Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 88 of 1373
Changsha is using real urban and production environments as testbeds for embodied AI robots. The approach emphasizes learning from real-world failures on production lines rather than relying solely on simulation.
钛媒体 · 23d ago
Alibaba’s Qwen team plans to open-source Qwen3.8-Flash-Next and an FP8 version on Aug. 26.
TechNode · 23d ago
Huawei has proposed building AI data centers in Egypt for military, surveillance, and other public-sector applications. The effort highlights Huawei’s ambitions to expand its AI chip business globally and has raised concerns in Washington.
Bloomberg Technology · 23d ago
A Tesla China insider said reports that the company emptied its Shanghai FSD data center and withdrew the related team are untrue. Public information indicates that Tesla continues work on data compliance and local adaptation for FSD in China.
TechNode · 23d ago
Huawei Noah’s Ark Lab introduced RoboHarness, an orchestration system that lets VLA, WAM, reinforcement-learning policies, and TAMP planners collaborate on complex robotic tasks. Without retraining the underlying policies, a Coding Agent decomposes tasks, routes subtasks, and dynamically switches strategies to support long-horizon execution, including zero-shot task completion.
雷峰网 · 23d ago
Agnes Video 2.5 Flash, described as a top-tier video AI tool, is now available for free. The article highlights its newly accessible capabilities through a polished recreation of the Chinese song-related work 《牛來》.
量子位 · 23d ago
A mysterious embodied AI model is demonstrated controlling or supporting robots associated with Unitree and Zhiyuan Robotics. The showcase reportedly runs for 10 minutes in a single continuous take, but the model’s name and technical specifications are not disclosed.
量子位 · 23d ago
OJO, founded by former Jianying head Zhang Qizhi, has secured a nearly 100 million yuan angel round. Backed by Shunwei Capital and Lenovo Capital, the company is building a collaborative workspace for design Agent teams.
钛媒体 · 23d ago
This analysis examines the grassroots segment of the robotics industry and how smaller players may bring robots closer to ordinary consumers' budgets. It highlights the role of entrepreneurs and companies across diverse industries in making embodied AI more accessible.
钛媒体 · 23d ago
The article explores AI experiments in senior-care communities. It argues that deploying AI in real industrial settings requires deep integration with business workflows, domain professionals, and organizational systems.
钛媒体 · 23d ago
DeepSeek is reportedly nearing a pre-IPO funding round valuing the company at approximately 500 billion yuan, or US$74 billion, before investment. The company is seeking about 50 billion yuan and may pursue a listing on Shanghai’s Star Market, potentially in 2027.
SCMP Technology · 23d ago
Leju Robotics has made 13 external investments across humanoid-robot components, embodied-AI software, data, and application scenarios as it prepares for an IPO. The strategy aims to reduce hardware costs, secure alternative AI capabilities, and turn Kuavo from a standalone robot into an integrated solution, though losses and uncertain investment synergies remain significant risks.
虎嗅 · 23d ago
Sam Altman says he underestimated how slowly companies and people would change their workflows after GPT-4, despite rapid improvements in model capability. He describes OpenAI’s preferred direction as a platform with a unified AI entry point and APIs for developers, while emphasizing compute infrastructure, practical deployment, and adaptive safety practices.
虎嗅 · 23d ago
The article argues that major technology companies are racing to control chips, data centers, models, agents, applications, and devices, but this breadth may create coordination and resource-allocation problems. OpenAI's platform-oriented stance and Google's AI reorganization are presented as signs that focused control points may be more sustainable than doing everything in-house.
虎嗅 · 23d ago
China’s Ministry of Industry and Information Technology says the new energy vehicle industry grew from a million-vehicle scale to more than ten million during the 14th Five-Year Plan. The next plan will emphasize vehicle quality, autonomous-driving safety, market access pilots, and vehicle-road-cloud integration.
