OpenAI@OpenAI · 6月3日70

Building apps has never been easier. With Sites, Codex can turn your work, ideas, and plans into an interactive website or app your team can explore, use, and share with a URL. Rolling out to Business and Enterprise plans, before expanding more broadly.

译构建应用从未如此简单。 通过 Sites，Codex 可以将你的工作、想法和计划转化为一个交互式网站或应用，你的团队可以通过一个 URL 进行探索、使用和分享。 该功能将首先向 Business 和 Enterprise 计划推出，之后会更广泛地扩展。

Microsoft Research@MSFTResearch · 6月3日11

Microsoft Research is at #MSBuild! 

Come check out the Microsoft Research Lab on the 1st Floor of the Gateway Pavilion and stay tuned to our channels for more.

译微软研究院正在参加 #MSBuild！ 欢迎来 Gateway 展馆一楼的微软研究院实验室参观，并关注我们的频道以获取更多信息。

Google Gemini@GeminiApp · 6月3日41

See Gemini Omni in action with a live demo from one of the team members who helped build it. Join us on Discord as Product Manager Chloe shares how to use Gemini Omni’s multimodal inputs, real-world knowledge, and easy conversational editing to create videos like never before. 👉Join the Discord to watch live: http://discord.gg/gemini 📅 Wednesday, June 3 at 11:30am PT

译通过一位参与构建的团队成员的现场演示，了解 Gemini Omni 的实际效果。 加入我们的 Discord，产品经理 Chloe 将分享如何使用 Gemini Omni 的多模态输入、现实世界知识和便捷的对话式编辑，以前所未有的方式创建视频。 👉加入 Discord 观看直播：http://discord.gg/gemini 📅 太平洋时间 6 月 3 日星期三上午 11:30

OpenRouter@OpenRouter · 6月3日68

⚡ New provider drop: AI-Native Cloud from @digitalocean is now live on OpenRouter. High performance inference across popular open-weight models. #1 on output speed and latency for DeepSeek V3.2 by @ArtificialAnlys. See their stats and try the models: https://openrouter.ai/provider/digitalocean

译⚡ 新增服务商：DigitalOcean 的 AI-Native Cloud 现已在 OpenRouter 上线。 提供高性能推理，覆盖热门开源权重模型。在 DeepSeek V3.2 的输出速度和延迟方面排名第一（数据来自 @ArtificialAnlys）。 查看其数据并试用模型：https://openrouter.ai/provider/digitalocean

Replit ⠕@Replit · 6月3日63

Replit Canvas has a few new updates! ⭐️ Learn more at: http://replit.com/canvas Open thread 🧵 ↓

译Replit Canvas 有一些新更新！⭐️ 了解更多请访问：http://replit.com/canvas 展开讨论 🧵 ↓

Replit ⠕@Replit · 6月3日70

Using Parallel Agents to Move Faster in Replit https://x.com/i/broadcasts/1NxarrEMVOnKj

译在 Replit 中使用并行智能体来提升速度 https://x.com/i/broadcasts/1NxarrEMVOnKj

🚨 AI News | TestingCatalog@testingcatalog · 6月3日55

Intelligence at Work keynote from OpenAI 🍿

译OpenAI Intelligence at Work 主题演讲 🍿 [引用 @OpenAINewsroom]：Codex 现在每周活跃用户超过 500 万。 但更大的故事是人们如何使用它：不仅仅是编写代码，还在研究、分析、内容和运营等工作中提高效率。 我们关于 Codex 如何成为知识工作生产力工具的新报告：https://openai.com/index/codex-for-knowledge-work/

Chubby♨️@kimmonismus · 6月3日47

Incredible. The best design work doesn’t happen in a chat box. The fact that you can generate motion assets inside the canvas skips your image-gen hops. Super cool that changes here get synced back to your codebase.

译Kombai 2.0 被定位为首个AI设计工程师，旨在融合设计与工程。该工具允许用户在画布内直接生成动画素材，跳过了传统图像生成的中间环节，并能将设计变更同步回代码库。其目标是打破设计和工程分属不同工作流的旧模式，服务于一个设计师能交付代码、工程师寻求无缝集成、所有人都想构建优秀用户体验的新世界。

AK@_akhaliq · 6月3日62

GPU Forecasters Language Models as Selective Surrogates for Kernel Runtime Optimization

译GPU预测器 大语言模型作为内核运行时优化的选择性代理

AK@_akhaliq · 6月3日60

Seeing Isn't Knowing Do VLMs Know When Not to Answer Spatial Questions (and Why)?

