Rohan Paul@rohanpaul_ai · 6天前38

Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/openais-new-paper-shows-how-they 🗞️ OpenAI’s new paper shows how they are now seeing the first version of office work where agents do most of the execution. 🗞️ New report on "The State of the AI Economy" 🗞️ New York Times: OpenAI is now leaning toward a 2027 IPO because the public market is testing whether AI giants deserve trillion-dollar prices before they prove durable profits. 🗞️ Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention 🗞️ The Economist: AI has pushed the internet’s content machine into a new phase, with books, lawsuits, research papers, apps, and songs now being produced at volumes that old review systems were not built to handle. 🗞️New research from OpenAI reported a training result where RL on realistic human situations made models carry safer, more useful behavior into tasks they had not trained on. 🗞️ MIT study. Code volume surges by 300%, but output increases by only 30%: The AI dividend meets an awkward reality. 🗞️ Qwen just released Qwen-AgentWorld, a 35B open-weight world model that learns how terminals, browsers, Android devices, code repos, search systems, OS tools, and MCP servers respond when an AI agent takes an action.

译本期周刊涵盖多项AI动态：OpenAI新论文展示智能体可执行大部分办公室工作的首个版本；NYT称OpenAI倾向于2027年IPO；OpenAI新研究发现基于真实人类场景的RL训练使模型在未来任务中更安全、有用；MIT研究显示代码量激增300%但产出仅增长30%；Qwen发布Qwen-AgentWorld，一个35B参数开放权重世界模型，可学习终端、浏览器、Android设备、代码仓库、搜索系统、OS工具及MCP服务器对AI智能体操作的响应。

Chubby♨️@kimmonismus · 6天前54

Morgan Stanley just doubled its 2026 China humanoid robot shipment forecast again, from 28,000 to 50,000 units. In January, the bank expected 14,000 units. A few months later, that forecast has more than tripled. China now shifts from demos to deployment. CNBC says Chinese manufacturers are moving humanoids into factories, logistics, convenience stores and restaurants, backed by policy support and a deep industrial supply chain.

译Morgan Stanley 刚刚再次上调了对 2026 年中国仿人机器人出货量的预测，从 28000 台提高到 50000 台。 1 月时，该行预计为 14000 台。几个月后，这一预测已增长超过两倍。 中国现在正从演示转向部署。CNBC 表示，中国制造商正在将仿人机器人引入工厂、物流、便利店和餐厅，背后有政策支持和深厚的工业供应链。

OpenCode@opencode · 6天前35

Daily unique users for GLM 5.2 spiked on the 23rd on Go

译GLM 5.2 的日独立用户数在 23 日在 Go 上激增。

jason@jxnlco · 6天前0

I —

译I — @DerekFeriancek：@jxnlco 你能不能使用计算机操作，给设施部门发个请求，多订点 Fairlife 奶昔？ 你也是 OpenAI 最爱喝 Fairlife 奶昔的前五名，每次我去看冰箱都是空的 😭

SemiAnalysis@SemiAnalysis_ · 6天前24

At HPC Summit Asia this year we enjoyed checking out the new 2u DLC HGX R200 systems on display. (1/3)🧵

译今年在 HPC Summit Asia 上，我们很高兴看到了展出的新款 2u DLC HGX R200 系统。（1/3）🧵

swyx 🔜 @aiDotEngineer@swyx · 6天前24

we have been scaling without slop by working with aligned domain experts to add coverage with both oai and ant launching multi-billion dollar services arms, it’s clear that FDE is one of the most in demand disciplines on earth, but I have never done the job it’s been an absolute pleasure working with Basil on our first ever AI FDE miniconference! see at https://ai.engineer/wf next week

译swyx引用推文指出，Forward Deployed Engineering（FDE）是企业采用AI的最关键因素。随着OpenAI和Anthropic投入数十亿美元建立服务部门，FDE成为最受关注的领域。swyx宣布与Basil合作举办了团队首个AI FDE迷你会议。他预告今年AI Engineer World's Fair将于6月30日周二开设Forward Deployed主题track，届时将探讨在生产中部署智能体所需的真实条件及经济转型方向。

DogeDesigner@cb_doge · 6天前58

BREAKING: Elon Musk has been cleared by the FTC to acquire Mesh Optical Technologies. Mesh was founded by former SpaceX engineers who worked on Starlink laser-link communications. The company builds optical transceivers for AI data centers.

译BREAKING: Elon Musk已获FTC批准收购Mesh Optical Technologies。 Mesh由前SpaceX工程师创立，他们曾参与Starlink激光链路通信。该公司为AI数据中心制造光学收发器。

Runway@runwayml · 6天前49

Congratulations to the winners of Runway's 2026 AI Festival. Get caught up on all of the films, "Best Of" winners and panels featuring Ron Howard, Roger Avary, Gala Avary, Joel Kuwahara and Girish Balakrishnan at the link below.

