TUJLTECH UPDATES, JUST LATEST

DAILY AI BRIEFING · 每日 AI 简报

让变化经过筛选,
再抵达你。

科技更新,只看最新。每天从可靠来源中挑选值得留下的 AI 动向;中英对照,链接回到原文,事实与判断分开。

READ THIS EDITION ↓

THE SIGNAL / 今日信号

五件值得知道的事

5 stories · 约 7 分钟

01

IBM 发布约 3.85 亿参数的 Granite 时间序列模型 r2

IBM releases the roughly 385-million-parameter Granite time-series r2 model

Granite Time Series PatchTST-FM-r2 支持零样本预测、概率预测与缺失值填补,并公开权重、推理流程及复现代码。IBM 报告其在 9 月 8 日的 GIFT-Eval 可复现零样本比较中排名第二,并强调商业友好的许可。排名来自团队所列筛选条件与基准,不能直接外推到任意业务数据。

Granite Time Series PatchTST-FM-r2 supports zero-shot and probabilistic forecasting plus missing-value imputation, with weights, inference pipelines, and reproduction code released. IBM reports second place in its September 8 comparison of replicable zero-shot GIFT-Eval models and highlights commercially friendly licensing. The ranking depends on the stated filters and benchmark; it does not establish performance on arbitrary business data.

阅读原始来源 / Read original source
02

Acrobat 新增文档转音频、交互报告与企业知识库能力

Acrobat adds document audio, interactive reports, and enterprise knowledge tools

Adobe 宣布 Productivity Agent 的新功能,可把文档转成音频摘要、播客、可视化报告和演示文稿;企业功能包括 Knowledge Base 与 Analyzer,用于文档集合问答和结构化信息提取。可用功能取决于订阅方案。Adobe 表示回答带可点击引用,读者仍应回到原文核对数字、限定条件和遗漏。

Adobe announced Productivity Agent features for audio summaries, podcasts, visual reports, and presentations. Enterprise capabilities include Knowledge Base and Analyzer for collection-wide questions and structured extraction. Availability depends on the subscription plan. Adobe says answers include clickable citations; readers should still check original documents for figures, qualifications, and omissions.

阅读原始来源 / Read original source
03

Alteryx 将已有分析流程接入对话智能体与 MCP

Alteryx connects existing analytics workflows to conversational agents and MCP

Alteryx 公告介绍 Ask Alteryx、Agent Studio、面向 OpenAI 的 Insights 和 MCP Server,让外部智能体调用已有数据、计算与流程,并继承现有权限。公告中的 token 和速度改善来自客户案例及内部测试,不能视为普遍保证;部分后续连接和调度能力仍在计划中。

Alteryx described Ask Alteryx, Agent Studio, Insights for OpenAI, and an MCP Server that lets external agents use existing data, calculations, and workflows under existing permissions. Token and speed improvements in the announcement come from customer examples and internal tests, not universal guarantees. Some expanded connection and scheduling capabilities remain planned.

阅读原始来源 / Read original source
04

Red Hat AI 3.5 正式发布,增加共享 GPU 管理与运行可见性

Red Hat AI 3.5 reaches general availability with shared-GPU controls and observability

Red Hat 宣布 AI 3.5 正式可用,整合多租户资源管理、使用情况观察和智能体开发工具。公告区分功能成熟度:部分 Kubernetes 推理支持已正式可用,Amazon EKS 支持属于技术预览,Kubeflow Spark Operator 属于开发者预览。平台提供的控制能力仍需按实际部署配置和验证。

Red Hat announced general availability of AI 3.5, combining multi-tenant resource management, usage visibility, and agent development tools. The release distinguishes maturity levels: some Kubernetes serving support is generally available, Amazon EKS support is a technology preview, and the Kubeflow Spark Operator is a developer preview. Controls still require configuration and validation for the actual deployment.

阅读原始来源 / Read original source
05

一项阿拉伯语评测复盘强调:分数也受完整调用流程影响

An Arabic evaluation study highlights how the full serving pipeline shapes scores

Omer Nacar 发布对 HUMAIN M3 预览接口的评测复盘,说明提示格式、输出预算、截断处理、响应解析和计分分母如何影响观察结果。实验在 9 月 5 日进行,文章在 9 月 9 日发表。这是作者研究,未确认同行评议;仅覆盖所选子集,不能当作整个模型或全部阿拉伯语能力的最终评价。

Omer Nacar published an evaluation retrospective on the HUMAIN M3 preview endpoint, examining prompt formatting, output budgets, truncation handling, response parsing, and scoring denominators. Experiments took place on September 5; the article appeared on September 9. This is author-reported research with peer-review status unconfirmed, covering selected subsets rather than a final judgment of the model or all Arabic capabilities.

阅读原始来源 / Read original source

EDITOR'S NOTE · 编者的话

今天的共同线索,是让结果带着可以检查的来处。

本期收录 9 月 9 日发布的模型、文档工具、分析平台和评测文章。事实是这些资源与产品公告已经公开;我们的判断是,来源引用、可重复的计算和运行记录,正在成为 AI 进入日常工作的基础。排行榜名次、节省成本和生产可用性各有适用条件,不能只凭一个数字替代具体任务的验证。

This edition covers models, document tools, analytics platforms, and an evaluation article published on September 9. Their publication is fact; our judgment is that citations, repeatable calculations, and operational records are becoming foundations for everyday AI use. Rankings, cost savings, and production readiness each have conditions and cannot replace testing on the actual task.

TWO ADDRESSES · 两个地址

一个保存我们,
一个观察世界。

RUJF.AIRecord Us Just Forever

保存属于我们、记忆与创作的东西。

TUJL.COMTech Updates, Just Latest

观察外面的世界,记录 AI 此刻正在发生什么。

站外推广狗狗加速