让变化经过筛选，
再抵达你。

Anthropic Fellows 发布 TASTE，以 92 组两两比较测试模型能否判断 AI 安全研究提案。团队估算人类研究者的一致率为 77%，表现最佳的 Fable 5 为 60%。这是规模较小的早期基准，置信区间较宽；结果表明模型仍落后于人类判断，而不是证明某个系统具备可靠的研究评审能力。

Anthropic Fellows released TASTE, a benchmark of 92 pairwise comparisons asking models to judge AI-safety research proposals. The team estimates human researcher agreement at 77%, while the best model, Fable 5, reached 60%. This is a small early benchmark with wide confidence intervals; it suggests a gap from human judgment rather than establishing reliable automated peer review.

AWS 为 SageMaker Feature Store 增加 BatchWriteRecord 与 ListRecords。前者单次调用最多可向多个特征组写入 25 条记录，支持部分成功、单条 TTL 与 EventTime 顺序；后者可枚举 Standard 和 In-Memory 层的记录 ID。这是官方产品更新，实际吞吐与成本仍取决于工作负载。

AWS has added BatchWriteRecord and ListRecords to SageMaker Feature Store. One call can write up to 25 records across multiple feature groups, with partial success, per-record TTL, and EventTime ordering; ListRecords enumerates record IDs in the Standard and In-Memory tiers. This is an official product update, while real throughput and cost will depend on the workload.

Decathlon 与 AWS 介绍其 Chronos-2 需求预测流程：每个区域最多覆盖 25,000 件商品，同时生成 12 周和 52 周预测；模型约每六个月微调，预测每周批量运行。这是单一企业与供应商联合案例，说明一种工程路径，不代表其他零售场景会得到相同结果。

Decathlon and AWS describe a Chronos-2 demand-forecasting workflow covering up to 25,000 products per zone with 12- and 52-week horizons. Models are fine-tuned roughly every six months and forecasts run in weekly batches. This is a joint customer-and-vendor case that illustrates one engineering path, not a promise of identical results elsewhere.

AWS 为 CreateInferenceComponent 新增 SchedulingConfig，可用 AvailabilityZoneBalance 在 SPREAD 与 BINPACK 之间选择，帮助团队在高可用与资源密度之间取舍。Salesforce 案例称通过共置模型将基础设施成本降低 8 倍；这一数字来自客户案例，应按其特定架构与流量条件理解。

AWS has added SchedulingConfig to CreateInferenceComponent, using AvailabilityZoneBalance to choose between SPREAD and BINPACK placement for different reliability and density goals. A Salesforce case reports an 8× infrastructure-cost reduction from model co-hosting; that figure belongs to the customer's specific architecture and traffic conditions.

Google Labs 调整 Gemini Notebook 的使用额度：计算相关上限从每日刷新改为每五小时刷新，并会依据提示复杂度、对话长度、来源数量与功能而变化；视频概览和幻灯片等任务可排队延后生成并发送通知。更新计划 9 月 2 日开始推出，属于产品公告。

Google Labs is changing Gemini Notebook's compute limits from daily resets to refreshes every five hours, with consumption varying by prompt complexity, chat length, source count, and feature. Video Overviews and Slide Decks can be deferred and notify users when ready. The product change is scheduled to begin rolling out on September 2.

AWS 发布一套用 AWS Transform custom、GitHub Actions 与 Dependabot 构建持续现代化流水线的技术方案，示例覆盖依赖修复、自动文档、跨仓库变换与知识复用。文章明确示例在非交互模式中使用信任全部工具的参数，并提醒生产环境先审查安全策略；这是技术演示，不是无人监督变更的普遍背书。

AWS has published a technical pattern combining AWS Transform custom, GitHub Actions, and Dependabot for dependency remediation, automatic documentation, cross-repository transformations, and knowledge reuse. The walkthrough explicitly uses a trust-all-tools flag for non-interactive runs and advises reviewing security policy before production use. It is a demonstration, not a general endorsement of unsupervised changes.