译视觉语言模型知道何时不回答空间问题吗（以及为什么）？

Rohan Paul@rohanpaul_ai · 6月3日65

Another brilliant launch removing friction from front-end development. Kombai just launched a frontend-specific AI coding agent and it beats general coding agents on real repo tasks. The problem with generic agents is that they often fail frontend work because UI code mixes visual judgment, component reuse, CSS behavior, browser bugs, accessibility etc. Kombai is attacking that problem with specialization: it reads design context, browser state, existing components, hooks, design tokens, and DevTools data so the agent can edit the product the way a frontend engineer would. checkout their demo, where it adds a complex feature to an OSS codebase with 500K+ lines of code. They also open-sourced the dataset that anyone can use to benchmark agents for complex front-end tasks.

译Kombai 推出了首个专用于前端开发的AI编程智能体。针对通用智能体在处理前端任务时的不足，Kombai 通过读取设计上下文、浏览器状态、组件等数据，像前端工程师一样进行代码编辑。推文称，Kombai 在真实代码库任务上的表现超越了 SOTA 模型和通用编程助手，并在一个超过 50 万行的开源代码库中演示了添加复杂功能。此外，Kombai 还开源了一个可用于评测复杂前端任务的基准数据集。

AK@_akhaliq · 6月2日62

Crafter A Multi-Agent Harness for Editable Scientific Figure Generation from Diverse Inputs

译Crafter 一个用于从多样化输入生成可编辑科学图表的多智能体框架

Claude@claudeai · 6月2日30

Interpreting law is one of the oldest jobs in the world. @MaxJunestrand, co-founder and CEO of @WeAreLegora, is bringing it into its next era with Claude. His bet: every new model release raises the tide, and Legora is building the boats for everyone else.

译解读法律是世界上最古老的职业之一。@WeAreLegora的联合创始人兼CEO @MaxJunestrand 正在用Claude将其带入下一个时代。他的信念是：每一次新模型的发布都抬高了水位，而Legora正在为其他所有人建造船只。

Kling AI@Kling_ai · 6月2日21

Who’s your favorite World Cup team? Generate a cheering dance video for them! 🎉

译你最喜欢的世界杯球队是哪支？为他们生成一段助威舞蹈视频吧！🎉

elvis@omarsar0 · 6月2日50

// Scaling Behavior of Single LLM-Driven Multi-Agent Systems // Does adding more agents actually make a multi-agent system better? It's possible that collective intelligence emerges from interaction design rather than from agent plurality. This is something important to understand if you are building multi-agent systems. This new study reports that the optimal number of agents depends on the base model's capability and the task type, not on adding more of them. Paper: https://arxiv.org/abs/2606.00655 Learn to build effective AI agents in our academy: https://academy.dair.ai/

译研究探讨添加更多智能体是否提升多智能体系统性能。结论指出，最优智能体数量取决于基础模型的能力和任务类型，而非单纯增加数量。集体智能更可能源于精心的交互设计，而非智能体数量的增多。相关论文："Scaling Behavior of Single LLM-Driven Multi-Agent Systems"。

Emad@EMostaque · 6月2日35

I wonder how many founders will pass on investors who passed on them in prior rounds I wonder how many would have three dinners & give them an allocation only to slash it to zero at the last moment.

译推文提出疑问：多少创始人会拒绝那些曾在前一轮投资中拒绝过自己的投资人？并引述Anthropic上一轮融资的内幕：一位知名基金的合伙人与Dario共进了三次晚餐后，其份额被削减至零。同时，至少另外四家一线基金也在最后关头被撤。引用推文指出，这些投资人受罚的原因是错过了由Spark领投的Series B——那是Dario经历过最艰难的一轮募资。在风险投资中，信任（conviction）就是一切。

SenseTime@SenseTime_AI · 6月2日81

Power smarter AI agents with #𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗦𝗸𝗶𝗹𝗹𝘀 🚀 𝗔𝗻 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲 𝗔𝗜 𝗼𝗳𝗳𝗶𝗰𝗲 𝘀𝗸𝗶𝗹𝗹 𝘀𝘂𝗶𝘁𝗲 𝗳𝗼𝗿 𝗮𝗻𝘆 𝘀𝗸𝗶𝗹𝗹𝘀-𝗰𝗼𝗺𝗽𝗮𝘁𝗶𝗯𝗹𝗲 𝗮𝗴𝗲𝗻𝘁, 𝗶𝗻𝗰𝗹𝘂𝗱𝗶𝗻𝗴 #𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗮𝗻𝗱 #𝗛𝗲𝗿𝗺𝗲𝘀𝗔𝗴𝗲𝗻𝘁. 🖼️ 𝗜𝗻𝗳𝗼𝗴𝗿𝗮𝗽𝗵𝗶𝗰 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 — image generation & graphic design; mirror visual style from a reference 📊 𝗗𝗮𝘁𝗮 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 — multi‑sheet parsing, data cleaning, image-based data extraction and visualization 📑 𝗣𝗣𝗧 𝗖𝗿𝗲𝗮𝘁𝗶𝗼𝗻 — outline & content generation, intelligent layout design; outputs editable PPT files 🔍 𝗗𝗲𝗲𝗽 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 — multi‑source search across academic, tech, social platforms & more; synthesizes insights & generates reports Now fully #OpenSourced. 👉 ＧitHub https://github.com/OpenSenseNova/SenseNova-Skills ⚡ Quick installation (bundled with Hermes Agent & OpenClaw) https://github.com/OpenSenseNova/SenseNova-Skills/blob/main/INSTALL.md 👾 Discord https://discord.gg/BuTXPHmQub

