Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 234 of 1375
Nvidia is reportedly seeking a Chinese base-station maker to combine its AI computing platform with wireless network equipment for 6G systems targeting overseas markets. Sources say Nvidia and Shenzhen-based Jiaxian Communications have been in contact since 2025, but the arrangement remains unconfirmed.
TechNode · 51d ago
Huanfang Quant's nine displayed products all fell more than 20% in July, with the worst product dropping 22.15%. The article attributes the broader quantitative-fund drawdown to market concentration, synchronized factor losses, strategy homogeneity, high exposure, and AI models encountering out-of-sample conditions.
虎嗅 · 51d ago
Researchers found AI-generated child sexual abuse imagery in Meta's ad library, including content posted after the company had been warned. The findings point to continuing weaknesses in Meta's systems for preventing and removing abusive AI-generated content.
Digital Trends · 51d ago
Baidu has reportedly started integrating its internal office agent dodo into Baidu Dazi, transferring dodo's research staff and resources to the Dazi team. dodo supports meeting summaries, document processing, email search, knowledge-base access, and code and requirements management.
TechNode · 51d ago
ByteDance founder Zhang Yiming reportedly told the Seed AI research team that it would not rely on AI distillation to improve its models. The team is willing to temporarily trail competitors rather than use outputs from larger models as a primary training shortcut.
TechNode · 51d ago
Huawei has launched the 798-gram MateBook Pro S, powered by the Kirin XE90 PC processor and running HarmonyOS. The 14.2-inch flexible OLED laptop starts at RMB7,999, with premium versions priced up to RMB14,999.
TechNode · 51d ago
MiniMax H3 is reported to have reached the top position in the open-source AI community. The headline presents it as setting a new performance threshold—or “kill line”—for video models, though the excerpt provides no benchmark details.
量子位 · 51d ago
Envision’s Wulanchabu Xinghe Base has begun operations as a major AI computing hub. The facility is described as the world’s largest AI compute supercenter, marking a significant expansion of large-scale AI infrastructure.
量子位 · 51d ago
DeepSeek announced that it plans to significantly increase prices for its API services in the near future. The announcement follows surging global demand for its low-cost AI models, though specific pricing details have not yet been released.
SCMP Technology · 51d ago
Siemens raised its earnings expectations for the second time this year. The company cited stronger data center spending and higher returns from industrial software sales as key drivers.
Bloomberg Technology · 51d ago
The report examines metal raw-material industries amid the expansion of AI infrastructure. It argues that global resource competition and strategic balancing are turning AI metals into a new focus of great-power competition.
钛媒体 · 51d ago
NVIDIA has begun offering its Alpamayo family of autonomous-driving AI models under an open license that permits commercial use. The new Alpamayo 2 Super model claims leading performance on inference benchmarks.
ITmedia AI+ (日本) · 51d ago
PI announced 2200V PowiGaN, raising the commercial voltage ceiling for GaN from 1700V to 2200V. The article argues this is the commercialization of a limit already demonstrated in a 2024 PCIM paper, signaling that GaN is entering applications traditionally dominated by SiC.
钛媒体 · 51d ago
Microsoft disclosed that OpenAI generated $24.1 billion of revenue in its latest fiscal year, representing more than half—and potentially nearly 70%—of its AI business revenue. The disclosure highlights Microsoft’s continued financial dependence on OpenAI despite efforts to diversify its AI partnerships and models.
IT之家 · 51d ago
Recruit has released its 2026 training materials for new graduate engineers for free. The 13 materials cover how engineers can use AI, how AI is changing software development, and knowledge needed for career development.
ITmedia AI+ (日本) · 51d ago
Cloudflare has launched Cloudflare OS, an open-source AI agent workspace that runs in the browser. It includes shareable, customizable micro-apps designed to give enterprise employees a secure and tailored environment for AI-assisted collaboration.
cnBeta (Full RSS) · 51d ago
DeepSeek announced plans to significantly raise API prices, with the final pricing scheme still pending. The analysis argues that the increase reflects commercial sustainability and may bring DeepSeek closer to other advanced domestic models in price-performance.
