Satya Nadella@satyanadella · 7小时前64

The future of the firm is a learning loop in which human capital and token capital compound. With our new Frontier Co., our ambition is to help every enterprise build its own AI capability, and to help create a frontier ecosystem where every organization can turn its knowledge, workflows, and judgment into its own AI systems that continuously improve. https://blogs.microsoft.com/blog/2026/07/02/microsoft-frontier-company-ai-engineering-that-amplifies-and-protects-your-intelligence/

译企业的未来是一个学习循环，人力资本与token资本不断累积。 借助我们的新Frontier Co.，我们的目标是帮助每个企业构建自己的AI能力，并帮助创建一个前沿生态，使每个组织都能将其知识、工作流程和判断力转化为自己的AI系统，并持续改进。https://blogs.microsoft.com/blog/2026/07/02/microsoft-frontier-company-ai-engineering-that-amplifies-and-protects-your-intelligence/

Microsoft Research@MSFTResearch · 2天前39

AI agents often fail because their instructions, or skills, are manually modified with no guarantee of improvement. Learn how SkillOpt turns skill editing into a training process, making agent behavior more reliable without changing model weights: https://msft.it/6012vsvEs

译AI 智能体常常失败，因为它们的指令（即技能）被手动修改，且无法保证改进。了解 SkillOpt 如何将技能编辑转变为训练过程，在不改变模型权重的情况下使智能体行为更可靠：https://msft.it/6012vsvEs

Rohan Paul@rohanpaul_ai · 2天前40

Satya Nadella just talked about a new AI Dev Box (RTX Spark Dev Box) with 20 CPU cores, 128GB unified memory, and petaflop of local compute for trillion-parameter-scale model runs. ---- From "Stanford Online" YouTube channel, (link in comment)

译Satya Nadella 刚刚谈到了一款新的 AI Dev Box（RTX Spark Dev Box），配备 20 个 CPU 核心、128GB 统一内存和千万亿次本地计算能力，用于运行万亿参数规模的模型。

Microsoft Research@MSFTResearch · 3天前46

AI agents can't remember past conversations. They must constantly reload or retrieve context, which grows less efficient as tasks get longer and more complex. Memora solves this with a scalable memory system separating what’s stored from how it's retrieved: https://msft.it/6018vs3gC

译AI智能体无法记住过去的对话。它们必须不断加载或检索上下文，随着任务变得更长更复杂，效率越来越低。 Memora通过一个可扩展的记忆系统解决了这个问题，该系统将存储的内容与检索方式分离开来：https://msft.it/6018vs3gC

ClaudeDevs@ClaudeDevs · 3天前53

You can now run Claude models in Microsoft Foundry, hosted on Azure. Claude Opus 4.8 and Claude Haiku 4.5 are available through the Messages API with capabilities like prompt caching, thinking, and more. https://x.com/claudeai/status/2071653958905467027?s=20

译你现在可以在 Microsoft Foundry（托管于 Azure）上运行 Claude 模型。 Claude Opus 4.8 和 Claude Haiku 4.5 通过 Messages API 提供，支持 prompt caching、thinking 等功能。

Claude@claudeai · 3天前55

Claude in Microsoft Foundry is now generally available, hosted on Azure. Azure customers get Claude Opus 4.8 and Claude Haiku 4.5, with Azure authentication, billing, and commitment retirement.

译Claude 现已在 Microsoft Foundry 中正式可用，托管于 Azure。 Azure 客户可获得 Claude Opus 4.8 和 Claude Haiku 4.5，并支持 Azure 身份验证、计费和承诺预留。

Rohan Paul@rohanpaul_ai · 5天前58

The U.S. AI buildout is running into a harder constraint than GPUs: permission. The Information’s new map finds 300+ state and local data-center bans or moratoriums since 2023, with 275+ passed this year and 75+ still under consideration; resistance is strongest in the Midwest and South, exactly where hyperscalers want cheap land and megawatt-scale power. But the backlash against data centers is outrunning the facts. The most current datacenter is increasingly designed to solve the two biggest fears: water and power bills. Start with water. Microsoft’s next-generation AI data centers use chip-level, closed-loop cooling that consumes zero water for cooling and can avoid more than 125 million liters per year per site. Its fleetwide water-use efficiency has already improved 39% since 2021, to 0.30 liters per kWh. Google reports that 86% of its freshwater withdrawals come from low- or medium-risk sources, and its global data-center fleet runs at a 1.09 PUE versus a 1.56 industry average—meaning far less wasted overhead energy. Electric bills are not automatically shifted to households, either. A recent causal study of U.S. retail rates from 2015–2024 found data centers modestly lowered average rates by spreading fixed grid costs across more electricity sales. And “bring your own power” is already the new trend - e.g. Google’s 500 MW nuclear deal, Microsoft’s 835 MW Three Mile Island agreement, and Meta’s 1,121 MW nuclear contract.

