2026 年 5 月 20 日 · 周三

oh no contextual got windsurfed

哦不，contextual 被 windsurfed 了

there's 4 parts to this AI SDLC 1. have ~50 tests in place, with instructions to add more, including "make a memory that whenever you do browser e2e tests, use computer vision to visually spot check design and ux issues as well on mobile/desktop/ipad/ultrawide resolutions" 2.

这个 AI SDLC（软件开发生命周期）有 4 个部分：1. 先准备好大约 50 个测试，并附上继续添加更多测试的说明，包括“建立一条 memory（记忆）：每当你做 browser e2e tests（浏览器端到端测试…

Still absolutely amazed I got to be on a pod with Rick Rubin right after Clown from Slipknot Code can be art too

仍然完全不敢相信，我居然能在播客里紧接着 Slipknot 的 Clown 之后和 Rick Rubin 同台。代码也可以是艺术。

Weird quirk of history: I worked at Microsoft in 2003-2005 when they were trying to make WinFS work I guess with LLMs, WinFS can happen now, and it's called GBrain

历史里一个很奇怪的小插曲：我在 2003 到 2005 年在 Microsoft 工作，那时他们正试图让 WinFS 跑起来。我想，有了 LLM（大语言模型），WinFS 现在终于可以实现了，而它叫作 GBrain。

Token costs will become a dominant topic in enterprises going forward with AI.

随着 AI 在企业中的推进，token 成本将成为一个主导性话题。刚参加完一场与许多 Fortune 500 企业 CIO 共进的晚餐，这就是现场讨论最激烈的话题。大家正在采用各种策略，但基本上没有人觉得自己已经找到了正…

Gemini 3.5 Flash is out, and it's a major jump over Gemini 3 Flash in model capability for knowledge work.

Gemini 3.5 Flash 已发布，相比 Gemini 3 Flash，它在知识型工作方面的模型能力有了重大跃升。我们在早期版本中，使用 Box AI Complex Work Eval 对它进行了评估，该模型在复…

Met a lot of great folks at Google IO @gregisenberg @alliekmiller @AlexFinn One thing I think about often living both the PM and creator life is I don’t want to be just an “influencer” wielding a selfie cam.

在 Google IO 见到了很多很棒的人：@gregisenberg、@alliekmiller、@AlexFinn。作为同时过着 PM（产品经理）和 creator（内容创作者）双重生活的人，我经常在想的一件事是，我…

The man just speaks to me with everything he says: “Just try a lot and build to learn” “You need to do 3-4 iterations to find what takes off and what doesn’t.” “We only have a 90 day roadmap and maybe if we are lucky it’s 120 days.” “I don’t know if we’ll ever go back to 1 year roadmaps.

他说的每一句话都特别能打动我：“就是要多尝试，在 build（构建）中学习。”“你需要做 3-4 轮 iteration（迭代），才能找到什么会起飞、什么不会。”“我们的 roadmap（路线图）只有 90 天，如果运气…

i am excited to see what will happen with tokenmaxxing startups, both for how they work internally and the products they can build. openai offered to invest $2M in tokens into every startup in the current yc batch. happy building!

我很期待看到 tokenmaxxing startups 会发生什么变化，不仅包括它们内部将如何运作，也包括它们能够构建出什么样的产品。OpenAI 提出，愿意向当前这一届 YC batch 中的每家 startup 投…

we will offer this until we sell out of our current allocation for this program. (we will make sure to leave enough capacity for ChatGPT, Codex, etc.) we plan to offer it again in the future; our intention remains to build as much compute as fast as we can.

我们会持续提供这项计划，直到这项项目当前分配的额度售罄为止。（我们会确保为 ChatGPT、Codex 等保留足够的 capacity（容量）。）我们计划未来再次提供这一计划；我们的目标仍然是以尽可能快的速度建设尽可能多…

i now use: Composer 2.5 for planning Composer 2.5 for building & iterations Composer 2.5 for debugging a great all-rounder, especially for UI work – gets you in flow with Design Mode in Cursor

我现在这样用：用 Composer 2.5 做规划，用 Composer 2.5 做构建和迭代，也用 Composer 2.5 做调试。它是个很全能的 all-rounder，尤其适合 UI 工作——配合 Cursor …

turn the backlog into reality just @cursor_ai in jira ☑️

只要在 jira 里 @cursor_ai，就能把 backlog 变成现实 ☑️

Introducing Gemini Spark!