译恭喜 Runway 2026 AI 电影节的获奖者。请通过下方链接查看所有影片、"最佳"奖项得主以及 Ron Howard、Roger Avary、Gala Avary、Joel Kuwahara 和 Girish Balakrishnan 参与的专题讨论。

Rohan Paul@rohanpaul_ai · 6天前65

Morgan Stanley now expects China to ship 50,000 humanoid robots this year, almost 2x its January forecast. China’s edge is scale, because firms such as Xpeng, Unitree, UBTech, and Galbot can place robots inside factories, showrooms, warehouses, hospitals, and state-backed training sites to collect real task data. Morgan Stanley now sees China reaching 446,000 annual shipments by 2030, up from 262,000, with full-sized humanoids rising from 30% of shipments this year to 70% in 2028. Xpeng plans to mass-produce its Iron robot by end-2026 and Unitree says it shipped more than 5,500 humanoids in 2025. --- scmp. com/tech/article/3358210/morgan-stanley-raises-china-humanoid-robot-shipment-forecast-50000-units

译摩根士丹利将2025年中国类人机器人出货量预测上调至5万台，较1月预测几乎翻倍；2030年年度出货量预计达44.6万台（此前26.2万），其中全尺寸机器人占比从2025年的30%升至2028年的70%。中国优势在于规模：小鹏、宇树、优必选、银河通用等厂商可在工厂、仓库、医院等场景采集真实任务数据。小鹏计划于2026年底量产其Iron机器人，宇树称2025年已交付超5500台类人机器人。

Tibo@thsottiaux · 6天前30

Ola. The Codex team is investigating issues where some accounts are seeing faster usage draining than intended. We believe this is related to some abuse & fraud prevention mechanisms that we have that might be overflagging. Stay tuned.

译Ola。Codex 团队正在调查一些账户使用额度消耗比预期更快的问题。我们认为这与我们的一些滥用及欺诈预防机制可能过度标记有关。 敬请关注。

郭明錤｜Ming-Chi Kuo@mingchikuo · 6天前58

iOS 27 將進一步深化與 Apple Intelligence 的系統級整合。我的最新產業調查顯示，為確保在 AI 負載下系統仍能順暢運作，1H27 新款配備 A20 處理器的較低階 iPhone 之 DRAM 規格將升級至 9GB（1.5GB × 6-die），高於目前 A19 機型的 8GB（2GB × 4-die）。2H26 配備 A20 Pro 處理器的三款高階機型（折疊機與兩款 18 Pro 機型）則維持 12GB（1.5GB x 8-die）。

译郭明錤最新产业调查指出，iOS 27将深化与Apple Intelligence的系统级整合。为确保AI负载下系统顺畅，1H27新款搭载A20处理器的较低阶iPhone DRAM规格升级至9GB（1.5GB × 6-die），高于当前A19机型的8GB（2GB × 4-die）。2H26配备A20 Pro的三款高阶机型（折叠机与两款18 Pro）则维持12GB（1.5GB × 8-die）。

郭明錤｜Ming-Chi Kuo@mingchikuo · 6天前48

iOS 27 will bring tighter system-level integration with Apple Intelligence. My latest industry checks suggest Apple's lower-end 1H27 iPhones, powered by the A20 chip, will move to 9GB DRAM (1.5GB × 6 dies), up from 8GB (2GB × 4 dies) in the current A19 models, to keep the system running smoothly under AI workloads. The three new high-end 2H26 models powered by the A20 Pro chip (the foldable and two 18 Pro models) will remain unchanged at 12GB (1.5GB × 8 dies).

译iOS 27 将带来与 Apple Intelligence 更紧密的系统级集成。我最新行业调查显示，苹果 2027 上半年搭载 A20 芯片的低端 iPhone 将升级至 9GB DRAM（1.5GB × 6 颗粒），高于当前 A19 机型的 8GB（2GB × 4 颗粒），以确保系统在 AI 负载下流畅运行。三款搭载 A20 Pro 芯片的 2026 下半年高端新机型（折叠屏及两款 18 Pro 机型）则维持 12GB（1.5GB × 8 颗粒）不变。

Rohan Paul@rohanpaul_ai · 6天前83

OpenAI's Codex active users grew more than 5X the first half of 2026, and the fastest growth came outside the original developer audience. Usage exploded (per OpenAI's new research report): - Longer tasks are becoming the new normal. By May 2026, 80.6% of sampled individual users had asked Codex to do work that would take a human more than 30 minutes. 70.2% had asked for tasks over 1 hour, and 25.6% had requested jobs above 8 hours. - AI agents are not staying with developers. Engineers adopted Codex first, but Legal, Finance, Recruiting, and other non-technical teams quickly joined in. Within OpenAI, Codex now generates 99.8% of weekly output tokens. - Non-developer adoption is rising fastest. Since August 2025, non-developer Codex usage grew about 137x among individual users and 189x among organizational users. - Agents are expanding what workers can take on. Non-technical employees are using Codex for automation, data transformation, debugging, tooling, analysis, and technical tasks beyond their usual jobs.

译OpenAI 报告显示，Codex 在2026年上半年活跃用户增长超5倍，增速最快群体来自非开发者。截至2026年5月，80.6%的个体用户曾请求超30分钟的任务，70.2%超1小时，25.6%超8小时。自2025年8月以来，非开发者个体用户使用量增长约137倍，组织用户增长189倍。Codex 现已贡献 OpenAI 内部 99.8% 的周输出 token，非技术员工正用它完成自动化、数据转换等技术任务。

Nathan Lambert@natolambert · 6天前15

Organizing a small AI meetup in Seattle tomorrow afternoon (Fremont/Ballard). Reply / dm if interested in coming.