译美国AI基础设施建设最大瓶颈已从GPU变为许可。《The Information》地图显示2023年以来有300多项州级和地方数据中心禁令或暂停，今年通过275项，还有75项在审，中西部和南部抵制最强。但现实数据反超担忧：微软新一代AI数据中心采用芯片级闭环冷却，每站每年避免超1.25亿升水，全舰队用水效率自2021年提升39%至0.30升/kWh；谷歌86%淡水来自低中风险源，全球PUE 1.09（行业均值1.56）。2015–2024年研究表明数据中心通过分摊固定电网成本适度降低了居民平均电价。趋势已转为“自带电源”：谷歌500 MW核电、微软835 MW三哩岛、Meta 1121 MW核电协议。

Microsoft Research@MSFTResearch · 6天前63

What do people actually do with AI at work? A new analysis of five million M365 Copilot conversations has answers. Scott Counts breaks it down in a new video. And dive into the analysis here: https://msft.it/6011vqpbL

译人们在工作场景中实际用 AI 做什么？对五百万次 M365 Copilot 对话的新分析给出了答案。Scott Counts 在一段新视频中进行了讲解。深入了解分析请戳：https://msft.it/6011vqpbL

Microsoft Research@MSFTResearch · 6天前41

Following up with the social copy I’ve drafted: What do people actually do with AI at work? A new analysis of five million M365 Copilot conversations has answers. Scott Counts breaks it down in a new video. And dive into the analysis here: https://msft.it/6015vUHsh

译跟进我起草的社交文案：人们在工作中的 AI 到底用来做什么？一项对五百万次 M365 Copilot 对话的新分析给出了答案。Scott Counts 在一段新视频中进行了详细解读。 点击此处深入了解分析：https://msft.it/6015vUHsh

🚨 AI News | TestingCatalog@testingcatalog · 6天前54

Microsoft announced Skills for Copilot in Excel! > Today, we’re introducing new features built for financial professionals to continue doing that with skills for repeatable workflows, new financial connectors for trusted data, and improved capabilities for traceability. We'll soon hear more announcements about Office integrations from Perplexity and OpenAI as well.

译Microsoft 为 Excel 中的 Copilot 推出 Skills 功能，面向金融专业人士。新功能包括：重复性工作流的技能（Skills）、新的金融数据连接器（financial connectors）以及可追溯性改进。此外，Perplexity 和 OpenAI 也即将发布 Office 集成相关公告。

Rohan Paul@rohanpaul_ai · 7天前60

Microsoft just turned Copilot in Excel into a finance workflow system So it can follow team-defined methods, pull institutional data, and leave an audit trail for every workbook edit. The new connectors bring finance data from FactSet, Morningstar, PitchBook, CB Insights, Daloopa, and S&P/Kensho into Excel, which means analysts can start from licensed market, private-company, filing, research, transcript, and fundamentals data instead of copying numbers by hand. For Finance work its so important for AI to show the source, the formula path, or the exact cells it changed, because closing books, updating forecasts, and building valuation models depend on traceable numbers. Microsoft’s answer is skills, where a team writes a SKILL.md file that tells Copilot how to run repeatable jobs like DCF models, 3-statement models, board packages, variance analysis, and monthly reporting. Personalization and workbook rules make Copilot follow preferred formatting, naming, structure, and formula conventions, so the AI behaves less like a generic assistant and more like a junior analyst trained on the team’s house style. Plan with Copilot forces the model to state which ranges, sheets, formulas, and assumptions it wants to touch before it edits, while Show Changes now labels Copilot’s edits beside human edits.

