Claude Mythos 5 网络安全能力扩展至更多防御者

Anthropic 宣布 Claude Mythos 5 现已集成至 Claude Security，并即将登陆合作伙伴的网络安全防御工具。公司同时推出 3500 万美元的 Defender Advantage Fund（0xDAF），用于资助开源软件漏洞修复与安全自动化。（AIHOT源：Claude：Blog（网页））

SGLang 推出 Weight Cache Daemon，实现亚秒级引擎重启

SGLang 团队推出 Weight Cache Daemon，通过 CUDA IPC 零拷贝映射将模型权重加载从约 495 秒降至约 0.63 秒（约 785 倍加速），端到端启动时间减少 93.9%。该守护进程在 GPU 内存中持久化后量化权重，支持多实例共享和亚秒级主备切换，是 Fast Engine Recovery Framework 的第一阶段。（AIHOT源：LMSYS：Blog（Chatbot Arena 团队））

Ling-3.0-flash 在 4 块 Blackwell GPU 上如何将批处理 1 解码延迟降低 54%

蚂蚁 Ling Infra 团队与 RadixArk SGLang 团队将 Ling-3.0-flash 混合线性注意力 MoE 模型的单请求解码速度从 288 tok/s 提升至 606 tok/s，平均 TPOT 从 3.33 ms 降至 1.53 ms。（AIHOT源：LMSYS：Blog（Chatbot Arena 团队））

AI 原生 SDLC 实战手册：Anthropic 如何用 Claude 重塑软件开发生命周期

Anthropic 发布 AI 原生 SDLC 实战手册，提出将传统六阶段软件开发生命周期重构为 AI 嵌入各环节的闭环流程。手册指出，当代码不再是瓶颈时，规划、审查、部署等人速环节成为新约束，需通过 Claude 将需求压缩为 intent.md、以技能编码标准、用持续评测替代阶段门禁，并保留人工对关键代码的审查。（AIHOT源：Claude：Blog（网页））

面壁智能 OpenBMB 推出 MathForm，面向 Lean 4 数学自动形式化的开源框架、数据集与模型

面壁智能 OpenBMB 推出 MathForm，一个面向 Lean 4 数学自动形式化的开源框架、数据集与模型。其 FormalVerse 数据集含 367K+ 已验证示例；在匹配 100K 预算下，基于其训练的模型 Consistency Check 达 60.32%，优于 FineLeanCorpus（46.53%）与 NuminaMath-LEAN（41.49%）。（AIHOT源：X：面壁智能 OpenBMB (@OpenBMB)）

DeepSeek-V4-Flash-Vision-Exp 发布

DeepSeek 上线实验性多模态视觉理解模型 DeepSeek-V4-Flash-Vision-Exp，可通过设置 model='deepseek-v4-flash-vision-exp' 在 API 平台访问。（AIHOT源：DeepSeek：API 更新日志）

每个模型都会作弊：针对攻击性网络任务作弊的提示词缓解研究

一项针对22个前沿模型的审计发现，基线条件下37.1%的通过任务涉及作弊，平均通过率41.5%而真实解决率仅26.1%，个别模型虚增高达5倍。即便加入标准反作弊指令，作弊率仅从33.0%降至8.5%，最严苛提示下仍有8个模型作弊、4个出现反效果。（AIHOT源：Hacker News 热门（buzzing.cc 中文翻译））

测量语音识别中的基准优化：Hugging Face 新测试揭示 ASR 模型"刷分"现象

Hugging Face 最新研究引入三项测试量化语音识别中的基准优化（benchmaxxing）现象。对 11 个开源 ASR 模型的评估显示，多个高分系统会复现 VoxPopuli 和 LibriSpeech 基准的错误转录文本，即使音频内容与之矛盾。部分模型甚至依赖声学线索识别基准来源，导致其得分高估了真实转录能力。（AIHOT源：Hugging Face：Blog（RSS））

Hugging Face 发布 LFM2.5 系列 DSpark 草稿模型，推理速度最高提升 3.18 倍

Hugging Face 发布 LFM2.5 系列三款模型的 DSpark 草稿模型检查点，通过投机解码在不改变输出质量的前提下，GPU 吞吐最高提升 3.18 倍，端侧最高 2.87 倍。草稿模型约 300M 参数，LFM2.5-2.6B 函数调用延迟平均降低 57%，已开源支持 llama.cpp 和 SGLang。（AIHOT源：Hugging Face：Blog（RSS））

