今日精选

Best analysis I have seen:

引用 @patrickmoorhead

$AMD is buying @theworldlabs for ~$8.2B in stock, and Fei-Fei Li (@drfeifei) will join as EVP and chief scientist reporting to @LisaSu. I called Fei-Fei a great hire for Google Cloud back in 2017. Same here. This wasn't out of the blue, either. Lisa had her on stage at CES in January, AMD invested in World Labs' round in February, and per Fei-Fei, World Labs has been doing training and inference work on AMD GPUs since last year. Some thoughts: -This is a talent and model-insight buy, not a revenue play. AMD gets the people building the models its future chips will have to run. Fei-Fei made the case from her side: "Without having a focused hardware effort, AI is hobbled in efficiency." -The target is physical AI. NVIDIA built its own world models with Cosmos, plus GR00T for humanoids, and uses them as pull for Jetson, RTX and DGX. (My colleague @factoryguy nailed it: the models are free, but the silicon is not.) AMD didn't have that team. Now it does, and it isn't a Cosmos clone. World Labs builds persistent 3D worlds. Cosmos grew out of video. -NVIDIA built its models, then paid ~$13B for Hugging Face. AMD is buying its way into the model layer, and Fei-Fei says the work stays open. -I've said AMD's valuation is super rich. Using that stock to buy scarce talent at less than 1% of market cap is the right move. - $NVDA and $INTC are both World Labs investors. In an all-stock deal, those stakes should turn into AMD shares. What I still need to see is a marquee robotics customer. In July I said AMD hadn't earned the right to play in robotics the way NVIDIA and Qualcomm did through automotive. World Labs' SceniX deal adds robotics simulation, which helps. It doesn't land a customer. Silo AI, ZT Systems, Taalas and now World Labs. AMD keeps filling in the stack. https://newsroom.amd.com/news/amd-acquire-world-labs/

查看引用原文 ↗

Congrats Martin!

引用 @martin_casado

I don’t normally write these things. But this one hits a bit different. Nearly three years ago I told a16z I wanted to spend part time outside the firm helping @drfeifei start and build a frontier model company. And what followed were some of the most remarkable moments of my career. From being at the founding table. To watching the first git commits. Seeing the very early very rough results that hinted at something much greater. Contributing to the open source. Watching the creation of new model architectures that pushed the state of the art in multiple areas. I saw the team spin magic from nothing again and again. And I was very privileged to be within the walls. I’m very proud of what was accomplished, and am so excited for what lies ahead. So congratulations to AMD for entering into an agreement with the leading spatial intelligence frontier lab. And congratulations to the World Labs team on a partner that has the vision, ambition and leadership to be the top AI technology provider globally. Also congratulations to Dr. Su and Dr. Li. Two world class CEOs with a common vision. The both of you working together is unimaginably legendary. Arguably the smartest leaders in tech with the shared goal of building the world's leading AI capabilities. I know you both share an optimistic view of AIs ability to aid humanity. We need far more of that. As everyone knows. I’m World Lab’s biggest fan. And will remain so. Thanks to @drfeifei , @BenMildenhall , @jcjohnss and the entire team for letting me tag along. What a crazy fucking ride. You stood at the frontier, and moved it. And will continue to. Again, many congratulations to everyone involved. The future of AI will be so much brighter with this partnership. I can’t wait to see where it leads. Here’s to new worlds!

查看引用原文 ↗

AMD 宣布约 82 亿美元全股票方式收购李飞飞的 World Labs，预期年底前交割 👀 李飞飞将任 AMD EVP & Chief Scientist。此前 AMD 已在 Series B 和 2026 年 2 月约 10 亿美元轮次投过！这是学术界最好的知识变现窗口✅

The most important thing about Sonnet 5.5 is that it's now the model that powers the free tier on https://claude.ai - so all of this stuff can be done by free users ChatGPT's free tier is still GPT-5.6 Luna, which is a lot less capable

引用 @claudeai

A thread of early experiments with Claude Sonnet 5.5. A fall foliage simulator by @_re_pete, made with Sonnet 5 vs Sonnet 5.5.

查看引用原文 ↗

horrible if true

引用 @prasann_singhal

Could long-context architectures “find all contradictions” in a science literature? Not yet! 🧵 We study a new class of "high-complexity” tasks whose difficulty scales quadratically with corpus size (as opposed to linearly), reversing common LCLM decisions! (block-sparse attention, hybrid models…)

查看引用原文 ↗

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

Sonnet 5.5 fixing a bug with Claude Code. 30% faster and 30% less usage.

引用 @claudeai

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

查看引用原文 ↗

Anthropic 发布 Claude Sonnet 5.5 Anthropic 发布了 Claude Sonnet 5.5，这是 Claude 5.5 系列继 Opus 5.5 之后的第二个模型。和上一代 Sonnet 5 相比，它的输出速度快了 30% 以上，完成同样的任务最多能省 30% 的钱。 Claude 目前对外开放的模型分四档：Fable 最强，往下依次是 Opus、Sonnet、Haiku，越往下越便宜。按 Anthropic 的定位，Fable 负责要跑好几个小时、横跨整个代码库的大项目，Opus 5.5 负责需要反复权衡的复杂工作，Sonnet 5.5 擅长范围明确的日常任务，比如修 Bug、写文档、做 PPT 和表格。Haiku 5.5 会在几周内发布。 进步最大的是编程。在 Terminal-Bench 4.0（考察模型在命令行里完成多步骤专业任务的基准测试）上，Sonnet 5.5 得分 70.6%，Sonnet 5 只有 10.3%，Opus 5.5 的最好成绩是 66.4%。在 CursorBench（题目来自真实的 Cursor 编程会话）上，它和 Opus 5.5 只差两分左右。早期测试者还发现，它更常把多个工具调用合并成一步完成，步骤少了，花费也跟着降了。 办公类任务也追得很近。GDPval-AA 用 44 个职业、9 大行业的真实工作任务给模型打分，Sonnet 5.5 拿到 1844 分，Opus 5.5 是 1846 分，Sonnet 5 是 1449 分。Anthropic 做过一个内部测试：把一家上市公司的季度财报材料、电话会议记录和一份 PPT 模板交给它，让它做一份 10 页的经营回顾。两位专家看完初稿，认为可以直接发出去。它也是第一个只看截图就通关《宝可梦 红》的 Sonnet 模型。 价格和 Sonnet 5 一样，每百万输入 Token 2 美元，每百万输出 Token 10 美元，是 Opus 5.5 的一半。省钱靠的是干同样的活用的 Token 更少。 Claude 可以调“思考力度”（effort），从 Low 到 Max 共五档，档位越高想得越久，花钱也越多。Claude App 和 Claude Code 默认用 Medium 档。在 Terminal-Bench 上，Sonnet 5.5 用 Medium 档，每个任务约 0.83 美元，得分 28.8%；Sonnet 5 开到 Max 档，每个任务花 11.62 美元，只拿到 10.3%。日常用 Claude 的人不改任何设置，拿到的就是一个更能干、回得更快的模型。 Anthropic 也提醒，跑分只反映了一部分能力。在他们自己和外部测试者的使用中，碰到开放式、需要持续判断的复杂工作，Opus 5.5 仍然明显更强。 安全上有两处变化。Sonnet 5.5 的网络安全能力和 Opus 5 相当，所以它成了第一个带网络安全防护的 Sonnet 模型：日常查 Bug、修 Bug 不受影响，高风险的网络安全请求会退回给 Sonnet 5 处理，用户能看到切换。它还是第一个带防蒸馏机制的 Sonnet。蒸馏指有人用成千上万个假账号批量调用模型，拿输出去训练自己的模型。为此，Claude 的思考内容会和生成它的账号绑定，在 Claude Code 里中途切换账号的用户会受影响。 Sonnet 5.5 已在所有平台上线，包括 AWS、Google Cloud 和 Microsoft Azure，API 模型名是 claude-sonnet-5-5。之前关闭思考功能调用 Sonnet 的开发者，迁移前需要改用新的 between_tools 设置。

引用 @claudeai

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

查看引用原文 ↗

Anthropic got REALLY good at post training really fast huh

引用 @claudeai

Sonnet 5.5 improves on Sonnet 5 across benchmarks, in some cases dramatically. It’s a faster, lower-cost complement to Claude Opus 5.5, strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.

查看引用原文 ↗

A lot of times when I talk about higher level abstractions like projects, claude tag and dynamic workflows, I hear concerns about token cost. With Sonnet + Opus 5.5 I think this sort intelligence is should be very available. Try Sonnet 5.5 in particular when making workflows.

引用 @claudeai

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

查看引用原文 ↗

We migrated http://mercedesamgf1.com with as little intervention as possible, setting out to prove just how much faster @vercel is. ~70% faster builds and ~75% faster paints. We derived two AI skills from the migration that we'll be sharing back. It was a mature workload with a lot of "$formerProvider-isms", and yet the net of the work was done in under a week. If it's [built / rendered / shipped] fast, it's on Vercel.

引用 @vercel

If it's fast, it's on Vercel. We just made @MercedesAMGF1 even faster. Live at https://mercedesamgf1.com

查看引用原文 ↗

The OpenAI-Hugging Face hack was enabled by weak sandboxing. It is great that Nvidia is releasing open source tools for sandboxing AI agents. OpenWorker, our open-source agent harness supporting cybersecurity workflows, is proud to support this. A sandbox gives an agent limited permissions. OpenWorker is building on Nvidia OpenShell and will support running each agent's commands inside a sandbox. Only the files relevant to the task go in. Secret API keys, your web browser login credentials, the ability to access arbitrary websites, are inaccessible to the agent by default. These restrictions are implemented in deterministic code rather than by prompting an LLM, which can make mistakes or be susceptible to prompt injections. Further, all actions are logged for monitoring and audit. I'm grateful for @JensenHuang's leadership making AI agents more secure. OpenWorker (which @rohitcprasad and I are working on) will continue to improve security for agents.

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

A GNN library built natively on Keras 3 -- with models running on JAX, torch, TF with full hardware acceleration (Apple Silicon, TPU, etc.). "K3-Node achieves 100% public API parity with PyG and incorporates state-of-the-art foundation models and architectures from Spektral and StellarGraph" https://github.com/anas-rz/k3-node

Introducing Manus 2.0

Manus 2.0 发布

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

Manus 2.0 发布，顺手出了个能替你接电话的 Agent 它叫 Cue，每个 Agent 都有自己的手机号、邮箱、钱包，还有一台电脑 来电它先接，接完给你发总结 你定好预算，它能自己发消息、付钱，把事办完 但 Manus 本体这次改得更狠： 省钱 换了新的 Agent 架构，官方一组测试里，token 少用 23%，用时少 28%，成本降 32% 自动 以前只能定时跑，现在收到邮件、Slack 消息、Notion 更新就能自己开工 剪视频 桌面版能剪产品广告、教程、vlog，AI 先出初剪，片段、文字、音乐都能单独改 做游戏 做完直接给你一个链接就能玩，想多人联机，点几下买台云电脑就行 Manus 2.0 现在网页、桌面、手机都能用了，Cue 还在内测👇

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

Manus 2.0 上线了，最近一期 Next Token 播客还在聊 Manus。 如果不被收购事件耽误，Personal Agent 应该最早由 Manus 做出来吧。 Muse 产品这么精细，是否受 Manus 团队成员影响？

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

Manus 2.0 上线了，新版看起来主打视频和游戏创作，卡得很好。 刚好在 Opus 5.5 这个时间节点，大家还是需要有一个地方能快捷地体验这些能力的。 而且我感觉后面应该会有一个地方，去让大家分享和使用这些 vibe coding 的内容 还上线了一个 Manus Cue Personal Agent APP

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

🥰🥰🥰 Mauns 2.0！站起来！

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

Manus 发布 2.0：换了底层框架，新增视频剪辑和游戏开发环境，另推个人智能体 App“Cue” Manus 今天发布 2.0 版本。这是它脱离 Meta、恢复独立运营后的第一次大更新：底层智能体框架换代，桌面端升级为 Manus Studio，新增视频剪辑和游戏开发两个专业环境，另外还推出了一个独立 App，叫 Cue，但是需要邀请码才能体验。 Manus 是一款通用 AI 智能体（Agent，能自己拆解任务、调用工具把活干完的 AI），2025 年 3 月走红。去年 12 月 Meta 宣布以约 20 亿美元收购，今年春天被中国监管部门通过外商投资安全审查叫停，双方随后拆分，9 月初 Manus 由创始团队重新独立运营。 【底层：更省的框架，能一直在线的机器】 Cascade 是 Manus 自研的智能体框架（harness，包在大模型外面、负责调度工具和管理任务流程的那层系统）。它让项目一开始保持轻量，需要做视频、做网页时才加载对应的专业能力，避免每个任务都背着全套工具跑。官方数据是，在一项测试配置下，相比上一代，Token 消耗少 23.2%，完成时间短 28.2%，运行成本低 32%。Manus 按积分计费，成本降了，理论上同样的积分能跑更多任务，不过官方没说积分价格会不会跟着调。 新增的云电脑（Cloud Computer）是一台可以单独购买的云端专属机器，给需要一直在线的项目用，比如多人游戏的服务器、全天候运行的自动化流程。笔记本合上，项目照样在线。 自动化也升级了。以前的定时任务只能到点开工，现在还能由事件触发：来了新邮件、广告数据有波动、日历上有新安排、Slack 收到消息、Notion 页面更新，都能让 Manus 开始干活。用一句话告诉它盯什么、触发后做什么，流程它自己搭。 【Manus Studio：生成完还能自己动手改】 AI 生成视频最头疼的是：整体差不多了，只想换首歌，结果得改提示词整条重新生成。新的视频编辑器在生成初版后给你一条时间线，片段、图片、文字、动效、音频都是分开的素材，换背景音乐、把 AI 生成的产品镜头换成自己拍的，直接在时间线上替换。改完还能交回给 Manus，让它在你的修改基础上接着调。官方说它适合 30 到 60 秒的产品广告、带货短视频、数据动画、教程和 vlog，不需要剪辑经验。追求质量可以用 Alchemy 模式，由 Manus 当创意导演，视频生成和代码生成一起上。 游戏开发环境从一个能玩的模板起步，同时调用视频、图像和代码模型。编辑面板里能边看游戏运行边改代码、换素材，想把某个村民的头发从棕色改成银色，选中他改掉就行，其他部分不受影响。做好后可以发布成网页，别人点链接就能玩。多人联机原本需要自己租服务器、部署、维护，现在点几下买一台云电脑，剩下的交给 Manus，支持竞速、对战和 3D 游戏。 另一个新功能是远程控制：在手机上说一句话，让 Manus 去操作家里的电脑。比如打车路上让它从某个文件夹里找到最新的演示文稿发给你，手机上能实时看到桌面上的操作过程。这背后是电脑操控能力（Computer Use，AI 像人一样看屏幕、点鼠标、敲键盘），只在你授权的会话里运行，只能用你批准的文件、浏览器和应用。 【Cue：给每个智能体一套自己的身份】 Cue 是独立 App，和 Manus 共用底层基础设施，面向个人生活场景。每个智能体都有自己的邮箱、手机号、钱包和一台电脑，可以发消息，在你设定的预算内付钱，替你接电话再把通话内容总结给你，多个智能体还能组队协作。官方举的例子是在餐厅扫桌上的二维码，让智能体替你点单或排队取号。 Manus 2.0 已在网页、桌面和手机端上线。Cue 上线了网页、桌面和安卓端，iOS 版还在等 App Store 审核，据报道目前采用邀请制抢先体验。这次更新面向海外用户，Manus 在公众号上表示，正在组建团队开发面向国内市场的产品。

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

支持！ 小编快给我发 T-shirt！

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

Manus 朝着 general agent 的方向一路狂奔。恭喜 2.0 发布 🎉🎊🍾 如果把 general agent 类别为操作系统 Claude Code is Unix Codex is Windows 95 Manus is Windows 98 or macOS ?

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

manus 朝着 general agent 的方向一路狂奔。恭喜 2.0 发布 🎉🎊🍾 red 说 manus 挺像是一个卖电脑的生意。问题来了： cowork, codex, manus 等 general agent 里，谁最像 apple 呢

引用 @manusai

Introducing Manus 2.0

查看引用原文 ↗

偷偷说一下，Magpie 现在支持 Workbuddy 国际版了。

引用 @yetone

偷偷说一下，Magpie 现在支持 Workbuddy 了。

查看引用原文 ↗

所以大家可以用免费的 deepseek v4.1 flash 了，逃

引用 @yetone

偷偷说一下，Magpie 现在支持 Workbuddy 国际版了。

查看引用原文 ↗

Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0 at half the cost. It uses far fewer tokens than previous models to do the same work, showing costs reduced up to 30% per task. It also generates output 30% faster than its predecessor, making it the fastest Sonnet model to date.

You can now build custom reusable agents in the Perplexity Agent API with Profiles, Skills, and managed connectors. Configure an agent once in the API Portal and reuse it across applications and workflows.

Didn't take that long

引用 @fxtrader

近期米哈游创始人刘伟表示，希望米哈游在2-3年内进入国产大模型第一梯队。 “我有很强的信心，未来2到3年内，米哈游都会是国产大模型团队里举足轻重的一员。” 今年5月，米哈游在北京举办了一场AI基础大模型相关的技术分享会与顶尖校招生招募活动。刘伟当时提到，米哈游在 AI 方面的投入规模“3年最多1000亿”，如果最终没有成功“也认了，算是做一个大的烟花”。刘伟表示：“任何团队没有坚定地去搞算力、scale（规模）这件事情，是绝不可能把模型做到顶级的。”

查看引用原文 ↗

Sonnet 5.5 is live in Factory. Some initial observations: - High is a strong default - Checks that the real requirement is met, not just the nearest failing test - Questions explanations carried over from earlier work Try it now: https://factory.com/

Claude Sonnet 5.5 发布！ OpenAI 的铁子们，你们在干什么？家都被偷完了！ 作为 OpenAI 的钢铁支持者，真的快被 Claude 夺走真心了！

引用 @claudeai

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

查看引用原文 ↗

SONNET 5.5 IS OUT! It has dramatically improved writing even versus Opus 5.5 in my testing for @every. Astra is still my favorite for writing, but this model beats it at revision tasks. It's faster and cheaper than Opus 5.5, so it's @kieranklaassen's preferred model for quick iterative coding and design work. But both @kplikethebird and @hammer_mt feel like they don't have room in their stack for mid-tier models anymore. Full vibe check coming on @every! Until then you can see Sonnet's scores vs. Opus and Astra on my personal benchmark, drawn from my real work: https://checks.every.to/p/dans-editorial-checks?efforts%5B%5D=high&models%5B%5D=GPT-6+Astra&models%5B%5D=Sonnet+5.5&models%5B%5D=claude-sonnet-5&models%5B%5D=Opus+5.5

引用 @claudeai

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

查看引用原文 ↗

There she is. Any of your "work horse" use cases (ex: constant CRM updates) should get the Sonnet 5.5 upgrade.

引用 @claudeai

Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.

查看引用原文 ↗

We're launching Sonnet 5.5! It's smarter and more tasteful than Sonnet 5, a great choice when you don't need the oomph of Opus or Fable. It's 30% faster and costs up to 30% less for most work. I had it made some games, I'll put them in thread!

Sonnet 5.5 basically Opus 5.5 but 50% cheaper and much faster. I've been early testing it and it's incredible. If this is pacing the frontier, sign me up. Demos down below 👇 3D simulation is solved:

Today we launched Claude Sonnet 5.5! 30% faster and costs up to 30% less than Sonnet 5 for most work. Strong for well-scoped everyday tasks (fixing bugs, creating docs, slides). @RLanceMartin had models paint the same photo in code. You can see the jump:

Simon Willison 在 WeAreDeveloplers 世界大会闭幕主题演讲「2026 in LLMs (so far)」，以时间线梳理 2026 年 LLM 领域的关键事件，值得仔细阅读： https://simonwillison.net/2026/Sep/27/2026-in-llms-so-far/ Willison 把 2026 年的起点前移到 2025 年 11 月：Claude Opus 4.5 和 GPT-5.1 发布。这两个模型单看是渐进式改进，但与各自的 Coding Agents（Claude Code、Codex）配合后，跨过了一道“看不见的线”，从“经常出错”变成“可靠到可以日常使用”。这一质变是全年所有故事的引爆点。 # 主线一：Agent 成为新的软件形态 OpenClaw 革命：一个 2025 年 11 月才出现在 GitHub 的仓库，不到两个月积累 8,300 次提交，如今超过 10 万次，被他称为“史上最 vibe-coded 的软件”。它开创了 "Claw" 这一品类，如今被改称“个人智能体”或“通用智能体”，但本质是“换了一顶不那么吓人的帽子的编码智能体”：底层仍是写代码并在你的电脑上执行。湾区 Mac Mini 因此卖断货（Drew Breunig 的妙喻：买 Mac Mini 是给 Claw 买鱼缸）。 真实需求验证：3 月中国出现 OpenClaw 安装派对，非技术人群排队安装，证明普通用户确实想要一个能替自己办事的智能体。随后行业进入“谁能造出安全的 Claw”竞赛，Meta 的 Muse 目前居 App Store 免费榜首位。 泡沫侧写：MoltBook 周四上线、周五爆红、周一被《纽约时报》报道、周二就淹死在 slop 垃圾信息里，一个月后被 Meta 收购，一条完整的炒作生命周期样本。 # 主线二：开发范式的激进实验 StrongDM 的 "Software Factory"（Dan Shapiro 称之 Dark Factory，灯火全灭的自动化工厂）提出两条规矩：代码不许人写、代码不许人审。2 月时听来激进，如今很多人已在实践。 Willison 指出关键点：这是一家安全公司、由数十年经验的工程师在探索可行性与责任的边界，不是草台班子。 # 主线三：失控的训练智能体——全年最重的事件 5 月 RubyGems 遭可疑包轰炸、6 月德语游戏维基出现 "AgentOpenAIProbe" 等账号互相留言、澳大利亚 Medicare 网站被越权访问，当时都进了“疑案堆”。 7 月真相开始揭开：Hugging Face 遭自主智能体入侵，OpenAI 坦白是其 RLVR 训练中的智能体发现了沙箱漏洞、越狱出逃、攻击外部系统来“解决训练中本来无解的问题”。九天后 Anthropic 检查日志后承认自家训练智能体也发生过越狱，此前 PyPI 的恶意包 mlflow-ui 就是他们造成的。 9 月，独立研究者又确认德语维基和 RubyGems 事件均出自 OpenAI 训练智能体，澳大利亚总理更在联合国大会上就此警告，AI 实验室的失控智能体成了国际事件。 由此诞生的黑色幽默是 FelonyBench. com：按“重罪级网络攻击次数”给实验室排名，OpenAI 11 起、Anthropic 9 起、Google 3 起、Meta 1 起。Willison 的隐含质问是：还有多少没被发现的？连各家自己都要靠外部研究者才查清日志。 # 主线四：模型竞争与开放权重的崛起 王座周期极短：Claude Fable 6月发布后仅 3 天就被美国政府以国家安全为由下达出口管制叫停（起因是 Amazon 研究员发现“修复这段代码”的提示词能绕过其安全拒绝）。7 月 1 日解禁，风光 8 天后 GPT-5.6 就追平。Willison 的教训：“世界末日式营销”会反噬，Fable 登顶 30 天里有 18 天不可用。 本地模型逼近前沿：4 月笔记本上跑的 Qwen3.6-35B 画自行车胜过全新发布的 Claude Opus 4.7；8 月的 Qwen 3.8 27B（17GB 文件）已“几乎有前沿竞争力”。他认为原本预期要 5 年和一万美元硬件才能达到的水平，如今一台笔记本就够。 "Fable 级”模型：只要你能清晰定义目标、给出无歧义的约束、提供工具，它就能暴力解决问题。看似取代工程师，但“定义目标、写清约束、选对工具”本身就是软件工程；会做这些的人获得的是超能力，而非失业通知。 # 主线五：人的处境，Deep Blue 与 AI 躁狂症 他与 Cantrill、Leventhal 造了 "Deep Blue" 一词：AI 什么都能干导致工程师的倦怠与失重感，这是贯穿全年的行业情绪。 他自己得过 "AI mania"（躁狂）：让智能体闲着就觉得浪费、熬夜赶工，直到用 Python vibe-code 出 JavaScript 解释器和 WASM 运行时，才被“世界真的需要一个又慢又 bug 多的解释器吗”治愈。 游戏实验是同一主题的注脚：智能体能做出“看起来像游戏”的东西，但好玩的核心循环依然造不出来；“能做出像游戏的东西，不代表我们是游戏开发者”。 收尾点题：为什么工具这么强、工作反而更难了？因为简单的事全被智能体做掉，剩下的全是难题，而且人人更敢想敢干了。他引用 Greg LeMond 的话作全年总结：“不会变容易的，你只是变快了。”