译微软为 Copilot in Excel 新增 FactSet、Morningstar、PitchBook 等数据连接器，支持团队通过 SKILL.md 定义 DCF 模型、三表模型等可重复任务。Copilot 可遵循团队设定的格式、命名和公式规则。Plan with Copilot 强制模型在编辑前声明操作范围与假设，Show Changes 则标记 AI 与人工编辑。CEO Satya Nadella 称此举可通过 skills 扩展团队专业知识。

Microsoft Research@MSFTResearch · 7天前30

Researchers introduce generative causal testing, which translates black box models into clear hypotheses and verifies them in the scanner, revealing what specific brain regions respond to in language. https://msft.it/6011vUtRd

译研究人员引入了生成式因果测试，它将黑箱模型转化为清晰的假设，并在扫描仪中进行验证，揭示了大脑特定区域对语言的反应。

Satya Nadella@satyanadella · 7天前42

Today we’re bringing skills to Copilot for Excel, giving teams a new way to scale their expertise across every workbook.

译今天，我们为Excel中的Copilot带来了技能，为团队提供了一种在每张工作簿中扩展其专业知识的新方式。

Artificial Analysis@ArtificialAnlys · 7天前68

Microsoft's MAI-Image-2.5 ranks #2 in Text to Image and #3 in Image Editing in the Artificial Analysis Image Arena, behind only OpenAI's image models The latest addition to the MAI Image family is capable of handling both text to image generation and image editing, up to a maximum output resolution of roughly 1MP at flexible aspect ratios with a 32K token context. MAI-Image-2.5 is one of the strongest image models we have tested, ranking #2 in Text to Image behind only OpenAI's GPT Image 2. Its Image Editing capabilities rank at #3 on our leaderboard, ranking just behind OpenAI’s image models, with performance comparable to Google’s Nano Banana 2. MAI-Image-2.5-Flash, the faster, lower-cost variant, lands at #8 in Text to Image and #6 in Image Editing. MAI-Image-2.5 is priced at $48 per 1k images and MAI-Image-2.5-Flash at $20 per 1k images on the Foundry API. Both are also available in the MAI Playground, and MAI-Image-2.5 is available to users in PowerPoint and OneDrive. Congratulations to @MicrosoftAI on the release! See below for comparisons between MAI-Image-2.5 and other leading models in the Artificial Analysis Image Arena 🧵

译微软 MAI-Image-2.5 在 Artificial Analysis Image Arena 排行榜上位列文本到图像生成第2名、图像编辑第3名，仅次于 OpenAI 的图像模型。该模型支持文本到图像生成与图像编辑，最大输出分辨率约 1MP，支持灵活宽高比，上下文窗口 32K token。更快、更低成本的 MAI-Image-2.5-Flash 变体在文本到图像和图像编辑中分别排名第8和第6。API 定价：MAI-Image-2.5 为 $48/千图，Flash 版 $20/千图。模型已上线 Foundry API 和 MAI Playground，其中 MAI-Image-2.5 还可在 PowerPoint 和 OneDrive 中使用。

Microsoft Research@MSFTResearch · 6月24日42

Talos was built to help resolve a major bottleneck in genomic medicine: human review time. The open-source system recovered 90% of in-scope diagnoses while surfacing just 1.3 candidate variants per patient for expert review. https://msft.it/6014vUBwK

译Talos 旨在帮助解决基因组医学中的一个主要瓶颈：人工审查时间。 这个开源系统恢复了 90% 范围内的诊断，同时每个患者仅呈现 1.3 个候选变异供专家审查。

Rohan Paul@rohanpaul_ai · 6月24日46

New Microsoft paper argues that transformers generalize better when they learn compact internal states, not just next tokens. The problem is that normal transformers can look back at every earlier token, so they do not have to squeeze the past into a clean summary. token prediction alone can reward shortcuts that do not become coherent world models. That can work beautifully on familiar data and still fail when the model has to plan, detour, reason, or carry a hidden structure forward. NextLat fixes this by adding a training task where the model must predict its next hidden state, not just the next word. A hidden state is the model’s private summary of what it has seen, so predicting the next one pushes the model to learn how situations change over time. The authors tested this on map-like world modeling, math reasoning, graph planning, story prediction, and regular language modeling. The main result is that NextLat often learned more compact and useful internal states, solved planning tasks better, and sped up generation by up to 3.3x. Overall, it gives transformers some of the useful memory behavior of recurrent models without changing the transformer architecture or slowing normal inference. ---- Link – arxiv. org/abs/2511.05963 Title: "Next-Latent Prediction Transformers Learn Compact World Models"

