Max Zhou

We are making our discount permanent! 🎉 Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀

@guolin_ke 哈喽Guolin, 看了你的Git和Twitter主页，真的非常吸引我。我正在做 ai agent 方向创业，方便认识一下，交流交流吗？

A simple yet effective idea: enforcing *hypersphere (manifold) constraint* unlock the potential of continuous-token AR generative models. Our next-token, raster-order AR model hits FID 1.34 on ImageNet-256, setting a new SOTA for AR image generation. check huggingface.co/papers/2509.24…

With Linq, your messages reach customers in the channels they're already in

a prompt I've been using a lot recently: implement <SPEC> and while you do, keep a running implementation-notes.html file (or markdown) with decisions you had to make weren't in the spec, things you had to change, tradeoffs you had to make or anything else I should know

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

Krishna Rao is the CFO of Anthropic, and this is his first podcast appearance. He joined the company two years ago when run-rate revenue was about $250M. Today it is $30B. He has helped raise ~$75B and is responsible for the procurement and allocation of compute. I feel lucky we get to hear what it is like to sit inside a company this consequential at a moment this pivotal. We discuss: - The cone of uncertainty - How he allocates compute across Trainium, TPUs, and GPUs - What investors misunderstand about model companies - Why the returns to frontier intelligence keep rising - Platform vs application and where Anthropic builds its own products - How Anthropic uses Claude internally I have asked my closing question about the kindest thing more than 500 times. Krishna's answer is one I have never heard before. Enjoy! Timestamps: 0:00 Intro 2:38 The Compute Canvas 6:51 The "Cone of Uncertainty" 11:58 Why the Returns to Frontier Intelligence Are So High 16:45 Recursive Self-Improvement 20:20 Scaling Laws 23:30 Sourcing $100 Billion in Compute 28:05 Platform vs. Application Strategy 32:52 Pricing Dynamics 38:48 How Anthropic’s Finance Team Uses Claude 43:24 Raising Capital & Overcoming Investor Skepticism 52:32 Public Perception, Risks, and Government Regulation 57:25 Mythos Release 1:12:33 What Could Derail the AI Revolution? 1:13:47 Biotech and Healthcare 1:15:31 The Kindest Thing

Joined a new AI-native company this week and it’s kind of wild how different it feels already. The laptop arrived, I logged in, and an agent basically took over from there. It set up my dev env, pulled repos, fixed dependency issues, got permissions approved, pointed me at the backlog, linked the architecture docs, and surfaced the Slack debates I actually needed to read before touching production. When I needed context on something, I asked the agent and it found the exact thread from months ago explaining why a decision was made, who owned it, the related Linear issues, and the PRs connected to it. I’ve only been here 3 days but it honestly feels like I’ve worked here for a year because the usual friction and scavenger hunt for context just isn’t there anymore. We should probably stop calling this “onboarding” and rename it to “mounting” because this feels a lot more like mounting a distributed filesystem called “institutional memory” than slowly getting drip-fed context over 6 months.

@quxiaoyin Basically doing the same thing! New era of tech CEO work :)

I've stopped trying to be smarter than AI. This might sound extreme, but I've completely changed my work philosophy. I used to be a PM who didn't code. Now I write code constantly because I let AI do most of the thinking. My approach has three principles: First, I assume AI is smarter than me. So instead of trying to prove my intelligence to AI, I focus on giving it more authority and better tools. Second, I minimize human-to-human communication in my company. When engineering and marketing need to coordinate, I require them to use Claude Code to check what features we've shipped. No more asking each other questions that AI already knows the answer to. Third, I design processes where AI empowers other AI, not where humans manage AI. Instead of people supervising AI work, I create workflows where different AI systems collaborate and check each other's output. This isn't about eliminating humans. It's about positioning humans as architects and AI as executors. We design the systems, set the boundaries, take responsibility for outcomes. AI handles the implementation details. The companies that embrace this philosophy first will have massive advantages. While everyone else is teaching people to use AI as an assistant, we're building businesses where AI operates autonomously and humans focus on strategy and accountability. SkillBoss AI is built on this principle - creating infrastructure where AI tools work together seamlessly, reducing human coordination overhead. #AIFirst #BusinessStrategy

@mortiest_ricky @OpenRouter 哈喽, 看了你的Git和Twitter主页，真的非常吸引我。我正在做 ai agent 方向，方便认识一下，交流交流吗？

fxxk you @OpenRouter I've spent nearly $20,000 on your platform and suddenly my account was banned without any clear explanation. #OpenRouter #unfair