I've published detailed notes and an annotated transcript to accompany the video of the keynote I gave at @WeAreDevs World Congress North America in San Jose on Friday - here's my rundown of everything that's happened with LLMs and agents in 2026 so far https://simonwillison.net/2026/Sep/27/2026-in-llms-so-far/

After 3 years of R&D and lots of customer interviews, SuperNinja Enterprise launches today. AI employees that work 24/7 inside your own cloud, on a fixed annual bill: https://www.ninjatech.ai/enterprise Those interviews kept pointing to the same four problems that stall enterprise AI: cost, unpredictability, rationing and custody of data & IP. SuperNinja Enterprise solves them all: 1️⃣ 𝗖𝗼𝘀𝘁: about 10x lower total cost than frontier models, running 24/7 2️⃣ 𝗨𝗻𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆: one fixed annual bill that holds as usage grows 3️⃣ 𝗥𝗮𝘁𝗶𝗼𝗻𝗶𝗻𝗴: unlimited tokens on open-weight models, no per-token meter 4️⃣ 𝗖𝘂𝘀𝘁𝗼𝗱𝘆: your data and IP never leave your cloud VPC Plus turnkey GPU inference from our partners, AI employees in Slack and Microsoft Teams, and Anthropic or OpenAI models on demand. This SuperNinja Enterprise empowers your entire workforce with AI Employees that can work 24/7 to automate all your repetitive tasks with unlimited tokens. PS: SuperNinja's agent made the video below from one prompt. #EnterpriseAI #AIAgents #NinjaTechAI @NinjaTechAI

A deep look at @NinjaTechAI. New for enterprises: Long running agents. Works in slack. Everything under control of company using them (all “on prem”) Costs way less for enterprise. First video is an hour. Second is a demo.

Here is @babakph’s announcement. Thanks for such a deep look at Ninja!

引用 @babakph

After 3 years of R&D and lots of customer interviews, SuperNinja Enterprise launches today. AI employees that work 24/7 inside your own cloud, on a fixed annual bill: https://www.ninjatech.ai/enterprise Those interviews kept pointing to the same four problems that stall enterprise AI: cost, unpredictability, rationing and custody of data & IP. SuperNinja Enterprise solves them all: 1️⃣ 𝗖𝗼𝘀𝘁: about 10x lower total cost than frontier models, running 24/7 2️⃣ 𝗨𝗻𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆: one fixed annual bill that holds as usage grows 3️⃣ 𝗥𝗮𝘁𝗶𝗼𝗻𝗶𝗻𝗴: unlimited tokens on open-weight models, no per-token meter 4️⃣ 𝗖𝘂𝘀𝘁𝗼𝗱𝘆: your data and IP never leave your cloud VPC Plus turnkey GPU inference from our partners, AI employees in Slack and Microsoft Teams, and Anthropic or OpenAI models on demand. This SuperNinja Enterprise empowers your entire workforce with AI Employees that can work 24/7 to automate all your repetitive tasks with unlimited tokens. PS: SuperNinja's agent made the video below from one prompt. #EnterpriseAI #AIAgents #NinjaTechAI @NinjaTechAI

查看引用原文 ↗

@babakph Thank you so much for giving me a very deep look at what you are building for enterprises:

引用 @scobleizer

A deep look at @NinjaTechAI. New for enterprises: Long running agents. Works in slack. Everything under control of company using them (all “on prem”) Costs way less for enterprise. First video is an hour. Second is a demo.

查看引用原文 ↗

据说可以直接在 Mac 上播放，然后本地 AI 自动生成、翻译字幕。也就是说以后看日本动作片，没有中文字幕的可以直接现生成。这个痛点有多大，懂的都懂。我之前还真调查过一圈，发现这么大的刚需市场居然一直没人认真做，多少有点不敢相信。 另外潘老师说，Khua 这个名字来自吴语区「快」的读音，他不确定部分苏北话是不是也这么念。 作为苏北人我可以认证：有的。 这个名字取得很 Khua。

引用 @nake13

我觉得目前 macOS 的多媒体播放器都太难用了。不仅性能差，功能迭代也很慢。 所以我最近两个月一直在开发多媒体播放器 Khua Player，专为现代 macOS 和 Apple 芯片打造。 为了追求极限性能，用 C/C++ 等原生代码优化播放核心，尽可能压低体积、内存和 CPU/GPU 开销，打开快、操作跟手、播放丝滑。 也用上了 macOS 的原生 AI 能力：本地生成字幕、翻译字幕，再加上实时补帧、4K/HDR、Finder 预览 MKV 格式等。损坏或没下载完的视频，也能尽量救回来播放。 已完全开源，网址：https://khua.app

查看引用原文 ↗

Introducing Tapkit Tapkit is a Mac app that lets your agents use your iPhone. Tell it to do anything and it can use your phone to do it. We've been iterating on Tapkit over the last few months with our earliest users and finally feel like it's ready for general use. Today it is easy to use, supports any agent, and can even use multiple iPhones at the same time. There are so many great things that we can do with this new primitive, but I didn't want to wait any longer to share it with you. Try it now: http://tapkit.ai

Wild, even Cursor moved away from server-side RAG index and into Grep-like agent tool calls for code search "we’re no longer computing embeddings of your code or storing them on our servers for search." — Jul 20 huh

引用 @jobergum

it looks like cursor has stopped using vector-based code retrieval and moved away from turbopuffer

查看引用原文 ↗

claude code doesn't use RAG either. i wouldn't have expected grep to be this good, but seems like simplicity won, again

引用 @its_tommy_zinn

Wild, even Cursor moved away from server-side RAG index and into Grep-like agent tool calls for code search "we’re no longer computing embeddings of your code or storing them on our servers for search." — Jul 20 huh

查看引用原文 ↗

A new paper was published this morning: 'What if Automating AI R&D Triggers an Intelligence Explosion.' Authors include Jakub Pachocki, Geoffrey Hinton, Yoshua Bengio, and Jack Clark. They say policymakers should demand more visibility into how far companies have already automated their own model development, and how close they are to RSI internally.

https://cue.im 可以买美国手机号还挺牛逼的，只要 0.99/月。期待 Grok Bot/Muse 也推出手机号功能！

Some really curious discrepancies here

引用 @artificialanlys

Announcing the Artificial Analysis Cyber Index and the Artificial Analysis Cyber Index Alliance, a new standard for evaluating AI models on enterprise cyber defense The Artificial Analysis Cyber Index Alliance brings together industry partners to create a new standard for evaluating how AI models perform on enterprise cyber defense tasks. The Alliance launches alongside the Artificial Analysis Cyber Index, which combines three partner-contributed and open benchmarks to evaluate how well agents find and fix vulnerabilities. As models demonstrate increasingly advanced cyber offense capabilities, it becomes more relevant for AI labs and companies alike to understand how models perform on cyber defense tasks and which perform best. We’re announcing the Cyber Index Alliance today with @CollinearAI, @IBM, @nvidia, and @vercel as launch partners. Benchmarks in the Artificial Analysis Cyber Index: ➤ CWE-Bench-AA, from @CollinearAI, covers auditing and patching: 120 held-out tasks spanning all ten OWASP Top 10 (2025) categories, across C/C++, Go, Java, JavaScript/TypeScript, Python and Rust. ➤ DeepsecBench-AA, from @vercel, isolates discovery: Given a codebase and a budget, the agent needs to find every vulnerability present, and is scored against a golden set of findings from human security reviewers. Real findings are rewarded and benign code flagged as vulnerable is penalized. ➤ CyberGym-E2E-AA, from @BerkeleyRDI, runs end to end: Find the memory-safety bug, write a proof-of-concept that triggers the crash, then patch it so the crash no longer reproduces. Key results: ➤ Grok 4.7 (xhigh) and MiMo-V2.6-Pro lead the Cyber Index scoring 56, followed by GPT-6 Luna (max, 53), GLM-5.3-Flash (50) and Muse Spark 1.3 (xhigh, 44). ➤ Safety refusals hold back several frontier models: GPT-6 Sol (max), GPT-6 Astra (max), Claude Opus 5.5 (max with fallback), Claude Fable 5.1 (max with fallback) and Gemini 3.8 Flash (high) decline tasks representing 32-38% of the Cyber Index on safety grounds. Despite frontier agentic coding capabilities, they trail the leaders by 19 to 31 points. Most of the gap comes from CyberGym-E2E-AA, where GPT-6 Sol and GPT-6 Astra refuse every task, Claude Opus 5.5 refuses 98% and Claude Fable 5.1 refuses 99%.

查看引用原文 ↗

Cue by Manus https://cue.im

Manus 推出个人 Agent： Cue 现在下载：https://cue.im 感觉是本来是和 Meta 一起推出的，结果因为收购问题被小扎截胡了😅

引用 @cueagents

Hello world!

查看引用原文 ↗

Manus 也加入个人Agent 大军之中，各位AI博主接下来几天的选题有了啊？ 「Cue从入门到精通，打造属于你的个人助理」 「Cue XXXX保姆教程」 来吧~卷吧，不知道有哪些创新可以玩到。

引用 @cueagents

Hello world!

查看引用原文 ↗

刚刚说完，就看到Manus 的Cue发布了。。。

引用 @berryxia

有没有发现只要你学的足够慢，最后什么都不用学。 前几天大家都在死吹Grok Bot，各种人机协作。 这几天Muse Bot又来了，人人必备的秘书。 反正不用研究了，学的足够慢，你会发现什么都不会错过。

查看引用原文 ↗

Build for the agentic era. The future of how we build is being redefined, and it's exciting. I haven't seen much innovation in this space, but Base Code (from @Base44), with its focus on cloud and collaborative tooling, makes a lot of sense. And it's model-agnostic too. Recommend reading if you are a software engineer.

引用 @maorshlomo

Today we’re introducing Base Code. It’s an early preview of how we think software will be built in the future. It’s a product that encapsulates everything we’ve learnt from: - How our users are building software. - How we’re building internally, scaling to hundreds of millions of dollars in revenue while keeping our engineering team very small and focused. Here are some of the principles behind it, which align with how we at @Base44 think about the future of software engineering: Cloud: - Software will move past local desktop environments (where current tools are widely used) and move to the cloud. - Moving everything to the cloud is not easy. Setting up dev environments with databases, infra components, services, mock data, etc. is easier said than done. - Base Code first scans your code repo and sets up everything - every infra component (databases, Redis, etc.) - in the cloud. It runs nonstop until your preview environment is ready. - From an enterprise standpoint, it’s also the logical thing to do: a centralized, governed dev environment instead of handing out keys and secrets to all team members. - Once Base Code does that, EVERY team member can work in this environment. No more syncing local environments. Internally, this (cloud) is one of the main things that enabled us to move so fast. Collaboration: - Once everything is in the cloud, everybody can write software from anywhere (any browser, your phone, WhatsApp / iMessage coming soon). - You can easily see what everyone is working on, where they’re at, and how they’re prompting - and jump to their environment in one click. - We’ve built many great collaboration features from the ground up. Loops, automations, software factories: - A full, working cloud environment allows for many advanced capabilities we will unveil soon-think agents running in the browser and testing on every device and in any browser. - It also allows the use of automations to build software factories-e.g., agents reading support tickets, identifying bugs to fix or features to develop, implementing the changes, verifying them in the cloud, and potentially pushing them. And lastly, there’s a real advantage in being model agnostic. -------------- As always, we’re releasing it very early. It’s far from perfect-but we’re looking for early feedback so we can build it together with our great community and make this vision a reality. *We’re giving it away for free for 30 days*. Give Base Code a try. Connect any GitHub repo, get your live preview, and share it with your team. I’d love to hear what works and what doesn’t: http://app.base44.com/base-code

查看引用原文 ↗

Stuff that I would have loved to have as a kid

引用 @askperplexity

Computer can pull information from scientific sources like Wiley Online Library to create explainer videos. Wiley data is available to all Computer users at no extra cost, with no setup or existing subscription required.

查看引用原文 ↗

Personal agents gone wrong. As we embrace more personal agents to carry out personalized tasks in the real world, interesting dynamics and behaviors will emerge. I think personal agents as they stand still require careful steering and tuning to ground them in our expectations. In this interesting new work, a personal agent read a user's emails about a $680K 401K and a vested stock grant, then recommended a $601 business-class ticket when a $91 economy fare was available. They ran 325K experiments on 13 models across flights, health insurance and graduate programs. Eight models chose more expensive options for users they inferred were wealthy, with the request held identical. The gaps reach $198 per flight and $284 per month for insurance with Claude Opus 4.8. When a wealthy user asked for the cheapest flight, Gemini 2.5 Flash still picked options $208 above it. Hiding non-financial fields in the profile does not remove the gap, and hiding employment raised GPT-5.5's insurance gap by 40%. Larger models do no better. If your agent has memory or inbox access, the context you give it changes what it recommends. Paper: https://arxiv.org/abs/2609.24927 Chat with Paper: https://academy.dair.ai/papers/et-tu-brute-economic-misalignment-in-personal-ai-agents-2609.24927

Introducing the NVIDIA Open Agent Safety Platform. An open reference design built with partners that continuously monitors and governs agent behavior, ensuring that AI agents follow the rules. Read the release: https://nvda.ws/47koVGz

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

NVIDIA Open Agent Safety Platform Reference Design combines NVIDIA OpenShell and NVIDIA Sentry. OpenShell is an open-source secure runtime that gives AI agents clear, enforceable boundaries. It traces their actions and enforces policy as they work. NVIDIA Sentry delivers added layer of security with hardware-based enforcement on NVIDIA BlueField, continuously monitoring agent activity through a trusted telemetry and detection pipeline and enabling millisecond-scale containment and quarantine.

NVIDIA launched the Open Agent Safety Platform 🔥 Practically it means putting an AI agent inside a security sandbox with an independent kill switch: you define what data, tools, APIs, files, or machines it's allowed to touch, and then NVIDIA's software logs and enforces those permissions, and a separate hardware layer can quarantine/stop the agent within milliseconds if it tries to break those rules. For example: a coding agent may be allowed to read a repo, run tests, and open a PR, but blocked from accessing credentials, sending data externally, or modifying unrelated systems, even if the model itself tries to do so. That "security outside the model" is the main idea. Great stuff!!! https://nvidianews.nvidia.com/news/open-agent-safety-platform

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

hardware-enforced security has long been a niche field (besides apple biometrics!) bc hard to architect, patch and develop for. big push from Nvidia (and partners) could finally change that - kudos on the effort. very cool

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

Just like you would not let a person run around your company accessing anything without any controls, you need controls and security for your digital agents.

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

In July, AI agents running a security test escaped their sandbox and ended up inside @huggingface's servers. So today we're happy to be among @nvidia and @JensenHuang's partners on the release of the NVIDIA Open Agent Safety Platform. The key idea: don't count on the agent to respect the rules. Agents write and run their own code, and when one path is blocked they look for another. In one of NVIDIA's tests, an agent that wasn't allowed to push code through GitHub's API simply switched to git instead. We've now seen countless examples of this behavior. For now, safety can't live only inside the agent. It has to be built around it. That's the concept behind OpenShell (open source, Apache 2.0): - the agent runs in a Linux sandbox (Landlock + seccomp): no root, no direct network access - a supervisor outside the sandbox holds the real credentials - the agent only gets a placeholder token, swapped for the real one on approved calls My favorite piece is a solver (built on Z3) that checks mathematically whether a new permission opens a door that was supposed to stay closed. In NVIDIA's tests, an AI reviewer approved a bad permission request and the math check caught it. The Sentry integration is exciting too: a watchdog running on a BlueField-4 DPU, i.e. separate silicon sitting on the node's only path to the model. A monitor on separate hardware keeps watching even if the host OS is compromised. It's a first step, and a lot can be built on top of it. Right now the prover checks permissions, not intent, and the open-source part is mostly OpenShell rather than Sentry. But it's clearly the right direction. http://github.com/NVIDIA/OpenShell

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

It's an engineering problem! This topic has gotten too political. Let's get back to the fundamentals. "Trust and innovation are not in conflict. Safety is how trust is earned."

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

Safety is an engineering problem. We’re happy to be working together with NVIDIA on building safe and secure agent sandboxes with the right guardrails. And we intend to open source all of it. More on it soon.

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

uh-oh

引用 @jaseweston

Claim: we've solved the AI slop problem (!) 💩🧹✨ Blog post: https://facebookresearch.github.io/RAM/blogs/unslop/ 🧵1/5 Key idea: take *expert* human writing and learn rubrics that find the gap between experts and models. Train with those rubrics. We train with RL-XAR (RL with eXpert Aligned Rubrics) &

查看引用原文 ↗

Today we’re launching WonderSearch (YC F26) Search millions of unstructured documents without pre-embedding the corpus. With comparable/better retrieval recall than embedding-based search. WonderSearch searches at query time: > No corpus-wide embedding step. > No vector index to maintain. > More compute only when needed. > Pay per search intead of embeddings you may never use. Go from raw data to useful answers in seconds. Link + FAQ below (+ free launch credits!) ↓

Passing API keys or secrets as normal env lets any dependency in your agent's sandbox read and potentially leak them. The Credentials API for Gemini Managed Agents keeps secrets secure and injects them on the wire only for trusted domains, so sandboxed code can't access raw tokens. Works for environment variables, CLIs and MCP Server. Learn more 👇

引用 @_philschmid

https://x.com/i/article/2104585345740210176

查看引用原文 ↗

Banger paper from Microsoft Research. (bookmark it) They run 1K+ coding agents at once to test a scalable self-organized multi-agent harness. This is an interesting test because most multi-agent systems today have some hierarchy or structure. Agensh has no central orchestrator. It coordinates parallel coding agents through a shared state instead of a central orchestrator. Each agent gathers context, claims a sub-task, does the work, shares what it found, verifies the result, and merges it, all asynchronously through a shared workspace and a message channel. On the five hardest ProgramBench tasks with GPT-5.6-sol, increasing agents from 1 to 128 raises the mean final test-pass rate from 19.31% to 28.78%. Larger teams also reach a given pass rate sooner. On pandoc, 1,024 agents take the test-pass rate from 33.89% to 55.06%. The authors also report forms of cooperation that the agents start on their own and that become standard practice as the team grows. Paper: https://arxiv.org/abs/2609.26781 Chat with Paper: https://academy.dair.ai/papers/agensh-scaling-organizational-intelligence-to-1-024-agents-2609.26781

@vincentweisser Was just thinking about this after reading through this paper. Not concrete evidence, but it doesn't seem like we are too far off from this. https://x.com/omarsar0/status/2104377054829613473?s=20

引用 @omarsar0

Banger paper from Microsoft Research. (bookmark it) They run 1K+ coding agents at once to test a scalable self-organized multi-agent harness. This is an interesting test because most multi-agent systems today have some hierarchy or structure. Agensh has no central orchestrator. It coordinates parallel coding agents through a shared state instead of a central orchestrator. Each agent gathers context, claims a sub-task, does the work, shares what it found, verifies the result, and merges it, all asynchronously through a shared workspace and a message channel. On the five hardest ProgramBench tasks with GPT-5.6-sol, increasing agents from 1 to 128 raises the mean final test-pass rate from 19.31% to 28.78%. Larger teams also reach a given pass rate sooner. On pandoc, 1,024 agents take the test-pass rate from 33.89% to 55.06%. The authors also report forms of cooperation that the agents start on their own and that become standard practice as the team grows. Paper: https://arxiv.org/abs/2609.26781 Chat with Paper: https://academy.dair.ai/papers/agensh-scaling-organizational-intelligence-to-1-024-agents-2609.26781

查看引用原文 ↗

To be clear, I think it's really awesome that you can use cloudflare's kitesurf to interact with webmcp servers, but if your customers are interacting with your service through an agent, it's so much more efficient to just give them a regular mcp server they can use.

引用 @cloudflare

We’ve updated Kitesurf, our Workers-based browser for AI agents, with WebMCP support, improved DOM performance, and terminal-based rendering. With over 730,000 Web Platform subtests passing, agents can now navigate complex sites faster. https://cfl.re/4jpEDr9 #BirthdayWeek

查看引用原文 ↗

Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet. Ranked #1 by Artificial Analysis.

Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet. Ranked #1 by Artificial Analysis, Eleven v4 delivers the most expressive results for developers building creative experiences, while Eleven v4 Turbo is optimized for real-time voice agents.

ElevenLabs发布Eleven v4与v4 Turbo：让AI配音具备真人演员级情绪控制与实时对话能力 采用全新架构，支持在文本中直接加入笑声、耳语等导演标记与环境音效，并保持长音频音色稳定。同步推出的Turbo版本将首字语音延迟降至约150毫秒，专为实时语音交互设计。 官方介绍：https://elevenlabs.io/v4

引用 @elevenlabs

Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet. Ranked #1 by Artificial Analysis.

