Want to try this out? Just send this URL to your agent: developers.notion.com/prompts/setup Akshay Kothari @akothari I've been personal agent-maxing lately, and loving it. The new products are proactive, useful, and so fun. They get even better when they know me well. So I built a portable context vault in Notion and connected every agent to it. They write back new memories, skills, and preferences, then sync each morning. Now, whenever I try a new agent, the first thing I do is connect my vault, and it instantly gets all my preferences in one shot. I can also start a project with one agent, and switch over to another seamlessly. I shot a quick video to walk through the setup. LMK if you have any questions! 🔗 View on Twitter 🔗 View Quoted Tweet
ARCHIVE
内容归档
正文、图片和 Markdown 均已本地保存，不包含回复区。
Agent or Workflow? The answer depends on the job. Too much autonomy can make AI harder to control. Too much structure can limit flexibility. So how do you know which one to use? And what if you need both? We explore this through a practical IT support scenario 👇 dify.ai/blog/ai-agent-… f #Dify y #AIAgents s #AIWorkflow w
We’ve integrated @morluto ’s Reverse-Engineer-Anything skill inside Computer. And, we’re giving up to 10000 credits as a bootstrapping credit balance for any one to use them to reverse engineer any app of their choice! Have fun! Computer @AskPerplexity REA (Reverse Engineer Anything) by @morluto is now a skill in Computer. Ask Computer to reverse-engineer an app and it inspects how the app works without its source code, then builds a working version. Tasks that use /reverse-engineer are free through October 31, up to 10,000 credits per user. Here are 5 apps Computer made with it, available to try for a limited time: 1) Figma: Design with layers and auto-layout, plus sign-in, cross-device sync, and multiplayer editing. Try it: ui-editing.pplx.app 🔗 View on Twitter 🔗 View Quoted Tweet
In 2022, Berkeley PhD student Woosuk Kwon thought he could make Meta's new open-weight model, OPT-175B, run faster in two weeks. Four years and 3,000 contributors later, that project is @vllm_project , one of the largest open-source inference engines, running on ~500,000 GPUs at any given time. He and fellow researchers Simon Mo, Kaichao You, and Roger Wang borrowed "virtual memory" from operating systems, handing out memory only as it's needed so each GPU can juggle requests of wildly different sizes. (The "v" in vLLM is for virtual.) Their first user, Berkeley neighbor @arena , cut its GPU needs in half. Today Google, Amazon, and Meta all run vLLM in production. In January, the core maintainers founded @inferact and raised $150M to give vLLM a full-time team. They're hiring across eng, marketing, and devrel in SF and Singapore: a16zjobs.substack.com/p/inferact-rai… @simon_mo_ @woosuk_k @KaichaoYou @rogerw0108 🔗 View on Twitter