@redoragd 哈喽Gorden, 看了你的Git和Twitter主页，真的非常吸引我。我正在做 ai agent 方向，方便认识一下，交流交流吗？

CLICK github.com/doragd AND BECOME FRIENDS 🌸🌸

On my way to Beijing in Air Force One

Peekaboo 3.0 is live. Biggest release since 2.0. ⚡ Action-first macOS computer use 👁️ Unified screenshot + UI detection 🧩 Cleaner JSON across CLI + MCP 🛠️ Better snapshots I started this last year, but the models just weren’t good enough. Now they are. peekaboo.sh

@LiuVaayne 哈喽Vaayne, 我看了你的Git和Twitter主页，真的非常吸引我。我正在做 ai agent 方向，方便认识一下，交流交流吗？

用 Go 从零写了一个 OpenClaw 类 AI 助理。 10MB，单二进制，不依赖任何运行时。git 操作直接内置，下载就能用。 Telegram Bot 流式输出、终端 TUI 对话、定时任务、持久化记忆、Skill 扩展、多 LLM 支持都有。 做这件事最好玩的部分，不是功能本身。 是一点一点往里加东西的过程。加什么、不加什么，全是自己说了算。有点像装修自己的房子——每一处都是你选的，住起来才顺手。 我建议每个人都试试，开发一个最适合自己的 AI 助理。不用很完整，但一定要是自己的。

Codex now works directly in Chrome on macOS and Windows. It’s even better at working with apps and sites in Chrome, and now works in parallel across tabs in the background without taking over your browser. To get started, install the Chrome plugin in the Codex app.

运营中转站这段时间是真没赚到钱，只能说勉强cover了我自己用ai的消费。 所以目前打算把开中转站的一切全部开源，包含如何建站+营销，门槛最低，让这个行业更卷一点。 首先整个系统由3个部分组成： • 第CN2 回国专线服务器：放在海外但回国速度极快的 VPS，作为运行核心。 • sub2api：核心程序，负责把网页账号转成 API 接口。 • Cloudflare：把流量再绕一道，提升国内访问速度，同时隐藏真实服务器 IP。 你需要准备： • 一台 CN2 GIA 或 CN2 GT 线路的海外 VPS（推荐配置：2 核 CPU、2GB 内存、20GB 硬盘以上）。 普通海外 VPS 在国内晚高峰几乎不可用，而 CN2 GIA 通过专线绕开了拥堵的公网节点，国内访问延迟一般在 150ms 以内。如果你买了不是 CN2 的服务器，国内用户体验会非常糟糕。 • 一个域名（建议在 Cloudflare 或 Namecheap 上购买，便宜的 .top 或 .xyz 也行，几块钱一年）。 • 一个 Cloudflare 账号（免费）。 • 号池：初期可以用 claude code pro 账户+ 注册大量gpt账户，货比三家去找到别的号商卡商，等后期你就可以搞claude code max kiro 反代 aws bedrock（去跟sales聊，基本能搞到7.2折），但是初期只需要保障claude code pro账号稳定即可，因为你需要养号，后期转max。 完整请求路径如下： 国内用户的客户端 → 解析到 Cloudflare 的 IP → Cloudflare 边缘节点 → CN2 专线回源到你的服务器 → 宝塔面板的 Nginx 反向代理 → sub2api 程序 → 你的号池 → ChatGPT 或 Claude 网页 → 数据原路返回。 购买并初始化CN2服务商 CN2 GIA 线路的常见服务商有 BandwagonHost（搬瓦工）、RackNerd、CloudCone、Lisahost。新手推荐搬瓦工的 CN2 GIA-E 套餐，稳定但价格略贵。预算紧的可以看 Lisahost 的香港 CN2 套餐。 如果你懂命令行搭建Nginx，手动部署SSL证书，那你就自己搞，如果你不懂可以使用中国程序员流行的宝塔面板，一键搭建Nginx、一键部署SSL证书、可视化配置反向代理，全程鼠标点击操作，新手也能轻松上手。 安装完Linux + Nginx + MySQL + PHP，就可以开始设置防火墙，够买域名，添加DNS解析。 最后去命令行输入ping.api.你购买的域名，返回服务器ip就行了。 搭建sub2api: sub2api 是一个开源项目，可以把 ChatGPT 网页版、Claude 网页版的 cookie 或者 session 转换成 OpenAI 兼容的 API 接口。 打开sub2api的官方教程，安装流程安装docker，拉取并启动sub2api的容器。 你需要把号池数据放到 /www/sub2api/data 目录下，sub2api 容器会读取这个目录。具体格式参考 sub2api 项目文档。 设置Nginx反向代理 添加完之后目标url是127.0.0.1:8080因为 sub2api 容器监听的就是这个地址。Nginx 收到外部请求后，转给本机的 8080 端口，sub2api 处理完返回给 Nginx，Nginx 再发回给用户。 后面你去问claude code 如何优化Nginx的配置，AI API 调用是流式响应（SSE），需要长连接 + 不缓存才能正常工作。默认 Nginx 配置在这种场景下会出问题，按照claude的提示优化，proxy_buffering 必须关闭，如果不关闭这个，AI 的回答会"卡一阵 → 一次性吐出"，而不是逐字流式输出。客户端会感觉非常慢甚至超时。 申请HTTPS证书： OpenAI 兼容客户端基本只信任 HTTPS。HTTP 明文会暴露 API Key 给中间网络。 申请好Let's Encrypt证书之后，回到 SSL 主界面，把"强制 HTTPS"开关打开。 优化Cloudflare配置 测试HTTPS-开启cloudflare代理-Cloudflare SSL 模式必须设为 Full (strict) AI API 是动态接口，Cloudflare 的某些"优化"会破坏流式响应。 Cloudflare → 你的域名 → 速度 → 优化。 全部关掉以下选项： • Auto Minify（自动压缩 HTML/CSS/JS）：关闭。 • Rocket Loader：关闭。 • Mirage：关闭。 • Polish：关闭。 设置缓存规则： Cloudflare → 缓存 → 配置。 Caching Level 选 Bypass，或者保持 Standard 但是后面用页面规则覆盖。 更彻底的做法：Cloudflare → 规则 → 页面规则 → 创建页面规则。 URL 模式：api.example.com* 设置：Cache Level = Bypass 设置防火墙规 Cloudflare → 安全性 → WAF → 自定义规则 → 创建规则。 规则一：限制单个 IP 频率 字段：IP source address，操作：Rate limiting，每 10 秒最多 30 次请求，超出后挑战或屏蔽 1 小时。 规则二：屏蔽明显恶意爬虫 字段：User Agent，运算符：包含，值：python-requests 启用 Cloudflare Argo Smart Routing，每月 5 美元，能在 Cloudflare 内部用最优路径路由你的流量。对国内用户访问海外服务器有 30% 到 50% 的速度提升。预算够推荐开。 测试上线 用 curl 测试 API，或者打开 CherryStudio 或 ChatBox，填写你的api地址和key做测试 使用Prometheus/Grafana，或者直接用宝塔面板做监控，可以看到 CPU、内存、流量实时数据。如果 sub2api 容器经常吃满 CPU，考虑升级服务器配置。