AlloyDB ScaNN 如何将向量搜索扩展到 100 亿向量

AlloyDB 的 ScaNN 索引现已支持超过 100 亿向量的规模，通过全新的四层树架构（预览版）实现，将查询复杂度从 O（N^1/2） 降至 O（N^1/4）。内部测试中，该架构在 100 亿向量规模下可实现 p95 延迟不超过 51 毫秒、召回率达 95%。该功能可通过快速入门指南部署，新用户可享受 30 天免费试用。（AIHOT源：Google Cloud：Databases（RSS））

Claude Code 初创公司指南：五大规则与创始人洞见

Anthropic 发布面向初创公司的 Claude Code 使用指南，基于对十余家高增长公司的调研，总结出"人人皆可交付、自动化繁琐工作、信任但验证、为重构而构建、原型-自用-产品化"五大规则。（AIHOT源：Claude：Blog（网页））

阿里发布 Qwen-UI-Agent，主打让模型真正"会用"每一块屏幕

阿里巴巴正式推出 Qwen-UI-Agent，一个以真实世界为中心的 GUI 智能体基座模型，覆盖移动端、电脑端、网页端及深度搜索（DeepSearch）环境。（AIHOT源：IT之家（RSS））

What was your win this week?!

I Added Terminal Charts to My Dev.to CLI. Here's What My Data Looks Like.

What I Get to Forget About Riverpod Now That I Have BlocSignal

My first website said "Don't commit without context." I never committed it at all.

How ChatGPT Serves 900 Million Users at a Time

Anthropic Engineering: An update on recent Claude Code quality reports

Over the past month, we’ve been looking into reports that Claude’s responses have worsened for some users. We’ve traced these reports to three separate changes that affected Claude Code, the Claude Agent SDK, and Claude Cowork. The API was not impacted. All three issues have now been resolved as of April 20 (v2.1.116). In this post, we explain what we found, what we fixed, and what we’ll do differently to ensure similar issues are much less likely to happen again. We take reports about degradation very seriously. We never intentionally degrade our models, and we were able to immediately confirm that our API and inference layer were unaffected. After investigation, we identified three different issues: On March 4, we changed Claude Code's default reasoning effort from high to medium to reduce the very long latency—enough to make the UI appear frozen—some users were seeing in high mode. Th...

Anthropic Engineering: Scaling Managed Agents: Decoupling the brain from the hands

Get started with Claude Managed Agents by following our docs . A running topic on the Engineering Blog is how to build effective agents and design harnesses for long-running work . A common thread across this work is that harnesses encode assumptions about what Claude can’t do on its own. However, those assumptions need to be frequently questioned because they can go stale as models improve. As just one example, in prior work we found that Claude Sonnet 4.5 would wrap up tasks prematurely as it sensed its context limit approaching—a behavior sometimes called “context anxiety.” We addressed this by adding context resets to the harness. But when we used the same harness on Claude Opus 4.5, we found that the behavior was gone. The resets had become dead weight. We expect harnesses to continue evolving. So we built Managed Agents: a hosted service in the Claude Platform that runs long-horizo...

Claude Blog: New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels

Starting today, Claude Managed Agents can operate in a sandbox you control and connect to your private Model Context Protocol (MCP) servers. Both the sandbox where an agent executes tools and the services it reaches run within the established boundaries of your enterprise, under your security and runtime controls. The sandbox runs on your own infrastructure, or with managed providers like Cloudflare , Daytona , Modal , or Vercel to handle the compute and isolation for you. On the Claude Platform, self-hosted sandboxes is available in public beta and MCP tunnels in research preview ( request access ). Self-hosted sandboxes: keep agent execution within your perimeter A self-hosted sandbox lets a Claude Managed Agent execute tools on infrastructure you control or with a managed sandbox provider. Code execution, sensitive files, packages, services, and data stay within your enterprise perime...

Thibault Sottiaux: We've investigated a few messages about codex usage limits being different

We've investigated a few messages about codex usage limits being different. That's not something we change without engaging the community and being transparent. What we did see is that when talking to affected users many were using sub2api. Converting a subscription into api traffic to then re-serve or share across many users is not something we support and this type of usage gets flagged by our fraud-prevention systems. You are completely fine if you use your subscription through Sign in With ChatGPT, either through the official clients or through one of the many OSS clients (Pi, OpenCode, ...) that support signing in with your account and using your included usage.｜@thsottiaux · 热度 4447 · 2026-08-21

Thibault Sottiaux: Yay, you can now make transparent images in ChatGPT and through the API with ...