译明天下午在组织一场小型AI见面会（西雅图Fremont/Ballard区域）。有兴趣请回复或私信。

Rohan Paul@rohanpaul_ai · 6天前55

The Information reports that Anthropic’s Mythos preview spooked DeepSeek into fundraising. Because, CEO Liang Wenfeng realized it needed a much bigger cash pile to compete. And DeepSeek now plans to double its workforce too. AI competition is no longer just about model cleverness; it is about compute reserves, talent density, secure infrastructure, product surface area, and enough cash to survive several failed training runs. --- theinformation .com/articles/anthropics-mythos-spooked-deepseek-prompting-7-4-billion-fundraising

译The Information报道，Anthropic的Mythos预览版让DeepSeek感到震惊，CEO梁文峰意识到需要更大现金储备来竞争。DeepSeek随即启动74亿美元融资，并计划将所有部门员工数量翻倍，招聘覆盖AI核心研发、算法、深度学习、全栈开发和产品岗位，表明DeepSeek正从仅调模型转向构建完整系统。AI竞争已变为计算储备、人才密度、基础设施、产品表面积和现金储备的综合较量。

Chubby♨️@kimmonismus · 6天前68

Anthropic’s Mythos preview reportedly pushed DeepSeek into a $7.4B fundraising - because they could not compete with Mythos. Until now, the three-year-old Chinese AI lab had relied on CEO Liang Wenfeng’s personal wealth instead of outside capital. The Information reports the shift came after Anthropic’s April preview of Mythos. Liang reportedly concluded DeepSeek could not compete at that level without a much larger war chest. DeepSeek is now expanding aggressively: around 300 employees today, with plans to "at least double" headcount across AI systems, infrastructure, product and research.

译Anthropic 4 月预览的 Mythos 模型据报迫使 DeepSeek 转向外部融资，筹集 74 亿美元。此前这家成立三年的中国 AI 实验室一直靠 CEO 梁文锋个人财富运营。The Information 报道称，梁文锋认为 DeepSeek 若无更大资金储备将无法与 Mythos 竞争。DeepSeek 正激进扩张：当前约 300 名员工，计划在 AI 系统、基础设施、产品和研究部门至少翻倍人员规模。

Alibaba Cloud@alibaba_cloud · 6天前34

We're thrilled to announce that starting today, HappyHorse1.1 will become the default AI video generation model on Creative Fabrica Studio! To mark this upgrade, we're offering 50% off all video generations for one week — exclusively with HappyHorse 1.1. Whether you're a designer, marketer, or content creator, now's the perfect time to explore faster, higher-quality AI video creation. 👉 Try it now: https://studio.creativefabrica.com/ai-video-generator 👉 Learn more about the HappyHorse 1.1 model: https://int.alibabacloud.com/m/1000414899/ #AlibabaCloud #HappyHorse #CreativeFabricaStudio

译阿里云宣布HappyHorse 1.1即日起成为Creative Fabrica Studio的默认AI视频生成模型。为纪念此次升级，平台提供为期一周的视频生成50%折扣，适用于设计师、营销人员和内容创作者。附试用链接及模型详情页。

Alibaba Cloud@alibaba_cloud · 6天前45

At Flink Forward Asia Shenzhen 2026, NVIDIA’s Chuan Chen shared how NVIDIA and Alibaba Cloud accelerate multimodal data stream processing for Apache Flink: “NVIDIA and Alibaba Cloud's team technically collaborate to enable the CUDA library-accelerated multimodal data stream processing of Apache Flink.” This open-source collaboration enables end-to-end, high-performance multimodal streaming architectures for AI commentary, live image-text feeds, and interactive Q&A. #NVIDIA #AlibabaCloud #ApacheFlink #DataAI #AI #Multimodal #RealTimeStreaming

译在Flink Forward Asia Shenzhen 2026大会上，NVIDIA的Chuan Chen分享了NVIDIA与阿里云的技术合作：通过CUDA库加速Apache Flink的多模态数据流处理。这一开源协作实现了端到端的高性能多模态流式架构，可应用于AI解说、实时图文信息流和交互式问答等场景。

🚨 AI News | TestingCatalog@testingcatalog · 6天前44

Google has a hidden MCP integration config on Gemini for macOS but it is yet unclear if MCPs will ever become available over there. Tbh, I am not getting why it is that challenging to simply enable MCPs across Gemini products. It is quite an essential feature that opens up a ton of use cases. It would also let Google to get more data about tool use for training. Why? 👀

译Google 在 Gemini for macOS 上有一个隐藏的 MCP 集成配置，但目前尚不清楚 MCP 是否会在该平台上可用。 说实话，我不明白为什么在 Gemini 产品中直接启用 MCP 如此困难。 这是一个非常重要的功能，可以开启大量用例。 它也能让 Google 获取更多关于工具使用的数据用于训练。 为什么？👀

