让变化经过筛选，
再抵达你。

OpenAI 公布其首款自研推理芯片 Jalapeño 的首批测量结果，并计划于年底前开始部署。公司称，在 InferenceX 的公开模型测试中，该系统同时改善了吞吐、功耗与延迟；1.5–1.9 倍每瓦工作量及 1.7–3.6 倍更低端到端延迟等数字来自 OpenAI 的测试，应视为厂商报告，尚待更广泛的独立复现。

OpenAI has published the first measured results for Jalapeño, its first custom inference chip, and plans to begin deploying it by year-end. On public-model tests using InferenceX, the company reports gains across throughput, power, and latency. Figures such as 1.5–1.9× more work per watt and 1.7–3.6× lower end-to-end latency are OpenAI-reported and await broader independent replication.

OpenAI 推出 Admin 插件，让管理员在对话中查看采用率与额度、管理成员和群组、排查权限，并处理部分用量请求。官方说明称，插件沿用用户现有角色与工作区策略，不扩大访问权限；较大影响的操作仍可在应用前复核。

OpenAI has introduced an Admin plugin for reviewing adoption and credits, managing members and groups, diagnosing access, and handling supported usage requests from ChatGPT Work or Codex. OpenAI says it preserves existing roles and workspace policies rather than granting wider access, with review available for broader-impact actions.

AWS 宣布 Amazon OpenSearch Service 支持 MCP Apps。工具调用可在 AI 助手内返回交互式追踪瀑布图、服务拓扑和日志模式，便于工程师核对智能体给出的根因判断；数据仍保留在客户自己的 AWS 账户中。

AWS says Amazon OpenSearch Service now supports MCP Apps, allowing tool calls to return interactive trace waterfalls, service maps, and log-pattern views inside an AI assistant. The goal is to make an agent's root-cause hypothesis easier to verify while keeping customer data in the customer's AWS account.

AWS 发布一套技术方案，把 Amazon Quick Desktop 与 FSx for NetApp ONTAP、S3 访问点和 Slack 组合，用于生成可追溯来源、遵循 IAM 权限的周报。文中“从数小时缩短到数分钟”属于方案方描述，实际收益取决于文件结构、权限设计与工作流。

AWS published a technical pattern combining Amazon Quick Desktop with FSx for NetApp ONTAP, S3 access points, and Slack to produce source-traceable reports under IAM controls. The claim that weekly preparation can fall from hours to minutes is solution-reported; real gains depend on document structure, permissions, and workflow design.

Mary Kay 与 AWS 介绍了一套由 AWS DevOps Agent、Amazon Bedrock、Strands SDK 和 AgentCore 组成的运维流程：系统调查事件、分类路由，再由专用智能体提出或执行受控修复。案例称部分事件可在两分钟内形成并合并拉取请求、单次成本约 0.16–0.50 美元；这些是单一企业案例数据，不代表普遍结果。

Mary Kay and AWS describe an operations pipeline built with AWS DevOps Agent, Amazon Bedrock, the Strands SDK, and AgentCore: events are investigated and routed before specialized agents propose or execute controlled fixes. The case study reports some pull requests reaching merge in under two minutes at roughly $0.16–$0.50 per action. These are single-company case results, not general benchmarks.