译微软新论文Next-Latent Prediction (NextLat) 提出一种自监督学习方法，在常规token预测基础上增加预测下一隐藏状态的任务，迫使Transformer学习紧凑的内部世界模型。该方法在地图式世界建模、数学推理、图规划、故事预测等任务上表现更优，生成速度通过自推测解码最高提升3.3x，且无需改变Transformer架构或减慢正常推理。

Microsoft Research@MSFTResearch · 6月24日48

On the Microsoft Research Podcast, Microsoft’s Jaime Teevan, Jenna Butler, Jake Hofman, and Rebecca Janssen explore how looking at AI as more than a tool for automating tasks people are already capable of might open the door to new possibilities. https://msft.it/6011vo8cX

译在微软研究播客上，微软的 Jaime Teevan、Jenna Butler、Jake Hofman 和 Rebecca Janssen 探讨了，将 AI 视为不仅仅是自动化人们已能完成任务的工具，可能会如何打开新的可能性。https://msft.it/6011vo8cX

elvis@omarsar0 · 6月23日65

Microsoft Teams just crossed a line I did not expect this year. It now runs an AI employee that does the work, not one that answers questions. Here is what is new:

译Microsoft Teams 刚刚跨过了一条我今年没预料到的界限。 它现在运行一个 AI 员工，这个员工能完成工作，而不是回答问题的员工。 以下是新内容：

Rohan Paul@rohanpaul_ai · 6月23日67

In his new interview Microsoft CEO Satya Nadella warned that AI power is becoming too concentrated for society to treat it as normal tech progress. His concern is not that AI models are getting smarter, but that the money, chips, data centers, and user access behind them are being controlled by too few firms. When only a handful of companies can train frontier systems, every business using AI becomes dependent on their prices, rules, failures, and product choices. --- firstpost. com/tech/you-cant-call-it-progress-microsoft-ceo-satya-nadella-warns-against-concentration-of-ai-power-14025004.html

译微软CEO萨提亚·纳德拉在最新采访中警告，AI权力正变得过于集中，不能被视为正常的技术进步。他担忧的不是AI模型越来越智能，而是背后的资金、芯片、数据中心和用户访问被少数公司控制。当只有少数企业能够训练前沿模型时，所有使用AI的业务都将依赖于这些公司的定价、规则、故障和产品选择。

Chubby♨️@kimmonismus · 6月21日28

I did a Microsoft Build Vlog. Check it out! :) https://www.youtube.com/watch?v=HxSbcXuPAnI

译我做了一个微软 Build 大会的 Vlog。来看看吧！:) https://www.youtube.com/watch?v=HxSbcXuPAnI

AYi@AYi_AInotes · 6月20日65

微软在干一件很骚的事，一边把 OpenAI 的模型卖给中国公司，一边把中国的 DeepSeek 模型准备卖给西方客户，两头通吃，可以说全球独此一家了。 这件事能跑通，靠的是微软跟 OpenAI 签的特殊合同， 微软拿到了全球自由转售权，OpenAI 管不着， 然后在中国市场，最大的客户是字节跳动，每年在 Azure 和 AI 服务上砸的钱超过 10 亿美元。 最有意思的是合规处理，模型不部署在中国境内，中国客户通过新加坡的数据中心访问。 OpenAI 那边一直催微软加强安全，特别怕中国客户搞模型蒸馏，拿 GPT 的输出训自己的模型。 微软对外说有自动监控，只服务企业客户不对个人开放。 但据彭博社的消息，实际上审查没那么严。 更微妙的是另一头，微软同时在测 DeepSeek 的 R1 和 V4，打算卖给西方客户。 美国模型卖中国，中国模型卖西方，微软自己变成了一个巨大的 AI 转换插头。 这件事背后的信号也很有意思， 中美 AI 在最高层面互相防着，限制出口、限制投资、限制芯片， 但在商业层，套利空间已经大到连微软这种体量的公司也愿意两头下注。 所以地缘归地缘，生意归生意哈哈哈

译微软凭借与OpenAI的特殊合同获得全球自由转售权，将OpenAI模型卖给中国客户（最大客户字节跳动每年在Azure和AI服务上投入超10亿美元），模型通过新加坡数据中心访问，同时监控防蒸馏。另一边，微软正在测试DeepSeek-R1和DeepSeek-V4，准备反向卖给西方客户。这一“双向AI模型贸易网络”凸显中美地缘壁垒下商业套利空间巨大。