IT之家 · 23d ago
Apple提前发布搭载M6、M5 Ultra的新款Mac mini与Mac Studio，借此展示2纳米芯片、多晶粒封装和本地AI算力路线。Mac mini被重新定位为全天候智能体计算工具，而高内存配置与Mac集群则瞄准开发者、小型团队及隐私敏感企业。
虎嗅 · 23d ago
A Confluent survey found that only 17% of Japanese companies are running autonomous AI such as AI agents in production, the lowest level among surveyed markets. The report examines why deployments remain stuck at the proof-of-concept stage and how companies can move toward operational adoption.
ITmedia AI+ (日本) · 23d ago
TRACE is a transition-aware residual control framework for multi-objective materials discovery with LLM agents. By learning from parent-edit-child transitions and property changes, it raises macro-average hit rate from 18.13% with LLEMA to 25.96%.
ArXiv AI · 23d ago
Okta Japan reported that approximately one in two companies uses multiple AI tools rather than relying on a single service such as ChatGPT. The report analyzes four years of anonymized access data from more than 20,000 organizations using Okta Platform.
ITmedia AI+ (日本) · 23d ago
STEP-KTODER is a code preference optimization framework that treats module-level functions as learning steps and labels them with automatically generated unit tests. Experiments across HumanEval(+), MBPP(+), BigCodeBench, and LiveCodeBench show gains over outcome-only KTO and DPO.
ArXiv AI · 23d ago
RENDER is a benchmark control that evaluates how the reader-facing format of memory and RAG inputs affects LLM answer quality. Across 500 LongMemEval questions and nine models, resolved memory packets substantially outperformed recency-truncated raw dialogue, while deployed-style templates produced large model-specific performance gaps.
ArXiv AI · 23d ago
MolEmb adapts multimodal large language models into general molecular embedding models conditioned on both molecular profiles and natural-language context. The framework performs competitively on molecular property prediction and enables cross-modal molecule–text retrieval.
ArXiv AI · 23d ago
A case-study audit found that 354 of 366 LLM-generated autobiographical entries failed scene-level verification against an independent record, a 96.7% failure rate. Grounding generation in the subject’s corpus improved results but still left an 83.3% verification-failure rate.
ArXiv AI · 23d ago
Researchers introduce a multi-agent framework that lets LLM agents design and run controlled experiments with scientific simulation models for pharmaceutical process optimization. By comparing simulated interventions and interpreting outcomes, the system produces more specific, actionable, and user-rated helpful recommendations than language-only reasoning.
ArXiv AI · 23d ago
FLARE is an uncertainty-aware framework for evaluating whether AI adoption in healthcare is financially and operationally worthwhile. A stroke-care case study found a break-even point of about 3,992 patients annually, with positive first-year ROI at roughly 5,000 patients.
ArXiv AI · 23d ago
ESQ-Bench is a new Oracle-first benchmark designed to test NL2SQL dialect generalization and silent semantic divergence in realistic enterprise schemas. Across three complexity tiers, even execution-passing queries frequently produced incorrect results, while Claude Sonnet 4.6 outperformed GPT-4o and local Llama 3.2 by a wide margin.
ArXiv AI · 23d ago
A study finds that AION-1 relies heavily on survey segmentation maps rather than image pixels, causing large shifts in predicted flux, size, ellipticity, and redshift when only metadata is changed. The inherited detection incompleteness can bias tomographic mean redshifts beyond LSST DESC requirements, while removing the detection channel eliminates the effect without measurable performance loss.
ArXiv AI · 23d ago
This arXiv study finds that measured AI preferences vary substantially depending on the prompt instrument used to elicit them. Across eight models, 15 welfare-related outcomes, and five instruments, rankings generalized weakly, suggesting that one preference measurement provides limited evidence about what another instrument would report.
ArXiv AI · 23d ago
This arXiv position paper argues that increasingly autonomous AI agents can undermine effective human oversight rather than simply benefiting from it. It proposes designing agent workflows and organizational protocols that preserve critical judgment and counteract skill atrophy caused by prolonged automation use.
ArXiv AI · 23d ago
This arXiv paper proposes a framework for determining when LLM-assisted scientific research remains epistemically legitimate and accountable. It argues that human verification and ownership matter more than the degree of machine involvement, and introduces an "epistemic audit" to document delegation, provenance, verification, and responsibility.
ArXiv AI · 23d ago