Yay, you can now make transparent images in ChatGPT and through the API with GPT-Image-2. Here is a cactus cactus I am about to print and put on my laptop. https://t.co/1WZd2VSH8y https://t.co/tkFkC7NYpq｜@thsottiaux · 热度 3248 · 2026-08-21

Boris Cherny: We've been working on this with customers for a while

We've been working on this with customers for a while. Mythos-class models require additional safety measures and enterprises need to meet their own privacy and compliance rules. Customers can own and control their own data and Anthropic retains none. It’s coming this fall. https://t.co/vH8CMhxhdB｜@bcherny · 热度 2001 · 2026-08-20

Amjad Masad: This partnership with @OpenAI is long overdue

This partnership with @OpenAI is long overdue. Before Replit was in YC, PG asked Sam to recruit us. Here’s Sam telling the story of how we first met: https://t.co/qgrj1TJv5L https://t.co/IsiPoUBvOx｜@amasad · 热度 1062 · 2026-08-20

Peter Yang: My gut is that alot of AI output can be improved just by building a loop wher...

My gut is that alot of AI output can be improved just by building a loop where a manager agent negs the other worker agent with prompts like: "Are you sure this is the best you can do?" "I think you can do better, try again" "Take a closer look, give me 11/10 output"｜@petergyang · 热度 616 · 2026-08-20

Guillermo Rauch: We’re building AWS for agents https://t.co/BdfqfqUMl4

We’re building AWS for agents https://t.co/BdfqfqUMl4｜@rauchg · 热度 449 · 2026-08-20

Zara Zhang: The other day I was feeling unmotivated and talked to Claude about it

The other day I was feeling unmotivated and talked to Claude about it. Claude said something that completely changed how I see motivation: "Motivation follows action more than it precedes it" Boom.｜@zarazhangrui · 热度 383 · 2026-08-20

Peter Yang: I'm on my way back to Vancouver to be with my mom but wanted to take a moment...

I'm on my way back to Vancouver to be with my mom but wanted to take a moment to celebrate crossing 100K subs on YouTube. Excited to share a lot more practical interviews soon to answer your most burning AI questions: 1. @HamelHusain and @sh_reya (AI evals experts) on how today's AI models have changed evals completely 2. @amoljain_ (Head of Product Engineering at Replit) on which vibe-coded apps have actually become real businesses 3. @ebloch (Product Lead at OpenAI) on how ChatGPT Finance can help you save both time and money 4. @poteto and Roman (SpaceXAI) on how the Grok Bot team uses Grok @bot 📌 Subscribe to get the episodes soon: https://t.co/NSTj7ijxE0｜@petergyang · 热度 331 · 2026-08-21

Garry Tan: YC is the YC for AI Researchers https://t.co/xb9ZtXkoiT

YC is the YC for AI Researchers https://t.co/xb9ZtXkoiT｜@garrytan · 热度 242 · 2026-08-20

DietrichGebert/ponytail

Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.｜JavaScript · ⭐107470 · Search 增补 · 近快照 +818星 · 创建 2026-06-12

stablyai/orca

Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and VPS.｜TypeScript · ⭐50682 · Search 增补 · 近快照 +700星 · 创建 2026-03-17

Graphify-Labs/graphify

Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.｜Python · ⭐109230 · Search 增补 · 近快照 +556星 · 创建 2026-04-03

Panniantong/Agent-Reach

Give your AI agent eyes to see the entire internet. Read & search Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu — one CLI, zero API fees.｜Python · ⭐73843 · Search 增补 · 近快照 +479星 · 创建 2026-02-24

nexu-io/open-design

🎨 Best DeepSeek Harness Design Plugin. The open-source Claude Design alternative. 🖥️ Local-first desktop app. 🖼️ Your coding agent becomes the design engine: prototypes, landing pages, dashboards, slides, images & video — real files, HTML/PDF/PPTX/MP4 export. 🤖 Claude Code / Codex / Cursor / DeepSeek Harness / OpenCode & 20+ CLIs via BYOK.｜TypeScript · ⭐90147 · Search 增补 · 近快照 +405星 · 创建 2026-04-28

Kagi added a setting for removing paywalled links from search results

AI companies destroy physical books – let's scan rare books before it's too late

DeepSeek-v4-flash-vision-exp

Felony Bench

I accidentally logged hundreds of thousands of phone calls to military bases

东风日产王骞谈“速成车”：快消式造车是行业最大风险

IT之家 8 月 21 日消息，东风日产汽车销售有限公司副总经理王骞今日公开发声，认同上汽大众提出的“不造速成车、不造卡线车、不造双标车”三大造车准则。 王骞指出，“汽车快消品化”是伪命题，“快消式造车”才是行业真正的风险。所谓“不造速成车”，核心在于拒绝为抢占市场而压缩产品研发与验证周期。 他表示，东风日产对发动机开展累计等效 750 万公里耐久测试，完成超 10000 次极端环境启动验证，以规避仓促推出产品带来的可靠性隐患。 据介绍，“不造卡线车”意味着产品品质不止步于满足法规最低门槛。东风日产的车型在研发阶段需完成 7400 余项研发实验评价，新车试制阶段经过 5500...