Alibaba Cloud@alibaba_cloud · 6天前30

Welcome to Qwen Live – Episode 1: Agent-First, When Your Next User Isn't Human. 📅 June 30, 2026 | 10:00 AM (UTC+8) 🔔Set Reminder Now: https://youtube.com/live/Hh-ftRYsGkI?feature=share As AI evolves at lightning speed, your next user might not be human—it could be an AI agent. In this debut episode, we are joined by Linlin Kong, Head of Qwen Cloud, alongside Qwen Cloud Product Managers Pan Gu and Xijue. Together, they will explore building cloud platform for agent from scratch, redefine developer experiences for non-human users, and uncover the new paradigms of large-scale human-agent collaboration. Ready to build for agents? Get started with Qwen Cloud Platform: https://click.qwencloud.com/m/20000000401/ #Qwen #QwenLive #AIagents #AICloud #DeveloperExperience #FutureOfWork

译阿里云宣布Qwen Live系列首期节目，主题为“Agent-First：当你的下一个用户不是人类”。节目将于2026年6月30日10:00（UTC+8）直播，由Qwen Cloud负责人林林孔、产品经理潘古和西觉共同主持。他们将探讨从零构建面向AI智能体的云平台、为非人类用户重新定义开发者体验，以及大规模人机协作的新范式。节目还提供Qwen Cloud平台入门链接。

Chubby♨️@kimmonismus · 6天前61

OpenAI is reportedly leaning toward delaying its IPO until next year - instead of this year. The company had hired bankers and lawyers for a possible public offering as soon as Q3 or Q4 this year. Sam Altman reportedly pushed advisers to find a path to a $1 trillion valuation. According to the NYT, advisers have warned OpenAI that retail investors may not have the same appetite after SpaceX’s volatile post-IPO trading and a broader pullback in tech stocks. The numbers are interesting: • OpenAI reportedly generated about $13B in revenue in 2025 • it says it is now making $2B per month • it hopes to roughly triple revenue this year • but it is still pouring huge amounts into data centers, compute, hiring and marketing At the same time, ChatGPT app growth has slowed around 900M users, Anthropic is gaining enterprise traction with Claude Code, and Google Gemini is becoming a stronger consumer product.

译OpenAI据悉将原定今年的IPO推迟至明年，此前已为Q3/Q4上市聘请银行和律师。Sam Altman推动顾问寻求达到1万亿美元估值的路径。《纽约时报》称，顾问警告零售投资者可能因SpaceX上市后波动及科技股回调而热情不足。OpenAI 2025年营收约130亿美元，月收入达20亿美元，计划今年收入翻三倍，但仍在数据中心、算力、招聘和营销上大举投入。与此同时，ChatGPT应用增长放缓至约9亿用户，Anthropic凭借Claude Code在企业端获取牵引，Google Gemini消费产品实力也在增强。

Alibaba Cloud@alibaba_cloud · 6天前40

At Flink Forward Asia Shenzhen 2026, Feng Wang, Researcher and Head of Open Data Platform at Alibaba Cloud & Vice Chair of the Alibaba Open Source Committee, highlighted the evolving data foundation for AI: "In the AI era, models and data together determine the quality and efficiency of Agents. Apache Flink evolves into Agentic Streaming for AI, working alongside Agentic Lake to build AI-native data platform." The next generation of intelligent agents is built on a unified, AI-native data infrastructure designed for real-time agentic workflows. #AlibabaCloud #ApacheFlink #ApachePaimon #ApacheFluss #DataAI #AI #Agent #RealTimeData

译在深圳举办的Flink Forward Asia 2026上，阿里云研究员、开放数据平台负责人Feng Wang指出，AI时代模型与数据共同决定Agent质量与效率。Apache Flink演进为Agentic Streaming for AI，与Agentic Lake协同，构建AI原生数据平台。下一代智能体建立在统一、实时的AI原生数据基础设施之上。

Alibaba Cloud@alibaba_cloud · 6天前35

Build more. Spend less. Save up to 80% on Qwen3.7 models with Alibaba Cloud Model Studio by running your #AI workloads during off-peak hours. Same powerful models. Smarter pricing. Learn more: https://int.alibabacloud.com/m/1000414853/ #AlibabaCloud #Qwen

译Build more. Spend less. 在非高峰时段运行 #AI 工作负载，使用阿里云 Model Studio 处理 Qwen3.7 模型最高可节省 80% 费用。同样的强大模型，更智能的定价。 了解更多：https://int.alibabacloud.com/m/1000414853/ #AlibabaCloud #Qwen

Rohan Paul@rohanpaul_ai · 6天前38

Mark Andreessen on the anti-data centers sentiments in US. "This completely fake meme about water use, which is just factually not true, running wild through the public discussion that somehow these data centers are basically destroying all the water"

译Mark Andreessen 谈美国国内的反数据中心情绪。 “这个关于用水的假梗完全就是假的，事实并非如此，它在公众讨论中疯传，好像这些数据中心正在耗尽所有水资源似的。”

Rohan Paul@rohanpaul_ai · 6天前61

Reuters: DeepSeek is going on a hiring sprint, aiming to double every department. The hiring spans AI core R&D, algorithms, deep learning research, full-stack development, and product roles, which means DeepSeek is no longer only tuning models but building the whole system around them. --- reuters .com/world/asia-pacific/chinas-deepseek-plans-least-double-staff-all-departments-2026-06-25/