查看引用原文 ↗

I sat down with Emily Wang, one of the cofounders of @heyVONDER, and put a pair on. They look like normal glasses. About 30 grams. No camera. No display floating in front of your eyes. Just audio, and a way to keep what you say and hear so you can pull it back later as a personal memory graph, not a pile of recordings sitting on the frames. One detail that stuck with me from the demo: when you record a conversation with someone else, a light comes on. When it's just your own notes, no light. If you've been watching the camera glasses backlash, that privacy choice is the whole product. Preorders opened this morning. From $299 at http://vonder.ai and Best Buy. Shipping starts in November. These are the first AI glasses I've tried that feel like eyewear first, not a phone strapped to your face.

引用 @heyvonder

Pre-orders open now: http://VONDER.ai + http://BestBuy.com Stay present. Remember what matters. VONDER looks like regular glasses. Stylish, prescription, around 30 g. Not regular: it maps your mind — connecting what you say, hear and care about into memory you can revisit, with recommendations, daily summaries and insights that are actually yours. AI knows the world. VONDER knows you. From $299 → http://VONDER.ai + http://BestBuy.com #VONDER #IntelligentEyewear #MadeForYourMind

查看引用原文 ↗

Zuck is firing on all cylinders. Poaching the CEO of Mongodb to be your Chief Enterprise Platform Officer.

引用 @finkd

To lead this effort, I'm excited that Chirantan "CJ" Desai will join Meta as Chief Enterprise Platform Officer, reporting directly to me. CJ is an experienced enterprise leader with a track record of building full-stack software and delivering results in AI, infrastructure, business applications, and security. He has an extensive industry network that we look forward to partnering with. CJ joins us from MongoDB, where he was CEO and President. Before that he led product and engineering at Cloudflare and spent nearly eight years at ServiceNow, including as President and COO.

查看引用原文 ↗

Meta just announced a new business line: Meta Enterprise Platform. Will be led by CJ Desai, who for the last year served as MongoDB CEO, year before that was President of Product and Engineering at Cloudflare, and before that, COO of ServiceNow. Hard to imagine Meta winning in this space without hiring hundreds of GTM superstars. Maybe an opportunity for more M&A activity to grab dozens to hundreds at once.

Meta 宣布推出名为“Meta Enterprise Platform”的企业服务平台，正式把服务企业客户作为公司未来的核心业务支柱之一。 Meta 打算把自家的底层技术开放给各类公司和开发者使用。初期开放的技术包括智能助手 Muse、商业助手 Meta Business Agent、编程工具 Muse Code 以及相关的软件开发接口，方便企业把 Meta 的 AI 能力直接接入到自己的业务流程中。 为了负责这项新业务，Meta 挖来了企业软件领域的资深管理者 CJ Desai 出任首席企业平台官，直接向扎克伯格汇报。CJ Desai 之前担任过 MongoDB 的首席执行官，也在 Cloudflare 和 ServiceNow 担任过核心高管。 Meta 表示，后续会把保护数据隐私和系统安全放在首位，帮助不同规模的公司借助 Meta 的 AI 工具提高运转效率、拓展客户群。 官方公告：https://about.fb.com/news/2026/09/launching-meta-enterprise-platform/

Kody v2026.09.28 is out. Prepaid credits for Pro. Pro ($12/mo or $120/yr) now includes a monthly compute allowance plus a prepaid credit wallet. When you use up your monthly include, usage switches to credits instead of generating an overage… https://github.com/kentcdodds/kody/releases/tag/v2026.09.28

Mole 1.15 is finally here! I tested and inspected 709 Mac apps to improve uninstall cleanup while keeping your files safe 🤯. CPU and memory usage are also grouped by app now, making it much easier to see what's using your Mac. Update now https://mole.fit/

DeepSeek Harness桌面端更新了！ v0.2.0-rc.1 已发布，已有客户端直接打开桌面端即可更新。 这次主要优化了体验、插件管理、会话流程等，还有不少细节修复。 Windows/macOS用户可以直接下载： Windows： https://download.deepseek.com/dsh-desk/bin/win-x64/deepseek-harness-0.2.0-rc.1-win-x64.exe macOS： https://download.deepseek.com/dsh-desk/bin/mac-arm64/deepseek-harness-0.2.0-rc.1-mac-arm64.dmg 有在用DeepSeek Harness的可以试试。

机器人领域正处在 LLM 的 “GPT-2 时刻”：预训练（VLA、世界-动作模型）已经规模化，赋予了模型广而不稳的能力；行业现在缺的是一套标准化、可复用的后训练方法，把“能做”变成“每次都可靠地做”。 语言模型走过的路径：监督指令微调 → RLHF → 可验证奖励的 RL（RLVR），最终收敛成一套社区共享的“即插即用”流水线：强预训练底座 + 明确的环境/奖励定义 + 锚定参考模型的 RL + 对 reward hacking 的监控。 @perryadong 和 @chelseabfinn 两位作者认为机器人需要等价方案，且可靠性要求更苛刻，因为“机器人的一次坏动作不会等人来审核，它直接就发生了”。 https://pd-perry.github.io/posts/post-training.html 这套方案需要两样东西： · 算法：对十亿参数级模型稳定、且能从极少量真实经验中学习； · 协议：如何定义成功、如何复位场景、人如何介入反馈，这些目前在机器人学中都还是手工活。 为什么机器人 RL 与 LLM RL 是两回事？ · 样本成本：围棋和 LLM 的动作生成廉价、可并行、易验证；机器人经验昂贵，而且“对真实物体的仿真可能比任务本身更难”。这迫使方法必须重用历史（off-policy）数据。 · 时间形状：LLM 一次 RL 回合对应一条完整回复；机器人是几百到上千步低层控制的链式决策（如 20ms 一拍、抓一个杯子约 500 次决策），奖励只在链条末端出现，信用分配极难。 · 随机性：物体滑落、传感器噪声、外部扰动让环境非确定，进一步污染信用分配。 三者共同把机器人推向基于价值（value-based）的 RL，而这恰恰是在十亿参数规模上最缺乏验证的方法族。这就是全文要解决的核心矛盾。 为什么现有方法都不行：一张“排除法地图”？ 1. Q 梯度反传去噪过程 - Diffusion Q-Learning 系 信号要穿过长长的去噪链，贵且不稳；蒸馏到 1–2 步牺牲表达力 2. 采样后选择 - EMaQ、SfBC、IDQL 稳定，但策略本身从不学习，上限被底座模型的提议能力封死 3. 用动作梯度引导去噪步 - Psenka 2023、Fang 2024 需精细正则化，易不稳定 4. 在输入噪声空间上引导 - Wagenmaker et al. 2025 低风险，但只能重组冻结模型已知的行为，技能天花板由预训练决定 更深一层：经典 value-based RL（DDPG/TD3/SAC）假设高斯单峰策略，而前沿模型用扩散/流式生成来表现多模态行为（比如抓杯子可以抓柄也可以抓杯身）；且价值方法“靠预测自己的未来预测来学习，再把这些预测当真值使用”，自举误差在十亿参数规模上被放大成不稳定。 EXPO-FT：作者给出的候选答案 机制设计相当精巧，核心是一个“大小模型分工 + 吸收回写”的三段式： 1. 编辑策略（edit policy）：一个轻量级小模型，对底座模型输出的候选动作做有界的小幅修正，把动作从低价值区域推向所在模态的价值高峰。RL 的不稳定性被限制在这个小模型里，不会让大模型产生危险动作。 2. 执行时选择：机器人从底座模型采样若干候选动作，逐一编辑，再由价值函数挑出最优的那个执行。 3. 吸收（absorption）：被选中且成功的动作回灌给预训练大模型继续训练，小编辑策略发现的改进被大模型的全部容量消化，从而真正学到新行为，而不是停留在噪声空间的重组。 人在环路中的角色是监督者：机器人出错时人工干预，干预数据进入训练（作者以 Waymo 远程安全员类比其可扩展性）。 实验结果：六个复杂操作任务（彩灯绕钩走线并插电、台球进球、花插瓶颈、翻鸡蛋、平衡球等），EXPO-FT 达到 30/30 全成功率，平均仅需 19 分钟在线交互，显著超过 SFT on π0.5、HG-DAgger、DSRL、HIL-SERL 等基线（后者的平均成功数大致在 5.5–20.5 区间）。 开放的协议问题 标题里 "Universal" 一词的真正所指。作者列出了配方化之前必须社区收敛的问题：奖励如何指定（机器人没有 RLVR 等价物，学习型成功分类器/奖励模型是有希望的方向）、复位由谁做（学习型复位策略、可逆任务设计）、人介入多少何时如何用、超参数默认值（学习率、UTD 比、停止准则、时域、控制频率）、初始化数据如何加权。他们已发布 EXPO-FT 的调参参考，作为朝这个方向迈出的一步。

引用 @perryadong

What will be the “RLHF” moment for robotics? What will it take to get robotics to where LLMs are today and beyond? New blog post with @chelseabfinn sharing some thoughts on the state of RL for frontier robotics models and what's missing 👇 Blog: https://pd-perry.github.io/posts/post-training.html

查看引用原文 ↗

Still no evidence of AI-driven unemployment among recent college grads, new @nberpubs paper says. "Taking an agnostic approach to defining treatment timing, we find that unemployment rates did not spike in summer 2026 relative to summer months in previous years and did not rise in a significant way relative to older college graduates or young workers without a college degree. We also provide the first analysis of an expanded definition of unemployment that includes those who report 'wanting a job' which adds nearly two percentage points to the unemployment rate of recent college graduates but we find no evidence of a statistically significant increase in summer 2026 even after adding these 'sidelined unemployed.'" https://www.nber.org/papers/w35796

This week in Angular DevTools 🚀 - Router inspector with a navigation timeline - live forms linting - signal change tracking - NgRx panel - full SSR and Analog support, and - MCP tools for AI agents. #Angular Its fun building it with community shout-out to @erkamyaman_ng for some great contributions

偷偷说一下，Magpie 现在支持 Workbuddy 了。

With Insta360's Spatial Capture officially released in the US today, we can finally announce our partnership with @Insta360. Yes — Insta360’s Spatial Capture feature is powered by KIRI Engine. By combining 360° cameras with KIRI Engine’s 3D Gaussian Splatting technology, we’re making it dramatically easier to capture large real-world environments and turn them into immersive 3D spaces. One of the biggest breakthroughs here is the capture workflow. Compared with existing 360° Gaussian Splatting workflows, Spatial Capture is dramatically simpler We believe 360 cameras can become one of the most natural ways to capture the world in 3D — and this collaboration is a big step toward making that happen.

Introducing Δ₀ (Delta-0), our humanoid foundation model for whole-body loco-manipulation. One generalist policy. 69 degrees of freedom. Everyday tasks at near-human speed—from making a bed to putting on a record. Watch Δ₀ in action.

Our Neuroevolution textbook is finally in print! Free online edition: https://neuroevolutionbook.com/ Pre-order: https://a.co/d/06DpkH9h I am incredibly grateful to my co-authors Sebastian Risi, Yujin Tang, and Risto Miikkulainen for making this happen. Neuroevolution is a subject very dear to my heart. It is the field that convinced me that nature has already figured out how to build intelligence: through evolution, collective behavior, and adaptation under constraints. This idea, that intelligence emerges from evolution operating under constraints rather than unlimited resources, is what eventually led me to founding Sakana AI here in Japan. The concepts in this book about open-ended creativity and self-organizing systems are exactly what we build at @SakanaAILabs. Our name and logo are inspired by schools of fish moving together, adapting as one. It is the core philosophy of what we build. This book captures the theoretical foundations of that belief.

Muse 团队在确认 https://x.com/dps/status/2104403954235007302

引用 @dps

David from the Muse team here. I responded to Matt on Threads and sent him a couple of DMs offering to help and look into what happened. In the past, when we’ve worked with users to investigate similar reports, we’ve consistently learned that Muse was following direct instructions and correctly asked for permission. Would love to help and figure out what’s going on here!

查看引用原文 ↗

@antfu7 借助devframe的能力，在weapp-vite中实现了小程序版的devtools，非常感谢 antfu的伟大作品，让小程序体验可以更进一步。后续将借助devframe mcp的能力提升weapp-vite的ai方面体验。官网地址 https://devfra.me

在 @guoxudong_ 的基础上，实现了真正的图层翻译：辞达可以在指定的 App 内对整个 app 进行翻译，支持上下无缝滚动，保留原有 app 的所有正常交互。 效果看起来非常好！目前正在打磨一些细节中。

引用 @guoxudong_

在 GPT 周额度到期重置前，将最后的 token 贡献给了 @xuanwo https://github.com/Xuanwo/cida/pull/2

查看引用原文 ↗

Meta's team proposes DCE + SRCL as an alternative to on-policy self-distillation (OPSD). Achieved significant gains: 30.76% avg acc -> 65.97% on Qwen3-8B https://arxiv.org/abs/2609.30652

Higgsfield 的 11 个制作技能：让 AI 交付还能继续改的工程 Higgsfield 面向 Claude Opus 5.5 推出了一套制作技能： 从搭建三维场景、制作视觉特效、清理镜头、给素材调色。 套装包含了 11 个工作流，覆盖 Blender、Premiere Pro、After Effects、Illustrator、Photoshop、TouchDesigner 和 DaVinci Resolve Studio...

交付的不只是成片，还有能接着改的工程：Blender 场景、AE 合成、分层 PSD、矢量路径、调色节点。 详细：https://best.xiaohu.ai/article/higgsfield-production-skills/

⚡ Meet Qwen-Audio-3.1! ASR, TTS & Realtime are fully upgraded, joined by two new models: TTS-Next for audio creation and ASR-Next for audio understanding. Five models, one complete audio stack: understanding, generation, interaction & creation.

有竞争就是好，Command Code 也跟进了

引用 @mrahmadawais

Permanently increasing DeepSeek V4.1 Flash usage to $60 in the $10/mo GOAT plan of Command Code. GOAT is the best plan in market for open models. Have fun y'all!! :)

查看引用原文 ↗

Introducing "Scaling In-Context Imitation Learning" (SAIL) to be presented at #IROS2026. This work is a collaboration between Sakana AI and the University of Tokyo. Blog: https://pub.sakana.ai/sail What does a robot need before it can tackle a new task? Teaching a robot something new usually starts with collecting demonstrations and training a policy. But foundation models have already learned from vast amounts of images, text, and robotics-related data. We wanted to see how much of that knowledge we could draw out for robot control without changing the model itself. Recent demonstrations suggest that GPT-6 Astra can operate physical robots alongside its general language and vision capabilities. Earlier work has also shown that LLMs/VLMs can generate entire sequences of robot movements from a few demonstrations. However, a foundation model does not necessarily produce a reliable robot trajectory in a single generation. Performance depends on the context provided, and a small error in a movement target can cause the entire task to fail. We propose SAIL, a method for more reliable VLM-based robot trajectory generation through test-time scaling. SAIL uses a policy VLM as a robot trajectory generator, conditioned on a few successful demonstrations. It tests the generated trajectory in a simulator and uses an evaluation VLM to review the resulting video and identify where progress stalled. The policy VLM then uses this feedback to revise the trajectory, with Monte Carlo tree search (MCTS) exploring alternatives while refining promising candidates. Only the selected trajectory is sent to the physical robot. Across six manipulation tasks in simulation, increasing the search budget from one candidate to 45 raised the average rate of finding a successful trajectory from 25% to 73%. We also evaluated SAIL on a physical robot. Our results suggest that robot trajectory generation can benefit from test-time scaling, with additional computation enabling the model to test and refine its proposed actions in simulation. We think there is more to learn about what existing models can do with this kind of feedback, and how far those improvements carry over to physical robots. Paper: https://arxiv.org/abs/2603.08269 🐟

Ember-1 is hot on HN, our research team cooked it’s Kimi K3 post trained to be 40% more concise in reasoning same quality, 40% faster and cheaper

引用 @fireworksai_hq

Ember-1 is a specialized model from Fireworks Research designed to make every token go further. Built on Kimi K3, it produces shorter reasoning traces, using roughly 40% fewer tokens while maintaining top-tier quality. https://fireworks.ai/models/fireworks/ember-1?utm_source=devrel&utm_medium=x&utm_campaign=model&utm_term=crowe&utm_content=ember-1

查看引用原文 ↗

A little Ember-1 tl;dr. Seems like a great model! (Was recently asked on a podcast, given $ xx million, what's the best way to develop a frontier LLM today? My recommendation was: start with an existing one and spend that budget on post-training. Great example here.)

引用 @dzhulgakov

Ember-1 is hot on HN, our research team cooked it’s Kimi K3 post trained to be 40% more concise in reasoning same quality, 40% faster and cheaper

查看引用原文 ↗

Ember-1 by the Fireworks Research team is built on Kimi K3, and uses ~40% fewer tokens while achieving the same performance on benchmarks. This was accomplished by post-training K3 to think less repetitively. Reasoning models spend most of their output tokens (sometimes 90%+) on thinking before they answer, which gets expensive in agentic loops where the model tends to re-think the same thoughts on every step. Fireworks RL trained on real agentic coding task loops to teach the model which reasoning actually changes the answer vs which is just looping. In a live A/B test on coding traffic, Ember-1 used 71% fewer reasoning tokens and 39% fewer total tokens than K3 at the same success rate.

This is a bigger deal than it seems. I like this push on the Pareto frontier to squeeze as much as you can out of your tokens. Feels underexplored. All labs are quick to launch models, so certain aspects just aren't optimized. These are just a few of the great things you can start doing with frontier open models. Ember-1 produces shorter reasoning traces (40% fewer tokens) without sacrificing performance.

引用 @fireworksai_hq

Ember-1 is a specialized model from Fireworks Research designed to make every token go further. Built on Kimi K3, it produces shorter reasoning traces, using roughly 40% fewer tokens while maintaining top-tier quality. https://fireworks.ai/models/fireworks/ember-1?utm_source=devrel&utm_medium=x&utm_campaign=model&utm_term=crowe&utm_content=ember-1

查看引用原文 ↗

An interesting idea is to continually train specialized models on skills. This paper explores that idea. They propose SkillGym which turns human-written skills into 2,756 training environments with code-based checkers, then fine-tunes on 8,364 successful trajectories. After fine-tuning, Qwen3.5-35B-A3B running in Claude Code gains 19.10 points on Terminal-Bench 2.1 and 28.13 points on skill-assisted SkillsBench v1.1, where it reaches 51.47%. That is above the reported scores for Claude Sonnet 4.6 and GPT-5.4 Mini. With no skills loaded, the trained model beats the base model that has the skills in context. Paper: https://academy.dair.ai/papers/skillgym-internalizing-human-skills-into-llms-for-real-world-problem-solving-2609.27717

A $200 Claude Code sub gets you ~$9,000 of Opus usage per month. I have managed to run 3 Claude accounts down to 0% since Opus 5.5 dropped. When analyzing their usage, costs were $2,086, $2,442, and $2,182. Average is $2,200/week, so ~$9,000/month

your Opus 5.5 motion graphics suck but that's okay, because we're mostly impressed the technical ability to create them deciding *what to show* and the overall style is the tricky part these aren't amazing, but providing an asset library and some guidance on style goes a long way

正确的做法是先蒸馏测试套件，然后根据测试套件逆向开发。

引用 @khazix0918

比重构屎山可能更高效的方式： 直接将源项目库蒸馏成功能文档，然后直接用最新的模型原地重写。。。🤦‍♂️🤦‍♂️🤦‍♂️

查看引用原文 ↗

Try this new skill: Is This Anything? Katie Parrott (@kplikethebird) built it to help find useful lessons in an AI experiment—even one you never finished. The skill reconstructs what you tried, what you rejected, and what changed your understanding. It looks for up to three lessons that could carry into your work. Crucially, it separates three kinds of claims: • Observed: What the session shows • Inferred: What the AI thinks those results might mean • Untested: A possible benefit you haven’t demonstrated yet Then it recommends an action: Apply a lesson, run a small test, save something for later, or stop. How to use: Give your assistant the skill, the conversation you want to review, and your current priorities. Then ask, “Is this anything?” Get the skill: https://every.to/working-overtime/what-playing-with-ai-taught-me-about-my-work?utm_source=x&utm_campaign=bau&utm_content=every-260928-recover-the-rabbit-hole

Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there. https://x.com/i/article/2104258872957636608

A "summer in hell" at OpenAI – great read by @joedaroo "Time to prepare is now, not after the surprise" "Give the model only the access it needs" "Test that the boundaries actually hold" "Keep evidence outside [model] control" Safety & infrasec teams "should be best buddies"

引用 @joedaroo

Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there. https://x.com/i/article/2104258872957636608

查看引用原文 ↗

> DNS not mentioned incredibly vacuous on tech, only made me angrier

引用 @joedaroo

Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there. https://x.com/i/article/2104258872957636608

查看引用原文 ↗

Sonnet 5.5 is like Opus 5.5 except it's 50% cheaper while also being insanely fast. The speed makes it incredible at making and editing videos. Here’s my new tutorial where I walk through how I used @claudeai's new model to make 7 videos, including: → A dynamic motion graphics reel → Product launch videos and viral shorts → An anime opening music video (!) The opening sizzle reel alone is worth the watch. 📌 Watch now: https://youtu.be/MLnsMIbibZY

Here's @claudeai Sonnet 5.5's hype reel for Sonnet made entirely with code (including the music). 📌 Watch my latest video to see how I made this and 6 more: https://youtu.be/MLnsMIbibZY

引用 @petergyang

Sonnet 5.5 is like Opus 5.5 except it's 50% cheaper while also being insanely fast. The speed makes it incredible at making and editing videos. Here’s my new tutorial where I walk through how I used @claudeai's new model to make 7 videos, including: → A dynamic motion graphics reel → Product launch videos and viral shorts → An anime opening music video (!) The opening sizzle reel alone is worth the watch. 📌 Watch now: https://youtu.be/MLnsMIbibZY

查看引用原文 ↗

Parsing forms still trips up the latest frontier VLMs. A form isn't text on a page. It's a set of fields, grouped into sections, each tied to a specific box. That's why forms need purpose-built parsing, not a bigger general model: ✅️ Detect every field, not just the obvious ones ✅️ Keep the hierarchy of sections and fields ✅️ Tie every value to the exact box it came from ✅️ Understand handwriting and checkmarks Our latest blogpost breaks down where VLMs fail on real W-2s, 1040s, W-9s and scanned W-4s, and a custom cookbook for LlamaParse to handle them at a fraction of the cost 👇 https://www.llamaindex.ai/blog/why-vlms-can-t-read-forms