AYi@AYi_AInotes · 6月20日75

还得是微软会闷声发大财啊，它现在已经成为全球最大的 AI 中间商，oh no ，是最大的中转站， 不仅把ChatGPT卖给中国企业，也把DeepSeek 反向卖给西方客户😁 以上来自彭博社的最新报道，绝对权威可信，看完让我意外的不仅仅是微软把 GPT 卖给中国那部分，后面那一句：微软同时在测试 DeepSeek-R1 和 DeepSeek-V4，准备把这些中国模型卖给西方客户。 好家伙，左手接 GPT 卖到中国，右手接 DeepSeek 卖到西方，这他么不是一家 AI 公司在卖模型，简直就是一个跨中美 AI 模型的双向贸易网络正在成形呀

译彭博社报道，微软已成为全球最大AI模型中转站，既将ChatGPT卖给中国企业，也反向将DeepSeek模型卖给西方客户。报道称微软正在测试DeepSeek-R1和DeepSeek-V4，计划向西方客户提供这些中国模型。这一模式构建起跨中美AI模型的双向贸易网络。

Rohan Paul@rohanpaul_ai · 6月20日47

New Microsoft + York Univ paper argues that LLMs should not be treated as human-like without clear tests and narrower claims. Many studies ask whether LLMs have things like understanding, empathy, anxiety, or self-awareness, but they often build those ideas into the test from the start. The author shows that, in principle, the old strategy game can implement logic gates, train a tiny perceptron, and serve as a substrate for computation. If the same language model could be rebuilt inside a game, with goats moving around as bits, would we still say it “understands,” “feels anxiety,” or “has empathy” when it produces the same sentence? The point is not that the game is secretly intelligent, but that the same computation can be represented in a very different form. If an LLM-like system were rebuilt inside that game, its answers might stay similar, but people would probably find its “feelings” or “understanding” much less convincing. The authors argue that this shows a big measurement problem: many human-like claims about LLMs may depend on the interface and the observer, not only on the system itself. The paper is not saying LLMs definitely lack human-like attributes, or that all talk of AI cognition is nonsense. It is saying that many experiments smuggle the conclusion into the setup: they assume the model has, or cannot have, a human-like property, then interpret behavior through that assumption. ---- Link – arxiv. org/abs/2605.31514 Title: "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II"

译微软与约克大学新论文指出，许多研究在未经严格测试的情况下就将理解、共情、焦虑等人类属性赋予LLM，往往一开始就把这些概念内嵌到测试设计中。作者论证，原则上老策略游戏《帝国时代II》也能实现逻辑门、训练小型感知机，作为计算基底。若同样的语言模型以山羊移动作为bit在游戏中重建，输出相似句子，人们将不再认为它“理解”或“有共情”。论文并非否定AI认知，而是揭示测量问题：许多关于LLM类人属性的声称依赖于界面和观察者的预设，而不是系统本身。

Berryxia.AI@berryxia · 6月19日32

Grok &Word 插件在Office 里面也可以起飞了😂 不过国内用wps的人感觉比Office多啊，尤其很多人估计还是盗版吧…

译Grok & Word 插件在 Office 里面也可以起飞了😂 不过国内用 wps 的人感觉比 Office 多啊，尤其很多人估计还是盗版吧…

Chubby♨️@kimmonismus · 6月19日69

I'm curious to see if agents like Viktor will increase the enjoyment of Microsoft Teams meetings ;)

译团队协作AI智能体Viktor正式登陆Microsoft Teams。此前已在Slack上线，仅靠单一应用实现2000万美元年化收入运行率（无销售团队、无推广）。Viktor主打零门槛：用户无需学习、无需提示，像@同事一样提及Viktor即可完成任务，甚至无需主动提及，价值自动送达。面向全球3.2亿Microsoft Teams用户，面向大公司一线运营与管理者。新用户获赠100美元启动积分，无需绑定信用卡。

Microsoft Research@MSFTResearch · 6月19日39

Can you remember a new password after listening to five hours of stories? A transformer can. AI researchers Subutai Ahmad and Nicolò Fusi explore the fundamental differences between human memory and machine intelligence with Microsoft CVP Doug Burger. https://msft.it/6016vllRa

译你能在听完五个小时的故事后记住一个新密码吗？ Transformer 可以。 AI 研究员 Subutai Ahmad 和 Nicolò Fusi 与微软 CVP Doug Burger 一起探讨人类记忆与机器智能之间的根本差异。https://msft.it/6016vllRa

🚨 AI News | TestingCatalog@testingcatalog · 6月18日67

EXCLUSIVE 🔥: DeepSeek was just the beginning. Microsoft is evaluating "many" open models for Copilot Cowork. > This is adding internal pressure on MAI teams, as the GLM, MiniMax, and Kimi models are evolving more quickly > Microsoft is aiming to make models "interchangeable" and separate the harness from the underlying models themselves. > As smaller models evolve, it is also possible that some of the tasks will be executed locally in the future.