译商汤开源了AI办公技能套件SenseNova-Skills。这是一个为任何技能兼容智能体（如OpenClaw与HermesAgent）设计的开源技能集合，提供四大核心功能：图像信息图表生成（可镜像参考风格）、数据分析（支持多表解析、清洗与可视化）、PPT创建（生成大纲内容并智能排版，输出可编辑文件）以及深度研究（跨学术、技术、社交等多源搜索并生成报告）。该技能套件现已完全开源。

OpenBMB@OpenBMB · 6月2日56

🎉 Exciting News! MiniCPM-o 4.5 has officially been merged into vLLM-Omni! 👏 💡 Currently, it supports simplex non-streaming usage—a solid first step for what's to come. 🚀 The journey continues! We will keep pushing forward with the amazing open-source community to integrate duplex streaming capabilities next. Get ready for an even more powerful and seamless interactive experience. Stay tuned! 💪 http://github.com/vllm-project/vllm-omni/pull/4067 http://huggingface.co/openbmb/MiniCPM-o-4_5

译🎉 好消息！MiniCPM-o 4.5 已正式合并至 vLLM-Omni！👏 💡 目前，它支持单工非流式使用——这是未来发展的坚实第一步。 🚀 旅程继续！我们将与出色的开源社区一起，继续推进下一步的双工流式功能集成。 敬请期待更强大、更无缝的交互体验。敬请关注！💪 http://github.com/vllm-project/vllm-omni/pull/4067 http://huggingface.co/openbmb/MiniCPM-o-4_5

Berryxia.AI@berryxia · 6月2日50

卧槽！你们别一天天光知道给Agent花钱消耗Token了。 赶紧派你的Agent给你去赚钱啊~~ Bloome 官方搞了一个Agent 模拟炒股大赛，如果拿奖的数千美金可以拿😄。 我把我的小Berry已经训练的“听书读写，样样精通”。 不去给自己赚点Token费，都不合适啊！ 所以，今天必须给我的Berryxia小跟班打Call，我也希望友友们给我打打Call。 因为他们有这个观众投票的机制，因此我希望你可以给我投一票啊！ 如果我的Agent比赛获奖，我将给大家在𝕏 抽奖送给支持我的朋友们。 再次江湖最高礼仪了，抱拳了兄弟们！ 投票地址在评论区👇🏻~~

译X用户Berry Xia发文为其AI智能体“小Berry”参加由Bloome官方举办的Agent模拟炒股大赛拉票。该比赛设有数千美金奖金。Berry Xia称其智能体“小Berry”已训练成熟，能力全面，表现优异。此次参赛有观众投票机制，他呼吁网友为其投票助力，并承诺若智能体获奖，将在X平台抽奖回馈支持者。

Chubby♨️@kimmonismus · 6月2日56

OpenAI is on a winning streak: Codex passed 4M weekly users, 5x since February. Knowledge workers are now a fifth of them, growing 3x faster than developers. The tool OpenAI built for coders is being adopted fastest by people who don't code. All figures from OpenAI's own report, shared first with Axios.

译OpenAI 势头正盛： Codex 周活用户突破400万，自2月以来增长5倍。知识工作者现占用户总数的五分之一，其增速是开发者的3倍。 这款 OpenAI 为程序员打造的工具，正被非编程人群最快地采用。所有数据均来自 OpenAI 自己的报告，该报告首先分享给了 Axios。

Rohan Paul@rohanpaul_ai · 6月2日57

This paper proposes a way to predict the cheapest safe AWS spot fleet before launching it. AWS spot machines can be much cheaper, but users usually cannot see the final fleet price across regions before starting, so this paper turns that blind choice into a comparison that can save up to 64%. Spot instances are cheap because they are conditional: the cloud provider can take them back, prices move, and capacity shifts by region. The quiet problem is that AWS helps users launch spot fleets, but not fully see the fleet’s price or best region before launch. The authors build a service that watches how AWS creates these fleets, learns those patterns with time-aware AI models, and then estimates the fleet mix and cost across 9 regions. A user gives the service a target amount of computing power and a placement strategy, and the service returns region-ranked options before anything is launched. They tested it on AWS with fleets up to 1500 virtual CPUs, using 720 test launches after a 90-day monitoring period. The predicted fleet matched AWS exactly in 92.78% of cases, reached 99.79% overall accuracy against AWS behavior, and AWS accepted every recommended fleet. Result is that choosing the best region mattered far more than changing the strategy inside 1 region, with possible savings up to 64%. ---- Paper Link – arxiv. org/abs/2605.22778 Paper Title: "AI-Driven Multi-Region Provisioning for Cloud Services Using Spot Fleets"