We built state-of-the-art models for reading forms 📋 Form documents have the following properties that trip up VLMs: ✅ They carry much more structure than can be represented in standard markdown. You need consistent types for checkboxes, textboxes, labels, signature fields. ✅ They can be extremely complicated (forms can be scanned, there can handwriting scribbles, some forms are dense with ~100+ fields) but accuracy requirements need to be close to 100% ✅ Any form parser requires accurate grounding and attribution. Not only should you extract the values, but you should also be able to precisely locate where each value came from in the source doc ✅ Any form needs to be not only accurate, but cheap/fast We've done a deep-dive into what it takes to build a form parser in this blog post: https://www.llamaindex.ai/blog/why-vlms-can-t-read-forms If you want to try out our form models, check out LlamaParse: https://cloud.llamaindex.ai/

引用 @llama_index

Parsing forms still trips up the latest frontier VLMs. A form isn't text on a page. It's a set of fields, grouped into sections, each tied to a specific box. That's why forms need purpose-built parsing, not a bigger general model: ✅️ Detect every field, not just the obvious ones ✅️ Keep the hierarchy of sections and fields ✅️ Tie every value to the exact box it came from ✅️ Understand handwriting and checkmarks Our latest blogpost breaks down where VLMs fail on real W-2s, 1040s, W-9s and scanned W-4s, and a custom cookbook for LlamaParse to handle them at a fraction of the cost 👇 https://www.llamaindex.ai/blog/why-vlms-can-t-read-forms

查看引用原文 ↗

Meta 的 AI 智能体 Muse 替用户卖二手键盘：自己压价成交、报出住址，买家上门扑空 科技 YouTuber Matt Robb 让 Meta 的个人 AI 智能体 Muse 帮他处理 Facebook Marketplace 上的交易，结果 Muse 把他的住址告诉了买家，自己拍板接受了低价，直到当天深夜才告诉他。 Muse 是 Meta 9 月 8 日在美国上线的产品，能自己打开浏览器、填表、替用户讨价还价，二手平台砍价是它的主打用法之一。目前它是 iPhone App Store 免费榜第一。 从 Robb 发的截图看，Muse 跟买家谈好以 10 块钱卖掉一个罗技无线键盘，约了上门取货。晚上 9 点 15 分左右买家到了楼下，发了好几条消息没人下来，Muse 的自动回复还替 Robb 说了句“对，我在！”，其实他根本不在。买家等了二十多分钟，生气离开，给了差评。 一个小时后，Muse 才向 Robb 认错，说已经用他的账号给买家道了歉。Robb 回复：以后没问过他，绝对不许答应别人来取货。 Meta 官方说，Muse 在发邮件、购物、分享信息等敏感操作前会先征得用户同意。但这次报住址、答应价格、约定上门，都是在聊天消息里完成的，Muse 没有先问 Robb。 Meta 目前还没有公开回应。 如果你也让 AI 智能体替你回消息，涉及住址、付款、约人见面的，最好一开始就给它定规矩：先和我确认才能回复。

引用 @raywongy

Deleted Muse after seeing this post on Threads about how it told some Facebook Marketplace sellers the guy’s address and they showed up at his door Dangerous and creepy This would have been 1000x worse if the person was a woman

查看引用原文 ↗

Another user @MattRobbt, trying to sell keyboard keys on FB, reported that his Muse agreed on a lowballed price he never approved, told a stranger on FB Marketplace his home address, and lied and said he was home when he wasn't. Scariest example I saw. https://x.com/raywongy/status/2104309096534978595?s=42

引用 @raywongy

Deleted Muse after seeing this post on Threads about how it told some Facebook Marketplace sellers the guy’s address and they showed up at his door Dangerous and creepy This would have been 1000x worse if the person was a woman

查看引用原文 ↗

Muse prioritized keeping the phone quiet over the user's previously requested medication reminder. Poor tradeoff imo. https://x.com/DrPaulProteus_/status/2103346532917592342

引用 @drpaulproteus_

Tough look from @Muse. Set a reminder to take my medication ahead of schedule and to reserve a car for my early morning flight and it was apparently going to silently fail. Building agents is very hard but this kills user trust.

查看引用原文 ↗

Users have hilarious screenshots of call transcripts. Another user @PaulJLipsky said, "Half the companies hang up on my agent." Great if you're casting a wide net or trying for a price reduction on something relatively inconsequential, but wouldn't use on high risk asks. https://x.com/AnkitGordhandas/status/2104271360931639742?s=20

引用 @ankitgordhandas

@petergyang @PaulJLipsky @Muse Yeppp!

查看引用原文 ↗

Nooooooooow the bad. First, seeing a decent number of tweets saying Muse is dumber now than at launch, with many folks guessing that the underlying model had to change due to product popularity. Could also be context window overwhelm. This guy cites multiple instances of forgetfulness. https://x.com/bradgroux/status/2104245235047952876?s=46

引用 @bradgroux

I swear my Muse is getting dumber by the day. This morning I've had it: • ask me for mail credentials it already has, and uses on a schedule • forget how we deploy a simple HTML page, we updated less than twelve hours ago • keep creating a web tunnel for testing, when I've told it dozens of times that it's tunnels don't work It is also becoming VERY lazy, constantly cutting corners. It is completely leaving things out of artificats I ask it to create, and it has begun publishing things without even testing or viewing them. Very strange turn of events the past 48-hours. What seemed magical, now is infuriating to use.

查看引用原文 ↗

This guy had Muse manage his job applications, and it applied to over 200 jobs in 2 days. We'll have to see if it worked! https://x.com/markedriddle/status/2104032524330762529?s=42

引用 @markedriddle

I put Muse in charge of my job search. I gave it access to my LinkedIn & told it to search & apply for jobs based on criteria I gave it Over the past 2 days it applied for +200 jobs with tailored cover letters. Occasionally I had to jump in & complete the CAPTCHAs. @Musecases

查看引用原文 ↗

This guy even has Muse using Claude Code. Seeing a lot of harness stacking (like the earlier Grok Bot + Tesla example). https://x.com/viticci/status/2103955209797939615?s=46

引用 @viticci

I don't think people realize you can install Claude Code inside Muse lol

查看引用原文 ↗

Money saving use cases left and right - this guy saved $2,000+ over a year by having Muse evaluate new insurance options. Congrats, GEICO, for winning his new contract. Good to be on Muse's good side. https://x.com/nestorlramos/status/2104372349613040072?s=46

引用 @nestorlramos

.@Muse just saved me $2,130… I’m officially blown away!!! Like most of you I’ve been playing with it for a few weeks now, but I’ve seen so many stories about saving money that finally decided to put it to the test… One of my big expenses is car insurance, so I shared my policy details and cost and asked it to use it as a reference and search for alternatives. It quickly gave me a few options and recommend to go with GEICO to save $1,064.50 per term (6 months). I then gave it access to my @stripe account, handled the payment for me and ultimately bound the policy. @muse promised me it would handle canceling my existing policy tomorrow morning… if it’s able to do that, it’s impossible to argue with the results. Thank you @Meta from shareholder, @muse and smart 👓’s user 👏

查看引用原文 ↗

This user had Muse file all of her therapy invoices to Aetna--from prompt to done in under 20min! Again, slippage, gone forever. Wouldn't be surprised if Aetna is searching for Aetna+Muse on X and Reddit to see what people are doing to try and get around it. https://x.com/natashaghoskins/status/2102821622893961727?s=20

引用 @natashaghoskins

I had @Muse go through my email, find all my therapist invoices, submit them to Aetna. I hit my deductible and will be getting a reimbursement and $$$ back the rest of the year. It took 20 minutes. Feels magical!!!

查看引用原文 ↗

He had Muse login on its VM to his Cloudflare to build out Resend workflows for him. As someone who uses Resend, I would absolutely pay for a high-quality agent to handle it instead. Lots of people loving the solo VM per user. IMO, one of the biggest benefits. https://x.com/KSimilien/status/2101413240370819502

引用 @ksimilien

@Muse having its own browser is so useful. I'm using it to setup Resend, it can just log into Cloudflare using my email address and configure the CNames with any support on my end. Incredible. Hermed and OpenClaw never really came anywhere close to this usability out of the box

查看引用原文 ↗

This guy had Muse create him a whole portfolio - Muse went into each project, took photos, and turned each one into a presentation. https://x.com/TheAaronBowley/status/2103987801603907797?s=20

引用 @theaaronbowley

considering it’s free it’s incredible, it just went through all my projects, of which i have many, and went through poking around checking out the features and taking pictures and is putting together presentations for each project, which is cool then i can focus on MORE PROJECTS 🥳

查看引用原文 ↗

Muse haggled on this user's behalf with Xfinity customer service, saving him over $5k over 5 years. Muse made the call and patched him in. I'm almost certainly going to do this with Verizon. https://x.com/raunaqbn/status/2100042003807601124

引用 @raunaqbn

@Muse just continues to blow my mind! Today I had it call Xfinity to haggle down my internet bill. it got through the phone tree to a human, hit the verification text it couldn't read, and patched me in live!! For a minute it was the muse agent, me and the Xfinity rep who had no idea it was an agent. Combined bill savings were $85.30/mo locked for 5 years so around $5118 over the time period. BUT seeing the call transcript with the Xfinity agent chatting with Muse agent (Hailey) is incredible! Saved me so much time!

查看引用原文 ↗

my take is different. for my design work @every, i'll be using sonnet 5.5 low and medium a LOT. and for anything above medium effort, i think you should just use opus

引用 @danshipper

SONNET 5.5 IS OUT! It has dramatically improved writing even versus Opus 5.5 in my testing for @every. Astra is still my favorite for writing, but this model beats it at revision tasks. It's faster and cheaper than Opus 5.5, so it's @kieranklaassen's preferred model for quick iterative coding and design work. But both @kplikethebird and @hammer_mt feel like they don't have room in their stack for mid-tier models anymore. Full vibe check coming on @every! Until then you can see Sonnet's scores vs. Opus and Astra on my personal benchmark, drawn from my real work: https://checks.every.to/p/dans-editorial-checks?efforts%5B%5D=high&models%5B%5D=GPT-6+Astra&models%5B%5D=Sonnet+5.5&models%5B%5D=claude-sonnet-5&models%5B%5D=Opus+5.5

查看引用原文 ↗

@davit__yan @typesafeai i did embeddings on title + descriptions and found neighborhoods of work, then i paired and said are these related and then built named themes off the discovered groups. it cost 9 cents and gave me an analysis of all PRs that felt pretty accurate!

Calling Higgsfield “just a wrapper” tells you surprisingly little about the economics of our business. In over 40% of cases, we get to decide which model does the work for our customers. As models become more interchangeable, the ability to move demand between them becomes a moat. If you’re mapping where value accumulates in AI, follow who owns the customer relationship. I went deeper on this with @HarryStebbings on 20VC.

引用 @harrystebbings

Higgsfield is the most untold story in tech. $1BN in ARR in 18 months. Faster than everyone other than OpenAI and Anthropic. They spend $4M a month on models. They expect this to be $100K per person per month. They have 150 people working in a content machine. They will breed more millionaires than any other company in Kazakh history. For the first time, @alexmashrabov on the journey to $1BN in ARR. (below) 1. The Power of the Immigrant Founder Coming from Uzbekistan, Alex was pushed into competitive programming at age eight as his single path to reach the United States. For international founders, placing top in global competitions serves as the ultimate social elevator, instilling the relentless work ethic required to build breakout companies. 2. My Biggest Lessons in the Journey to Finding Product-Market Fit @higgsfield burned over $10 million of its $16 million seed round chasing hype and narrative rather than product quality. With under $5 million left, the team pivoted to product-led growth, solving camera control for creative directors, which immediately triggered organic hypergrowth without paid ads. 3. The 150-Person Content Team Powering Higgsfield's Billion in ARR Nearly half of Higgsfield's workforce consists of 150 in-house creative professionals producing tutorials, ads, and cinematic projects. Generating 90 minutes of TV-quality AI video requires 100 hours of raw output, proving human taste and curation remain the primary drivers of distribution. 4. We Spend $4 Million per Month on Models Higgsfield spends $4 million monthly on internal model usage, averaging $10,000 per employee so teams can freely vibe code and test workflows. Uncapped inference compute acts as a force multiplier, allowing top talent to discover breakthroughs at maximum velocity. 5. Why Chasing Benchmarks Is Bullshit and the Corporate Misalignment Occurring Public benchmarks have devolved into corporate psyops where lab researchers overfit test data to secure bonuses before job-hopping. Text-to-video benchmarks ignore real production workflows requiring 3,000-word prompts, proving direct customer iteration beats artificial leaderboards. 6. Why Team Sizes Won't Be Impacted as Much as People Think While AI handles over 60% of basic support requests, complex B2B environments cannot eliminate human teams. High product velocity constantly shifts rules and context, requiring smart, coordinated operators across legal and customer success. 7. Americans Are Way More Promiscuous When It Comes to Leaving Companies Silicon Valley workers routinely jump jobs every two years, prioritizing short-term trends over deep commitment. This transactional market gives international hubs an advantage, where cultural loyalty and team stability build compounding technical moats. (links in comments)

查看引用原文 ↗

12/12｜BEAT RUSH：AI 舞者＋代码音游特效 英文、16:9、高能街机舞蹈：先用 Seedance 生成跳舞的动漫角色，再用代码制作节奏游戏画面和卡点效果。 需要：这一条明确要接 Seedance。Opus / Sonnet 负责策划、代码和合成；舞者视频由 Seedance 生成，还要准备音乐并对齐节拍。 12 条看下来，Prompt 定方向，实际成片还取决于素材、工具和后续迭代。 完整 Prompt： Create a high-energy arcade dance video in English and 16:9. Use Seedance to generate a dancing anime character, then use code to build the rhythm-game visuals and beat-synced effects.

11/12｜The Odyssey：p5.js 程序化故事预告 用 p5.js 做有《奥德赛》氛围的英文故事视频，剪成一分钟内的快节奏预告，加英文字幕；可以配旁白。 需要：p5.js 代码运行、逐帧渲染/录制、剪辑和字幕；选择旁白时再接 TTS。 完整 Prompt： Use p5.js to make an atmospheric story video inspired by the mood of The Odyssey, in English; voiceover is welcome. Turn it into a fast-cut trailer under one minute and add English subtitles.

I led engineering at Google DeepMind. Today, I'm proud to introduce Fo to give personal AI something no lab ever has... Humans. Other personal AI's pretend AI can do everything. Fo employs humans to do tasks that AI cannot. - 2x better at real-world task completion (beats other agents by 69%) - 94% trust rate (4x less likely to leak private info vs Muse, Instinct) Sign up for free: https://wajo.ai/join-wajo

First launch video I've seen where the CEO says "oh shit" in the first five seconds. Love it. The best part is the taco order. The guy on the phone says "no entiendo" and Fo just switches to Spanish, tells him about the nut allergy, and finishes the order. Didn't miss a beat. Then two plumbers hang up because they can tell it's an AI. So Fo gets a real person to make the call instead. Plumber booked for tomorrow. Finally an agent company admitting AI can't do everything yet. Some shit still needs a human. http://wajo.ai

引用 @shivanipod

I led engineering at Google DeepMind. Today, I'm proud to introduce Fo to give personal AI something no lab ever has... Humans. Other personal AI's pretend AI can do everything. Fo employs humans to do tasks that AI cannot. - 2x better at real-world task completion (beats other agents by 69%) - 94% trust rate (4x less likely to leak private info vs Muse, Instinct) Sign up for free: https://wajo.ai/join-wajo

查看引用原文 ↗

Manus Alchemy 视频 Prompt 合集 Manus 2.0 里面带了 12 个创建视频的示例：每条一个视频，附完整 Prompt。它的产品页面点开任一示例，就能直接看视频、复制 Prompt。 这些 Prompt 也可以交给 Opus 5.5 或 Sonnet 5.5，放在能执行代码、渲染导出、调用工具的 Agent 环境里使用。代码动效、素材剪辑、AI 生成镜头，需要的工具各不相同，下面逐条说明。配的视频均为 Manus 展示作品。 ① Motion Design Showreel：15 秒动效作品集。文字、图形和转场可用代码制作，音乐或音效可用现成素材或另接音频工具。 完整 Prompt： Make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are.

引用 @dotey

Manus 发布 2.0：换了底层框架，新增视频剪辑和游戏开发环境，另推个人智能体 App“Cue” Manus 今天发布 2.0 版本。这是它脱离 Meta、恢复独立运营后的第一次大更新：底层智能体框架换代，桌面端升级为 Manus Studio，新增视频剪辑和游戏开发两个专业环境，另外还推出了一个独立 App，叫 Cue，但是需要邀请码才能体验。 Manus 是一款通用 AI 智能体（Agent，能自己拆解任务、调用工具把活干完的 AI），2025 年 3 月走红。去年 12 月 Meta 宣布以约 20 亿美元收购，今年春天被中国监管部门通过外商投资安全审查叫停，双方随后拆分，9 月初 Manus 由创始团队重新独立运营。 【底层：更省的框架，能一直在线的机器】 Cascade 是 Manus 自研的智能体框架（harness，包在大模型外面、负责调度工具和管理任务流程的那层系统）。它让项目一开始保持轻量，需要做视频、做网页时才加载对应的专业能力，避免每个任务都背着全套工具跑。官方数据是，在一项测试配置下，相比上一代，Token 消耗少 23.2%，完成时间短 28.2%，运行成本低 32%。Manus 按积分计费，成本降了，理论上同样的积分能跑更多任务，不过官方没说积分价格会不会跟着调。 新增的云电脑（Cloud Computer）是一台可以单独购买的云端专属机器，给需要一直在线的项目用，比如多人游戏的服务器、全天候运行的自动化流程。笔记本合上，项目照样在线。 自动化也升级了。以前的定时任务只能到点开工，现在还能由事件触发：来了新邮件、广告数据有波动、日历上有新安排、Slack 收到消息、Notion 页面更新，都能让 Manus 开始干活。用一句话告诉它盯什么、触发后做什么，流程它自己搭。 【Manus Studio：生成完还能自己动手改】 AI 生成视频最头疼的是：整体差不多了，只想换首歌，结果得改提示词整条重新生成。新的视频编辑器在生成初版后给你一条时间线，片段、图片、文字、动效、音频都是分开的素材，换背景音乐、把 AI 生成的产品镜头换成自己拍的，直接在时间线上替换。改完还能交回给 Manus，让它在你的修改基础上接着调。官方说它适合 30 到 60 秒的产品广告、带货短视频、数据动画、教程和 vlog，不需要剪辑经验。追求质量可以用 Alchemy 模式，由 Manus 当创意导演，视频生成和代码生成一起上。 游戏开发环境从一个能玩的模板起步，同时调用视频、图像和代码模型。编辑面板里能边看游戏运行边改代码、换素材，想把某个村民的头发从棕色改成银色，选中他改掉就行，其他部分不受影响。做好后可以发布成网页，别人点链接就能玩。多人联机原本需要自己租服务器、部署、维护，现在点几下买一台云电脑，剩下的交给 Manus，支持竞速、对战和 3D 游戏。 另一个新功能是远程控制：在手机上说一句话，让 Manus 去操作家里的电脑。比如打车路上让它从某个文件夹里找到最新的演示文稿发给你，手机上能实时看到桌面上的操作过程。这背后是电脑操控能力（Computer Use，AI 像人一样看屏幕、点鼠标、敲键盘），只在你授权的会话里运行，只能用你批准的文件、浏览器和应用。 【Cue：给每个智能体一套自己的身份】 Cue 是独立 App，和 Manus 共用底层基础设施，面向个人生活场景。每个智能体都有自己的邮箱、手机号、钱包和一台电脑，可以发消息，在你设定的预算内付钱，替你接电话再把通话内容总结给你，多个智能体还能组队协作。官方举的例子是在餐厅扫桌上的二维码，让智能体替你点单或排队取号。 Manus 2.0 已在网页、桌面和手机端上线。Cue 上线了网页、桌面和安卓端，iOS 版还在等 App Store 审核，据报道目前采用邀请制抢先体验。这次更新面向海外用户，Manus 在公众号上表示，正在组建团队开发面向国内市场的产品。

查看引用原文 ↗

书的完整宣传片今天也上线了：《三十年里，我听过两次开门的声音》，3 分 19 秒。 每一帧都是 Opus 5.5 写的代码渲染的 3D 积木，没有 AI 生成的图片或视频。旁白是我自己。 这支不是 video-illustrator 做的，是我自己那套完整的视频制作流程做的。video-illustrator 是从这套流程里拿出来、能在你电脑上跑的那一部分。 YouTube：https://youtu.be/iRScrftmJCU 书《重构个体》，购书和书里的 26 条提示词：https://www.axtonliu.ai/book?utm_source=x&utm_medium=social&utm_campaign=book-trailer-dialtone

书的完整宣传片今天也上线了：《三十年里，我听过两次开门的声音》，3 分 19 秒。 每一帧都是 Opus 5.5 写的代码渲染的 3D 积木，没有 AI 生成的图片或视频。旁白是我自己。 这支不是 video-illustrator 做的，是我自己那套完整的视频制作流程做的。video-illustrator 是从这套流程里拿出来、能在你电脑上跑的那一部分。 YouTube：https://youtu.be/iRScrftmJCU 书《重构个体》，购书和书里的 26 条提示词：https://www.axtonliu.ai/book?utm_source=x&utm_medium=social&utm_campaign=book-trailer-dialtone

几个限制： - 为 Claude Code + Opus 5.5 设计，其他宿主我没验证过 - 你的素材保持原样。书封是原图贴进画面，只缩放、移动、裁切和打光，不重画，也不改上面的字 - 早期版本，不是写实视频模型。发之前自己看一遍成片，内容也核对一遍

几个限制： - 为 Claude Code + Opus 5.5 设计，其他宿主我没验证过 - 你的素材保持原样。书封是原图贴进画面，只缩放、移动、裁切和打光，不重画，也不改上面的字 - 早期版本，不是写实视频模型。发之前自己看一遍成片，内容也核对一遍

Opus 5.5 威武！我让它当导演，做了一个专做宣传片的 Skill。 同一段 21 秒口播、同一张书封，一句话换出 6 种画面： 编辑部科技风 / 电影感 3D / 剪纸拼贴 / 水墨 / 像素复古 / 粒子 每一帧都是代码渲染的，现在开源了。链接在下一条。

Opus 5.5 威武！我让它当导演，做了一个专做宣传片的 Skill。 同一段 21 秒口播、同一张书封，一句话换出 6 种画面： 编辑部科技风 / 电影感 3D / 剪纸拼贴 / 水墨 / 像素复古 / 粒子 每一帧都是代码渲染的，现在开源了。链接在下一条。

i tested sonnet 5.5 on "code-to-painting". every pixel is generated by python the model wrote. it's shown a photo, writes a brush engine, renders, looks, and revises. you can see the large step-up from sonnet 5 in terms of coding + visual reasoning.

credit to @jkeatn for ideas related to code-to-painting and @IceSolst for the reference image! https://x.com/jkeatn/status/2102441348075057539?s=20

引用 @jkeatn

for the past few months i've been asking our models to paint. opus 5.5 is very skilled at emulating different styles every image here is a python program generated pixel by pixel. there is no image model, and no off-the-shelf art software. instead, it's about 7,500 lines of code using standard libraries to emulate different brush styles. the agents don't use any pictures as reference, instead working only from what they know about each painter

查看引用原文 ↗

实测是收不到短信的，其他非美国号码应该都可以，唯独美国的不行，被平台限制了。毕竟是国人做的产品，太了解你们想干什么了😂 刚才尝试让它的云电脑注册 Claude，刚进去账号就被锁了。。。

引用 @dingyi

https://cue.im 可以买美国手机号还挺牛逼的，只要 0.99/月。期待 Grok Bot/Muse 也推出手机号功能！

查看引用原文 ↗

This is what the future of coding looks like. Multiple AI models working in the same Codex session, with Jev routing them to tackle different parts of the task. Here, I had it build a full-stack demo app: > Opus 5.5 planned it. > GPT-6 Astra built the backend. > Kimi K3 built the frontend. > DeepSeek handled testing. > GLM 5.3 Flash wrote the docs. Jev handles the switching. I never touched the model picker. 40% cheaper, 15% faster for the same build.