译微软正评估GLM、MiniMax、Kimi等多款开源模型用于Copilot Cowork，旨在降低推理成本。据Axios报道，微软考虑托管DeepSeek V4作为更便宜选项，同时将Copilot Cowork从无限定价转为按使用量计费。微软称用户每周执行数百任务，成本可能很高。若采用DeepSeek，模型将是可选、经微调和安全防护，并完全托管在Azure上。微软推动模型“可互换”策略，未来部分任务可能移至本地执行。

elvis@omarsar0 · 6月18日68

Microsoft Teams just got its first AI employee. I tested it. A real AI employee that lives in the channel, does the work, and proposes the next move. Not another prompt box. Worth a look. @viktor__com

译Viktor AI智能体正式入驻Microsoft Teams，直接嵌入频道，用户@提及即可获得完成的工作，无需学习或提示。Viktor已在Slack实现2000万美元年化经常性收入，现扩展至拥有3.2亿用户的Teams。新用户可获100美元免费积分，无需绑定信用卡。其目标是以零门槛让每个职场人获得AI价值。

X.PIN@thexpin · 6月18日61

US and Chinese AI are quietly swapping places in the most awkward way possible. Per Bloomberg: Microsoft sells OpenAI services to ByteDance for $1B+/year. Ant Group, Meituan and Tencent are all on Azure. Azure's AI revenue in China grew ~3x in the fiscal year ending June 2025 — faster than anywhere else on earth. Meanwhile, US developers are quietly switching to Chinese models to cut costs. Cursor (the AI coding tool Musk reportedly wants to buy) has used Qwen and Kimi. One hour of coding on Claude: ~$10. Same job on DeepSeek: under 50 cents. Everyone thinks the neighbor's cooking smells better. The Chinese food is cheaper — and they deliver. https://restofworld.org/2026/when-americans-choose-chinese-ai/

译微软向字节跳动出售OpenAI服务（每年超10亿美元），蚂蚁、美团、腾讯使用Azure。截至2025年6月财年，Azure中国AI收入增长约3倍。同时，美国开发者转向中国模型降本：Cursor（马斯克据称想收购的AI编程工具）使用Qwen和Kimi；Claude一小时编码约10美元，DeepSeek不到50美分。

Microsoft Research@MSFTResearch · 6月18日38

ResNet has received the Longuet-Higgins Prize at CVPR 2026, recognizing research with proven, lasting impact. A decade after its publication, residual connections remain foundational to how modern AI systems are built, with over 320,000 citations and growing. https://msft.it/6017vl6bl

译ResNet 在 CVPR 2026 上获得了 Longuet-Higgins 奖，表彰其已被证实、具有持久影响力的研究。发表十年后，残差连接仍然是现代 AI 系统构建的基础，引用量已超过 32 万次且仍在增长。

meng shao@shao__meng · 6月17日66

微软 Copilot Cowork 正式 GA，考虑引入 Azure 托管的 DeepSeek V4 作为低成本模型选项，按算力/用量计费 token maxxing 已经被证实商业模式不可行！ Copilot Cowork 等 Agent 无法再用「包月无限用」的模式卖，因为 Agent 会在一个任务里反复调用模型（读文件、写代码、调工具、自我纠错），token 消耗因此急剧放大；用户每周跑几百个任务时，生产力上去了，账单也会失控。简单任务也被丢给最贵的 frontier 模型，进一步推高成本。 DeepSeek 进入 Copilot 栈？ · 正在测试 微调版 DeepSeek V4，作为 Anthropic / OpenAI 模型的低成本替代 · 预计数周内公布最终选择 · 若落地：可选、非强制，完全托管在 Azure 上，数据不出 Microsoft 云，走现有企业安全/合规/数据驻留体系 · 已做微调，并加入减少偏见等安全层