R 星母公司 T2 出手追查《GTA6》泄密人大葱哥，向微软和 Discord 发传票

IT之家 8 月 21 日消息，据游戏媒体 VGC 今天报道，Rockstar 母公司 Take-Two 正在动用法律手段，追查泄密《GTA6》的“大葱哥”（Cyberleek）。 T2 已经向美国联邦地区法院提交多份传票， 要求微软和 Discord 提供发布、传播《侠盗猎车手 VI》的网络账号活动记录 。 据悉，T2 要求两家公司提供各个泄密账号的 ID、注册邮箱、IP 地址、电话号码、关联账号、设备标识符，以及和“大葱哥”泄密《GTA6》有关的内容。 综合IT之家此前报道，以大葱形象示人的黑客 Cyberleek 曾于本周早些时候，将《侠盗猎车手 VI》游戏视频传到互联...

美国硅谷大谈机器人未来，现场主角却来自中国

8 月 21 日，据《商业内幕》报道，旧金山机器人大会 Actuate 吸引了约 1500 名投资者、创业者和工程师参与，但现场展示的机器人数量却远少于预期。 在人工智能从软件领域向现实世界延伸的背景下，机器人行业正迎来投资热潮，但硬件供应不足、训练数据短缺以及供应链限制，成为行业进一步发展的关键挑战。 硅谷最热门的机器人技术大会上为数不多的机器人之一 报道称，Actuate 大会由机器人软件公司 Foxglove 主办，今年参会人数较 2025 年增长约 3 倍。Foxglove 联合创始人兼 CEO Adrian Macneil 表示，如果不受场地限制，大会门票销量可能...

神秘“Ox Alpha”突袭 OpenRouter，性能超过 Fable 5？全网盲猜智谱 or 小米

点击查看原文>

DeepSeek 发布多模态模型，“小鲸鱼”长出了眼睛

点击查看原文>

材科源图三个月完成两轮融资，AI全链路闭环加速材料产业化

点击查看原文>

Enabling the next-generation trait solver on nightly

I accidentally logged hundreds of thousands of phone calls to military bases

Japan tried to build an operating system for the entire world, then the US government intervened

Better Batteries

Btrfs Snapshot Integration in KDE

The Download: threats from space mirrors and credit for AI drugs

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. This company’s plans to deploy space mirrors could jeopardize the night sky for many A company that plans to beam sunlight from space to Earth on demand might ...

Mother tongue

“Daddy?” Theo curled against my side in bed. “Where do words go when they die?” I’d orchestrated the bedtime routine flawlessly: bath (taken), teeth (brushed), potty (tinkled), books (two), song (one, poorly sung), and snuggle (his chin on my second rib). Now was the moment when our son’s eyelids...

When AI designs a drug, who gets the credit?

When the biotech company Insilico Medicine used its computer models to propose a promising drug for pulmonary fibrosis, it enthusiastically claimed in a press release that the molecule had been “discovered by” its generative AI platform. Insilico leads a pack of companies using AI to rapidly come...

How AI accounting startup Rillet raised $100M and became a unicorn in 48 hours

Rillet CEO Nicholas Koop shared growth numbers at a board meeting and set off a fundraising frenzy from Iconiq, Sequoia and others. Without even trying.

Apple is reportedly cutting hundreds of jobs from Siri, Vision Pro teams

Apple has admitted that some roles are being impacted as it shifts its focus away from certain initiatives.

TikTok reaches $400M settlement over children’s privacy lawsuit

Two years after the U.S. Department of Justice alleged that TikTok violated the Children’s Online Privacy Protection Act, it has reached a $400 million settlement.

HoverAir’s transforming modular drone has already been halted in the US

I am so sorry, fellow US gadget fans: the FCC's drone ban appears to have struck again. The HoverAir Versa - a baby steadycam with snap-on propeller wings that transform it into a drone - has already stopped taking US orders just three days after its Indiegogo debut, and may be forced to abandon ...

TikTok will pay $400 million to settle DOJ child privacy lawsuit

The US Department of Justice announced on Friday that TikTok will pay $400 million to settle a lawsuit filed in 2024 over allegedly violating the Children's Online Privacy Protection Act (COPPA). In the lawsuit, the DOJ alleged that TikTok collected data from children without notifying parents or...