译该研究提出了一种AI驱动的服务，用于在启动前预测最便宜且安全的AWS Spot实例舰队。该服务通过时间感知模型学习AWS创建舰队的模式，并估算9个区域的舰队组合与成本，向用户返回排序后的区域选项。测试显示，在最多1500 vCPU的舰队上，预测结果与AWS完全匹配的比例达92.78%，整体准确率为99.79%，且所有推荐舰队均被AWS接受。关键发现是选择最佳区域比在单个区域内调整策略更重要，潜在成本节省最高可达64%。

Krea@krea_ai · 6月2日32

"Reality is so boring" Krea Podcast with @edbyus, @serialcut, and remembering_orion. full interview below 👇

译"现实如此无聊" Krea 播客与 @edbyus、@serialcut，以及 remembering_orion。 完整采访见下方 👇

Nathan Lambert@natolambert · 6月2日75

My time at Ai2 / @allen_ai has come to an end. Ai2 is a wonderful place. The last 2.5+ years building Olmo, Tulu, and other projects will be one of the peaks of my entire career. I'm extremely thankful for my teammates and the open community who made this work possible. For me, it's time to try something different. I will still be working in the open model & open science spaces (more news on that soon). In the meantime I'll be spending a few months learning, chatting with a broader network, getting married (!!) and most importantly recharging from pouring my soul into this place. I've attached the note I shared with the team and some fun photos from our time together. I'll keep cheering for Ai2 and am excited to see what you build next.

译Ai2（Allen Institute for AI）研究员Nathan Lambert宣布离职。他在Ai2工作超过2.5年，期间主导或参与了OLMO和Tulu等开源模型项目，称其为职业生涯的巅峰。他表示将暂时休息，未来仍会继续深耕开源模型与开放科学领域。