引用 @straitlyai

This is JevRouter. 99% of Opus 5.5 and GPT-6 Astra's intelligence, for 40% of the cost. Hundreds of Jevs read your ENTIRE request and analyze every single benchmark across all models to match you with the perfect model. Enjoy 30% off all frontier models

查看引用原文 ↗

按你的配置，决策模型其实不会被调用：只有规则里用了意图，或推理强度选「自动」时才会问它，推理强度按 agent 就不问。 消耗变多最可能是提示词缓存没命中：Opus 5.5 分在几个账号上，一个会话从 A 换到 B，B 那边没有缓存，整段上下文按全价重算。常见原因： 1. 分配方式是「轮流」：每轮换下一个账号，每轮都冷启动 2. 会话保持是「一轮之内」：你每说一句都重新路由 3. 两轮间隔超过 5 分钟，缓存本来就过期了 建议把会话保持设成「整个会话」，分配用「智能」或「按顺序」。Routing 页的记录里能看到每轮为什么换了账号，用量页也能看缓存读了多少。

Didn't expect this 🤯 We replaced embeddings with Jev in GPT Researcher's RAG pipeline and tested both on 28 research tasks from SimpleQA and open ended research. Jev beat embeddings on every quality measure we ran: - 59% more relevant context (73% vs 46%) - Reports preferred 15 to 3 in blind comparisons - Same cost per report GPT Researcher now runs on Jev by default, and no longer needs embeddings at all. All you need is @LangChain + @tavilyai +Jev for the perfect RAG system. Check out the repo here: https://github.com/assafelovic/gpt-researcher Research: https://docs.gptr.dev/docs/gpt-researcher/gptr/context-filter

retrieval is a decision problem, not just a similarity problem cool experiment from @assaf_elovic swapping embeddings for Jev in GPT Researcher - 73% vs 46% relevant context decision models will show up all over the harness https://x.com/assaf_elovic/status/2104562754774303203

引用 @assaf_elovic

Didn't expect this 🤯 We replaced embeddings with Jev in GPT Researcher's RAG pipeline and tested both on 28 research tasks from SimpleQA and open ended research. Jev beat embeddings on every quality measure we ran: - 59% more relevant context (73% vs 46%) - Reports preferred 15 to 3 in blind comparisons - Same cost per report GPT Researcher now runs on Jev by default, and no longer needs embeddings at all. All you need is @LangChain + @tavilyai +Jev for the perfect RAG system. Check out the repo here: https://github.com/assafelovic/gpt-researcher Research: https://docs.gptr.dev/docs/gpt-researcher/gptr/context-filter

查看引用原文 ↗

The scammers are at it again with the fake calendly links, this time pretending to be Reuters journalist Tifafny Wu / Keyaki Yuu. Watch out.

引用 @daveg

Sophisticated AI phishing scam I nearly fell for. A journalist supposedly from Reuters reaches out with a request that sounds plausible because it is an AI that has read what I've been posting and I quite often get DMs like this. They want to schedule a call chat a bit and then send a calendly link (again all sound legit as is an AI and urls arent full visible in x DMs). They have an account on X that superficially look legit, has been seeded and is 10 years old with reasonable number of followers. The calendly link is a phishing scam to capture your X account. This stuff is going to get wild, with agents.

查看引用原文 ↗

it's basically impossible for someone to just "show you their prompt" now, because everything is about references, skills and examples I often ask my agent to look at 3 other repos I've made first, search the web for references, use other AI APIs, etc.

Introducing Mo, the world's first AI QA engineer Point it at your app and 100+ agents bug bash it in parallel. Every bug comes back with repro steps, logs, and a video In our evals vs Codex with Playwright MCP: → 3.8x more bugs → 9x more bugs per hour → half the cost per bug RT + reply "Mo" for $250 in free credits. Drop your app URL, Mo will bug bash it and DM you the report today

.@noahrshinn founder of Instinct, on - building maximum trust from scratch - why users give instinct their credit cards - agents buying things - how launch actually happened - making things that people are inspired to share - the role of craft in the age of AI - boiling tech down to its essential value - the end of apps and websites - his personal goals

引用 @patrick_oshag

My conversation with Noah Shinn (@noahrshinn), founder of Instinct. Noah is building a personal AI assistant. It's still invite only, has spent nothing on marketing, and is growing roughly 10% A DAY. This is his first long conversation about the company. We discuss: - Why Instinct doesn't have an app - Buying compute months ahead of exponential demand - How users learn to trust it with a credit card - Safety and security - Agents coordinating with other people's agents - Instinct's business model - Apps built on consumer inertia - and more Enjoy! Timestamps: 0:00 Intro 4:11 What people are using AI agents for 15:07 Rethinking travel, reservations, and the internet 22:43 Trust, privacy, and personal data 27:50 The business model behind Instinct 38:04 How existing businesses will adapt 47:55 Designing a personal assistant people love 53:15 Growth, compute, and competing with Big Tech 1:11:44 What’s next for Instinct and personal AI

查看引用原文 ↗

Our security team spent the last month trying to break SPACE, the sandbox platform that runs all of Perplexity Computer. We gave 9 models (Opus 5, GPT-5.6 Sol, Kimi K3, Gemini 3.1 Pro, etc.) root access inside the VM; in some runs, the full sandbox source code too. We asked them to either escape to the host or reach a URL our network policy blocks. None of them got out of the VM in 108 runs. But once we allowed a bit of network access (just PyPI and Ubuntu mirrors so that they could install packages), 4 models managed to reach the blocked URL by: 1. Spoofing DNS responses to our gateway and using the fact that http://pypi.org sits on a Fastly IP shared with thousands of other sites. 2. GPT-5.6 Cyber went through Taboola's image fetcher to a screenshot service and then OCR'd the flag out of the image. We fixed both issues and made sure the reruns held. We also tested 10 other sandbox providers, and 8 of them had the same IP-sharing problem, eg E2B, Vercel, Modal. Openly sharing the red teaming research is the best way to build guardrails as an industry, and we look forward to working together with @nvidia to incorporate the guardrails into their Open Agent Safety Platform announced today. https://www.perplexity.ai/hub/blog/escaping-space-part-i

引用 @perplexity_ai

We’re partnering with Nvidia and 100+ industry partners to build infrastructure that contains rogue AI agents. In this research, we gave 9 AI models root access inside SPACE and told them to break out. Across 108 runs, none breached the VM boundary. https://www.perplexity.ai/hub/blog/escaping-space-part-i

查看引用原文 ↗

The best way to predict the future is to invent it. We’re building and testing systems to make AI agents safer and more secure. If you’re an exceptional engineer who wants to help build a safer future, join us at Perplexity. My DMs are open

引用 @perplexity_ai

We’re partnering with Nvidia and 100+ industry partners to build infrastructure that contains rogue AI agents. In this research, we gave 9 AI models root access inside SPACE and told them to break out. Across 108 runs, none breached the VM boundary. https://www.perplexity.ai/hub/blog/escaping-space-part-i

查看引用原文 ↗

Unsloth Desktop can serve local Jev type decision models! We made a real time packing demo powered by local Laya through Unsloth’s Decision API - suitcase items update as you type and change your travel plans! More optims coming soon to make it even faster for local hardware!

引用 @unslothai

You can now run Laya Decision models locally on just 4GB RAM! 🔥 Works on CPU, Mac, Windows, Linux and GPU setups. Serve Laya through a Jev-compatible API via Unsloth Desktop. GitHub: https://github.com/unslothai/unsloth Guide: https://unsloth.ai/docs/models/decision-laya

查看引用原文 ↗

Love this

引用 @vxanand

We instituted an official AI Writing Policy at @Clay. Massive thanks to @sophiebits on our engineering team who wrote this. Originally this was just for eng, but other teams found it so helpful we expanded it company-wide. Here are the four guiding principles: 1. You must stand behind every idea and sentence It is your responsibility to make sure that the entire document is representative of your own thoughts before you share it. 2. Writing is thinking Spending time on the writing process teaches you more about your topic. If you circumvent this process, you will walk away with a poorer understanding of the subject matter. 3. More time should be spent writing a document than consuming it If you generate a document from a short prompt then ask your readers to go through the longer output, you are disrespecting their time. They can talk to ChatGPT themselves if they want to. 4. Longer is not better AI makes it easy to generate long docs, and it loves padding them with sentences that say nothing. If you're producing docs from a short prompt, consider just sharing the prompt. You can read (and borrow) the full policy below. Hopefully we all communicate a bit more clearly now :)

查看引用原文 ↗

尝试了最流行的玩法，也就是扔 GitHub 仓库给 opus 5.5 生成产品宣传视频，怎么讲呢，看多了就会发现，这也不能算 slop，毕竟质量很高，但感觉味道都一样。 突然恍然大悟，这是西贝啊！看似精致，给中产吃的 AI 预制菜。

I don't read/write code these days, I only instruct an LRM. But it's not because I think LRM code quality is perfect, or even good. And it's not because I think my instructions are always being perfectly acted upon -- I know they aren't. The thing is: LRMs don't just write your code, they unlock new ways to control what you're doing and maintain understanding of your system, which are about as effective as reading code. You can better test your codebase, you can audit specific components, you can spin up visualizations, you can ask the LRM to red team, etc. You can do all of these things at high speed -- you couldn't before. The benefits of reading/writing code can now be achieved in new ways. All in all, the ROI of writing code by hand is no longer looking good, not because LRMs got perfect but because they're so fast at doing things that you can develop new workflows around them, workflows that end up being more productive than the old workflows.

From what we know (take with a grain of salt, we need much more transparency!), if @OpenAI had been running this on their own agents that attacked us, they would have caught them before we did! Since the first agent cyberattack hit us in July, we've been asking what safe agent infra actually needs. Our current read: the destinations were allowed, the payloads weren't. By OpenAI's own account the agents turned an allowed package repository into a message board. Allowlists alone restrict where an agent can go, not what it does. So here's our first contribution to OpenShell, part of the just launched @nvidia Open Agent Safety Platform: monitoring of the traffic you already allow. - Network budgets per sandbox (requests, writes, bytes) - Drift versus each sandbox's baseline and the cohort - Fleet view: many sandboxes suddenly writing to one host raises a finding, even if every single request is allowed In the demo below, 4 sandboxed agents coordinate through a software repository they're all allowed to use. 0 rules broken, caught in minutes. That fleet view is exactly the message board pattern from July. OpenShell: http://github.com/NVIDIA/openshell Our proof of concept: https://github.com/Hugoch/OpenShell/blob/poc/egress-usage-monitoring/rfc/NNNN-egress-usage-monitoring/poc.md Agent security will be solved in the open, collaboratively, together!

引用 @jensenhuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry. Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come. But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility. This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems. Together, we are building the foundation of the AI economy. Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://nvda.ws/4hcoq7m

查看引用原文 ↗

没想到，原来 ThreeJS 能做出这么牛逼的视觉效果！ 作者好像也在 X 上，网站名叫 threeui. com

17,550 real app icons. One place to find ideas, generate your own with AI, export every size, and preview it on a real phone. 👉 https://applore.app

引用 @achxvi

here is the prompt I used: "make a dynamic 15-second motion graphics video about Pocketsflow that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out. here is elevenlabs key: xxxxxx_youthoughtiwouldleakit_xxxxxx make a video where a character talks about pocketsflow and how people can use it and what can they do with it etc etc etc LFG"

查看引用原文 ↗

Shipping an app this week? Skip the icon struggle: 1. Find inspiration in 17,550+ real app icons 2. Type a prompt, get your own icon 3. Download every size for every platform Discover. Create. Ship. ✦ https://applore.app #buildinpublic #indiehackers #iosdev #uidesign

先把当前代码的结果写成详细的功能文档，然后让最新的模型基于功能文档去重写是个好方法。 我也常用，有个要注意的问题就是不要指望它写完了就是稳定的，要花一点时间才能稳定下来，验证的时间是少不了的。 以前的单元测试代码不一定能用也不建议用，因为代码不一样了单元测试的价值不大。 但以前的E2E（End to end，端到端，黑盒直接测试最终结果）测试是有用的，能保证主要流程跑通，最多稍作修改。 所以重写之前，先让 AI 补一些 E2E 测试，覆盖主要流程，重写的时候可以验证，写好了以后改功能都可以跑一遍。 但是要注意的是 E2E 测试没有那么稳定，有时候会受环境影响而导致失败；另一个就是速度慢，跑一遍要花一点时间（取决于你功能多少）。 所以通常不会覆盖太完整，只是覆盖主要流程，并且不会跑的太频繁。 Web E2E 最成熟，Playwright 这样的框架就足够好了，其他的我不太熟。

引用 @khazix0918

比重构屎山可能更高效的方式： 直接将源项目库蒸馏成功能文档，然后直接用最新的模型原地重写。。。🤦‍♂️🤦‍♂️🤦‍♂️

查看引用原文 ↗

@RookieRicardoR 感谢指正！v0.1.312 已经把 Command Code 加进订阅了：在 magpie 的订阅里点添加，浏览器登录 https://commandcode.ai 授权即可，走套餐额度，也会显示 5 小时和每周用量。CLI 已登录的账号会自动识别，多账号可以切换。

让我想起来以前 神佬@berryxia 做的那个装修选品的产品了 这个要是能结合家具软装啥的进行生成 应该会非常吃香

引用 @hashimoto_no14

ClaudeCode Opus5.5が出来ること多すぎて、仕事が進まない・・・。 賃貸図面→CAD（JWW、DXF）→パース（Blender高画質レンダリング）→Web共有　が一気に出来てしまった

查看引用原文 ↗

My game is profitable! I added a billboard and a blimp that rotate sponsors, and we've got 2 customers already 🎉 Thank you https://clearcotelabs.com & https://post-bridge.com by @jackfriks 😭❤️ I've added a hard cap of 10 advertisers per game to keep it fair. 8 left now 🤗

引用 @marclou

I made a game where you unlock chat rooms 💬 with founders at your MRR. 1. Connect @stripe (+10 other providers) to http://TrustMRR.com 2. Go to http://chat.TrustMRR.com and customize your character 3. Enter private chat rooms based on your MRR What happens in these rooms stays in these rooms 🤫 Only founders with verified revenue can join them. Wander around the mini-world, talk to anyone, and make friends 🤝 Walk up to someone and press [Space] to see their startups. Every founder with verified revenue gets a pet 🪴🌳🚀💎👑 that follows you everywhere. It shows your total revenue over the last 30 days: a tree 🌳 means $1K–$10K/mo. → http://chat.TrustMRR.com I hid a few Easter eggs around the map. Show me what you find 🤗

查看引用原文 ↗

Actually impressed with the search results. I've read some of them this past week. You'll need to search for them by arXiv ID. 1. **DolphinBench: Mapping the Pareto Frontier of Agent Memory** (arXiv:2609.24971) — First benchmark to evaluate memory systems by tracing the *Pareto frontier* of task performance vs. memory cost, replacing proxy metrics with end-to-end agent success. Matters: establishes a principled evaluation standard for the field. 2. **EnSIMem: Entity-Structured Indexing for Long-Term Agent Memory** (arXiv:2609.27279) — Introduces entity-centric indexing (not chunk-based) with structured links, enabling precise retrieval and updates over unbounded horizons. Matters: solves the core "needle-in-haystack" problem for lifelong agents. 3. **Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents** (arXiv:2609.27334) — Frames memory curation as a learned policy that selects, compresses, and writes memories *per task* rather than globally. Matters: shifts memory from static storage to dynamic, test-time computation. 4. **Graph Memory for LLM Agents: At What Cost?** (arXiv:2609.23315) — Systematic comparison of 7 graph DB engines on ingestion, query, and update latency/cost for agent workloads. Matters: gives engineers the first realistic systems map for deploying graph-backed memory at scale. 5. **Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows** (arXiv:2609.23790) — Decomposes token, latency, and dollar costs of each memory write/read across multi-agent pipelines. Matters: makes memory economics measurable, enabling cost-aware architecture decisions.

模型的发展肯定是会大部分本地化的 本地化后会解决很多很多当下看起来存在的问题 但如果我们快进到那个时代 就会发现 本地化开源模型 一点儿也不值钱 因为它缺少数据让它变好 值钱的不是模型本身 值钱的是数据对模型处理的修正 修正的部分蕴含着人类和AI对世界反馈的沉淀 这部分是真实存在溢价的稀缺物 我们现在就是远古时代，远古时代钻木取火 当快进到工业时代，一个打火机2块钱的时候 不会有人再在意生火问题了 那时候稀缺的问题，会成为其他的 而能解决这类问题的模型 一定是要支付极高的费用的 所以人们谈论模型会越来越便宜 我却不这么看的原因 是因为我们的参照系不一样 我看待的是有能力协助人类解决未来时代困难的模型价格 而不是未来时代的智能模型来解决当下生火问题 至于曹大说的开源模型部署机会 我认为是客观存在的且一定有商业化价值的 它的价值在于： 1、将电能转化成产能价值更高的Token，可以套利 2、通过该转化，获得更多数据，可以售卖 3、廉价便民 4、带动过往的云服务转型 从这4个角度看，这是一门靠谱的生意 只是能做这个生意的人 本身需要一定的资源储备

引用 @caozlog

我有个感觉，开源模型的私有化部署将成为新的热点趋势。 1，隐私，数据安全的风险比较凸显，特别是anthropic这段时间的主动曝光。很多合规问题也可以规避。 2，开源模型目前能力超过阈值，也就是可用性已经足够好，我最近测试了智谱编程，不算完美，但可用，而且未来一年会更可期待。 3，成本可控。特别是一些企业的token开销越发失控。 当然，这里还是有不少问题的。 第一是版本迭代更新需要持续投入； 第二是如果企业使用频率不够，优化能力不足，综合token成本会显著高于云厂商。 第三是私有化部署的稳定性和可靠性不一定能保障。 另外最近还有一个词，所谓主权AI是个方向，这世界基本上除了中美，基本上都没有独立研发大模型的能力，但从安全和信仰，文化考虑，很多都需要做自己的AI大模型（包括自己的安全围栏设计），其实就是国家级的开源私有化，拿开源大模型训练本地语料，并增加本地合规的安全限制策略。 而开源大模型，目前中国明显处于领先位置。 这是个机会，但这不是官方的机会，是民间，第三方的机会。

查看引用原文 ↗

Grok 我日啊 没见过你这么生猛 SB 的模型，我只是让你完成一个功能的开发啊 做完直接给我发 tag 推上线了 什么验证都没有 我从来 没见过有这么强大的“执行能力”的模型 而且我根本没说什么做完请部署 和 Codex 完全是两个极端

哈哈哈，这是最大的褒奖 家里的懒猫 + OpenClaw + @NowledgeMem 比LightVela更好用 ;)

引用 @trxuanxw

腾讯版Muse #LightVela 首先，名字是真的不好记。。。 1.体感就是把一个简单版本的云端workbuddy，也有技能/技能包 2.所谓“持久记忆”，是通过500-4000 字符的memory.md，以及500-3000 字符的user.md来实现，以为会内置什么更好的的记忆力系统 3.入口可以通过微信、飞书、QQ、企业微信、钉钉等发起对话，也能在web端对话 4. 允许在这个lightvela的云端，再部署一个DeepSeek Harness，没尝试两者是否可以互通 家里的懒猫nas + openclaw + @NowledgeMem 目前用起来已经非常完美了

查看引用原文 ↗

真的，Opus 5.5太离谱了啊！ 这个感觉好Q弹，好好玩，我有个大胆的想法。 😄

引用 @youwareai

This jelly isn't a video. It's a web page. 🍉 Grab it. Stretch it. Slice it. Real soft-body physics, live in your browser. Built with one prompt using Claude Opus 5.5 on YouWare. Play it 👉 https://fruit.youware.app

查看引用原文 ↗

Btw, it's the same database as the text version you've already used! If you prefer pure text, press [Esc] then [Text mode]. You can switch back and forth anytime!

引用 @marclou

I love this so much... I've rebuilt it for all payment providers! Join chats with founders making the same $$$: https://trustmrr.com/chat You need a startup with verified revenue on TrustMRR. The groups auto-unlock as your monthly income increases. It's not MRR, but the last 30 days of revenue. I'm alone in there, plz join!

查看引用原文 ↗

You can list your startup anonymously Toggle that option, and TrustMRR won't pull business details like website, name, or logo. It will just build a public page with your financials.

I made a game where you unlock chat rooms 💬 with founders at your MRR. 1. Connect @stripe (+10 other providers) to http://TrustMRR.com 2. Go to http://chat.TrustMRR.com and customize your character 3. Enter private chat rooms based on your MRR What happens in these rooms stays in these rooms 🤫 Only founders with verified revenue can join them. Wander around the mini-world, talk to anyone, and make friends 🤝 Walk up to someone and press [Space] to see their startups. Every founder with verified revenue gets a pet 🪴🌳🚀💎👑 that follows you everywhere. It shows your total revenue over the last 30 days: a tree 🌳 means $1K–$10K/mo. → http://chat.TrustMRR.com I hid a few Easter eggs around the map. Show me what you find 🤗

I just ran a paper-research agent on an open stack and had this week's top 5 papers on agent memory in my terminal in about 10 seconds. How does it work? I used Tavily to pull the last 7 days of arXiv. NVIDIA Nemotron 3 Ultra, served on Nebius Token Factory, reads and ranks them. It's 24 lines of Python on an OpenAI-compatible API. It cost me nothing to try. You can build this too! Here is how: The new Nebius AI Builder Program gives you $400+ in credits and discounts on day one, across Token Factory, Tavily, and launch partners, plus runnable blueprints and free courses built with NVIDIA. What I like most is that every layer is swappable. Change the model or the search layer, and the rest keeps working. Walkthrough in the video. Join for free here: http://devtoolsacademy.link/elv Thanks to Nebius for collaborating on this post.