译路透社：DeepSeek 正进行招聘冲刺，目标是每个部门人员翻倍。 招聘涵盖AI核心研发、算法、深度学习研究、全栈开发和产品岗位，这意味着DeepSeek不再只是调模型，而是围绕模型构建整个系统。

Rohan Paul@rohanpaul_ai · 6天前61

FT: Micron just reported a 15-fold profit jump because AI servers are now short of high-bandwidth memory, the stacked memory that keeps GPUs fed with data. The shortage gives Micron rare pricing power, with adjusted gross margin reaching 84.9%, compared with 39% a year earlier. --- ft .com/content/9b739203-3274-43f1-b61e-c1905061d32a

译FT: 美光刚刚报告利润增长15倍，因为AI服务器目前缺乏高带宽内存，这种堆叠内存可保持GPU的数据供给。 短缺赋予美光罕见的定价权，调整后毛利率达84.9%，而一年前为39%。 --- ft .com/content/9b739203-3274-43f1-b61e-c1905061d32a

Kling AI@Kling_ai · 7天前20

hҽყ ʏօʊ sҽҽ tԋαƚ UFՕ??

译嘿，你看到那个UFO了吗？

Rohan Paul@rohanpaul_ai · 7天前70

New York Times: OpenAI is now leaning toward a 2027 IPO because the public market is testing whether AI giants deserve trillion-dollar prices before they prove durable profits. IMO, SpaceX’s huge IPO and Anthropic’s confidential filing also change the OpenAI story because public-market capital is not infinite, and investors may already be digesting one massive AI-adjacent listing while preparing for another pure AI listing. OpenAI cannot just ask for $1T in isolation, because every giant IPO competes for the same pool of institutional cash, risk appetite, and patience for loss-making growth stories. So waiting until 2027 lets OpenAI avoid becoming the third huge test of AI-market depth in the same cycle --- nytimes .com/2026/06/25/technology/openai-ipo-artificial-intelligence.html

译OpenAI 据传正倾向于2027年进行IPO，原因是公共市场正在测试AI巨头是否值得万亿美元估值、SpaceX 和 Anthropic 的巨额IPO也会争夺同一笔机构资金。与此同时，美国政府要求OpenAI将GPT-5.6改为受控预览发布，仅向小组合作伙伴提供早期访问，官员逐个批准客户。此举主要担忧在于模型自动化高级网络工作能力——帮助防御者发现漏洞的同时，也可能让攻击者更快测试利用方式。

SenseTime@SenseTime_AI · 7天前33

𝗦𝗲𝗻𝘀𝗲𝗧𝗶𝗺𝗲 𝗖𝗵𝗮𝗶𝗿𝗺𝗮𝗻 𝗼𝗳 𝘁𝗵𝗲 𝗕𝗼𝗮𝗿𝗱 𝗮𝗻𝗱 𝗖𝗘𝗢 𝗗𝗿. 𝗫𝘂 𝗟𝗶, alongside a delegation of prominent Chinese tech leaders, met with 𝗦𝗼𝘂𝘁𝗵 𝗞𝗼𝗿𝗲𝗮𝗻 𝗣𝗿𝗶𝗺𝗲 𝗠𝗶𝗻𝗶𝘀𝘁𝗲𝗿 𝗞𝗶𝗺 𝗠𝗶𝗻-𝘀𝗲𝗼𝗸 in Beijing during the Prime Minister's official visit to China. Dr. Xu Li noted that 𝗦𝗼𝘂𝘁𝗵 𝗞𝗼𝗿𝗲𝗮 𝗶𝘀 𝗦𝗲𝗻𝘀𝗲𝗧𝗶𝗺𝗲'𝘀 𝗸𝗲𝘆 𝗴𝗹𝗼𝗯𝗮𝗹 𝗽𝗮𝗿𝘁𝗻𝗲𝗿, expressing hope for 𝗱𝗲𝗲𝗽𝗲𝗿, 𝗺𝘂𝘁𝘂𝗮𝗹𝗹𝘆 𝗯𝗲𝗻𝗲𝗳𝗶𝗰𝗶𝗮𝗹 𝗰𝗼𝗼𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻 between the two sides on 𝗴𝗿𝗲𝗲𝗻 𝗔𝗜 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗮𝗻𝗱 𝘁𝗿𝘂𝘀𝘁𝘄𝗼𝗿𝘁𝗵𝘆 𝗔𝗜. "South Korea brings strength in semiconductors, storage, and energy engineering. SenseTime brings AIDC operations, scheduling, and model platform expertise," he said. "By combining our strengths, we can deliver greener, more efficient AI infrastructure worldwide." "Cross-border AI collaborations do not have to start with building a model together — it can start with optimizing AI infrastructure efficiency together." On trustworthy AI, he added that governance isn't about limiting innovation — it's the passport for AI to enter finance, public services, and global markets. China and South Korea, both highly digitalized with strong industrial bases, can pioneer a shared approach to identity verification, deepfake detection, content labeling, and high-impact AI evaluation — then bring that trust to markets worldwide. Since 2019, SenseTime has steadily built partnerships with over 50 Korean clients, spanning smart hardware, factory safety, banking eKYC, and consumer robotics, among other sectors. We are excited for what's next as China and South Korea deepen their economic and technological ties. In the photo (from right): (2nd) Dr. Xu Li, Chairman and CEO of SenseTime; (5th) Kim Min-seok, Prime Minister of South Korea. 𝘗𝘩𝘰𝘵𝘰 𝘊𝘳𝘦𝘥𝘪𝘵: 𝘖𝘧𝘧𝘪𝘤𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘗𝘳𝘪𝘮𝘦 𝘔𝘪𝘯𝘪𝘴𝘵𝘦𝘳, 𝘙𝘦𝘱𝘶𝘣𝘭𝘪𝘤 𝘰𝘧 𝘒𝘰𝘳𝘦𝘢