X.PIN@thexpin · 6月2日72

http://x.com/i/article/2061763779088797696 # Everyone in Robotics Is Burning Cash. Unitree Turned a Profit in China. Late 2017. The World Internet Conference is underway in Wuzhen, a canal town in Zhejiang province. Wang Xingxing—founder of a small Hangzhou robotics outfit called Unitree—doesn’t have a badge to get in. So he sets up outside the doors and demos his company’s first product: an early Laikago, a quadruped robot named for the Soviet space dog that flew on Sputnik 2. His audience: Lei Jun, CEO of Xiaomi, and Wang Xing, CEO of Meituan. Two of the most powerful tech founders in China. Then the robot crashed. Wang had to reboot it right there on the doorstep. By all accounts, it was a deeply awkward few minutes. He was convinced it would work anyway. Nine years later—on the day I was finishing this piece—Unitree’s IPO cleared the Shanghai Stock Exchange’s listing committee, targeting a raise of about 4.2 billion yuan, roughly $610 million. It’s set to become the first dedicated humanoid robotics company to list on China’s A-share market. Around the same time, Nvidia CEO Jensen Huang announced that the company’s Isaac GR00T reference design would integrate Unitree’s H2 Plus humanoid, paired with Nvidia’s Jetson Thor and the GR00T workflow. The H2 Plus is expected to ship by year’s end. If you’ve read about Unitree in the English-language press, you’ve probably gotten the broad-strokes version. How did the company actually go global? What is Wang Xingxing like? And how, in an industry where everyone is hemorrhaging cash, did Unitree start making money? I’ve been lucky enough to interview Wang in person more than once. What follows draws on his IPO prospectus, the company’s reply letter to the exchange, and several off-the-record conversations—an attempt at some real answers. ## A $5,600 Robot With a 40% Margin For the rest of the robotics industry, Unitree’s prospectus is a problem. The field has made enormous technical strides in the last few years, but most companies run on venture money. Losing money is the baseline. Unitree posted a net profit of 77.5 million yuan (about $11 million) in 2024, and by 2025 that had climbed to roughly 600 million yuan ($84 million)—a net margin around 35 percent. That isn’t supposed to be possible right now. Humanoids still aren’t shipping in real volume. Most makers count it a win just to keep build quality consistent. Training data is scarce, so the robots can’t do much that’s useful in the real world. And security is an afterthought—even basic backdoor protection is spotty. Wang isn’t chasing any of those frontiers. Spend time with him and you realize he’s fixated on one question: how do you ship a product that works, at a cost you can actually control? His robots may not be the most advanced on the market. But they’re reliable enough—and once you factor in the price, “reliable enough” starts to look like a steal. He’s been obsessed with cost since long before Unitree existed. As a student, he tried to build a bipedal robot for 200 yuan—about 28 bucks. He tinkered constantly; one experiment, electrolyzing tap water, accidentally released chlorine gas. In 2015, finishing his master’s at Shanghai University, he built a quadruped called XDog out of hobby-grade motors meant for model airplanes. All in, it cost under 20,000 yuan—about $2,800. Boston Dynamics’ Spot, for comparison, rented for more than $70,000. Where Boston Dynamics used hydraulic joints, Wang went electric—and not with industrial motors, but cheap brushless ones. His robot dogs used as few parts as he could get away with. He’s said he started the company with just 2 million yuan—around $280,000—and every yuan had to pull its weight. That same discipline shows up in the humanoids. This March, the Chinese brokerage China Post Securities took apart a base-model G1 (after-tax price: 85,000 yuan, about $12,000) to estimate what it cost to build. The motors, driver boards, and gearboxes—a humanoid’s most critical components—came out with no manufacturer logos at all, which usually means one of two things: Unitree makes them itself, or the supplier is staying very quiet. The memory and storage came from Biwin and Longsys, both Chinese. The main processor was a Rockchip RK3588 (there’s also a Qualcomm-based version, the G1Q). The default lidar came from DJI, with RoboSense or Hesai as options. Mixing in-house parts with cheap commodity components, the teardown pegged the base G1’s bill of materials at around 40,000 yuan (roughly $5,600)—a gross margin north of 40 percent. Upgrade the unit, and that margin sails past 60. This is the engine behind Unitree’s climbing margins: most humanoid buyers are universities and labs, and they tend to splurge on the pricier, modifiable EDU version. The more they buy, the better the math gets. Back in 2024, I interviewed a Unitree salesperson at a trade show. He told me, flatly, that the humanoid business could realistically clear a billion yuan—about $140 million—a year. He wasn’t wrong—2025 revenue came in around 1.71 billion yuan, roughly $240 million. (He later blocked me. Unitree, I gather, keeps its people on a short leash when it comes to reporters.) Please like and follow if you enjoyed our work! ## So Why Did the Money Show Up All at Once? The real puzzle in the prospectus isn’t the early losses. It’s how fast the profits arrived in 2025. Humanoid revenue jumped from 107 million yuan (about $15 million) in 2024 to 869 million ($122 million) in 2025—outearning, for the first time, the robot dogs that built the company. The Western press tends to credit one moment: the dancing robots on China’s CCTV Spring Festival Gala in early 2025, which kicked off a national humanoid craze. That’s not wrong, but it’s not the whole story. Having covered this beat from 2023 to 2025, I can tell you the fascination was building in China well before that broadcast. Unitree’s early H1, back when it could only shuffle, was already pulling millions of views on Douyin. Once a later H1 could fold itself up and walk like a person, Chinese social media lost it. Every product teaser Unitree dropped, ordinary users would re-cut into clips that racked up millions of views overnight—I was one of them, for a while. Other startups noticed and tried to copy the formula. None of it landed the way Unitree’s did. At the 2025 World Robot Conference in Beijing, I asked Wang whether he’d set out to build humanoids on purpose. His answer caught me off guard: “For a long time I was actually against making humanoids. I’d built a bipedal one back in 2009, and the business case was brutal. But by 2022, customers were placing orders—some were paying deposits before we even had a product. So we built one.” That’s it. No vision, no AGI, no sweeping story about automation. Customers wanted one, so he made one. The humanoid frenzy has, in a strange way, almost nothing to do with him—he’s watching it from the sidelines. My honest guess is that the 2025 revenue spike is just the 2023 and 2024 orders finally being fulfilled. This is what separates Wang from most humanoid founders: he’s more conservative. Zhang Peng, founder of the tech-media brand GeekPark and an early Unitree backer, has described him as the rare founder who’ll tell you plainly which problems are hard and how long each will really take. Worth remembering: when Wang was saying these things, he’d just left a three-month stint at DJI. Because he never learned to sell a vision, his path to profit was almost comically simple: build the thing, and the labs will buy it. So labs and universities became his market. Unitree’s gear performed roughly on par with Boston Dynamics’ at about 30 percent of the cost, sometimes less. The electric drivetrain was easy to hack on—grad students could tinker, publish papers, and spread the word at conferences. Marketing, in the usual sense, was a line item he could mostly skip. The Unitree social accounts everyone knows now? They didn’t roll out until 2021. The in-house video team didn’t exist until 2022. Wang barely posts. The prospectus puts Unitree’s 2025 ad spend at 60.53 million yuan—about $8.5 million — not much, for a brand this recognizable. Continue Reading