译微软 Copilot Cowork 正式全球可用，支持多模型。为控制成本，正评估引入微调版 DeepSeek V4 作为 Anthropic/OpenAI 模型的低成本替代，按算力/用量计费。模型完全托管于 Azure，数据不出微软云，已加入安全层，数周内公布。同时指出，Agent 任务反复调用模型致 token 消耗大幅增加，包月无限用模式已不可行。

Satya Nadella@satyanadella · 6月17日59

New Azure milestone. The fastest time to train yet at the largest reported scale for this leading AI training benchmark. A great example of what is possible when we bring together full-stack innovation across silicon, systems, networking, and software, along with our deep partnership with @nvidia, to advance the frontier of AI infra. https://techcommunity.microsoft.com/blog/AzureHighPerformanceComputingBlog/azure-sets-a-new-performance-record-for-llm-training-benchmark-at-extreme-scale/4523077

译新的Azure里程碑。在最大规模报告下，针对这一领先的AI训练基准，实现了最快的训练时间。 这是当我们整合硅片、系统、网络和软件的全栈创新，并与@nvidia深度合作，推进AI基础设施前沿时所能实现的绝佳范例。 https://techcommunity.microsoft.com/blog/AzureHighPerformanceComputingBlog/azure-sets-a-new-performance-record-for-llm-training-benchmark-at-extreme-scale/4523077

DogeDesigner@cb_doge · 6月17日80

Grok is now available inside Microsoft PowerPoint. I created this entire presentation using just one line of text. Grok built the whole deck for me in minutes. Absolutely amazing.

译Grok 现在可在 Microsoft PowerPoint 中使用。 我仅用一行文本就创建了整份演示文稿。Grok 在几分钟内为我构建了完整的幻灯片。 绝对令人惊叹。

Rohan Paul@rohanpaul_ai · 6月17日48

Microsoft’s 2026 Work Trend Index shows Singapore workforce ahead on AI adoption. 66% of Singapore AI users say they now produce work they could not have made a year ago, compared with 58% globally. Among Frontier Professionals, Microsoft’s term for advanced AI users, that rises to 82%

译微软的2026年工作趋势指数显示新加坡劳动力在AI采用方面领先。 66%的新加坡AI用户表示他们现在能产出一年前无法完成的工作，而全球比例为58%。 在前沿专业人士（微软对高级AI用户的称呼）中，这一比例升至82%。

Chubby♨️@kimmonismus · 6月17日75

Axios reports that Microsoft is considering a Microsoft-hosted version of DeepSeek V4 as a cheaper model option for Copilot Cowork. Microsoft says Copilot Cowork can’t work on unlimited pricing. “We have users who do hundreds of tasks a week… but the consequence is the costs can go very high,” Charles Lamanna told Axios. So Microsoft is moving Copilot Cowork to usage-based pricing, and exploring cheaper open-source model options. If Microsoft really goes with DeepSeek, it would be optional, fine-tuned, safeguarded, and hosted fully on Azure. Still: Microsoft adding a Chinese AI model to an enterprise Copilot product would be huge.

译微软正考虑为 Copilot Cowork 提供微软托管的 DeepSeek V4 版本，作为更便宜的模型选项。Copilot Cowork 将放弃无限定价，转向按使用量计费，原因是成本过高（用户每周执行数百项任务导致费用激增）。若采用 DeepSeek，该模型将是可选的、经过微调与安全防护，并完全托管于 Azure。Axios 报道称微软已微调了一个可用模型，最终决定待定。

Satya Nadella@satyanadella · 6月17日70

Copilot Cowork is now generally available worldwide, now with multi-model support! Every organization can put long-running agents to work on complex, multi-step tasks, grounded in your organization's unique knowledge and know-how. https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/?v=15

译Copilot Cowork 现已全球正式可用，并支持多模型！ 每个组织都可以让长期运行的智能体处理复杂的多步骤任务，基于你组织的独特知识和专有技术。 https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/?v=15

Rohan Paul@rohanpaul_ai · 6月16日50

Microsoft CEO Nadella explains how the real power of AI is transformation. Docs become websites, websites become apps, and firms scale decisions that once took hours into minutes. "what’s the real difference between a document, a website, and an application? It’s the lack of software that can transform itself. Interestingly enough, AI finally gives us that. I can write a document and say, “No, I don’t want it as a document. I want it as a website,” and it will transform that document, using code, into a website. Then I can say, “I don’t like the website. I want an app,” and it will write more code to transform it." --- Microsoft CEO Satya Nadella and BlackRock Chairman & CEO Larry Fink at the World Economic Forum in Davos. From 'DRM News' YT Channel (link in comment)

