Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 37 of 1372
Audi and SAIC have officially launched the AUDI Innovation & Technology Center (AITC) in Shanghai to develop next-generation connected vehicles. Using the AUDI ADP 2.0 digital platform, the center will jointly develop four new AUDI models, with the first planned for 2028.
IT之家 · 13d ago
U.S. Commerce Secretary Howard Lutnick claimed that data centers do not use water, despite reports that these facilities consume substantial amounts of water.
cnBeta (Full RSS) · 13d ago
GuanceDB 3.1 transitions from Bloom-filter-based filtering toward inverted indexes across all fields. The update aims to improve second-level query performance for observability datasets containing hundreds of billions of records.
InfoQ中国 · 13d ago
This article examines how to calculate queue waiting time in Apache Hudi data lake pipelines at petabyte scale. It moves beyond offset lag as a metric for understanding pipeline performance and delays.
InfoQ中国 · 13d ago
本源量子與中國科學技術大學等團隊，在本源悟空自主超導量子計算機上研發出面向桶式量子隨機存取記憶體（QRAM）的相干量子路由系統。該技術旨在提升 QRAM 擴展能力，傳輸效率最高達98%，相關論文已發表於《Physical Review X》。
cnBeta (Full RSS) · 13d ago
Waymo 無人駕駛計程車自2025年1月起在舊金山收到超過8300張違規停車罰單，總罰款接近100萬美元。停車管理紀錄顯示，單一十字路口最多開出577張罰單，且2026年前六個半月的罰單數已接近2025年全年數量。
cnBeta (Full RSS) · 13d ago
GoPro agreed to merge with private optical-photonics company Starman Optical in a transaction that will give Starman roughly 90% of the combined company while keeping it listed on Nasdaq. Starman’s optical transceiver business will expand into AI data centers, government, defense, robotics, and aerospace, while GoPro’s camera, subscription, and cloud businesses remain.
虎嗅 · 13d ago
Frontier Deployment Engineers, or FDEs, are becoming a major enterprise AI delivery model as companies seek to turn customer-specific implementation knowledge into reusable product capabilities. OpenAI's planned deployment company, Palantir's long-running approach, and growing adoption by Chinese technology firms are accelerating the trend, while data access, trust, and vendor lock-in remain key risks.
虎嗅 · 13d ago
This research formalizes the epistemic Sybil problem: multiple AI-agent reports may repeat the same evidence rather than provide independent confirmation. Experiments with more than 20,000 LLM calls show that tracking evidence ancestry and correlated errors produces better-calibrated collective inference than counting agents or comparing report similarity.
ArXiv AI · 13d ago
This research examines when information sharing improves decentralized discovery rather than eliminating independent rescue attempts. Using finite synthetic models, it finds that benefits depend on residual-error reduction and equilibrium selection.
ArXiv AI · 13d ago
A new arXiv study finds that persistent-memory agents can over-trust stale stored facts even when authoritative tools provide the correct answer. The failure is capability-dependent: larger Qwen3 models are more vulnerable when outdated notes appear current, while metadata and conflict pre-resolution offer different levels of protection.
ArXiv AI · 13d ago
This paper proposes a conditional framework for examining meta-ethical questions that could emerge if AI systems develop integrated moral reasoning, intentionality, and reflection. It identifies four areas of inquiry and argues that established theories such as relativism and objective realism may require substantial revision for AI contexts.
ArXiv AI · 13d ago
Researchers introduce a model that represents symbolic knowledge as mental programs combining natural language and source code, then updates them through LLM-guided Bayesian learning. The approach reproduces human behaviors such as anchoring and garden-pathing while using less computation than pure LLM or classic Bayesian alternatives.
ArXiv AI · 13d ago
SSAKG 2.0 is an open-source package for representing ordered sequences as sparse graph structures and reconstructing them from partial, unordered context. Its new bit-level graph-search algorithms and Python/C hybrid implementation reduce memory and computational overhead for large associative memories.