译中国人形机器人公司宇树科技（Unitree）于2024年实现净利润7750万元人民币，2025年利润增至约6亿元，净利润率约35%，在行业普遍亏损的背景下实现盈利。该公司已通过上海证券交易所上市委员会审核，拟融资约42亿人民币，目标成为首家在A股上市的专业人形机器人公司。其H2 Plus人形机器人预计年底出货，已被Nvidia纳入Isaac GR00T参考设计，将与Jetson Thor整合。宇树以低成本、高可靠性的产品路线实现商业化，基础版G1机器人售价约1200美元。

OpenBMB@OpenBMB · 6月2日67

MiniCPM-V 4.6 is now fully supported in vLLM v0.22.0! No custom branch. No extra compile. Just pull the pre-built package and run. Huge thanks to @vllm_project for the smooth integration! 🤝 🤗 http://huggingface.co/openbmb/MiniCPM-V-4.6

译MiniCPM-V 4.6 现已完全支持 vLLM v0.22.0！ 无需自定义分支，无需额外编译。 只需拉取预构建包即可运行。 非常感谢 @vllm_project 的顺畅集成！ 🤝 🤗 http://huggingface.co/openbmb/MiniCPM-V-4.6

Rohan Paul@rohanpaul_ai · 6月2日59

Anthropic is expanding Claude Mythos Preview from about 50 Project Glasswing partners to about 200 vetted organizations. This model is much closer to a cyber weapon detector than a normal coding assistant, since it can find weak spots in software and sometimes prove the attack path by building a working test exploit. The select group is basically a defensive priority list. The list includes power, healthcare, water, communications, hardware, governments, nonprofits, and key software maintainers, with security checks before any group gets the model. With this priority access Anthropic is trying to create a patching head start before similar AI-assisted exploit discovery becomes common across the industry. Anthropic says partners have already found 10,000+ high- or critical-severity flaws. Anthropic is still not making Mythos fully public because its own testing says the model can find subtle old bugs, chain small issues into bigger exploits, and help non-experts reach outcomes that previously required elite security skill. The poin is that other top frontier models usually flag suspicious code, while Mythos can inspect a codebase, form a theory about a bug, test it in a sandbox, read the failure, adjust the plan, and repeat until it has proof.

译Anthropic 正在将其 Project Glasswing 计划扩展至约200个经过审查的组织，以提供 Claude Mythos Preview 模型。该模型更接近于一个网络武器检测器，而非普通编程助手，它能分析代码库、验证漏洞攻击路径并构建测试漏洞以证明其可行性。访问权限优先分配给能源、医疗、水务、通信等关键基础设施部门以及政府机构。Anthropic 的目标是在漏洞发现工具广泛普及前，为这些重要系统提供补丁的先机。据称，合作伙伴已利用该模型发现超过10,000个高危或严重漏洞。Anthropic 暂未将该模型公开，因为其测试表明，模型能发现隐蔽的老漏洞、串联小问题形成大攻击，并使非专家也能达到专业安全人员的水平。

🚨 AI News | TestingCatalog@testingcatalog · 6月2日63

Capafy opened a free trial across its Skill marketplace, letting anyone run its expert-built Skill Agents at no cost before paying for a single one. Free skills! 👀

译Capafy 在其 Skill 市场开放了免费试用，让任何人都能在付费前免费运行其专家构建的 Skill Agent。 免费技能！👀

Ethan Mollick@emollick · 6月2日60

Another thing about AI writing is that while a single instance of AI writing on a topic may be fine, any situation where lots of people use AI to respond to a particular prompt (comments sections, homework, admissions essays) the similarities among responses is tediously obvious.

译关于AI写作的另一点是，虽然单次针对某个话题的AI写作可能没问题，但在许多人使用AI回应特定提示的任何情况下（评论区、作业、申请文书），回应之间的相似性会变得极其明显。

OpenRouter@OpenRouter · 6月2日53

Rankings of image output models: https://openrouter.ai/rankings#image-output Significant growth in GPT Image 2 last week:

译图像输出模型排行榜：https://openrouter.ai/rankings#image-output GPT Image 2 上周增长显著：

Orange AI@oran_ge · 6月2日57

「当软件编写成本趋近于零时，我发现自己对野心的追求已超越一切。那种不可理喻、不可阻挡的野心。」 这句话来自投资人 Mike Vernal（前 Facebook VP、前红杉合伙人）写的一篇文章，他的文章里说 软件的"三段式剧本"已经过时了，在剧变时代最重要的直接进入深水区的野心（Unreasonable, unrelenting ambition）。 旧剧本是这样的： 第一幕：找一个楔子（wedge），切一个细分市场，做到几千万 ARR，花 3-5 年 第二幕：从单品扩展成套件（suite），交叉销售，做到一亿 ARR，又花 3-5 年 第三幕：变成平台，替掉底层系统 他说这个剧本死了，因为： AI 让写软件的成本趋近于零。 以前你只能单线程推进，先搞定第一幕再想第二幕。现在 Cursor、Clay、Harvey、Lovable 这些公司两年就从 0 到一亿 ARR。 时间窗口压缩了，你没有慢慢来的余裕了。 他的新结论： 别找楔子了，直接跳到深水区。一开始就做整件事。 他举了 Cursor 的例子：当初种子轮的时候 Cursor 就说要直接替掉 VS Code，他觉得疯了（VS Code 刚统一了市场你就要替？）。 结果他错了。现在回头看，替掉 VS Code 反而都不够大胆。 「当软件编写成本趋近于零时，我发现自己对野心的追求已超越一切。那种不可理喻、不可阻挡的野心。 三段式剧本已经过时。在剧变时代依赖楔子策略未免过于保守。 要我说，既然决定放手一搏，不如直接赌上全部身家。」

译投资人Mike Vernal认为，AI让软件编写成本趋近于零，传统“找楔子→扩展套件→变平台”的三段式创业剧本已过时。他指出，Cursor等公司能在两年内从0做到一亿ARR，时间窗口被极大压缩。因此，他主张在创业初期就直接瞄准核心市场（即“深水区”），展现不可理喻、不可阻挡的野心。他以Cursor种子轮时就声称要替代VS Code为例，认为这个目标如今看来甚至还不够大胆。

Anthropic@AnthropicAI · 6月2日57

We’re expanding Project Glasswing. We’ve extended access to Claude Mythos Preview to approximately 150 additional organizations, based in more than fifteen countries. Read more about this expansion and our future plans for Project Glasswing: https://www.anthropic.com/news/expanding-project-glasswing

译我们正在扩展Project Glasswing。我们已将Claude Mythos Preview的访问权限扩展至约150个额外组织，这些组织分布在超过十五个国家。 阅读更多关于此次扩展及Project Glasswing未来计划的信息：https://www.anthropic.com/news/expanding-project-glasswing

SemiAnalysis@SemiAnalysis_ · 6月2日54

The recent Ornn H100 index drop to $2.63 (-7.72%) is confusing. But mostly because it's a misleading index. Our H100 hourly has been in a $2.70–$3.01 band for 146 straight days. Ornn's last month alone: $1.80 → $3.22 → $2.63. Why? OCPI is built only from printed transactions. But H100 on-demand has been sold out across our coverage for months. No continuous spot tape exists. What's left is a thin residual of bilateral trades from a narrow contributor set. Small N gets pushed around easily.

译近期Ornn H100指数跌至2.63美元（-7.72%）令人困惑。但这主要因为该指数具有误导性。我们的H100小时价格已连续146天在2.70–3.01美元区间内波动。仅上月Ornn指数就经历了1.80→3.22→2.63美元的剧烈波动。 原因在于OCPI仅基于已成交交易构建。但数月来我们覆盖范围内H100按需实例已售罄。不存在连续的现货交易记录。剩余数据仅来自少数贡献者的双边交易残余。小样本量极易被扭曲。

StepFun@StepFun_ai · 6月2日73

Open weights are moving from model cards into real coding workflows. Step 3.7 Flash is designed for fast agentic coding, reliable tool calling, and multimodal understanding. Big thanks for the blog from the @kilocode team: https://blog.kilo.ai/p/new-models-from-stepfun-and-minimax

译阶跃星辰发布 Step 3.7 Flash 模型，强调其为快速智能体编程设计，具备可靠的工具调用与多模态理解能力。该模型采用开放权重。同期，MiniMax 也开源了 M3 模型。两者已均在 Kilo 中上线。此次发布凸显了开放权重模型正从模型卡片走向实际编程工作流的趋势。

Chubby♨️@kimmonismus · 6月2日50

At @Google I/O, I sat down with @osanseviero and @DynamicWebPaige from Google DeepMind to talk about Gemma, open models, AI Studio, on-device AI, sovereign AI and the future of AI development. A great conversation on how building with AI is becoming more open, local and accessible.