虎嗅 · 51d ago
A Chinese automaker’s highly automated factory is retraining production workers in data systems, AI models, digital twins, cloud platforms, and OpenClaw. The story shows how smart-factory investment is reducing traditional operator roles while shifting pressure onto workers who lack digital skills.
虎嗅 · 51d ago
Meta’s Ray-Ban smartglasses are reportedly gaining widespread adoption, with more than 7 million pairs sold in 2025. Supporters praise their photography and accessibility benefits, while critics worry about nonconsensual filming, privacy and data collection.
The Guardian Technology · 51d ago
This arXiv paper introduces a self-verifying instrument for long-horizon agents, where a deterministic Executive controls beliefs and LLMs can only submit typed proposals. Experiments show that removing the commitment mechanism causes goal abandonment without increasing binding error, while the system’s practical task efficacy remains zero on ARC-AGI-3.
ArXiv AI · 51d ago
SafeCommit is a risk-control layer that determines whether memory-grounded agents have enough evidence to safely execute actions with external side effects. It uses calibrated plausible-world sets, conformal action certificates, low-impact probes, and conservative fallbacks to manage stale, conflicting, incomplete, or corrupted memory.
ArXiv AI · 51d ago
Researchers introduce a domain-knowledge-free neurosymbolic method for fusing imperfect ViT-based perception models under distribution shifts and coordinated label-flipping attacks. Using Label Vector Pools and exact or heuristic abductive inference, the system achieves 0.42 F1 at a 90% flip rate, outperforming MV-Plurality's 0.35.
ArXiv AI · 51d ago
The paper introduces RAIL—Reasoning, Assurances, Interfacing and Learning—as a unified framework for designing reliable neurosymbolic AI systems. It applies the principles to areas including physics-aware machine learning, Alpha-style neuro-guided search, causal learning and tool-augmented LLMs.
ArXiv AI · 51d ago
NeuMoSync is a neuromodulatory architecture designed to reduce plasticity loss and improve knowledge transfer in continual learning. It uses neuron-specific signals informed by current inputs and network history, outperforming existing methods across several continual-learning benchmarks.
ArXiv AI · 51d ago
This arXiv study introduces an active-inference foraging agent that dynamically reallocates limited interoceptive precision toward its most urgent bodily need. In the AffectWorld environment, the approach more than doubled learning-phase survival compared with uniform precision, while also accelerating learning for the attended channel.
ArXiv AI · 51d ago
MCTS-Report uses Monte Carlo Tree Search to progressively build reports from tabular data, jointly optimizing factual accuracy, visual quality, and narrative coherence. On the new MMRBench benchmark, it achieves an overall score of 77.9 and outperforms strong baselines.
ArXiv AI · 51d ago
MatrAIx is a population-scale simulated-user evaluation infrastructure for testing AI systems and digital products with heterogeneous persona agents. It includes 8.3 billion persona records, four interaction environments, 1,010 application tasks, and validation showing 91.5% behavioral adherence in controlled trials.
ArXiv AI · 51d ago
FinProBench introduces Role-Grounded Rubric Construction (RGRC), a method that builds evaluation criteria from deliverables created by professionals in the same role. Across 57 occupations, RGRC delivered its largest gains in specialized roles, achieving 99.1% versus 78.0% for prompt-only rubrics.
ArXiv AI · 51d ago
FinPerMA is a theory-informed benchmark for testing whether LLM agents update personalized user models after significant events. Across 2,994 questions and 276 personas, seven frontier models and multiple memory configurations showed low accuracy, with simple retrieval sometimes outperforming specialized memory systems.
ArXiv AI · 51d ago
BrainBench is a unified benchmark for evaluating whether large language models can follow instructions, analyze EEG and physiological signals, and produce scientifically grounded reports. It covers four assessment areas across 17 datasets and evaluates models through more than 100,000 executions using CodeAct and BrainAgent.
ArXiv AI · 51d ago