it's time to invent math 2.0 (sounds on) 🔗 View on Twitter SAIR @SAIRfoundation The math community deserves its own AI system. Introducing Open Math OS: the infrastructure for Math 2.0. Open models and AI agents. 24/7 cloud compute with 100+ tools. The community controls its data. Compute is free for approved teams. Apply now: openmathmodel.org 🔗 View Quoted Tweet
Anthropic @AnthropicAI New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re… 🔗 View Quoted Tweet
"Code only" vs. "Code with explanation"—similar questions don't always mean identical requirements. That's why traditional semantic caching often breaks. To solve this, Zilliz CTO James Luan tested 𝗝𝗲𝘃, a budget-friendly model that acts as a gateway guard before serving cached outputs. Jev evaluates 5 critical dimensions in real time: • 𝗧𝗮𝘀𝗸 𝗙𝗶𝘁: Does the answer directly solve the current query? • 𝗖𝗼𝗻𝘀𝘁𝗿𝗮𝗶𝗻𝘁 𝗠𝗮𝘁𝗰𝗵: Does it meet specific parameters (e.g., region, version, limits)? • 𝗙𝗼𝗿𝗺𝗮𝘁 & 𝗖𝗼𝗺𝗽𝗹𝗲𝘁𝗲𝗻𝗲𝘀𝘀: Is it fully complete in the requested style and language? • 𝗧𝗶𝗺𝗲 𝗦𝗲𝗻𝘀𝗶𝘁𝗶𝘃𝗶𝘁𝘆: Does it rely on outdated or changing data? • 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗔𝘃𝗮𝗶𝗹𝗮𝗯𝗶𝗹𝗶𝘁𝘆: Is all required context present to support the response? The test results on 1,152 real-world examples: • 𝗭𝗲𝗿𝗼 𝗶𝗻𝗰𝗼𝗿𝗿𝗲𝗰𝘁 𝗿𝗲𝘂𝘀𝗲𝘀 recorded. • 𝟮𝟮𝟳 𝘃𝗮𝗹𝗶𝗱 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲𝘀 𝘀𝗮𝘃𝗲𝗱, avoiding costly re-generations. • 𝗠𝗲𝗱𝗶𝗮𝗻 𝗹𝗮𝘁𝗲𝗻𝗰𝘆 𝗽𝗲𝗻𝗮𝗹𝘁𝘆: Only 0.79s. Semantic caching is a tradeoff between LLM costs and response accuracy. Jev shows that adding milvus.io/blog/jevs-chea… that tradeoff significantly safer. Check out the complete benchmark breakdown (feat. Laya and GPTCache reranker): https://t.co/mHFX3wo3PF
Gemini 4 Argon is a frontier model and has set a new SoTA on Wide-And-Deep-Research (WANDR) benchmark. Available now to all Perplexity users both on fast search and Computer (agent) modes. Perplexity @perplexity_ai Gemini 4 Argon sets a new record on WANDR. It is now available in Perplexity for Pro and Max subscribers. 🔗 View Quoted Tweet
At Advertising Week New York 2026, Kling AI joined WPP and Adobe to explore how AI video is moving from creative experimentation into real-world production. The next phase of AI video is not just about better generation. It is about connecting models, tools and workflows so creative ideas can scale with greater consistency and control. Kling 4.0 is now fully available on kling.ai , bringing more refined motion, more precise creative control, and professional-grade audiovisual output, opening up new possibilities for brand and advertising production. Swipe for highlights. ✨
