv AI · 10d ago
SetupAIResearch
Research
Iris Search Agents Reach New Open-Source Frontier
Iris-mini and Iris-pro are new search agents trained at 35B-A3B and 397B-A17B scales. Using SFT-RL climbing, context management, and live-search optimization, they achieve leading open-source results across BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE.
ArXiv AI · 10d ago
SetupAIResearch
Research
HarvestBench Puts a Price on AI Compassion
HarvestBench is a reproducible benchmark that measures whether LLM agents will spend resources to avoid killing animals during a simulated corn harvest. Across nine models, kill rates ranged from 0.4% to 98.8%, while moral briefings dramatically reduced harmful behavior.
ArXiv AI · 10d ago
SetupAIInfrastructure
Infrastructure
Harbor Standardizes Large-Scale Agent Evaluation
Harbor Adapters provides a unified infrastructure for evaluating arbitrary agents across more than 80 agentic benchmarks. The accompanying Harbor-Index distills 82 challenging tasks across 29 benchmarks, enabling affordable and reproducible evaluation of agent capabilities and failure modes.
ArXiv AI · 10d ago
SetupAIResearch
Research
From Matching Models to Recruiting Agents
This systematized narrative review examines the evolution of AI recruitment, from profile matching and ranking to LLM-powered, tool-using recruiting agents. Reviewing 40 representative works, it highlights evaluation gaps involving fairness, privacy, security, workflow failures, and real-world outcomes.
ArXiv AI · 10d ago
SetupAIResearch
Research
EXAONE Finance Targets Missing, Long-Horizon Markets
EXAONE Finance is a financial time-series foundation model designed for long, multivariate, intermittently observed market data. Its attention-free architecture and finance-specific pretraining reportedly achieve state-of-the-art results across accuracy, asset ranking, and portfolio profitability on FinVerse.
ArXiv AI · 10d ago
SetupAIResearch
Research
CLM Turns Enterprise Knowledge into Governed Action
The paper introduces Corporate Language Model (CLM), a foundation for converting fragmented and tacit enterprise knowledge into ontology-grounded, auditable, and executable intelligence. Its architecture combines a Neurosymbolic Mesh, Skill Graph, Living Digital Twins, deep security controls, and Spec-as-Code, with a hospital case study demonstrating three maturity stages under Brazil’s LGPD.
ArXiv AI · 10d ago
SetupAIResearch
Research
Better Models, Riskier AI Systems
This arXiv study finds that improving individual LLM capabilities can worsen system-level outcomes when agents behave more similarly. In financial-market simulations, correlated decisions reduce risk when reasoning is accurate but amplify losses under shared misinformation.
ArXiv AI · 10d ago
Business
Game Companies Go All-In on Hard Tech
Chinese and overseas game companies are sharply increasing investments in AI chips, embodied intelligence, XR, brain-computer interfaces, advanced packaging, and other hard-tech fields. Beyond direct investments, they are acting as LPs, technology acquirers, AI customers, and buyers of compute and data.
虎嗅 · 10d ago
Image
When AI Detection Becomes a Witch Hunt
Artists are increasingly pressured to prove that their work was created by humans through livestreamed AI-authorship challenges. The article argues that these public trials are vulnerable to mob psychology, harassment, and fundamental limits in reliably distinguishing AI-generated work from human-created work.
虎嗅 · 10d ago
Applications
Hitachi Unveils Four Physical AI Services
Hitachi has announced four new services focused on implementing Physical AI. Rather than treating FDE as a single all-purpose expert, Hitachi proposes a team-based approach to address the complexity of real-world AI deployment.
ITmedia AI+ (日本) · 10d ago
Research
Gartner Flags New Technology for Peak Hype
Gartner has published its 2026 Hype Cycle, showing that agentic AI has moved beyond the 2025 Innovation Trigger stage. A technology appearing for the first time has already entered the Peak of Inflated Expectations, highlighting the gap between attention and practical maturity.
ITmedia AI+ (日本) · 10d ago
Models
GPT-6 Astra Low Beats GPT-5.6 Sol High
OpenAI representative Tibo stated on X that GPT-6 Astra’s low setting performs better than GPT-5.6 Sol’s high setting. The comment effectively recommends Astra’s low configuration over Sol’s high configuration.
ITmedia AI+ (日本) · 11d ago
Applications
Nubia NaviX Ultra AI Phone Opens Reservations
Nubia has opened reservations for the NaviX Ultra, described as the world’s first AI agent smartphone, with a launch scheduled for September 16. Developed with ZTE and ByteDance, it integrates the Doubao assistant and is expected to feature flagship hardware and high-speed domestic LPDDR5X memory.
IT之家 · 11d ago
Business
SenseTime Finds Profit in Practical Generative AI
SenseTime is pursuing profitability by focusing on generative AI that helps enterprises complete specific business tasks rather than competing solely on model size. The strategy contributed to a net profit of 617.3 million yuan in the first half of 2026.
SCMP Technology · 11d ago