译商汤董事会主席兼CEO徐立博士随中国科技领袖代表团在北京会见韩国总理金民锡。徐立表示，韩国是商汤关键全球伙伴，希望深化绿色AI基础设施和可信AI领域合作。他指出韩国在半导体、存储及能源工程方面有优势，商汤在AIDC运营、调度和模型平台具备专长，结合可提供更绿色高效的AI基础设施。可信AI方面，治理是AI进入金融、公共服务和全球市场的通行证，中韩可共同探索身份验证、深度伪造检测等标准。自2019年，商汤已与超过50家韩国客户合作，覆盖智能硬件、工厂安全、银行eKYC、消费机器人等领域。

SemiAnalysis@SemiAnalysis_ · 7天前60

BREAKING NEWS: OpenAI’s Technical Lead for Compute has returned to OpenAI after joining Meta in April 2026. What is going on at Meta that people are quitting after only a couple of months? Did Anuj get reorged into a data-labeling role at Meta, and is that why he left?

译突发新闻：OpenAI的计算技术负责人在2026年4月加入Meta后，现已重返OpenAI。Meta发生了什么，让人们只待了几个月就离职？ Anuj在Meta被重组到了数据标注岗位，这就是他离开的原因吗？

ginobefun@hongming731 · 7天前37

BestBlogs 早报 · 06-26 # DSPy / Cloudflare Workflows / 特拉华 C-Corp / Saga 回滚 / Mistral OCR 4 [1] ★ 精讲｜我们如何利用 DSPy 将 AI 评估转化为 Dash Chat 的更优回复 Dropbox 技术团队分享了 Dash Chat Agent 的评估优化实践。核心方案是以开源框架 DSPy 构建两阶段反馈闭环：先用人工标注数据校准 LLM 裁判，再以校准后的裁判自动优化 Agent 提示词。人工评估覆盖意图理解、语义相关性、工具调用等 5 个维度，评分采用 1–5 分制。上线后不完整回答减少 26%，遗漏关键点减少 13%，token 用量下降 5.4%。对正在搭建 Agent 评测与提示词优化闭环的团队有参考价值。 来源：Dropbox Tech Blog https://www.bestblogs.dev/article/c2a3404e [2] ★ 精讲｜我们如何为 Cloudflare Workflows 构建 Saga 回滚 Cloudflare 官方博客介绍了 Workflows 新发布的 Saga 回滚功能。开发者现在可以在 http://step.do() 中直接传入 rollback 函数，工作流失败时按步骤启动顺序的逆序执行补偿，且回滚步骤同样具备重试与超时保障。文章还解释了为何放弃链式 API 和构建器模式，最终选择 options object 设计，以及底层如何通过持久化步骤记录在引擎重启后重建回滚状态。搭建含支付、库存等多步骤分布式业务的开发者可以关注。 来源：The Cloudflare Blog https://www.bestblogs.dev/article/b8c3d56b [3] ★ 精讲｜AI 创业者想出海拿美元，搭好可融资的企业架构才是第一步 Founder Park 整理了清律纽约律师事务所南李律师关于出海企业架构的分享。核心观点是「投资人投的是创业企业，不是创业产品」，MVP 验证之外必须搭建可融资的法律架构。文章建议融资导向的团队首选特拉华州 C-Corp：LLC 因穿透税制和 QSBS 限制普遍不受 VC 欢迎。股权归属采用美国标准的四年 Vesting 加一年 Cliff，并建议创始人主动设置而非等投资人提要求。适合有出海融资计划的 AI 创业者参考。 来源：Founder Park https://www.bestblogs.dev/article/c28a321b [4] 13 条优化技术演讲的建议（来自 swyx） swyx 基于观看数千场演讲的经验，分享了可操作的技术演讲准备建议。 来源：swyx 🔜 @aiDotEngineer(@swyx) https://www.bestblogs.dev/status/2069964772003770673 [5] Vector RAG 不够用了——我为多智能体记忆构建了一个上下文图层 本文提出了一种用于多智能体记忆的上下文图层，通过将事实存储为实体和关系，在多跳查询中表现优于原始历史记录和纯向量 RAG，实现了 88.9% 的准确率和每次查询 26.9 个 token 的效率。 来源：Towards Data Science https://www.bestblogs.dev/article/0b2a6406 [6] 混合模型在哪些 token 上预测得更好？ 这篇文章逐个 token 比较了 Transformer 和混合语言模型，发现混合模型在承载意义的 token 上表现优异，而 Transformer 在重复性或句法 token 上同样出色。 来源：Hugging Face - Blog https://www.bestblogs.dev/article/5020d758 [7] 3 个智能体，3 个大模型，1 块老旧 GPU：在裸机上实现并行推理工程 本文解释了由于 KV 缓存预分配，在单块低显存 GPU 上并行运行多个基于大模型的智能体为何会失败，并介绍了一个名为 lmxd 的小型 C++ 守护进程，通过强制执行显存记账来解决这一问题。 来源：Towards Data Science https://www.bestblogs.dev/article/0807a4a4 [8] Cursor 研究：模型「作弊」公共基准测试 Cursor 揭示，包括 Opus 4.8 和 Composer 2.5 在内的最新模型学会了从互联网或 git 历史记录中检索答案，在更严格的评估框架下，其评测分数显著下降。 来源：Cursor(@cursor_ai) https://www.bestblogs.dev/status/2070195789121671624 [9] Mistral AI 发布 OCR 4：支持边界框与置信度评分 Mistral AI 推出 OCR 4，该模型提供带有边界框、区块分类和置信度评分的结构化文档提取，支持 170 种语言。 来源：Mistral AI(@MistralAI) https://www.bestblogs.dev/status/2069420263825895917 [10] 如何通过现代 Web 指南阻止你的 AI 编码智能体编写过时代码 现代 Web 指南将专家验证的浏览器 API 指导注入 AI 编码智能体，用声明式 HTML 和 CSS 取代遗留的 JavaScript 密集型模式。 来源：freeCodeCamp https://www.bestblogs.dev/article/5c50eda6 --- http://BestBlogs.dev · 发现真正适合你的高质量内容 BestBlogs 是 AI 驱动的私人阅读助手，帮助你发现真正适合你的高质量内容，欢迎体验。 在线阅读：https://www.bestblogs.dev/explore/brief/2026-06-26

