📌 X/TWITTER
OpenAI CEO Sam Altman
Altman shared two major updates. First, a general-purpose AI model solved a significant open problem in mathematics — he calls it “a kinda big milestone” and frames it as a preview of what’s coming: “we’ll be saying this a lot over the coming years.” Second, he outlined OpenAI’s three most exciting priorities: AGI accelerating scientific research, AGI accelerating companies, and personal AGI accelerating everyone in achieving their goals.
Altman 分享了两条重要动态。通用 AI 模型解决了一个重大的数学开放问题——他称之为「一个重要里程碑」，并预示「未来几年我们会经常这样说」。他还列出了 OpenAI 最兴奋的三个方向：AGI 加速科学研究、AGI 加速公司发展、个人 AGI 帮助每个人实现目标。
OpenAI 前 CPO Kevin Weil
Weil amplified Altman’s math breakthrough, calling it “the next in a series of firsts for AI and mathematics.” No additional details, but his signal-boosting adds weight to what appears to be a genuine research milestone.
Weil 转发了 Altman 的数学突破消息，称之为「AI 与数学领域一系列首创中的最新一个」。
Y Combinator CEO Garry Tan
Tan made a strong endorsement of Exa, the search API: “Exa is what I trust for all my agents. We use it at YC. We use it in all my OpenClaw and Hermes Agents. There is no other option that is as fast, as reliable, and as complete.” He also warned that while AI’s benefits are real now, the US risks ceding its leadership if it can’t overcome the “mass smokescreen by forces that want to slow it down.” On a lighter note, he mused about “idea fusion with LSD (lateral synaptic drift).”
Tan 强力背书 Exa 搜索 API：「Exa 是我所有 agent 的默认选择。YC 在用，我的 OpenClaw 和 Hermes Agent 也在用。没有其他选项能这么快、这么可靠、这么全面。」他还警告，AI 的好处已切实可见，但如果美国无法突破「那些想拖慢它的力量制造的烟雾弹」，将失去领导地位。
- https://x.com/garrytan/status/2057202833251000503
- https://x.com/garrytan/status/2057168328226230520
- https://x.com/garrytan/status/2057238298805129383
Vercel CEO Guillermo Rauch
Rauch teased a major announcement: “This will bring AI to 42% of the web. Every model, every provider, every modality (text, image, video, audio).” No further details yet, but the scope — 42% of the web — suggests an infrastructure-level AI deployment play from Vercel.
Rauch 预告了一项重大发布：「这将把 AI 带到 42% 的网络上。所有模型、所有提供商、所有模态（文本、图像、视频、音频）。」42% 这个数字暗示 Vercel 正在酝酿基础设施级别的 AI 部署动作。
Every CEO Dan Shipper
Shipper broke the news that Anthropic has acquired developer tools startup Stainless for a reported $300 million. Stainless’s biggest customers included OpenAI and Google. He also linked to his AI & I podcast interview with Stainless founder Alex Rattray.
Shipper 爆料 Anthropic 以约 3 亿美元收购了开发者工具初创公司 Stainless，后者的最大客户包括 OpenAI 和 Google。他还附上了对 Stainless 创始人 Alex Rattray 的播客访谈链接。
- https://x.com/danshipper/status/2057122805657821240
- https://x.com/danshipper/status/2057123430776902031
Box CEO Aaron Levie
Levie recommended a deep-dive post on FDEs (Future Development Engineers — a new job category emerging around AI-native development). “This is a job that is going to be around as long as AI keeps changing rapidly, which it inevitably will.”
Levie 推荐了一篇关于 FDE（未来开发工程师，AI 原生开发催生的新岗位类别）的深度文章：「只要 AI 还在快速变化——这是必然的——这个岗位就会一直存在。」
Google Labs: Project Genie 全面开放
The official Google Labs account announced that Project Genie — the AI game creation tool — is now “fully available to all Google AI Ultra subscribers globally (18+).” Users describe characters, set the scene, and Genie generates a playable game in minutes.
Google Labs 官方宣布 Project Genie（AI 游戏创作工具）现已面向全球 Google AI Ultra 订阅用户全面开放（18+）。用户只需选择角色、设定场景，Genie 就能在几分钟内生成可玩游戏。
- https://x.com/GoogleLabs/status/2057179491693470166
- https://x.com/GoogleLabs/status/2057218835074437573
Builder Zara Zhang
Zhang shared two sharp observations on AI-native teams. From Google I/O: a T-shaped developer model where AI handles the breadth of skills while humans go deeper — applicable not just to developers but every job function. Additionally: “In an AI-native team, ICs should start thinking like managers (how to delegate tasks to agents, set standards, verify output) and managers should start thinking like ICs (how to go deep on the most critical building blocks).”