Codex 5.5 hack: "Are you 100% confident in this strategy? If not, find all possible loopholes, suggest proper fixes and run this loop until you are factually 100% confident in the new startegy" This works like charm. It makes Codex 5.5 high perform even better than codex 5.5 extra high. Why? Codex 5.5 is the only model i noticed that is self aware. It never makes high claims unless the model verifies everything. This doesn't work with Opus 4.7 cuz that's a very insecure model. You can paste this prompt over and over again, the model keeps saying "you're absolutely right,....." But with codex, after 2-3 iterations you'll notice yourself it actually patched all loopholes and this genuinely sounds like a good strategy. Try this out, thanks me later.

People get mad when someone says this out loud, but it’s true, and @eladgil is doing a lot of people a huge favor by saying it. Co-sign.

software engineers before vs after agents

Codex + Remotion 插件这个组合有点无敌 我一个视频剪辑小白，不到 1h 做出了下面这个电影票房排行榜短视频 Codex 负责找素材、下载片段，Remotion 负责衔接动画和画面编排 只需要表达清楚目标效果，就能做出完成度还不错的动画成片🤩 视频创作的门槛又一次被拉低了

别傻傻手剪视频了！Codex装上这个插件，直接“写代码”生成MP4！普通人也能批量做号！ 还在用剪映一帧帧调？太慢了！今天教你个降维打击的野路子：Codex + HyperFrames。不用懂代码，不用会剪辑，HTML直接变视频，效率提升10倍！ 1⃣ 安装与开启： 在Codex的插件市场找到“HyperFrames by HeyGen”，一键启用。它能把Codex变成你的“视频导演+剪辑师”。 2⃣ 下达指令（参数锁定）： 直接告诉Codex你要做什么。公式：用途 + 尺寸 + 时长 + 风格 + 平台。 例：“用HyperFrames做一个3:4竖屏教程视频，45秒，Swiss Pulse风格，用来发抖音。” Codex会自动拆结构、写分镜、安排时间轴。 3⃣ 全自动执行流： 你只需看着它跑。Codex会自动生成HTML源文件、GSAP动画、同步字幕、甚至配音。 核心逻辑：HTML就是视频源文件。文字、卡片、时间轴全用代码精确控制，改一个参数，整个视频自动更新。 4⃣ 一键导出： 跑完流程，直接预览并渲染成MP4。特别适合做工具教程、产品宣传片、数据图表视频。 别把时间浪费在重复劳动上。这套SOP跑通了，你一个人就是一个MCN。