译微软CEO萨提亚·纳德拉在达沃斯世界经济论坛与贝莱德董事长拉里·芬克对话时表示，AI的真正力量在于转换：文档可通过AI生成代码变成网站，网站可进一步转为应用；企业将原本耗时数小时的决策规模化为几分钟。纳德拉强调，AI首次赋予了软件自我转换的能力。

Microsoft Research@MSFTResearch · 6月16日27

30x faster analytics, GPU kernels generated automatically from SQL, AI matched to lab-grown tumor models for cancer treatment, and LLMs that learn across tasks without retraining. Dive into the latest issue of Research Focus: https://msft.it/6010vcYZ4

译30倍更快的分析，从SQL自动生成的GPU内核，AI与实验室培育的肿瘤模型匹配用于癌症治疗，以及无需重新训练即可跨任务学习的大语言模型。深入探索最新一期Research Focus：https://msft.it/6010vcYZ4

AYi@AYi_AInotes · 6月15日51

微软CEO Satya Nadella这篇文章 是我今晚睡前必须看完的， 我觉得他把AI时代公司到底该怎么活这件事，说得很透。 他说，以前的软件工具，本质上是“帮人干活的辅助工具”， 你用Word，Word不会因为你用了十年就变得更懂你， 但AI不一样——它能跟人形成一个实时循环：你给它经验，它反过来让你更强。 他用两个词来拆解这件事， 第一个叫“人类资本”， 就是公司里人的判断力、创造力、人脉、那些在走廊上聊出来的经验， 第二个叫“token资本”， 就是公司自己拥有、持续改进的AI能力。 然后他给了一句最关键的话： 人类的价值不会因为AI变强而降低，反而会变得更重要。 我当时读到这句，脑子嗡了一下， 因为它跟很多人的直觉是反的，大家都在焦虑AI会不会取代人，他说不会，因为总要有人指方向，有人做跨领域的连接，有人在模糊地带拍板， 没有人的引导，AI就是在空转。 但真正让我耳目一新的地方，是他对公司护城河的定义。 他说公司真正的护城河，不是一个更好的大模型，而应该是一个“学习循环”，就是把公司自己的工作流程、领域知识、做了十年才攒下来的判断经验，喂给AI，让它在这个过程中不断变聪明。 这个系统越用越强，慢慢变成公司独有的“机构记忆”， 别人挖不走，复制不了， 因为底层模型可以换，但这个长期积累下来的东西，谁也拿不走。 然后他说了一句让我意识到他真正在担心什么的话， 他说，如果少数几个大模型把所有行业的知识都吸走，普通公司会越来越没价值，最终只剩调用权限——就像当年全球化把制造业外包后，很多国家只剩组装线一样。 所以我反复看他这段话，发现他讲的不仅仅是技术战略， 更像是一个CEO在坦白自己最深的恐惧：别让一个行业，最后只剩下调用按钮的权力。

译微软CEO Satya Nadella撰文称，AI时代公司真正护城河是“学习循环”——将工作流程、领域知识和多年判断经验喂给AI，形成独有的“机构记忆”。他提出“人类资本”（人的判断力、创造力）和“token资本”（公司持续改进的AI能力），强调人类价值不会因AI变强而降低，反而更重要。他警告：若少数大模型吸走所有行业知识，普通公司只剩调用权限，失去价值。

凡人小北@frxiaobei · 6月15日57

纳德拉给微软定调，别卷最强模型，去做模型之上的生态。 逻辑是模型早晚商品化，生态锁定才能更持久。 但他说平台让长在上面的价值多于自己攫取的。这是手里没有 SOTA 的人，最舒服的世界观。 要是微软自己有 SOTA，这篇大概率换个写法。 不过抛开动机这个框架是很有用，把闭环建在模型之上，IP 留自己手里，模型随时可换。

译微软CEO纳德拉明确表示，微软不追求最强AI模型，而是聚焦模型之上的生态建设。其逻辑是：模型终将商品化，生态锁定才能更持久。同时他指出，平台让生长其上的价值多于自身攫取的。这一框架被评论为“没有SOTA”的公司最舒服的世界观，但背后动机是将其作为策略：把闭环建在模型之上，IP留自己手里，模型随时可换。