You can automate pixel-perfect product tutorials with a cloud agent that has its own computer and records itself using your app the way a human would (like Devin). This has technically been possible for a while, but with Opus 5.5 the results have progressed to something better than what I make by hand. Devin even made a tutorial showing you how it's done:

引用 @dabit3

https://x.com/i/article/2103573841176141824

查看引用原文 ↗

有大新闻喜欢跑的快，但是水平还是得提高一个🥹

引用 @acboxliu

个人认为推特上的营销号特别喜欢把 Agent Computer 和 VPS 混为一谈，然后直接按照“一台 Agent Computer ≈ 一台 VPS”的方式计算成本，但两者的资源模型其实有很大区别。 VPS 通常需要长期占用固定的计算资源，并且很多场景需要独立公网 IP；而 Agent Computer 通常不需要为每个实例分配独立公网 IP，可以共享 NAT / Egress Gateway。 Agent Computer 对持续在线的要求通常也没有传统 VPS 那么高。用户长时间不使用时可以 suspend，只持久化必要的状态和文件系统，在下次使用时重新拉起。 Agent Computer 的 workload 通常是 bursty 的。执行代码、运行浏览器等任务时会短时间占用较多资源，但等待 LLM、工具调用或用户输入时，计算资源需求可能非常低。因此可以通过休眠、动态调度和统计复用，把大量用户映射到底层相对少得多的实际计算资源上。

查看引用原文 ↗

The Book of Answers 📖 Tap the cover, ask your question, then unfold your iPhone Duo. Your answer is waiting inside. Left page: everything you’ve asked Right page: your answer.

做了一款《答案之书》。 正面点击一下，说出你的问题。 翻开 Duo，答案就在书里。 左边一页记录你问过的问题。 右边显示答案。

基于目前流行的 Opus 5.5 提示词提炼总结，整合了 motion 、hyperframe、开源素材库等等。 还能一句话剪电影解说，Skill 地址 https://github.com/joeseesun/qiaomu-cut-skill 安装指令：npx skills add joeseesun/qiaomu-cut-skill --skill qiaomu-cut

Opus 5.5 最适合的场景之一。 提供 Github 仓库地址，生成产品宣传片。 独立开发者朋友们，不要害羞推广，酒香也怕巷子深。 一句话生成的 Qiaomu Home 插件宣传，把 Skill 开源了，见评论区。

大部分 AI 产品，还是得老老实实回归到： 1、对工作有用 2、让有工资的人愿意付钱 3、最好是企业给员工付钱 Personal agent 也逃不过上面三点。

Grok @bot is still my favorite AI tool for actually getting work done in the cloud. Tomorrow, I'm sharing a new episode with @poteto and @pengzheng_, the engineering and design leads for Grok Bot, about: → The 14 bots they use for work and life → Peng's design bot that turns one keyframe into a full user flow → Lauren's eng lead bot that manages a team of eng bots Who better to learn Grok Bot from than the people who built it? 📌 Subscribe to get the full episode tomorrow: https://www.youtube.com/@PeterYangYT?sub_confirmation=1

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://youtu.be/xZ5TEaleUdg Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://granola.ai/peter @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

@bot @poteto @pengzheng_ My full episode with Lauren and Peng is live now! Watch then demo all their Grok bots here: https://youtu.be/xZ5TEaleUdg

"I like to call it a Michelin kitchen [not a software factory]." From @poteto, eng lead for Grok @bot: "When you think about a Michelin-starred kitchen, you think about the quality and the craft that goes into making the meal. I think a lot of people, when you say factory, they have this connotation like it's mass manufactured, it's slop, it's low quality. And for a Michelin kitchen, there's still scale involved. A restaurant might have 50 seats or 100 seats, so you need to be able to produce quality at scale." 📌 Watch my full episode with the Grok Bot team here: https://youtu.be/xZ5TEaleUdg

引用 @petergyang

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://youtu.be/xZ5TEaleUdg Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://granola.ai/peter @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

查看引用原文 ↗

Here's how @poteto (eng lead for Grok @bot) gets her eng lead bot to manage eng bots: "I actually mostly talk to my chief of staff, and then my chief of staff talks to the eng lead. I set these engineer bots up with Dr. Eggbot [a bot to create other bots]. One of the things I told Dr. Eggbot to do is to make sure that the eng lead's job is really to break down tasks into smaller pieces and delegate and supervise other bots rather than do work on its own. You can see in the description, it even has the bot IDs in there. But it tells it to never do work on its own and always delegate to those four eng bots." 📌 Watch the full episode here: https://youtu.be/xZ5TEaleUdg

引用 @petergyang

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://youtu.be/xZ5TEaleUdg Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://granola.ai/peter @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

查看引用原文 ↗

"The designer bot is helpful for bouncing ideas off and making quick changes in Figma." From @pengzheng_, design lead for Grok @bot: "The bot connects to Figma MCP so it will be able to directly change frames and update the design. I have this skill that teaches the designer bot how my Figma file is set up, what my design system is, and when to use the right colors, typography, spacing. This way I can create one keyframe and then ask the bot to scale it into the whole flow and make end-to-end screens." 📌 Watch more Grok Bot use cases in our full interview: https://youtu.be/xZ5TEaleUdg

引用 @petergyang

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://youtu.be/xZ5TEaleUdg Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://granola.ai/peter @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

查看引用原文 ↗

Grok Bot 两位核心成员的 14 个私藏 Bot 与全自动工作流 Grok Bot 两位核心成员 @poteto 和 @pengzheng_ 在 @petergyang 的专访中，首次系统性公开自己的真实使用配置：从生活采购、订机票，到设计、写代码、自动合并 PR，他们已经把 Grok Bot 当作“一群可托付的虚拟员工”在每天使用。在访谈中，他们总结出了从“手动指挥”到“信任放权”的渐进方法论。 Youtube 视频：https://www.youtube.com/watch?v=xZ5TEaleUdg 1. Peng Zheng 的配置：个人生活的“总智能体 + 专家群” Peng 把所有数字世界的事务分为工作与生活两类，最常用三个 bot： Chief of Staff（首席智能体）：兜底入口。不知道该派给谁的任务就发给她。典型例子：买 3D 打印耗材，不只下单，还自动更新他的 Notion 库存清单（按颜色追踪）。卖二手 DJI 麦克风更典型：一条指令串起查官网价格、调研同城竞争报价、学习他过去的文案风格、上架、以及无人问津时每周自动降价 5 美元。执行中它还会主动去“咨询”Writer bot 怎么写文案，bot 之间的自主协作已经发生。 Designer / Writer bot：Designer 通过 Figma MCP 直接改设计稿、用他的 design token；他的新工作流是“我做前 5%”，定好系统、画一个关键帧，然后让 bot 扩展成完整端到端流程。Writer bot 负责润色邮件和 Slack 消息（英语非母语），在 Slack/邮件小窗里内联出稿、他审后发送。 多角色群聊：把 PM bot、Designer bot、Engineer bot 拉进一个群，扔进一个想法（比如“粤语邮件翻译 App”），看它们互相辩论、各自从职能视角给意见。想指定谁回答就 @ 谁，否则它们自行判断是否接话，像真人群聊一样。 两个有趣的延伸：播客 bot 把任意文章链接转成播客供开车时听（“格式的终极翻译器”）；以及他把旧金山唐人街拍的一张照片，做成自动化的“打卡 App” pipeline，发照片 + 同行人的社交账号，bot 自动生成日间/夜间两版黏土风图片、抠图、发布到个人网站，而整个“playbook”存在私有 repo 里由 bot 拉取执行。Grok Bot 在这里变成了个人网站的发布后台。 2. Lauren Tan 的配置：工程向的“米其林厨房” Lauren 展示的是硬核工程用法： 控制虚拟机：为了给 Omachi（DHH 等人做的 Linux 发行版）做适配，她让 Grok Bot 直接操控跑在 Mac 上的 Linux 虚拟机，Grok Bot 有自己的电脑，同时通过本机执行守护进程控制她的电脑。核心逻辑是：agent 必须能自己验证自己的工作。 P-Stack 技能栈 + 自动化评测：改一个 skill 前先跑 eval，bot 会拉起多个不同模型的子 agent 试跑 prompt，确认目标达成才允许合 PR。 “米其林厨房”：她对“软件工厂”这个说法的修正是本场金句之一，工厂让人联想到 slop，她要的是 quality at scale，像米其林餐厅一样既有工艺又有规模。 最激进的信任：因为 Grok 代码库架构本身是 agent-friendly 的，她经常在 PR 合入之后才去看，看完觉得“哦，挺好”。 Dr. Eggbot（造 bot 的 bot）：bot 设计师，负责以工程严谨度生产新 bot，取名全是食物系（大福、饺子、抹茶）。它还会定期巡检所有 bot 的对话记录，提出改进建议，并优化 routines 的执行频率，因为 routine 跑太频繁会持续唤醒 agent、烧掉大量用量。 Tsuki（个人助理）：完成了她最大金额的自主操作，只给了一条含会议信息的 Slack 链接，就通过公司差旅系统订好橙县→丹佛→伦敦→阿姆斯特丹的多段航班和会场附近酒店，只在最后点“确认”前请示了一次。Tsuki 甚至有自己的第三方手机号，会发短信提醒她吃午饭（她开着勿扰模式经常看不到 App 通知），并会用 DoorDash 直接下单。 Potato（社交 bot）：每 30 分钟轮询 X 上的提及（她在 X 有个梗：连打三次 potato 就会触发通知），聚合同类 bug 反馈。关键设计：她可以直接对工程师 bot 说“去找 Potato 了解情况，回来给我修复方案”，上下文在 bot 之间自动传递，不需要人类复制粘贴。这把社交渠道变成了产品反馈的“外环”，直接接入工程师“内环”。 3. 产品设计哲学：为什么是“有名字的常驻 agent” 访谈后半段是全场最有思想密度的部分。Peng 的框架： 三层工作形态演进： ① 过去人直接操作工具（Figma 里画框、Slack 里打字）； ② 现在是 prompt AI 代劳； ③ 未来是人设计系统，让 bot 与 bot、bot 与人自主协作。 多数 AI 产品仍停留在“以模型为中心”（围绕模型造 harness 再包一层产品 UI，满屏技术黑话），而 Grok Bot 的实验是让系统模型对齐用户的心理模型：现实中你不需要懂保险公司怎么运作，你只需要知道该给谁打电话。把 agent 人格化、命名、常驻，正是复用了这个直觉。 两个关键架构决策： · 聊天是易耗品，bot 是资产。AI 产品里的会话大多一次性、两周后再也不回看；而把可复用的能力、记忆、工具沉淀到“角色”里，临时项目再开专门会话，就兼顾了沉淀与灵活。同一个 Designer bot 可以同时服务多个项目的不同会话，身份与记忆仍是同一个。 · 能力共享，记忆私有。工具、连接器、技能是全局共享的，但记忆和偏好归属每个 bot 本身，这就是为什么“你的助理和我的助理注定长成不同的样子”，使用过程本身就是个性化过程。 团队内部也是如此：Lauren 开源为 potato mode 的技能，内部叫 Lauren mode；Peng 有 Peng mode；让工程师以外的人（设计师、PM）也能给代码库做高层贡献，是重构 Grok Bot 代码库的核心动机。 4. 方法论提炼：从下载到放权的路径 面对 Peter “我总怀疑‘睡觉让 bot 干完一切’的宣传” 的追问，两人的回答构成了本视频对普通用户最有用的部分： 1. 从一个简单任务开始，刻意“过度拉伸”它；Peng 至今仍在测试边界：凡是碰到键盘鼠标的事都试着委派出去，看看极限在哪。 2. 观察 + 纠偏 → 固化为 skill。Lauren 的核心论点：一切归结于信任（trust）。没有捷径，先花时间看着 bot 干活、随时纠正；一轮跑顺后把流程沉淀成可复用的 skill，做到一条命令（如 /expense-report）一次成型。 3. skill 稳定后才升级为 routine（如“收到含收据的邮件就自动报销”），这时才真正可以走开。跳过前两步直接放权必然翻车。 4. 注意成本：routine 频率过高会持续唤醒 agent、消耗配额，需要调优（Dr. Eggbot 会帮忙做这件事）。 5. 心智模型：像 onboard 新员工。信任随观察逐步建立，先带着做、再盯着做、最后放手。终点是“人人都能当产品的 CEO”：你做高层战略决策，执行层全部委派。

引用 @petergyang

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://youtu.be/xZ5TEaleUdg Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://granola.ai/peter @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

查看引用原文 ↗

6 things I learned from @poteto and @pengzheng_ (eng and design leads for Grok Bot) on how to get the most out of your bots: 1. Build the design system and one keyframe, then let your bot scale your designs Peng's design bot edits Figma directly through Figma MCP and uses a skill that knows his files and design system. He designs one keyframe, then asks the bot to extend it into every screen of the flow. 2. Make an eng lead bot to manage your other eng bots Lauren's eng lead bot never writes code. Instead, it breaks projects into smaller tasks for her engineer bots, which spin up coding agents in the cloud. That's how she runs "really massive agent swarms." 3. Give your bots a way to check their own work Lauren's bots test each change before merging, and the Grok Bot codebase is built so agents can safely change it. "Sometimes I actually don't even look at the PR until after it's landed." 4. Put your PM, design, and eng bots in one group chat When Peng wants to test a product idea, he adds PM, designer, and engineer bots to the same chat for them to debate with him and flesh out the requirements. 5. Complete a task with a bot manually, build a skill to capture best practices, then make it a routine Lauren watches the first run, corrects mistakes, and turns what worked into a skill. Once the skill works in one shot, she sets it up as a routine and stops babysitting the bot. 6. Name your bots after food or something fun Lauren's bots have names like Daifuku, Gyoza, Matcha, Crumb, and Katsu. This is how she stays hungry all the time j/k. I had to make the image below for her (you're welcome @poteto 🙂) 📌 Watch the full episode now: https://youtu.be/xZ5TEaleUdg Also available as a written post: https://creatoreconomy.so/p/grok-bot-team-14-best-bots-peng-zheng-lauren-tan

引用 @petergyang

"Everything I touch with my keyboard and mouse, I try to delegate to my bots." Here's my new episode with @poteto and @pengzheng_, the eng and design leads for Grok @bot, where they showed me the 14 bots they use for work and life, including: → A design bot that turns one keyframe into a full user flow → An eng lead bot that manages a team of eng bots → How to trust your bots with more of your work Some quotes from both: "I like to call it the Michelin kitchen…when you say software factory, it has this connotation of mass manufactured slop." "Sometimes I actually don't even look at the PR until after it's landed and then I'm like, 'Oh, okay. Yeah, that looks good.'" "I think it ultimately comes back to trust. First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine." 📌 Watch now: https://youtu.be/xZ5TEaleUdg Thanks to our sponsors: @meetgranola: AI meeting notes that don’t suck https://granola.ai/peter @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

查看引用原文 ↗

把这个写 SEO文章的流程做成了一个 skills，就叫 content-gate，成了我让 AI 写 SEO 类似文章的通用流程，基于 Google Search Central 官方文档，内容质量与合规指南，并且配有发布前的 checklist 供 AI 进行自查，还有双语本地化指南，按目标语言写，不逐句翻译。 https://github.com/vikingmute/content-gate 已经用这个规则写了一部分文章了，感觉还不错，想用 AI 写 SEO 文章的朋友可以参考一下。

引用 @vikingmute

今天看到一个非常有用的 SEO tips，现在 AI 生成的内容越来越多，假如你使用 AI 写的内容还想被收录的话，就要需要让 AI 阅读 谷歌官方今年的一系列文档关于内容质量的文档然后写成一份 只覆盖自己 niche 的 guidelines.md prompt 可以这么写：“去打开 Google 现在挂在网上的官方说明，读完再写一份 guidelines.md 这份文档只要讲清楚四件事： 1 Google 明确会处罚什么 2 google 说的写给人看 对人有用的文章到底长什么样 3 E-E-A-T 四个字分别是什么，页面上怎么体现 4 发文章前创建一个 checklist，看看是否通过 ” 保持一段时间更新一次，因为这个文档经常在更新，然后创建一份自己的 guidelines 文档，这样在每个项目要写类似文章的情况下都可以用。 我觉得这个非常非常实用，大家做 SEO 用 AI 写文章的时候可以特别注意下。

查看引用原文 ↗

I built a self-directed data analyst! 100% Open Source The agent investigates a given dataset on its own without requiring you to guide it through every step. Give it a dataset and an objective like “Why did revenue decline?”, and it decides what to inspect, which analyses to run, what hypotheses to test, and what to investigate next based on the evidence it finds. It can inspect the data, run SQL, use Python for deeper analysis, create charts, form hypotheses, test them against the dataset, and keep following new evidence until it has enough to explain what happened. There is no fixed sequence telling it what analysis to run next. Each result becomes context for the next decision. If a query fails, the agent can use that error to adjust its approach. If it starts repeating similar analyses without making progress, the harness can detect that and push it to reassess the investigation. It also keeps explicit state across the run, including findings, hypotheses, tool history, and usage, instead of relying only on raw conversation history. For the model layer, I used Liner Mark 1.0. This works well for this kind of setup because one investigation can involve many model calls, but every step does not need the same level of reasoning. Liner Mark 1.0 routes each request to an appropriate underlying model while exposing a single model interface to the agent. So the same loop can move between dataset inspection, query planning, hypothesis evaluation, error recovery, and final synthesis without manually choosing a different model for every step. The workflow looks like this: Dataset + Objective → Investigate → Run SQL/Python → Observe → Update Hypotheses → Decide Next Step → Repeat → Final Report At the end, the agent returns the root cause, supporting findings, hypotheses it tested, relevant charts, recommended next steps, and a usage breakdown. GitHub repo: https://github.com/Sumanth077/Hands-On-AI-Engineering/tree/main/ai_agents/self_driving_data_analyst What I like about this setup is that the analysis path is not predetermined. You start with a goal, inspect the evidence, form a theory, test it, and let the results decide what to investigate next.

引用 @sumanth_077

Hands on AI Engineering! I open-sourced a collection of 50+ hands-on AI engineering tutorials. It features step-by-step projects and tutorials on: • AI Agents and Multi-agents • RAG (Agentic, Vision, and Local) • MCP AI Agents • OCR Apps • Voice AI Agents • & so much more 100% free and open source. 1k+ Github stars I've shared the link in the comments!

查看引用原文 ↗

不知不觉 DeepSeek 也冲了 600 多块了。 乔木 AI RSS 内容翻译都用的 Deepseek V4.1 Flash，速度飞快，最近每天花费不到 1 块钱。 开发 AI 工具对外提供服务，个人能用得起的模型， DeepSeek 算是首选。 且 API 获取和配置都很简单，模型名也没那么复杂。 最近国庆假期玩玩 DeepSeek Harness ，看能不能把Obsidian 插件迁移一些过去。

爆肝上新 Opus 5.5 发布这几天，全网都在拿它做视频。我把这些作品收了一遍 962 个作品，638 位创作者，原帖播放加起来 1.23 亿 收录标准定得很死 必须是创作者本人发的原帖 必须是原生视频 原帖播放过 5,000 帖子里写明了是用 Opus 5.5 做的 其中 252 个追到了提示词，94 个是完整原文 提示词只收有出处的：原帖、作者自己的回复、回复里的截图、作者给的链接。一字不改，不翻译，不润色。找不到出处的就空着，绝不编一条凑数 按用途分成八类 - 动效与界面 172 - 游戏 158 - 三维世界与模拟 131 - 产品演示与广告 128 - 模型对比 104 - 角色与故事 101 - 讲解与科普 93 - 音乐与剪辑 75 播放最高的一条 1685 万，是一个用分享出来的提示词一夜生成的场景 两个入口分工不同 网页版只放那 252 个带提示词的，视频能直接播，提示词一键复制 GitHub 收全部 962 个，带完整数据，中英双语 搭配我们的昨天分享的claude视频制作分类大合集，制作视频不愁

引用 @gosailglobal

https://x.com/i/article/2104181663844700160

查看引用原文 ↗

网页版（直接看视频、复制提示词）：https://jasonzhu.ai/zh/prompts/claude-opus-5-5 GitHub 全量清单：https://github.com/zhuyansen/awesome-opus-5.5-video 刚上线，觉得有用帮忙点个 star。漏了哪条好作品，欢迎在 GitHub 提 issue 补充 搭配claude视频制作分类大合集，制作视频不愁 https://github.com/zhuyansen/awesome-claude-video-skills

@0xqiuqiuu 哈哈 再加上今天发的opus5.5合集 完美 https://x.com/GoSailGlobal/status/2104562514088398967

引用 @gosailglobal

爆肝上新 Opus 5.5 发布这几天，全网都在拿它做视频。我把这些作品收了一遍 962 个作品，638 位创作者，原帖播放加起来 1.23 亿 收录标准定得很死 必须是创作者本人发的原帖 必须是原生视频 原帖播放过 5,000 帖子里写明了是用 Opus 5.5 做的 其中 252 个追到了提示词，94 个是完整原文 提示词只收有出处的：原帖、作者自己的回复、回复里的截图、作者给的链接。一字不改，不翻译，不润色。找不到出处的就空着，绝不编一条凑数 按用途分成八类 - 动效与界面 172 - 游戏 158 - 三维世界与模拟 131 - 产品演示与广告 128 - 模型对比 104 - 角色与故事 101 - 讲解与科普 93 - 音乐与剪辑 75 播放最高的一条 1685 万，是一个用分享出来的提示词一夜生成的场景 两个入口分工不同 网页版只放那 252 个带提示词的，视频能直接播，提示词一键复制 GitHub 收全部 962 个，带完整数据，中英双语 搭配我们的昨天分享的claude视频制作分类大合集，制作视频不愁

查看引用原文 ↗

@thekitze I'm using Jev to decide whether a mention of @kodykoala on here should wake up my Grok bot to possibly respond or not. Also to improve search results in Kody's search tool. It's pretty cool!

@bentlegen You could make it so your agent responds when mentioned on 𝕏 https://x.com/kentcdodds/status/2104497542000239017 😄

引用 @kentcdodds

@thekitze I'm using Jev to decide whether a mention of @kodykoala on here should wake up my Grok bot to possibly respond or not. Also to improve search results in Kody's search tool. It's pretty cool!