Engineering leaders told Grepr founder and CEO Jad Naous they spend 5 to 20% of their infrastructure budget on observability, and it's buying less and less reliability. Jad, a former a16z partner, joins a16z's Martin Casado to discuss whether AI can run production systems as well as it writes code. Grepr sits in a company's log stream, can scan a petabyte of logs a day, cuts observability costs by about 90%, and is building AI agents that use those signals to troubleshoot production issues. How AI is changing engineering at Jad's 10-person company: - A data lake feature that once took one engineer about six months took Jad 10 days - The gap between okay and great engineers on his team "just multiplied 10X" - He runs 3-4 agent sessions at once, with a "tech lead" agent reviewing every PR Why running live systems is a harder problem for AI: - Jad: AI can almost tune databases, but can't yet keep servers running on its own - Martin: every bit of code doubles the ways a program can behave, which is why humans still own systems - The real world is messier than AI's training data, self-driving took 20 years and $170B 0:55 From Stanford to a16z to Grepr 7:00 Why observability spend stopped paying off 11:00 From autocomplete to AI that trains engineers 13:10 Why AI widens the engineer skill gap 14:55 Syntax vs semantics 18:00 Inside Jad's AI workflow 22:55 Six months of work in 10 days 24:50 Can AI reason about how systems really run? 30:10 Why AI SRE tools start in the wrong place 32:30 Can AI run a distributed cluster yet? 37:35 Why engineers buy extra capacity over predicting 39:15 How AI could trace failures in production 43:45 Why AI for ops is harder than for code 49:55 Can AI do a VC's job? 52:05 Why self-driving took 20 years and $170B 57:05 Customers skipped the AI pitch for the 90% savings YouTube: youtu.be/blkIGLlZePg @jadtnaous @grepr_ai @martin_casado 🔗 View on Twitter
Presenting the Generative AI Fizzle™ *Walter Bloomberg @DeItaone ANTHROPIC’S IPO VALUATION EXPECTATIONS DROP FROM $2 TRILLION TO $1.5 TRILLION Some investors are considering a $1.5 trillion valuation for Anthropic’s upcoming IPO, down 25% from earlier expectations of $2 trillion. Anthropic reported $16.8 billion in Q3 revenue, up 45% quarter-over-quarter, but growth slowed sharply after revenue more than doubled in Q2. The company also reported adjusted profitability for the quarter. The key concern: Slowing revenue growth and rising AI infrastructure costs could weigh on one of the most anticipated IPOs. Source: Reuters Exclusive 🔗 View Quoted Tweet
更新一个新的设计模型： 设计中添加： 出设计图使用 ChatGPT 的 images2，用来用去出图还是 images2，好用，grok自带的 imagine 那个出图效果不行，汉字还是搞不明白，我用的 codex cli ，不知道大家觉得目前生图模型哪个最好？我还是觉得 images2 不错。 之后会把这个私有 skills 发布 供大家使用 Viking @vikingmute 最近把技术开发基本都挪到 Grok Bot 上了，分享一下现在的工作流： 1 一个产品一个 bot 每个项目各有专属 bot，记自己的 repo 测试规则和上下文，不搞写作 设计 review 测试拆一堆 2 写代码不用 Cursor cloud agent，慢没必要，直接在 bot 电脑上跑 Cursor CLI，我是老的 plan，bot 和 grok model 和 other models 是分开算的，还有 Codex CLI 作为辅助。 3 按任务换模型，安装到了全局 skill 当中 代码： 规划大任务用 Opus 5.5 写具体实现用 Grok 4.7 review 用 Sol 6.1 文案/文档： 用 Gemini 3.8 设计： 定首页方向大设计用 Opus 5.5 日常改设计用 Sonnet 5.5 设计落代码再回 Grok 产品更新视频用 Opus 5.5 这样按照 Cursor 三种不同的额度可以比较平均的分配 4 Bot 负责协调拆任务 盯 CLI 看 diff、跑项目测试（TinyShip 现在每次改完都跑全量的 E2E + smoke），最后给我总结。代码基本不看了，靠其他模型 review，大改动我会用 review-forge 选两个模型分别看，验收全靠全量 E2E。 5 自动化任务包括： * 扫描 Issue 并且提出方案 * 周计划可以全自动定时跑，现在需求都是 AI 替我想的，保证产品持续迭代，每周至少都会上 2-3 个点，所以最近项目升级都很频繁 现在感觉整个流程非常舒服，挺喜欢 Grok Bot 的，觉得是现在做的最好的，可惜就是 Linux 主机，做网站最好，对于一些Mac 的原生应用就没法测试了。 🔗 View Quoted Tweet