Zhang 分享了关于 AI 原生团队的两条洞察。Google I/O 展示了 T 型开发者模型：AI 负责技能广度，人类负责深度——这不仅适用于开发者，也适用于所有岗位。她还提出：「AI 原生团队中，IC 要学会像管理者一样思考（如何给 agent 派活、定标准、验产出），管理者要学会像 IC 一样思考（深入最关键的基础模块）。」
- https://x.com/zarazhangrui/status/2057267931025957348
- https://x.com/zarazhangrui/status/2057324988009685208
FPV Ventures Partner Nikunj Kothari
Kothari used Claude to distill SpaceX’s 308-page S-1 filing into an accessible summary of key findings. He also reflected candidly on founder life: “Being a founder is so freaking hard man… chewing glass every day. If you fail, no one cares. If you succeed, you are given way more problems.”
Kothari 用 Claude 把 SpaceX 308 页的 S-1 招股书提炼成了一份重点摘要。他也坦率地反思了创业者处境：「当创始人太难了，每天都在嚼玻璃。失败了没人管你，成功了问题反而更多。」
AI/DevRel Swyx
Swyx drew a connection between Sam Altman’s “build a business that gets better when models get better” thesis and his own concept of “Agent Labs” — seeing a direct correlation with model improvement trajectories. He also praised Exa, noting his team did a competitive bake-off and “it took all of 1.5 hrs for the team to unanimously converge on Exa.”
Swyx 将 Altman「建立一个模型变强、业务就变强的生意」的论点与自己的「Agent Labs」概念联系起来，认为这与模型进步轨迹直接相关。他还称赞 Exa：团队做了一次竞品对比，「仅用 1.5 小时就一致选择 Exa」。
Peter Yang (Roblox)
Yang offered a pointed observation on mental health in tech: “Maybe not working at a company that’s doing layoffs and PSC every few months is better for mental health.” A quiet but resonant signal about the current state of big tech.
Yang 对科技行业的心理健康状况给出了直接评论：「也许不去一家每隔几个月就裁员和绩效评估的公司，对心理健康更好。」一针见血。
OpenClaw Creator Peter Steinberger
Steinberger gave a strong endorsement to Cotypist, an autocomplete-everywhere tool: “Can’t recommend @cotypist enough. Autocomplete everywhere.” A practical tool pick from a developer who lives in the terminal.
Steinberger 强烈推荐了 Cotypist（全局自动补全工具）：「再怎么推荐 @cotypist 都不为过。无处不在的自动补全。」
South Park Commons GP Aditya Agarwal
Agarwal cheered on Feross Aboukhadijeh’s startup journey: “It was amazing to have Feross do his -1 journey at South Park Commons and to have been an early partner. Very proud of the progress the company has made. Lots to build!”
Agarwal 为 Feross Aboukhadijeh 的创业之路喝彩，称其在 South Park Commons 完成 -1 到 0 的旅程，为公司进展感到骄傲。
🎙️ 播客
AI & I by Every: Inside Stainless — The Developer Tools Startup Anthropic Just Bought for $300 Million
The Takeaway: MCP 目前的「一堆工具」模式在大规模 API 面前根本走不通。真正的解法是：让 LLM 写代码，让 API 提供商在沙箱中执行代码——模型只需要两个工具：写代码和查文档。
Dan Shipper interviews Alex Rattray, founder and CEO of Stainless — the API SDK company just acquired by Anthropic for a reported $300 million. Rattray built SDKs for OpenAI, Anthropic, Stripe, and many others. The conversation ranges from the nuts and bolts of MCP to a provocative vision for the future of AI agents.
Dan Shipper 采访了 Stainless 创始人兼 CEO Alex Rattray。Stainless 为 OpenAI、Anthropic、Stripe 等公司构建 API SDK，刚刚被 Anthropic 以约 3 亿美元收购。对话从 MCP 的技术细节延伸到 AI agent 未来的颠覆性愿景。
The MCP scaling problem. Rattray frames the issue vividly: if you wanted to expose Stripe’s entire API as MCP tools, you’d burn through the context window before the model even started working. Stripe has hundreds of endpoints. “Today’s models not only can’t handle that amount of context, it’s a poor use of context… it’s also confusing to the model. It’s just too much to hold in your brain at one time.” And in a real business scenario, one task might span five different SaaS tools — each with its own massive API surface.