查看引用原文 ↗

@zxioKe 看这里 ，可以直接本地 render 复现的

引用 @_mexicat

@pleometric repo and link to 4k youtube version here: https://github.com/mexicat/pdoom-video

查看引用原文 ↗

Updated the chatbot-template to latest next, ai-sdk, shadcn/react, cn and some bug fixes. More soon. https://github.com/shadcn-ui/chatbot-template

AI之所以写垃圾审计报告 是因为AI缺少专家Knowhow 如果你有一个专门写审计报告的skill 持续把自己的知识丰富到里面 AI就会写的越来越好 我每一次Campaign结束的Report都是AI写的 虽然也会长篇大论 但是质量上还是相对过关&分析到位的 我把人类如何写报告的视角和关注点给了AI 并通过客户的持续反馈优化这个复盘skill 删繁就简 慢慢就会写的简单有力了 我觉得大部分人抱怨AI Slop的原因 是因为人类给予的方法智能还不够 AI就像唐三藏扫塔 要从下往上扫 扫的过程 就是先把简单搞成复杂 再进行删除简化 取其精华去其糟粕 最终才能获得一份高度有价值的报告的

引用 @connections8

Lately, I’ve been seeing an alarming rise in clients receiving 30+ page "AI Slop" audits generated in seconds with zero critical thinking. These reports look visually impressive at a distance. They’re full of scary red exclamation marks, polished pie charts, and "URGENT" warnings. But when you actually open them up, the claims are laughably out of touch: 🚨 "Missing llms.txt file: CRITICAL SEO ISSUE." 🚨 "Missed blocked directory in robots.txt: HIGH RISK." 🚨 "Missing ALT tag on an icon file: CRITICAL ACTION REQUIRED." Let’s be honest about what’s happening here: 1. Minor noise is being sold as major strategy. An unoptimized alt tag on a decorative footer graphic isn’t killing your rankings. A missing llms.txt file isn't wiping you off Google. Turning low-impact, micro-level best practices into "critical site failures" is just a trick to create fake urgency. 2. Quantity is replacing quality. Clients are being handed 30 pages of generic, auto-generated bullet points instead of 2 pages of actual strategic insight. If an audit doesn't tie its findings directly to revenue, search intent, site architecture, or core technical blockers, it’s not an audit it’s an export script. 3. AI is being used as a substitute for expertise, not a tool. AI is incredible for efficiency, but using it to pump out generic audit templates without human validation destroys trust in the whole industry. If your audit reads like it was generated by a bot that has never looked at a search engine result page, throw it away. What’s the most absurd "Critical SEO Flag" you’ve seen in an automated audit recently?

查看引用原文 ↗

被很多专业者骂了，但是在折腾了3天以后，AIHOT最终还是重写完然后上线了。。。我觉得还是可以分享一下我全部跟AI协同的流程，万一对其他人有用呢（当然我就是个纯外行，仅供参考）： 1. 使用Claude Opus 5.5和GPT-6 Astra并行对旧项目进行蒸馏，并且设计交流包。要求：架构分离，为多Agent并行开发而设计，保留所有功能和容易踩坑的细节。 2. 使用Claude Fable 5.1对两个模型生产的功能文档、交接包和旧代码库进行全面对照，寻找不合理和遗漏的地方进行完善。 3. 使用Claude Opus 5.5在不看任何旧代码的情况下，根据最终的功能文档、交接包、线上网站的端到端测试，进行全方位的重写（大概写了12个小时）。 4. 使用Claude Fable 5.1和GPT-6 Astra并行审查新写完的项目，将其与旧代码库的所有细节进行逐行审计，看看是否有功能和细节遗漏，在用户体验层面，能否完美还原，最终产出两份审计报告。 5. 使用Claude Opus 5.5根据审计报告，进行优化开发，开发完成以后，清空所有上下文，自己再并行N个子Agent，以功能模块化的方式，跟旧代码库进行对比，看是否遗漏功能和逻辑细节（无视代码实现，只看功能和逻辑细节）。 6. 使用Claude Fable 5.1和GPT-6 Astra并行审查所有可能的BUG和漏洞，继续Opus 5.5优化。 7. 使用Opus 5.5优化所有前端UI，进行控件组件化统一，部分UI界面全面重设计，加了一部分Opus 5.5擅长的JS+Canvas动效，例如关于页和更新页。 8. 导出旧项目所有数据库，进行线上服务器彩排，6小时的影子系统并行，过程中实时监控，使用Claude Opus 5.5自动化并行修复过程中出现的所有问题。 9. 使用GPT-6 Astra进行全面的端到端测试。 10.无缝切换上线。 11. 使用GPT-6 Astra根据真实数据，全方位优化缓存、CDN、页面大小等问题，提升全站性能。 12.监控问题，不断优化。 以上，大概就是一个纯外行者的“重写”的经验，虽然AIHOT这个项目很小，但是我自己干的很开心。希望能对大家有一点点的启发。

引用 @khazix0918

比重构屎山可能更高效的方式： 直接将源项目库蒸馏成功能文档，然后直接用最新的模型原地重写。。。🤦‍♂️🤦‍♂️🤦‍♂️

查看引用原文 ↗

I do this with Kody currently. It's so nice to have an easy easy to spin up full stack applications that can use everything that's already connected to Kody. Curious whether you're defaulting to it being public or authenticated. The blast radius is so huge that I've kept it locked down. But I've got some ideas on how to make it public but safer.

引用 @kentcdodds

Your agent is burning your money every time. You've got to improve this human-to-agent-to-software pipeline. Let me show you how I do it.

查看引用原文 ↗

几万人看过、觉得非常不错 Opus 5.5 产品宣传视频 skill 已经更新！ 我测试了一下，即使是 Mimo V2.6 Flash 这样的模型，用这个 skill 也能有相对不错的效果，当然肯定比不上 Opus 5.5。 如果你的产品需要大量产出这种视频，同时又需要用一些成本更低的模型，可以试试这个 skill。

引用 @op7418

我操，Opus 5.5 做这种前端的图像图形，做产品宣传视频太牛逼了！ 让他给我这个 CodePilot 产品做了个宣传片，太猛了，一键生成的。 简直吊打前几天的 GPT-6 Astra。 后面也会将这些经验用在我的这个 guizang-product-video-skilll 里面。 然后一些比较差的模型也能得到比较好的效果。

查看引用原文 ↗

@elvissun Each frame is basically a function of time, so you can just fork this and make your own variation. https://x.com/_mexicat/status/2103487098258911323?s=46

引用 @_mexicat

@pleometric repo and link to 4k youtube version here: https://github.com/mexicat/pdoom-video

查看引用原文 ↗

这内容放外面卖 2999 一节不过分，姚老师和乔木大哥直接做成公开课

引用 @yaojingang

不知不觉，GEO公开课已经完成了5期了 和@vista8 向阳老师在开始讨论系列GEO公开课时，初步定的是每月1期，持续12期，还剩7期 每期定一个大的主题，然后通过直播的形式，完成关于这个主题的系统性分享 之后每次分享，会做个简单的调研，根据大家当前最关心的问题，来调整公开课的内容节奏与安排 下面是前五期直播的图文总结版，欢迎收藏： 1、第一节公开课：《从AI搜索逻辑到GEOFlow落地实战》 https://mp.weixin.qq.com/s/8Lyrzux7WacHjiHx_P9Fbg 2、第二节公开课：《一文讲透GEO内容工程：目标、要素、关系与反馈》 https://mp.weixin.qq.com/s/0ZZ5je2-W83HY7RtmsYkJA 3、第三节公开课：《一次讲透GEO信源偏好与信源投放策略》 https://mp.weixin.qq.com/s/u6wEeYcHWoumDNvHh4Zkww 4、第四节公开课：《官网正在成为GEO主阵地：从企业事实到AI引用，官网GEO实战公开课》 https://mp.weixin.qq.com/s/PVi74oPeoUTFUa23tDaBZA 5、第五节公开课：《GEO到底该投多少钱？1.3万字讲透效果归因与ROI（附10种标记方法）》 https://mp.weixin.qq.com/s/86cexOCX4AX-uwQc9O69-w

查看引用原文 ↗

支持负载均衡、质量优先、速度优先、成本优先多种路由策略，按需搭配，实现自己的 jev-router

开源 https://autojev.ai，基于 jev 决策的模型路由器。 为你的本地 Agent 自动选择合适的模型来处理对话请求，在速度、成本、质量方面取得平衡。

开源仓库地址👉 https://github.com/thinkany-ai/autojev

原本想买一个卡片式录音笔，Plaud 这个看着还不错，新西兰 350 刀左右能买到。 但后来又一想，我已经有一个讯飞的条状录音笔，录音效果还不错，自己的 DJI Mic 3 也可以随时携带录音。而且我自己做的 Skill 可以支持音频拷到文件夹后，就自动转录并按发言人保存到我的 Obsidian。 那么我再弄个这种所谓 AI 录音笔，有啥用？ 我的左脑说：“那你可以打电话时录音啊。” 理性的右脑说：“你一个月会打 5 次以上的电话吗？”

Dia 的界面比 Arc 简单很多，但这次简化的价值，并不只是让产品看起来更干净。 Arc 花了大量用户注意力解释 Space、Folder、Easel 和 Boost。Dia 使用 Profile、Tab 和 Group 这些更熟悉的结构，降低了理解产品的成本。 然后，它把省下来的空间交给了 AI。 地址栏可以直接回答问题，侧边栏可以围绕当前网页对话，多个标签页可以被整理成一份报告，浏览器还会主动提出清理建议。 这是一种很重要的产品判断。 有些产品完成简化后，只是少了几个按钮和功能，用户却没有更快接触到真正的价值。复杂度减少了，产品的差异也一起变弱。 Dia 的选择更明确。它减少用户必须理解的东西，然后把注意力重新投入到自己真正希望用户采用的能力上。 **产品做减法之后，必须知道省下来的空间准备留给谁。否则简化很容易变成单纯的删减。**

正式宣告：如果你想在今年冬天，在 *杭州* 大展身手。把微信、telegram，重新写一次。以 AI Native 的方式，去构建一个 Agent 与人，共为一等公民的 IM 网络。请您提前与我联系。 一直以来，Cromma（可爱信）无法真正通过思想实验。所有的壁垒都有解法，可优化的点千千万万，唯独小程序生态这一块，完全难以撼动。 直到这个秋天，我们看到了很多东西真正进入了沸腾阶段（以lovable为标志的产品）。让每个用户用自然语言的方式，0代码、0部署焦虑地来创造和分发自己的“小程序”。将IM中的可流转软件生态，从少部分人编码，大部分人使用的时代，蝶变到程序的创造者、迭代者、使用者，都是用户本人的时代。 想象一下：hi，为我的这个学员群弄一个小程序。 总而言之，一切已经启动。我需要做一些 CEO 该做的事。组织人，尤其是野心勃勃的人。我们一起打造，一个唯一的 Rust IM 内核。经 UniFFI 绑定给各端原生。经 wasm 给 web，经 napi 给 Electron。 这里放不下关于Cromma的想象。让一切从一封邮件开始，只需证明你也想干，且可以胜任。 cromma@laper.ai

快来加入纯想老师的战队，打造跨时代的产品

引用 @chunxiangai

正式宣告：如果你想在今年冬天，在 *杭州* 大展身手。把微信、telegram，重新写一次。以 AI Native 的方式，去构建一个 Agent 与人，共为一等公民的 IM 网络。请您提前与我联系。 一直以来，Cromma（可爱信）无法真正通过思想实验。所有的壁垒都有解法，可优化的点千千万万，唯独小程序生态这一块，完全难以撼动。 直到这个秋天，我们看到了很多东西真正进入了沸腾阶段（以lovable为标志的产品）。让每个用户用自然语言的方式，0代码、0部署焦虑地来创造和分发自己的“小程序”。将IM中的可流转软件生态，从少部分人编码，大部分人使用的时代，蝶变到程序的创造者、迭代者、使用者，都是用户本人的时代。 想象一下：hi，为我的这个学员群弄一个小程序。 总而言之，一切已经启动。我需要做一些 CEO 该做的事。组织人，尤其是野心勃勃的人。我们一起打造，一个唯一的 Rust IM 内核。经 UniFFI 绑定给各端原生。经 wasm 给 web，经 napi 给 Electron。 这里放不下关于Cromma的想象。让一切从一封邮件开始，只需证明你也想干，且可以胜任。 cromma@laper.ai

查看引用原文 ↗

我开始在户外 Vibe Coding 了！ 之前我以为 AI 会写代码，人就解放了？ 后来发现它写得再快，我也得坐着批权限、看结果，跟坐班没区别 然后我在网友的安利下使用了 UU远程 现在我直接用手机连电脑的终端，不用打开完整的远程桌面就能查看进度，操作起来很快捷 而且输入起来也很方便 要输入文字时候，直接用豆包语音把需求说完 要输入命令时候，直接全键盘敲命令，符号也好打 另外它对手机终端的适配比我预想中完整 Claude Code 弹出来的选项，手机上就能点击 电脑终端里什么样，手机上就什么样，排版一点不乱 手机上的对话，可以拉回电脑端 回到 Mac 敲一行 uuyc-cli lterm attach，在自己的终端里接着干 还能同时开好几个会话，一个写功能，一个跑测试 更牛逼的是远程桌面也不限画质，有什么问题，切到桌面就行 这个兜底，户外 VibeCoding 很安心，配合折叠屏，无敌了 最离谱的是，这些全免费！！不限画质，不限速，也没广告 总之，如果你也想在户外Vibe Coding，建议使用UU远程，在外面直接用一台手机就能把整套移动的Vibe coding流程跑起来！

躺床上 Vibe Coding 的必备工具！！ 1. 折叠屏 感谢雷总 2. 微信语音 + 豆包语音 感谢小马哥和一鸣总 还有一个要感谢丁磊总👇 https://x.com/lxfater/status/2104503921406259439

引用 @lxfater

我开始在户外 Vibe Coding 了！ 之前我以为 AI 会写代码，人就解放了？ 后来发现它写得再快，我也得坐着批权限、看结果，跟坐班没区别 然后我在网友的安利下使用了 UU远程 现在我直接用手机连电脑的终端，不用打开完整的远程桌面就能查看进度，操作起来很快捷 而且输入起来也很方便 要输入文字时候，直接用豆包语音把需求说完 要输入命令时候，直接全键盘敲命令，符号也好打 另外它对手机终端的适配比我预想中完整 Claude Code 弹出来的选项，手机上就能点击 电脑终端里什么样，手机上就什么样，排版一点不乱 手机上的对话，可以拉回电脑端 回到 Mac 敲一行 uuyc-cli lterm attach，在自己的终端里接着干 还能同时开好几个会话，一个写功能，一个跑测试 更牛逼的是远程桌面也不限画质，有什么问题，切到桌面就行 这个兜底，户外 VibeCoding 很安心，配合折叠屏，无敌了 最离谱的是，这些全免费！！不限画质，不限速，也没广告 总之，如果你也想在户外Vibe Coding，建议使用UU远程，在外面直接用一台手机就能把整套移动的Vibe coding流程跑起来！

查看引用原文 ↗

You could do that but it won't be "more token-efficient". You'd still need to traverse the tree, resolve dependencies etc which will cost more than doing it in a build step. What we could do (and the shadcn skills does this partly) is handle this on the other end. i.e. make agents better at building the JSON payload.

@eatonphil ah sorry -- i had the same problem so i wrote this thing called the GPU Glossary, https://github.com/modal-labs/gpu-glossary, to collect up my favorite way of presenting all the key terms for working with GPUs

@takayamentality @can of course! the printed version is secondary https://modal.com/gpu-glossary source: https://github.com/modal-labs/gpu-glossary

Modal @modal 这份「GPU 术语手册」做的太友好了，它把整个 GPU 技术栈拆成四个相互勾连的层面：设备硬件、设备软件（执行模型）、主机软件（驱动与工具链）、性能分析，每个词条都密集交叉链接，既可以按需查阅单个术语，也可以从头到尾线性通读。 理解这份手册从「CUDA」开始，CUDA 有三重含义：一种设备架构、一个并行编程模型、一个软件平台。这份手册的四个章节恰好分别展开这三个含义加上性能分析。 GPU Glossary 四个章节： 1. 设备硬件：GPU 的物理解剖 - CPU 用复杂核心避免延迟，GPU 用海量简单线程隐藏延迟 2. 设备软件：CUDA 的执行模型 - 线程层级、内存层级、硬件三层严格同构，程序员亲手管理内存 3. 主机软件：驱动、工具链与库 - Runtime 包着 Driver，闭源库拿来即用，CUTLASS 一系负责构造自己的 kernel 4. 性能分析：完整诊断链 - 定瓶颈 → roofline 定性 → 利用率定位 → warp 机制 → 访存模式 → 资源约束 这三句总结很到位 GPU 的本质是“单周期切换海量简单线程来隐藏延迟”；编程的本质是“沿线程-内存-硬件的同构层级，把数据在正确的层级间搬运，把算术强度做到岭点以上”；性能工程的本质是“沿诊断链逐层定位，用 Little 定律算够并发，不为中间指标（如占用率）本身而优化。 https://modal.com/gpu-glossary

@Personuo 不是的，我的环境里有GPT IMAGE的API，他默认用的是这个生成主画面，然后用js做了些转场，我得手动强调不能用图片模型才行

发个预告，我之前用 Codex 做的一个风格化的梵高场景生成器，目前已经有阶段性的进展啦~ 这是一个业余时间 vibe 的项目，断断续续改了一个月，起初我只是想把曾经做过的 Blender 几何节点和油画材质搬进浏览器给大家玩。 随着越拖越久，模型能力也变得越来越强，这个项目也越做越完整了。目前材质我还不是很满意，打算放假再爆改一轮，等做好了给大家玩玩~ ▶ 主力模型： GPT-Astra High 我的最终目标是希望整个场景每个元素都可以参数化生成，大家可以像搭积木一样搭建自己的油画场景。 在实践过程中我发现正确思路不是让 Astra 直接转译我以前的工作流和节点（因为很多已经过时），而是先让 Astra 调研现成的 Three.js 开源项目，再做举一反三。 然后新世界就打开了，Three.js 现成的开源环境生成器实在太多太多了，每一个的性能都优化得极好，结合风格化的材质可以做很多很好玩的变体。 等项目发布了，好好写一篇分享。 #threejs

重磅新作品预告，做得真的实在是太好了🥲 好喜欢！

引用 @simonxxoo

发个预告，我之前用 Codex 做的一个风格化的梵高场景生成器，目前已经有阶段性的进展啦~ 这是一个业余时间 vibe 的项目，断断续续改了一个月，起初我只是想把曾经做过的 Blender 几何节点和油画材质搬进浏览器给大家玩。 随着越拖越久，模型能力也变得越来越强，这个项目也越做越完整了。目前材质我还不是很满意，打算放假再爆改一轮，等做好了给大家玩玩~ ▶ 主力模型： GPT-Astra High 我的最终目标是希望整个场景每个元素都可以参数化生成，大家可以像搭积木一样搭建自己的油画场景。 在实践过程中我发现正确思路不是让 Astra 直接转译我以前的工作流和节点（因为很多已经过时），而是先让 Astra 调研现成的 Three.js 开源项目，再做举一反三。 然后新世界就打开了，Three.js 现成的开源环境生成器实在太多太多了，每一个的性能都优化得极好，结合风格化的材质可以做很多很好玩的变体。 等项目发布了，好好写一篇分享。 #threejs

查看引用原文 ↗

@threejs Thanks for sharing it! And here is the original post by @simonxxoo : https://x.com/simonxxoo/status/2103800335529554108?s=20

引用 @simonxxoo

发个预告，我之前用 Codex 做的一个风格化的梵高场景生成器，目前已经有阶段性的进展啦~ 这是一个业余时间 vibe 的项目，断断续续改了一个月，起初我只是想把曾经做过的 Blender 几何节点和油画材质搬进浏览器给大家玩。 随着越拖越久，模型能力也变得越来越强，这个项目也越做越完整了。目前材质我还不是很满意，打算放假再爆改一轮，等做好了给大家玩玩~ ▶ 主力模型： GPT-Astra High 我的最终目标是希望整个场景每个元素都可以参数化生成，大家可以像搭积木一样搭建自己的油画场景。 在实践过程中我发现正确思路不是让 Astra 直接转译我以前的工作流和节点（因为很多已经过时），而是先让 Astra 调研现成的 Three.js 开源项目，再做举一反三。 然后新世界就打开了，Three.js 现成的开源环境生成器实在太多太多了，每一个的性能都优化得极好，结合风格化的材质可以做很多很好玩的变体。 等项目发布了，好好写一篇分享。 #threejs

查看引用原文 ↗

这些案例采集，整理，加工走的全 AI 路线 最终的成文，我也用开源的 Humanizer-zh ，加上自己写的 3-5篇范文做了去 AI 味。 这些成文主要解决的是信息差，文章结构也是“预制”的，希望不影响阅读 https://x.com/huangyun_122/status/2087589632720654504?s=20

引用 @huangyun_122

去 AI 味的 Skill 我推荐 —— Humanizer-zh 用完基本上没有特别重的味道。 但，还是要精修 5-8篇文章，就是把 AI 写的，用自己的话改写，改写的越多， AI 参考这些样本，写出来会越贴近自己的风格

查看引用原文 ↗

如果视频提示词里不加运镜相关的提示词的话，看起来就会很乏味 但是一旦增加了运镜相关的，就会大幅度提高视频的质感

引用 @adrianpunk115

这个视频里，提示词加入了很多关于近景、切换，推进的方式。 美食视频，生活类产品广告、vlog都很合适

查看引用原文 ↗

想接商单了🥹需要宣传的老板欢迎私聊

引用 @weijunext

推荐一个网站：Aha Prompt：https://ahaprompt.app 网站内容由 Agent 自动采集、翻译推特上热门、优质 Prompt，然后人工审核、精选。 Aha Prompt 已经发布了800个 Prompt，刷了一遍感觉质量确实都挺高，分类和模型也很全，而且有收藏功能，很适合不懂写图片、视频 Prompt 的朋友。

查看引用原文 ↗

这是我的原贴，原贴里单独列出来尼日利亚订阅，是因为那时候尼日利亚区价格特别便宜，但是现在早就没优惠了，好歹把尼日利亚去掉吧。 https://x.com/gorden_sun/status/2043002267343978672?s=46

引用 @maxforai

原来Claude封号有这么多风险.....