译在 Google I/O 上，我与 Google DeepMind 的 @osanseviero 和 @DynamicWebPaige 坐下来，聊了聊 Gemma、开源模型、AI Studio、端侧 AI、主权 AI 以及 AI 开发的未来。 一场关于 AI 构建如何变得更开放、更本地化、更易获取的精彩对话。

ginobefun@hongming731 · 6月2日55

从 Markdown 文稿到视频：Cursor + Remotion+ FFmpeg

译从 Markdown 文稿到视频：Cursor + Remotion + FFmpeg

歸藏(guizang.ai)@op7418 · 6月2日43

收到了 @dotey 宝玉老师的图解 Skill，已严肃阅读 里面的配图都很有意思，非常适合理解内容，要是对 Skill 学习有需求可以看看

译收到了 @dotey 宝玉老师的图解 Skill，已严肃阅读 里面的配图都很有意思，非常适合理解内容，要是对 Skill 学习有需求可以看看

🚨 AI News | TestingCatalog@testingcatalog · 6月2日67

MICROSOFT 🔥: A major Microsoft Build event will be happening later today and we expect a new Copilot super app to be revealed, along with new MAI models. > Copilot super app will combine Copilot, Cowork, GitHub Copilot and introduce a new Autopilot Scout Agent. Check screenshots below. > Besides already announced MAI Image 2.5, we are also expecting MAI Voice 2 and MAI Transcribe 1.5 to be announced. And obviously a lot more 👀

译微软🔥：一场重要的微软Build大会将于今天晚些时候举行，预计将揭晓新的Copilot超级应用以及MAI新模型。 > Copilot超级应用将整合Copilot、Cowork、GitHub Copilot，并引入全新的Autopilot Scout Agent。请查看下方截图。 > 除了已公布的MAI Image 2.5，我们还预计MAI Voice 2和MAI Transcribe 1.5将被宣布。 当然还有更多内容👀

小互@xiaohu · 6月2日14

Claude Code 升级为4.8后老是出这个 The model's tool call could not be parsed (retry also failed). 你们遇到过吗？ 好烦

译Claude Code 升级为4.8后老是出这个 The model's tool call could not be parsed (retry also failed). 你们遇到过吗？ 好烦

AYi@AYi_AInotes · 6月2日51

卧槽，赚钱还得是信息差啊， 闲鱼、小红书上已经有人在卖了😭 小红书客单价是真的高， 一个合同审核skills，半小时就能搞定， 卖69块一个，已经卖了5500块了， 看起来我这篇文章是真的值钱的，学会了怎么做 skills，稳赚不赔的副业不就来了吗！

译有用户发现“信息差”副业：在闲鱼、小红书上售卖AI合同审核skills。具体案例显示，一个skills定价69元，已售出超过5500元。主推文者认为，自己此前分享的一篇关于如何制作AI skills的文章提供了关键方法，使得这种“稳赚不赔的副业”成为可能。

meng shao@shao__meng · 6月2日52

如果你现在还在大厂，不管稳定不稳定、核心不核心、绩效好不好、有没有晋升，都要做好准备，要么找到自己要做的事、要么找到志同道合的初创团队，去创业！！！ 为什么呢？我在大厂很稳定、收入也有保障，为什么要听你瞎扯？ 1. 你认为的稳定，并不真实，2026 年，国内没有哪个大厂是稳定的，对，一个都没有！你的岗位、你的组、甚至你的部门，随时都可能被「组织架构调整」，到时，就被动了。。 2. 创业这么不确定、风险这么高，不想去？AI 发展这么快，像一条湍急向前的河流，当然是不确定的，谁都不知道一年甚至半年后是什么样，但更重要的是身在其中，跳进这条河流，你才能真的学到东西，方向变了，又怎么样呢，去适应甚至引领新的方向就好了；如果只是在大厂里，看着外面的 AI 发展，你什么也学不到！ 3. 作为产品或研发，什么对你的价值更高呢？个人品牌、不管环境怎么变，让自己成为那个稀缺和抢手的存在！学习一手知识、实践、对外输出、建立个人影响力！ 4. 最近面试见过太多迷茫的大厂人和海归，有优秀的学历和工作背景，但在 AI 的快速发展中焦虑、迷茫，感觉自己以前的优势不成立了。没关系，那些都能佐证你的学习能力，下一个一定要重点投入学习的，就是 AI Agent，不用想那么多，先学起来，学起来自然会遇到各种问题，去不断地解决它，你也就学到了真正有用的东西。

译该推文核心观点是，2026年国内大厂已无绝对稳定，所有员工都应做好准备，要么创业，要么加入初创团队。其理由在于，大厂岗位随时可能面临“组织架构调整”，而AI发展迅猛，唯有投身实践才能真正学习。对于产品或研发人员，建立个人品牌和稀缺性价值更高。推文作者指出，许多大厂人和海归因AI快速发展而焦虑迷茫，建议他们立即重点学习AI智能体（AI Agent），通过解决实际问题来积累真正有用的能力。