让变化经过筛选，
再抵达你。

Google 发布面向实时语音交互的 Gemini 3.5 Transcribe，可把原始音频直接整理为带格式的文字，并针对噪声、专业术语与口语停顿进行处理。Google 称这是其目前最精确的语音转文字模型；该表述与相关质量结果来自厂商评测，仍需结合语言、口音和场景独立验证。

Google has introduced Gemini 3.5 Transcribe for real-time voice interactions. It converts raw audio into formatted text while handling noise, specialist vocabulary, and spoken disfluencies. Google calls it its most precise speech-to-text model to date; that description and the supporting quality results are vendor evaluations that still need independent testing across languages, accents, and settings.

OpenAI 发布一份关于学生和教育者如何在课堂之外使用 ChatGPT 的报告。其隐私保护分析称，各年龄段用户每周约有 7,000 万次对话用于检验知识；美国学年期间，与课业相关的消息每周峰值超过 4.6 亿条。这些统计由 OpenAI 基于自身平台数据得出，反映使用规模而非学习成效。

OpenAI has released a report on how students and educators use ChatGPT beyond the classroom. Its privacy-preserving analysis reports roughly 70 million weekly conversations across age groups devoted to testing knowledge, while U.S. classwork-related messages peak above 460 million per week during the school year. These are OpenAI platform statistics and measure usage, not learning outcomes.

在线旅游公司 loveholidays 介绍了将 Codex 扩展到数据、产品与其他业务团队的做法。案例称一年内 AI 辅助代码变更占比从 7% 升至 79%，部署频率提高 73%，同时工程团队规模未增加。这些数字来自客户与供应商联合案例，不能视为普遍基准。

Online travel company loveholidays describes expanding Codex across data, product, and other business teams. The case study reports AI-assisted code changes rising from 7% to 79% in a year and deployment frequency increasing 73% without engineering-team growth. These are customer-and-vendor case figures, not general benchmarks.

Hugging Face 发布使用 Sentence Transformers 训练与微调多向量嵌入模型的实践指南，涵盖数据格式、损失函数、评估器与索引优化。作者还在医学问答检索数据上比较了六个起点；这些结果是单一配方和数据集下的实测，不代表所有检索任务。

Hugging Face has published a practical guide to training and fine-tuning multi-vector embedding models with Sentence Transformers, covering data formats, losses, evaluators, and index optimization. The author also compares six starting points on a medical question-retrieval setup; those measurements belong to one recipe and dataset, not every retrieval task.

AWS 发布监督微调数据准备系列的第一篇，区分持续预训练、监督微调与强化微调，并给出格式一致性、质量筛选和训练/评估拆分建议。这是一篇以 Amazon Nova 与 Bedrock 文档为例的技术指南，而非新模型发布或效果保证。

AWS has published the first part of a data-preparation series for supervised fine-tuning. It distinguishes continued pre-training, supervised fine-tuning, and reinforcement fine-tuning, then covers format consistency, quality filtering, and train/evaluation splits. This is technical guidance illustrated with Amazon Nova and Bedrock documentation, not a model launch or a performance guarantee.

AWS 发布一套跨账户架构，让一个账户中的 AgentCore 智能体访问另一账户里由 Amazon Redshift Serverless 支撑的 Bedrock Knowledge Base，而无需复制源数据。文章重点是权限边界、请求流程和两种通用编排模型的选择，属于可复用的技术方案。

AWS has published a cross-account architecture in which an AgentCore agent accesses a Bedrock Knowledge Base backed by Amazon Redshift Serverless in another account without copying source data. The article focuses on permission boundaries, request flow, and choosing between two generally available orchestration models; it is a reusable technical pattern.