让变化经过筛选，
再抵达你。

OpenAI 称，ChatGPT Ads 上线不到 200 天，年化收入运行率达到 10 亿美元，已有数万家广告主使用；Ads Manager 的自助购买范围扩展至印度、欧洲、中东和北非。公司同时表示广告会明确标注、与回答分离，广告主无法访问私人对话。收入、用户与体验指标均为 OpenAI 自行报告。

OpenAI says ChatGPT Ads reached a $1 billion annualized revenue run rate less than 200 days after launch and is used by tens of thousands of advertisers. Self-service Ads Manager buying is expanding across India, Europe, the Middle East, and North Africa. The company says ads are labeled and separated from answers and that advertisers cannot access private conversations. Revenue, usage, and experience figures are OpenAI-reported.

社区项目 VLANeXt 发布面向机器人研究的代码与实验总结，从专用策略模块、动作分块、连续动作建模、多视角输入到世界模型，逐项比较设计选择，并在 LIBERO、LIBERO-plus 与真实机器人实验中测试。它提供可复现实验起点，但仍是作者团队发布的研究项目，基准与实机结果不能自动推广到其他机器人和环境。

The community project VLANeXt publishes code and an experimental recipe for robotics research, comparing dedicated policy modules, action chunking, continuous action modeling, multi-view inputs, and world models across LIBERO, LIBERO-plus, and real-robot experiments. It offers a reproducible starting point, but remains an author-published research project; benchmark and hardware results do not automatically transfer to other robots or environments.

一篇社区实践文章提出：智能体让改写段落、重做图表和交互解释变得便宜，真正稀缺的是决定什么值得关注、证据如何呈现，以及对生成文本逐项核对。作者建议把最强结果前置、围绕视觉组织文章，并像审查代码一样追踪逻辑、术语、数字和引用。这是个人工作方法，不是受控研究结论。

A community practice essay argues that agents make rewriting prose, rebuilding charts, and trying interactive explanations cheap; the scarce work is deciding what deserves attention, how evidence should appear, and how generated text is checked. It recommends leading with the strongest result, writing around visuals, and reviewing logic, terms, numbers, and citations as carefully as code. This is a personal workflow, not a controlled research finding.

《Nature Human Behaviour》评论指出，AI 会加速数字化、可扩展且往往集中于特定人群的行为科学研究，但关于人类行为的证据并不总能规模化。作者主张把快速自动化研究与真实环境中的慢速研究连接起来。文章属于同行评议期刊评论，提出的是方法论判断，而非新的实验效果。

A Nature Human Behaviour commentary says AI will accelerate behavioral research that is digital, scalable, and often concentrated in particular populations, while evidence about human behavior cannot always be scaled. The authors call for linking rapid automated work with slower research in real settings. This is a journal commentary making a methodological argument, not reporting a new experimental effect.

第五届 ECMWF—ESA 机器学习研讨会报告汇总 57 场专家报告与 88 份海报，覆盖地球数字孪生、物理—机器学习混合系统、地球观测、端到端数据同化与预测，以及高性能计算。报告强调 AI 应成为物理建模与数据同化的补充层，而不是替代它们；这是研讨会共识与趋势总结，不是单一系统的性能验证。

A report from the fifth ECMWF–ESA machine-learning workshop synthesizes 57 expert talks and 88 posters across Earth digital twins, hybrid physics–ML systems, Earth observation, end-to-end assimilation and prediction, and high-performance computing. It frames AI as an added layer for physical modeling and data assimilation rather than a replacement. This is a workshop synthesis and trend report, not a performance validation of one system.

系统综述与网络荟萃分析纳入 53 项研究、超过 700 万次患者入院记录；表现最佳的机器学习模型合并 AUC 为 0.88，敏感度 77.2%、特异度 84.7%，但阳性预测值只有 34.2%。作者还报告极高异质性、较高偏倚风险，并指出现实患者结局是否改善仍未证实，广泛采用前需要标准化外部验证和前瞻性试验。

A systematic review and network meta-analysis covers 53 studies and more than seven million patient admissions. The pooled AUC for the best-performing ML models was 0.88, with 77.2% sensitivity and 84.7% specificity, but positive predictive value was only 34.2%. The authors also report extreme heterogeneity and high risk of bias, with improved real-world outcomes still unproven; standardized external validation and prospective trials are needed before broad adoption.