MCP 的规模问题。 Rattray 用一个生动的场景说明：如果你想将 Stripe 的全部 API 暴露为 MCP 工具，模型还没开始干活就已经把上下文窗口烧光了。Stripe 有数百个端点。「今天的模型不仅处理不了那么多上下文，而且这也是对上下文的浪费……还会把模型搞混。一次要记住的东西实在太多了。」而在真实业务场景中，一个任务可能横跨五个不同的 SaaS 工具，每个都有自己的庞大 API 接口。
Code execution is the answer. Rather than giving the model 50+ finely crafted tools, Rattray argues for a radically simpler approach: give the model a code execution sandbox with the API’s TypeScript SDK, plus a tool to search documentation. “The model will write stripe.transactions.list, write three nested for loops… and only at the end when it found the right thing, it’ll console.log ‘found Dan, customer ID…’ The context hit coming back from all of this is going to be like 10 lines of text.” This approach keeps context usage to ~1,000 tokens upfront and runs at CPU speed in a server next to the API, not round-tripping the LLM for every step.
代码执行才是答案。 Rattray 认为，与其给模型 50 多个精心设计的工具，不如给出一个更激进而简洁的方案：给模型一个能运行 API TypeScript SDK 的代码沙箱，再加一个查文档的工具。「模型会写 stripe.transactions.list，写三个嵌套 for 循环……最后找到目标时才 console.log ‘找到 Dan，客户 ID…’，返回给模型的上下文只需 10 行文字。」这种方式将初始上下文压缩到约 1,000 token，且代码在 API 旁边的 CPU 服务器上飞速运行，不需要每一步都和 LLM 来回通信。
The Cyborg Future. Rattray’s long-term vision: “The future of AI is cyborgs.” Not human-machine hybrids in the sci-fi sense, but AI agents that are part neural net and part traditional CPU code. He sees two tracks — one-off operational tasks handled by code-executing agents, and enduringly useful code that gets committed to production repos. The same code that an AI writes to handle a one-off customer refund could eventually become automated infrastructure.
Cyborg（半机械人）未来。 Rattray 的长期愿景：「AI 的未来是 cyborg。」不是科幻中的人机混合体，而是神经网络+传统 CPU 代码的 agent。两条路径并行——代码执行 agent 处理一次性运营任务，而真正有价值的代码会被提交到生产仓库。AI 为处理一次性退款写的代码，最终可能变成自动化基础设施。
关键引语： “APIs are the dendrites of the Internet. Dendrites are where your neurons connect and actually exchange information… If all these servers in the cloud aren’t talking to each other, there is no Internet. There’s nothing going on.”
「API 是互联网的树突。树突是神经元连接和交换信息的地方……如果云端的所有服务器不互相通信，就没有互联网。一切都不存在了。」
Security model. Rattray argues security must happen at the API layer itself — OAuth with granular permissions and proper scopes — not by restricting what’s exposed through MCP. “At the end of the day, you could do anything that’s in the API under the hood. What people should be doing is using OAuth with granular permissions.”
安全模型。 Rattray 认为安全应该在 API 层本身实现——通过 OAuth 细粒度权限和正确的 scope——而不是靠限制 MCP 暴露的工具数量。「说到底，你可以在 API 层面做任何事。人们应该用带细粒度权限的 OAuth。」
Current usage at Stainless. Rattray uses Claude Code daily for business intelligence: querying Postgres via MCP, cross-referencing HubSpot, searching Notion notes and Gong transcripts. He maintains a private Git repo as an AI knowledge base — customer quotes, SQL queries, analytics — where Claude Code reads and writes markdown files. One practical pattern: iterating on a SQL query with Claude Code for board prep, then saving it to an analytics folder for reuse next quarter.
Stainless 内部的真实使用。 Rattray 日常用 Claude Code 做商业智能：通过 MCP 查询 Postgres、交叉引用 HubSpot、搜索 Notion 笔记和 Gong 通话记录。他维护一个私有 Git 仓库作为 AI 知识库——客户引用、SQL 查询、分析数据——Claude Code 读写 markdown 文件。一个实用模式：用 Claude Code 迭代 SQL 查询准备董事会材料，然后存入分析文件夹供下季度复用。
通过 Follow Builders 生成: https://github.com/zarazhangrui/follow-builders