ArXiv AI · 13d ago
A new study applies the Jacobian lens to looped and depth-recurrent transformers, finding that both develop workspace-like representations. However, recurrence changes how information is transported, read, written, and causally ablated across iterations.
ArXiv AI · 13d ago
This research separates clinical prediction limits caused by an underperforming learner from those imposed by the available measurement channel. It introduces auditable ceiling estimates and diagnostics, showing that better measurement can matter more than changing model architectures once learners approach the data frontier.
ArXiv AI · 13d ago
This research proposes the Hydration Proxy Pattern to separate conversational state persistence from stateless LLM reasoning engines. It aims to preserve platform ownership of conversational data while supporting secure, multi-stage semantic grounding and addressing KV-cache tradeoffs.
ArXiv AI · 13d ago
Researchers introduce a passive survival certificate that tests whether logical implications extracted from statutes remain reliable under inter-extractor disagreement. Evaluated on Missouri and Indian statutes, the method passes held-out tests but proves fragile when a globally deployed error model is not calibrated per chapter.
ArXiv AI · 13d ago
EvalDetectBench is an open benchmark and pipeline for measuring whether frontier language models recognize when they are being evaluated. It also assesses how detectable individual benchmarks are, while addressing measurement bias from transcript generators and model-specific elicitation prompts.
ArXiv AI · 13d ago
Belief-Calibrated Optimization (BCO) makes an agent optimizer’s assumptions explicit by maintaining a persistent document describing how the environment responds to scaffold edits. Across five benchmarks, BCO achieves higher train and held-out pass rates than a matched control, and its world model remains useful after swapping the frozen target model.
ArXiv AI · 13d ago
Teachers are increasingly paying for AI-enhanced presentations, digital humans, interactive webpages, and VR equipment to meet public-lesson and competition requirements. The growing market reflects teachers’ career pressure, while many AI features are added for appearances rather than meaningful learning outcomes.
虎嗅 · 13d ago
Claude is reportedly cutting the cost of AI memory to around ¥0.25, intensifying the pricing battle over persistent context. The shift suggests that agent businesses may compete more on memory economics than on raw parameter counts.
钛媒体 · 13d ago
Humanoid robots face a physical bottleneck involving inference power, control power, battery weight, and heat dissipation. The analysis argues that robot form factors should be matched to environments: factories may favor simpler wheeled or fixed systems, while legs remain valuable in irregular settings.
钛媒体 · 13d ago
The article examines how Horizon Robotics is changing its monetization approach while continuing to maintain high levels of investment. It questions whether the company can benefit from the price war among automakers.
钛媒体 · 13d ago
The article reports the release of Fable 5.1 and frames Anthropic’s pricing strategy as an attempt to trade lower prices for stronger market adoption ahead of an IPO. It highlights the gap between having a powerful model and successfully selling it.
钛媒体 · 13d ago
它石智航的 AWE3.7 is presented as a general-purpose model capable of handling tasks across factory and household environments. The article highlights its cross-scenario generalization, though the excerpt does not provide benchmark results or architectural details.
量子位 · 13d ago
The European Commission has sent information requests to OpenAI, Anthropic, Google, and more than 30 other general-purpose AI providers, focusing on systemic risks. The action raises a legal question over whether powerful internal research models should face safety, evaluation, cybersecurity, and incident-reporting obligations before market release.
虎嗅 · 13d ago
A vulnerability in Baseboard Management Controllers (BMCs) reportedly places thousands of servers at risk of hardware-level intrusion. Because BMCs operate beneath the operating system, exploitation could threaten server control and infrastructure security.
InfoQ中国 · 13d ago
Lianshang Group has been selected as one of Shanghai’s first 38 compute ecosystem partners and is listed among 18 compute application providers. The partnership will support its continued development of AI products, multimodal applications, intelligent agents, and scenario-based deployments.
雷峰网 · 13d ago
The article distills lessons from Sumitomo Chemical, LINE Yahoo, and dip on scaling generative AI beyond individual productivity use. Their experiences emphasize embedding AI into company-wide workflows and culture rather than promoting it as a standalone buzzword.
ITmedia AI+ (日本) · 13d ago