隆重推出 Gemini Spark！这是我们的 24/7 个人 AI agent（智能代理），旨在在你的指挥下主动管理任务，并帮助你应对数字生活中的各种事务。本周将向受信任的测试者推出，下周将以 Beta 版面向美国 G…

Glad you like it!

很高兴你喜欢它！

Have you explored Project Genie yet? 🌎 We just launched a huge set of updates!

你已经体验过 Project Genie 了吗？🌎 我们刚刚发布了一大波更新！现在，你可以基于 @GoogleMaps Street View 模拟世界，在新的 library（资料库）中管理你的创作，并将它们对外分享…

Happy anniversary, @FlowbyGoogle!

@FlowbyGoogle，周年快乐！从 text-to-video 到 Omni、agent 和 Tools——这一路真是精彩非凡。🚀 继续突破边界！迈向第二年。⚡

Even in the AI pilled crowd of the Bay Area, I think VERY few people have priced in the fact that we have already gone from assistant -> coworker -> and soon entering autonomous workers..

即使在 Bay Area 那群对 AI 极度上头的人里，我也认为，真正把这样一个事实计入预期的人 VERY few：我们已经从 assistant（助手）→ coworker（同事）→ 很快就要进入 autonomous…

Earthquake.. well not literally, but wow 🔥

Earthquake……当然不是字面意义上的，但哇，太炸了 🔥

Now also on the Claude Blog:

现在也已发布在 Claude Blog：

the future is bright, lets get to work

未来一片光明，让我们开始工作吧

Genuinely impressive release by Google today (remember when they were behind?) Gemini 3.5 Flash perf: * Building on prior strengths (83.6% of MMMU-Pro for multimodal), * big jump on agentic coding (76.2% on Terminal-Bench for agentic coding and 56.5% on Toolathon for real world tasks) * progress and expert tasks (57.9% on Finance Agent 2... we are cooked) * leading scores across SWE-Bench, OSWorld etc. (also, elegant to bold the top scores in the chart below even if when it's not Google leading) Ofc, just benchmarks, and also not cheap (~$9/M output), but Google is cookin'... we are all so spoiled to have the 3 labs compete

Google 今天发布的东西确实令人印象深刻（还记得他们以前还落后吗？）Gemini 3.5 Flash 的性能：* 在既有优势上继续提升（多模态方面，MMMU-Pro 达到 83.6%），* agentic codin…

Breaking: Anthropic attains sainthood, officially annointed by AI Jesus

突发：Anthropic 荣登圣坛，已获 AI Jesus 官方加冕为圣

From The Problem Solvers, a new series featuring founders taking on hard problems with Claude:

来自 The Problem Solvers——一个全新系列，聚焦那些借助 Claude 攻克难题的创始人：

Scott Wu (@ScottWu46) runs @cognition, the team behind Devin, an AI software engineer built on Claude.

Scott Wu (@ScottWu46) 运营着 @cognition——打造了 Devin 的团队；Devin 是一位基于 Claude 构建的 AI software engineer（AI 软件工程师）。他的目标…

We’re shipping a CDN pricing model that “smooths over” traffic spikes and viral events.

我们正在推出一种 CDN 定价模型，用来“平滑处理”流量高峰和爆火传播事件。这样一来，我们就能在丝毫不妥协质量与网络性能的前提下，消除压力和意外情况。实现可预测性，而不需要慢速路由、性能降级的 PoP，或分级优先带宽。

Claude Managed Agents 🤝 Vercel Sandbox

Claude Managed Agents 🤝 Vercel Sandbox

Alternate take

另一种看法

Very excited to speak at this!

非常高兴能在这个活动上发言！

what did karpathy see

karpathy 看到了什么

Personal update: I've joined Anthropic.

个人近况更新：我已加入 Anthropic。我认为，未来几年处在 LLMs 前沿的发展阶段将尤其具有塑造性。我非常高兴加入这里的团队，并重新回到 R&D（研发）工作中。我依然对教育怀有深厚热情，并计划在适当的时候继续这方…

The future is bright.

未来一片光明。我们将在 AI+Atoms 的交汇处迎来令人惊叹的事物。