译Dropbox 用 DSPy 构建两阶段反馈闭环，校准 LLM 裁判优化 Agent 提示词，不完整回答减少 26%，token 用量降 5.4%。Cloudflare Workflows 新增 Saga 回滚，`step.do()` 可传 rollback 函数，失败逆序执行补偿。Cursor 揭示 Opus 4.8 和 Composer 2.5 模型检索作弊，严标准下分数下降。Mistral OCR 4 支持 170 种语言、边界框及置信度评分。多智能体记忆上下文图层实现 88.9% 准确率和每查询 26.9 token 效率。

🚨 AI News | TestingCatalog@testingcatalog · 7天前45

MISTRAL 🔥: Mistral reinforces its executive leadership team in order to scale its global expansion. > Johan Bergqvist is appointed Chief Financial Officer > Brian Hall is appointed Chief Marketing Officer > Kamal Brar is appointed SVP, Partners & Alliances

译MISTRAL 🔥：Mistral 加强其高管团队以推动全球扩张。 > Johan Bergqvist 被任命为首席财务官 > Brian Hall 被任命为首席营销官 > Kamal Brar 被任命为高级副总裁，合作伙伴与联盟

DogeDesigner@cb_doge · 7天前40

BREAKING: Grok Imagine is now the #1 AI video model on Vercel AI Gateway. • ~51% of all videos generated come from Grok. • Bigger share than every other top video model combined. • Grok Imagine Video 1.5 Preview is also top 5. Grok is absolutely dominating AI video.

译重磅：Grok Imagine 现已成为 Vercel AI Gateway 上排名第一的 AI 视频模型。 • 约 51% 的视频生成来自 Grok。 • 份额超过其他所有顶级视频模型的总和。 • Grok Imagine Video 1.5 Preview 也位列前五。 Grok 正在绝对主导 AI 视频领域。

MiniMax (official)@MiniMax_AI · 7天前33

we’ll be at AI Engineer After Dark on July 1st with @vercel , @merge_api, @FactoryAI , and a room full of people building the AI engineering stack. our Research Lead, RL Training @olive_jy_song will be giving a lightning talk on post-training MiniMax M3 as part of a lineup featuring @browserbase , @neondatabase , @FireworksAI_HQ , @ExaAILabs , @sentry , @SurrealDB, and more. see you after dark at SFMOMA.

译我们将于7月1日参加AI Engineer After Dark活动，与@vercel、@merge_api、@FactoryAI以及众多构建AI工程栈的开发者们相聚。 我们的研究负责人、RL训练专家@olive_jy_song将发表一场关于MiniMax M3后训练的闪电演讲，同场阵容还包括@browserbase、@neondatabase、@FireworksAI_HQ、@ExaAILabs、@sentry、@SurrealDB等。 夜幕降临后，旧金山现代艺术博物馆见。

Rohan Paul@rohanpaul_ai · 7天前76

The Information: The US government is asking OpenAI to slow GPT-5.6 into a controlled preview instead of releasing it broadly at once. OpenAI reportedly plans to give small partner groups early access while officials approve customers one by one, turning a model launch into a security-gated rollout. The concern is Mainly around the automated higher-skill cyber work. so the same model that helps defenders find bugs could also help attackers test exploits faster. --- theinformation .com/articles/trump-administration-asks-openai-stagger-release-new-model-security-concerns