NTT Docomo Business is trialing a voice agent for older residents of Aizuwakamatsu, Fukushima, built on ElevenAgents. Heavy snow can make it hard for older residents to leave home in winter. With Takeda Health Foundation and the cooperation of Aizuwakamatsu City, NTT Docomo Business deployed CONN, its AI assistant, to share local information and hold everyday conversations. Read the full story: elevenlabs.io/blog/ntt-docom…
昨天学习了一下 handwired，做了一些例子，比如做了几个倒计时、贪吃蛇这种 claude.ai/artifact/VSxUQ… 最后做了个教学视频：把代码变成神经网络：不训练、一个数都不差，这有什么用 视频1: 英文版 视频2: 中文版 有人把普通代码原封不动地变成了神经网络，一次没训练，结果一个数都不差。 这期从期末总评和停车费讲起，说清神经网络的两种零件：加权求和和开关，再看一条“及格线”怎么变成 7 个权重、1 层的网络。 接着用闰年、FizzBuzz、阶梯电价、生命游戏、贪吃蛇等例子，回答几个问题： 权重和层到底是什么？ 和 ChatGPT 背后的 Transformer 差在哪？ 网络能不能在浏览器里现场训练，刷新还在吗？ 出了错怎么查？ 变成网络会更快吗，CPU 和 GPU 差在哪？ 这个思路还能用在哪？ --- 视频制作的 Skill 参考： github.com/JimLiu/baocut/… 模型：Opus 5.5 配音：Gemini 3.8 TTS 🔗 View on Twitter 🔗 View on Twitter 宝玉 @dotey Cheng Lou 开源 Handwired：手写权重，把程序逻辑变成不用训练的神经网络 Midjourney 工程师、前 React 核心团队成员 Cheng Lou 开源了一个自称“玩具项目”的 Handwired：把程序里的 if-else 和循环改写成神经网络的权重。权重是手写的，不用训练，输出和对应的 JavaScript 程序完全一致，在浏览器里就能跑。 原理一行代码就能讲清。比如 result = a > 4 ? 10 : 12，在 a 是整数时可以改写成： above4 = relu(a - 4) - relu(a - 5) result = above4 × 10 + (1 - above4) × 12 relu 就是 Math.max(0, x)，神经网络里最常见的激活函数。a 大于 4 时 above4 正好是 1，否则是 0。两个分支都算出来，再用它挑一个。神经网络本质上就是一堆乘法、加法加上 relu，所以程序里的判断和分支，都能收进一个大矩阵乘法里。 也许你会好奇，神经网络的输出不是带概率、不精确吗？ Cheng Lou 的回答是：训练出来的才这样，按程序逻辑手写的权重，同样的输入永远得到同样的输出。 他放了三个 demo。 第一个是 Mini OS，一个有三个窗口和程序坞的小桌面，拖动、缩放、层叠、滚动、文字输入、动画都有。打开浏览器开发者工具看，页面就是普通的网页元素，没有大模型，也没有扩散模型。驱动它的是一个 130 层、9909 个权重的“神经网络”，训练次数为零，在浏览器里用 CPU 实时计算。 第二个是实时搜索 10 万份 JSON 文档，每敲一个键重搜一遍。正则匹配、JSON 解析、高亮全是这种手写网络，再转成 WebGPU 代码放到显卡上并行跑。项目文档里的实测（Apple M5 Max + Chrome）：每次搜索 11 到 15 毫秒；浏览器自带的正则引擎一般 8 到 28 毫秒，遇到容易回溯的写法要 2 秒多。文档写明两者谁快说不准，所以页面把两个时间都列出来。浏览器不支持 WebGPU 时会退回 CPU，搜一遍要 13 秒。 第三个是会“学习”的相册。整个程序可微分，也就是能用训练神经网络的反向传播来调参数。你在左边画几个矩形当草图，右边的照片布局由一个小网络算出，它的 244 个参数随着你拖动实时训练。损失函数里加了一项对齐惩罚，照片会自己往一条线上靠，没有任何代码专门去对齐它们。参数很少，训练全在 JavaScript 里跑。 Cheng Lou 还设想，把一段 Handwired 程序接到 nanoGPT（Andrej Karpathy 的一个小型 GPT 训练项目）这样的模型上一起训练，模型既能写莎士比亚，又能精确执行那段程序。这个目前只是 README 里的设想，仓库里没有现成实现。 代码以 MIT 许可开源，demo 直接在网页上打开就能玩。想自己写，需要装 Bun，克隆仓库后从 template 目录的开关按钮示例开始，文档里有完整教程。 推文说“把普通 JS 代码变成神经网络”，容易让人以为能把现成代码丢进去自动转换。项目文档写的是：没有工具能把已有的 JavaScript 源码转成网络。你要用它提供的构建函数，把逻辑一句句写成比较、选择、加权求和这类操作。网络每次都要把所有操作跑一遍，不能提前退出，也不能跑运行时才知道次数的循环，所以常见的循环得改写成计数、树形合并之类的结构，文档专门用一章讲怎么改。所有数都是 32 位浮点数，每个值的取值范围要事先声明。 项目主页： chenglou.me/handwired/ GitHub： github.com/chenglou/handw… 原推文： x.com/_chenglou/stat… 🔗 View Quoted Tweet