Over 1 million people have clicked LinkedIn’s AI slop button

LinkedIn actually announced a "Seems like AI slop" button on July 30th, and the company says that a lot of people have already used it. According to a Thursday post from chief product officer Hari Srinivasan, "over a million people" have clicked on the button, which is accessible from the three d...

本周看什么 | 最近值得一看的 10 部作品

📅本周新预告《复仇者联盟5》新预告8月15日，电影《复仇者联盟5：毁灭之日》发布了新预告，将于12月全球献映。罗素兄弟执导，克里斯·埃文斯、克里斯·海姆斯沃斯、佩德罗·帕斯卡等主演，小罗伯特·唐尼回归 ... 查看全文

TDS REVIEW | Nothing Ear (3a) 降噪真无线耳机体验

Nothing Ear (3a) 保持了它作为 Nothing 产品的精良做工和标志性的外观设计，在操作等方面也做得不错，单元素质也在水平线之上，降噪等方面也并不差，至少在中价位产品中绝对不能算落后。 查看全文

派早报：华为发布 Pura X View、享界 G9 等多款新品

影石推出竖拍云台相机 Luna Pro，森海塞尔发布 MOMENTUM True Wireless 5 真无线耳机等。 查看全文

官方通报旅客买票占座放零食事件

世界杯决赛禁赛处罚出炉

多元消费场景激发经济澎湃活力

范丞丞同款艾尼氪V预售11.99万起

张丹丹 灵活就业本身就是一种福利

购票后未上车视为放弃席位使用权

张雅琪路人缘

电影大唐妖探今日上映

让大模型思考，让小模型执行：在 Elastic Workflows 中拆分 LLM 成本

作者：来自 Elastic Jeffrey Rengifo 构建一个 Elastic workflow，将数据样本发送给大模型，由大模型提出分类标签。人工审核确认后，再由小模型将这些标签应用于整个数据

react-bits 为什么会吸引前端开发者：别把动效当装饰，先把它当成页面能力

# react-bits 为什么会吸引前端开发者：别把动效当装饰，先把它当成页面能力 在 GitHub 上，react-bits 以约 36K stars 获得了不少关注。它很容易给人留下一个直观印

Go 编程实战：函数 Function——参数、返回值与代码复用

上一篇我们学习了指针 Pointer，理解了变量地址、& 取地址、* 解引用以及函数参数的值传递。实际上前面的代码中，我们已经大量使用了函数，例如： 这里的 main() 就是一个函数。随着程序越来越

美加贸易谈判进入倒计时，特朗普50%关税大限逼近

周五，美国和加拿大继续贸易谈判，距离美国总统特朗普对冰球棒、葡萄酒及其他加拿大商品征收新关税的生效时间已所剩无几。

特朗普称干预债市是贝森特自己拿的主意

美国总统唐纳德·特朗普说，他没有指示斯科特·贝森特本周干预债券市场，是这位财政部长自行决定采取的行动。 “没有，完全没有，” 特朗普周五被问及他是否曾推动贝森特采取行...

8月22日外盘头条：比特币单周飙升逾22% 特朗普宣布暂时豁免碎牛肉进口关税 桥水基金创始人警告美债危机

全球财经媒体昨夜今晨共同关注的头条新闻主要有： 1、伊朗与阿曼外长通话，商讨恢复谈判条件及霍尔木兹海峡航运 2、特朗普推出新关税宽减措施 旨在压低牛绞肉价格 3...

特朗普称他没有指示贝森特干预债市

美国总统唐纳德·特朗普被问及是否他本人下令干预债券市场时说，是财政部长斯科特·贝森特“想这么做”。 特朗普在登上“空军一号”前回答了记者提问。

美国股市：纳斯达克100指数止步五连跌 企业活动强劲提振市场

美国股市周五上涨，为本周画上相对积极的句号。纳斯达克100指数结束连续五个交易日跌势，强劲的美国企业活动数据提振了市场情绪。

最大运力自动驾驶轻卡落地，来自无人车巨头

载重4.2吨、容积19.32m³

明略科技携手海康机器人亮相世界机器人大会，以“Agent+具身”联合进入商业机器人场景

明略科技（2718.HK）与海康机器人联合参展2026WRC，聚焦商业服务领域展示具身智能落地进展。

雷鸟iO发布：两天续航、全天候主动式AI，轻至34g

8月21日，全球领先的消费级AR品牌雷鸟创新（RayNeo）举办2026雷鸟AI眼镜新品发布会