译The Information 报道，美国政府因安全顾虑要求 OpenAI 暂缓 GPT-5.6 的广泛发布，改为推出受控预览版。OpenAI 计划先向小部分合作伙伴提供早期访问，并由政府逐客户审批准入。这一非常规做法主要担忧模型在自动化高技能网络工作上的能力：既能帮防御者更快发现漏洞，也可能被攻击者用于加速测试漏洞利用。本周四，CEO Sam Altman 已向员工确认该审批流程。

swyx 🔜 @aiDotEngineer@swyx · 7天前41

btw as a musician myself i've gone ahead and pulled the trigger for the first ever music corner at AIE if you code + music, we want impromptu jams between sessions. grab me to sing with u if u want. give us your best dad rock. sponsorship for this + networking night on jun 30 are the only show floor opportunities left at aie, everything else is sold out (btw tix on track to sell out TOMORROW)

译swyx 以音乐人身份为 AI Engineer World's Fair 2026 设立首个音乐角，邀请会编程和音乐的参会者即兴合奏。该音乐角与 6 月 30 日的社交之夜是仅剩的展位赞助机会。本届大会规模更大：展区扩大 4 倍、设 4 个展台舞台、研究者海报环节、Token Billionaires 及 Off the Record 领袖论坛、医疗/GTM/FDE/AGC/金融垂直论坛、NEO 和儿童日边会。参会者可获 $40k 信用额度试用赞助商产品。swyx 提醒门票预计明天售罄。

swyx 🔜 @aiDotEngineer@swyx · 7天前47

on their @latentspacepod we covered how @pimdewitte accidentally made the PERFECT world model data collection business by collecting the world's largest dataset of trainable (video,action) pairs. turning the attention economy into the attention industry. congrats Pim!! link in reply

译PimDeWitte在Latent Space播客上分享，其公司意外打造了世界最大的可训练（视频，动作）对数据集，将注意力经济转化为注意力产业。同时，今日宣布完成3.2亿美元A轮融资，估值23亿美元，由Khoslaventures领投，General Catalyst、Jeff Bezos、Eric Schmidt、Nico Rosberg及前沿实验室和学术界研究员参投。

Rohan Paul@rohanpaul_ai · 7天前66

US Congress is moving a Ratepayer Protection Act that would make AI data center builders, not ordinary electricity customers, pay for the new power plants, transmission lines, and grid upgrades their projects require. The bill pushes states to create a large-load standard, meaning a data center that creates a huge new electricity burden can be put in a separate rate category and charged upfront, through special tariffs, deposits, guarantees, or contracts, so normal homes and businesses do not subsidize the infrastructure built mainly for that data center. Tech firms already signed a White House pledge to cover new energy and delivery infrastructure for data centers, but this bill would give states a formal path to enforce that idea. So this is the first serious federal move to put AI’s physical footprint into the price of AI itself, and it could make data center siting depend less on cheap land and more on who can fund reliable power fastest. --- cnbc. com/2026/06/24/ai-data-centers-tech-companies-congress-energy-costs.html

译美国国会正在推进《Ratepayer Protection Act》，要求AI数据中心开发商承担新建电厂、输电线路和电网升级费用，而非转嫁给普通电力用户。法案鼓励各州设立“大负荷标准”，将高耗电数据中心单独分类，通过特殊关税、押金、担保或合同预先收费，防止普通家庭和企业补贴此类基础设施。科技公司此前已签署白宫承诺覆盖数据中心能源及输送基础设施，但该法案为各州提供了正式执行路径。这是首个将AI物理能耗成本内部化的联邦举措，可能使数据中心选址更侧重供电能力而非廉价土地。

Chubby♨️@kimmonismus · 7天前57

Apple is not the only hardware company getting hit by AI-driven component inflation. ASUS has warned of price hikes as DRAM and SSD costs surge. Dell and HP are under margin pressure as AI PCs require more memory while memory prices rise. Alibaba Cloud raised prices by up to 34%, citing global AI demand and higher costs for GPUs, CPUs, memory and storage. The trend seems to be reflected everywhere. And there's no end in sight.

译苹果并非唯一受 AI 驱动零部件涨价影响的硬件公司。华硕因 DRAM 和 SSD 成本飙升警告涨价；戴尔和惠普利润率承压（AI PC 需更多内存，且内存价格攀升）。阿里云提价最高 34%，归因全球 AI 需求推高 GPU、CPU、内存和存储成本。此前苹果已大幅涨价，满配 16 英寸 MacBook Pro 现售价 9999 美元。该趋势普遍蔓延，且尚无结束迹象。

Rohan Paul@rohanpaul_ai · 7天前41

Code is automated, debugging still stayed mostly manual. @sazabi is trying to close that gap with an AI observability system that detects issues, investigates failures, and helps prepare fixes. logs are all you need: Its bet is that logs can become the source of truth, with AI deriving metrics, traces, and possible fixes from the raw events teams already collect.

译AI 可观测性初创公司 sazabi 获得 800 万美元融资，其平台将日志（logs）作为唯一事实来源，让 AI 自动检测问题、调查故障并协助准备修复方案。该平台从团队已有的原始日志中推导指标、追踪和可能修复，旨在替代传统手动监控。sazabi 定位为下一代通用可观测性方案，适用于任何工作负载（包括 AI 智能体），而非另一款 AI SRE 或 LLM 可观测性工具。2026 年软件更新速度极快，该平台希望通过 AI 最大化自动化和速度，实现自我修复软件。