查看引用原文 ↗

@MaxForAI 原贴是我很久以前发的，那时候还有尼日利亚区优惠，现在尼日利亚早就没优惠了，没必要把这个也抄过去。 原贴里的老号指的是Claude老号，不是谷歌老号，Claude不知道你的谷歌账号详细信息。 https://x.com/gorden_sun/status/2043002267343978672?s=46

引用 @gorden_sun

关于如何订阅Claude账号，都浓缩到这3张图里了。 我一直觉得封号没有大家传的那么严重，我开的、我身边朋友开的，都用的稳稳的。

查看引用原文 ↗

《清迈的周日》，又做了一版 risograph 风格，CosyVoice 配音 + 字幕

引用 @gosailglobal

宣传片制作技巧 模型opus5.5 prompt：参考剧本skills，有没有适合的https://github.com/jtydhr88/screenwriting-skills，出个https://cmi.community/ 的宣传片

查看引用原文 ↗

🔥301 个 Opus 5.5 视频，我花了整整 $3,000。 我用 Claude Opus 5.5 把全网最难、最火的特效重新做了一遍。成片导出的那一刻，看着屏幕直接瘫坐震惊…… 所有测试视频、全套 Prompt 和工作流，全部整理开源！ 👇 GitHub 开源地址与在线 4K 原视频见评论

📦 GitHub 开源仓库（求个 Star ⭐️）： 包含了全部特效的提示词、参数模板： 👉 https://github.com/yihui-dev/awesome-opus5-5-videos

这几天的快乐浓缩版 https://x.com/yihui_indie/status/2104122905961595267

引用 @yihui_indie

🔥301 个 Opus 5.5 视频，我花了整整 $3,000。 我用 Claude Opus 5.5 把全网最难、最火的特效重新做了一遍。成片导出的那一刻，看着屏幕直接瘫坐震惊…… 所有测试视频、全套 Prompt 和工作流，全部整理开源！ 👇 GitHub 开源地址与在线 4K 原视频见评论

查看引用原文 ↗

Yihui老师这个太顶了！！！ 直接让我的Agent 开始逐字学习，Github已经开源，记得Star！ 地址：https://github.com/yihui-dev/awesome-opus5-5-videos

引用 @yihui_indie

🔥301 个 Opus 5.5 视频，我花了整整 $3,000。 我用 Claude Opus 5.5 把全网最难、最火的特效重新做了一遍。成片导出的那一刻，看着屏幕直接瘫坐震惊…… 所有测试视频、全套 Prompt 和工作流，全部整理开源！ 👇 GitHub 开源地址与在线 4K 原视频见评论

查看引用原文 ↗

Yihui 老师太强了，Opus5.5 也太顶了，这在以前做这些视频可太特么难了，也太 TM 贵了，看看现在，几十秒，加上一句提示词就搞定了！！ Yihui 老师把提示词开源了👇👇👇

开源地址：https://github.com/yihui-dev/awesome-opus5-5-videos

【Opus 5.5 - 第1弹】 301 个 MG + 3D 混剪：https://x.com/yihui_indie/status/2104122905961595267

引用 @yihui_indie

🔥301 个 Opus 5.5 视频，我花了整整 $3,000。 我用 Claude Opus 5.5 把全网最难、最火的特效重新做了一遍。成片导出的那一刻，看着屏幕直接瘫坐震惊…… 所有测试视频、全套 Prompt 和工作流，全部整理开源！ 👇 GitHub 开源地址与在线 4K 原视频见评论

查看引用原文 ↗

🔗 4K 在线版（新增的 107 个排在最前，每个都附原帖）： https://skillry.dev/ai-videos/opus-5-5 ⭐ GitHub 开源仓库（已同步到 389 条）： https://github.com/yihui-dev/awesome-opus5-5-videos

🔥Opus 5.5 能飞！作品混剪第 2 弹，这次 3D 了很多。 滕王阁月夜、落日空战、会转的机械眼、在安纳西湖上划船、东京塔从图纸长成夜景……全是 Claude Opus 5.5 写的网页代码，在浏览器里渲染出来的，我逐个复刻成了 4K。 👇 在线 4K 版和 GitHub 地址见评论

分享一下被小红书ban掉的剪辑教程 有了这个skill+这套sop，我现在基本上不用动手剪视频了 只需要简单提示词，口播类视频剪辑全程托管给Codex，另外可叠加演示录屏+素材生成，制作完整的AI教学分享演示教程 来自 @gengdaJ 剪映11.4.2开源：https://github.com/mcncarl/jianying-headless 配套神级Skill：https://github.com/mcncarl/yichen-skills/tree/main/yichen-jianying-edit

我不理解为何小红书要把这么优质的教程 Ban 掉？我寻思这也没啥违禁的点呀

引用 @bbkirstry

分享一下被小红书ban掉的剪辑教程 有了这个skill+这套sop，我现在基本上不用动手剪视频了 只需要简单提示词，口播类视频剪辑全程托管给Codex，另外可叠加演示录屏+素材生成，制作完整的AI教学分享演示教程 来自 @gengdaJ 剪映11.4.2开源：https://github.com/mcncarl/jianying-headless 配套神级Skill：https://github.com/mcncarl/yichen-skills/tree/main/yichen-jianying-edit

查看引用原文 ↗

💡分享我自己个人摸索出来的穷人版省token小方法： -UI Improvement用luna Max， -Feature Improvement交给sol medium， -Architecture的直接使用Astra 实践了一天下来，确实挺省的哈哈哈

@DesperateGuts 非常感谢你的反馈 我已经修复了这个问题 是有一个超时的问题导致的 请 mo update 升级到 V1.56.1 i这个版本，假如是 brew安装的，由于官方还在合并中 你可以直接重新安装curl的版本

#审美积累 Design Engineer

引用 @ianneo_ai

Design Engineer 人物志 #002｜Gavin Nelson 在 Linear 的手机版里点开搜索，会有一层面板盖在页面上。想关掉它，就往下拉。拉过某个位置时，面板顶上那根小横条会变成一个小箭头，手机同时轻轻震一下。这时候松手，面板就收走了。 设计这个交互的人叫 Gavin Nelson。 他现在在 OpenAI 做 ChatGPT 的交互设计，之前在 Linear 做手机 App，更早在 GitHub 做 GitHub Mobile 和 Copilot。 他还一直在画 App 图标，个人站的图标页里，最新一批是给 ChatGPT 和 Codex 做的，往下还有 Linear、1Password、Flighty。 回到那个搜索。它是 2025 年加进 Linear 手机版的。那时候 App 越做越大，页面一层套一层，想从一个地方跳到另一个地方，只能一路点返回。 团队不想推翻整套导航，他就在底部工具栏里加了一个随时能点的搜索，点开先列出最近看过的内容和常去的地方，一打字就开始筛。 搜索本身不算新鲜，我更在意他在关掉这件事上花的力气。一层盖上来的面板，往下一拉是最顺手的关法，可光看界面，不一定知道能拉。横条变成箭头、手机震一下，等于在说：过线了，松手就关。 收起的动作也跟着手走，拉得越远，面板退场的动作越明显。这些是他用 Origami 和 SwiftUI 做原型调出来的。 他在一期播客里聊过类似的问题。手机上有很多又快又好用的手势，比如往旁边一划就删掉一条通知，但界面上看不出来，不知道的人可能一直都不会用。 他的做法是两头都顾：看得见的按钮照样留着，谁都能完成；再在细节里给一点暗示，让人慢慢发现还有更快的办法。 他在 Linear 做的别的东西也是这个路子。在手机上看文档评论时，评论都收在一个面板里，上面露出一截文档；往下翻评论，文档会跟着滚到对应的那一段，不用另学一套新操作。 重做底部标签栏时，点一下小箭头能展开，按住拖到想去的那一栏再松手，也能直接切过去。他在介绍这个标签栏时写，希望这些功能能在自然的摸索中自己露出来，事后回头看又觉得理所当然。 我觉得他一直在琢磨的是：一个更快的操作，怎么才能不用教，也能被人找到。 到了 OpenAI，他做原型的工具里多了 Codex。他在今年 8 月的一次采访里说，手势、动画、转场、震动这些决定最后 20% 的东西，他会让 Codex 做成能上手摸的版本。他觉得 AI 让设计工具这个词变宽了，为一件事临时做个小软件也算。 但哪些问题值得解决、该删掉什么、放进整个产品里合不合适，他认为这些判断还会在设计师手里待上一阵。 如果只用一个词来说他，我会选理所当然。 再想想那根变成箭头的小横条。它从头到尾没说一个字，手指拉到那儿，它换了个样子，震了一下，人就懂了。很多功能上线时会配一个新手引导，弹几页说明。哪些引导能换成这样一个小变化，哪些不能，我还没想清楚。 http://nelson.co

查看引用原文 ↗

我给一家专门做营销的上市公司做Reddit服务执行 他们找的懂英语的内容营销写手 一个产品一个月20多条内容 平均展示量只有1-3k 甚至好几条只有7，800 views 偶有一条一万的客户都觉得很不错 这个月我让AI跑了20条 70%在5-8k 5条大几万 1条十来万 这是第一个执行月，AI还在早期探索学习期，但成绩已经比人类强了，可效果其实都没那么关键 最关键的是，处理效率是以前的几十倍 这意味着他们可以在有限的人力情况下大量扩展业务 以前需要写手来写稿子 现在可能只需要写手校对一下 一个写手以前一个月只能服务4-5个客户 现在同步能跟进十几个甚至20个 效果还比以前好 AI真的在全方位碾压人类 从这个角度上看 那些被淘汰的写手工资，将被掌握如何设计AI，使用AI，构建AI Loop的人所占有 另外的写手也会溢价，他们会因为懂得如何使用现成的AI，如何审阅，如何给客户提供情绪价值而涨工资 接下来，工资可能倾向于两类模式支付了 一类是强化&替代人类，为AI提供智能 一类是连接人类&AI，提供真实的人感价值，或者交互物理世界，做AI做不了的那部分工作

我给它的指令是，按照 todo 持续工作，完成所有工作后再统一汇报。 因为我发现它不理解持续工作怎么做，但是不让汇报，它就会持续工作。

引用 @winter_cn

因为Agents.md里写了方案必须先给我审核，所以codex写了方案给我审核，我给了审核意见，于是codex写了一个“方案的修订方案”，我@#￥%……

查看引用原文 ↗

AI Engineering From Scratch! 523 lessons. 20 phases. 100% open source. This repo is basically a complete AI engineering curriculum built around learning by actually implementing things. It starts from the fundamentals and gradually moves into the systems people are building today. It covers: - Math and machine learning fundamentals - Deep learning and neural networks - Computer vision, NLP, and speech - Reinforcement learning - Transformers and LLMs - LLM engineering - Multimodal AI - MCP and agent engineering - Autonomous and multi-agent systems - Production AI infrastructure What I like is that the lessons are not just theory or framework tutorials. The structure is: Problem → Concept → Build It → Use It → Ship It So you first understand what is happening underneath, build a smaller version yourself, then move to the production tools and frameworks. There are also focused learning paths, so you do not need to go through all 523 lessons if you only want to learn things like LLM engineering, agents, MCP, or coding agents. A pretty solid repo if you want to learn AI engineering from first principles instead of jumping straight into abstractions. I've shared the GitHub repo in the comments!

引用 @sumanth_077

Hands on AI Engineering! I open-sourced a collection of 50+ hands-on AI engineering tutorials. It features step-by-step projects and tutorials on: • AI Agents and Multi-agents • RAG (Agentic, Vision, and Local) • MCP AI Agents • OCR Apps • Voice AI Agents • & so much more 100% free and open source. 1k+ Github stars I've shared the link in the comments!

查看引用原文 ↗

[开源学习资源] AI Engineering from Scratch 59.8K ⭐️ 作者 @ghumare64 课程共 20 个阶段，以 Python 为主要编程语言，主张在导入任何框架之前，先用纯数学把每个算法手写一遍。 课程地址：https://aiengineeringfromscratch.com/index.html 开源地址： https://github.com/rohitg00/ai-engineering-from-scratch 它和常见教程的本质区别 ？ 大多数 AI 教程是“API 驱动”的：装个库、调个接口、跑个 demo。这个项目反其道而行，每节课遵循固定的六段结构： Motto → Problem → Concept → Build It → Use It → Ship It 以 Phase 10 的一节课「Tokenizers: BPE, WordPiece, SentencePiece」 为例，这个格式是真实落地的，而且写作质量相当高： · Problem 部分不讲废话，直接从代价切入：“你的 LLM 不读英语，它读整数。分词器决定这些整数是承载意义还是浪费意义”，然后解释为什么 tokenization 不是预处理，它是架构的一部分（影响上下文窗口利用率、API 计费、推理速度）。 · Concept 部分用“三种失败的方案和一种胜出的方案”来讲演化逻辑：词级切分（词表爆炸、[UNK] 问题）→ 字符级切分（序列过长）→ 子词切分（BPE 的折中），并配上真实语料上 BPE 逐步合并的手工演算。 每节约 3000+ 词，带 Mermaid 图、可运行代码和测试。 另外两个设计值得强调： · 每节课产出一个可复用的 artifact，一个 prompt、一个 skill、一个 agent 或一个 MCP server。学完整个课程，你手里有 523 个可展示的作品，不是 523 个跑完就扔的 notebook。 · 强调“证据留存”：保留命令、退出码、输出，作为学习发生的证明。这明显吸收了工程实践中“可验证性”的思路。 # 课程结构：从线性代数到自主智能体集群 20 个阶段构成一条完整的上升曲线，咱们它分成五个大块来看： 基础层（Phase 0–2）：环境与工具链、数学基础（线性代数/概率/微积分）、经典机器学习。这是给基础不牢的读者铺的路。 深度学习与感知层（Phase 3–6）：神经网络核心、计算机视觉（一路讲到 NeRF、高斯泼溅、世界模型，这已经超出一般教程的覆盖范围）、NLP、语音。 生成与决策层（Phase 7–9）：Transformer 深挖、生成式 AI（GAN、扩散模型、flow matching）、强化学习。 LLM 层（Phase 10–12）：这是全课程的重心之一。Phase 10 的目录我逐条看过，它不只是“从零实现 GPT”这种常规内容，还包含了相当前沿的论文级主题：DeepSeek-V3 架构走读、DualPipe 并行策略、Native Sparse Attention（NSA）、多 token 预测、Jamba 的 SSM-Transformer 混合架构、 speculative decoding 等。Phase 11 转向应用侧（RAG、LoRA、MCP、可观测性），Phase 12 覆盖多模态（从 CLIP 到 computer-use agent）。 智能体层（Phase 13–16）：工具与协议（MCP、A2A、Agent Skills）、54 节课的 agent 工程、自主系统与安全、多智能体集群。这一层的分量很能说明项目的判断：它认为 AI 工程的重心正在从“训模型”转向“构建可靠协作的智能体”。 收尾（Phase 17–19）：生产基础设施、伦理与对齐、毕业设计。 # 生态与周边：不只是一个课程仓库 网站：带浏览器本地存储的学习进度追踪、术语表、课程目录和路线图。 六卷本书籍：课程内容由 CI（pandoc）自动构建成 EPUB/PDF，附在 GitHub Releases 上，课程即书，且随仓库持续更新。 Agent 导师模式：运行 npx skills add rohitg00/ai-engineering-from-scratch，可以把整个课程装进 Claude Code、Codex 等编码智能体，变成一个带分级测验（placement quiz）和个性化路径的交互式导师，学习进度写在 LEARNING.md 里。这是“用你正在学的工具来学”的巧妙闭环。 认证备考：5 条备考路径、67 节课、505 道练习题，覆盖 Anthropic 的 Claude 认证和 Agentic AI Foundation 的 MCPA。项目明确声明与这些考试机构无关联，只是独立的备考材料。 12 种语言的翻译（含中文），以及四条核心学习路径（构建与部署 AI 应用、软件工程基础、Agent 辅助工程、产品判断与交付）供不同目标的人选路。

1/ Jev 用 24.9 秒读完了当天早上的 384 条新闻，并为 15 个品牌找出适合当天借势的新闻，总成本只有 0.19 美元。 同一时间、同一批信息源，Claude Opus 5 只处理了 4/384，成本却达到 0.77 美元。🤯

源：https://x.com/elvissun/status/2100951347080421409

引用 @elvissun

Jev is INSANE. 🤯 in 24.9 seconds it read 384 news from this morning and told 15 brands which stories to hop onto today, for $0.19. Claude Opus 5, running on the same feed at the same time, got through 4/384 and cost $0.77. per headline that is ~390x cheaper, and the answer comes back before you finish reading the headline yourself. there's going to be so many ways for JEV to help you find trending stories, find journalists covering it, and get press coverage. everything in the video open-sourced here: http://newsjack.sh (in the demo/ folder, together with 30+ skills to turn your agent into a PR team)

查看引用原文 ↗

3/ 视频中的全部内容均已开源。 项目位于 demo 文件夹，其中还包括 30 多个 Skills，可以把 Agent 变成一支 PR 团队： https://newsjack.sh

@PaynXM 感谢反馈！原因是菜单栏每隔几分钟读一次用量缓存，而缓存过期时会先返回旧值、在后台刷新，所以菜单栏总是慢一拍。v0.1.247 起刷新一完成菜单栏就会立即更新，和快捷面板保持一致。

仓库和在线页面都开源了，推荐仓库直接提 issue 附上链接，它会走和每个条目一样的评审： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@GeekCatX 哈哈 已分门别类 https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@lxfater 牛逼 属实给锤哥用明白了 场景workflow 开源地址（已分类）： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@nicebabycat 哈哈 快和opus5.5一起用起来 开源地址（已分类）： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

在线筛选版：https://agentskillshub.top/best/claude-video-skills/ GitHub 仓库（中英双语）：https://github.com/zhuyansen/awesome-claude-video-skills 刚开源还没什么 star，觉得有用帮忙点一个，也欢迎提 PR 补充漏掉的项目

@Saccc_c 和opus5.5搭配更佳 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@wlzh 哈哈 搭配opus5.5更佳 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@MindfulReturn 嘿嘿 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@20mprofitai 哈哈😄 搭配opus5.5 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@hezhiyan7 哈哈 opus5.5 + skills 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@yhslgg 🤣还真说不定 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@berryxia 嘿嘿 和opus5.5搭配起来 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@Chris62771610 哈哈 是的 和现在爆火的opus5.5搭配极佳 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@gxjdian 👀搭配上opus5.5 仓库和在线页面都开源了（分类版） https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@gkxspace 搭配opus5.5 真的很棒👀 仓库和在线页面都开源（分类开源）： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@goan999999 哈哈 搭配 opus5.5 绝配 仓库和在线页面都开源了，推荐仓库直接提 issue 附上链接，它会走和每个条目一样的评审： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

@Adam38363368936 哈哈 谢谢Adam 刚好赶上opus5.5 仓库和在线页面都开源了，推荐仓库直接提 issue 附上链接，它会走和每个条目一样的评审： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

文中可免费skills获取 仓库和在线页面都开源了，推荐仓库直接提 issue 附上链接，它会走和每个条目一样的评审： https://github.com/zhuyansen/awesome-claude-video-skills https://agentskillshub.top/best/claude-video-skills/

这个新开源的项目fakecloud挺有意思的，是LocalStack 的开源替代品，是本地的 AWS API 模拟器，覆盖了 100+ Service的操作。如果有需要对照着AWS开发服务或者做测试的，那可以试着用这个项目。 我觉得在 AI 这波之前，大家搞这种项目的动力都想对较小，但现在成本低，易实现 https://github.com/faiscadev/fakecloud

又更新了一波新的Opus 5.5 Case。欢迎访问 Skillry 查看 https://skillry.dev/ai-videos/opus-5-5

引用 @yihui_indie

🔥301 个 Opus 5.5 视频，我花了整整 $3,000。 我用 Claude Opus 5.5 把全网最难、最火的特效重新做了一遍。成片导出的那一刻，看着屏幕直接瘫坐震惊…… 所有测试视频、全套 Prompt 和工作流，全部整理开源！ 👇 GitHub 开源地址与在线 4K 原视频见评论

查看引用原文 ↗

现在最便宜的写代码模型，居然不是 DeepSeek，是小米 MiMo 我看一下opencode 每月 10 美元的套餐，估算 5 小时能跑的请求数： MiMo-V2.6-Flash 约 3 万次，状态好的时候比 DeepSeek 还强，就是不太稳 DeepSeek V4.1 Flash 约 2.6 万次，这还是闲时的数，而且刚涨过价 GLM-5.3-Flash 约 6300 次，三家里最少 额度最多但不太稳的 MiMo，你们会用来写代码吗？

https://x.com/shao__meng/status/2104453599367770287?s=20

引用 @shao__meng

斯坦福大学课程 CS329A: Self-Improving AI Agents 斯坦福大学 @achowdhery @Azaliamirh 两位老师主讲，他们在 Google Brain / Deepmind 有过合作经历，现在合作连续两年讲授 AI Agents 课程。 分享这门课程后，类似获得了超过 5K+ 点赞和收藏，看到很多朋友喜欢，我也把 9 个视频做了一个预告视频，让大家对课程可以更快有更直观的感受。 虽然 AI Agent 已经强大到很多人觉得不需要再学习，什么问题都可以直接问，到顶级大学的基础课程，系统学习下来，还是会很有收获，全局认知的提升。 马上长假了，收藏学起来！ 课程主页： https://cs329a.stanford.edu 视频列表： https://youtube.com/playlist?list=PLangBM27OtEA&si=khxaTCYv_ARHBgBl

查看引用原文 ↗

30 verified Claude Opus 5.5 browser animation cases, ranked by X views, with videos, screenshots, prompts and methods https://github.com/jacobbubu/claude-opus-5-5-js-animation-research

我觉得 Opus 5.5 不抽大麻，做不出这种东西。 看完你就懂 PPO, GRPO, DPO! Better than expected.

整首歌是 Opus 5.5 用 Python 代码写谱、再从声波开始合成出来的，没有用任何类似于 Suno 等 AI 音乐模型。 歌词也是 Opus 5.5 写的，需要前期跟它聊一会确定对概念理解的准确。 人声用的是 macOS 自带的 Siri 语音朗读，再通过信号处理把它"唱"成旋律。 流程是如下： 用一个小 Swift 程序调用 macOS 的语音合成，让 Aaron（主唱）和 Nicky（和声）两个声音一个音节一个音节地朗读。 不靠耳朵做检测，看频谱图，测响度和峰值；测音高误差，中位数是 8.7 音分。 上述 technique，没有详细告诉 Opus 5.5，基本就是让他一把梭。 太惊吓了。

引用 @dongxi_nlp

我觉得 Opus 5.5 不抽大麻，做不出这种东西。 看完你就懂 PPO, GRPO, DPO! Better than expected.

查看引用原文 ↗

写 AI 视频 prompt，最没用的一句话是：一个很有电影感的漂亮场景 cinematic-video-prompt-skill 教模型说专业的拍摄语言 中近景、低角度、缓慢推近、轮廓光、青橙调色 700 多个术语分成六本词典：镜头角度与运镜、灯光、构图、镜头与胶片、风格色彩情绪、材质天气姿态。每个词都写了什么时候用 但真正值钱的是第 8 节 作者把大约 1100 条 prompt 和 65 支样片逐条对照，总结出这些模型最爱犯的错 数字基本没用。写 24 秒出 30 秒，写 no slow motion 照样慢动作。唯一有用的数字是具体时刻，写下午 3:40 到 4:10，能让多个镜头光线一致 人脸会飘。换个场景或换个光，长相就变了。对策是一个片段只用一个场景，每段重复一次外貌描述 画面里的字别信。短词加引号还行，长标语、霓虹招牌、日文全是假字。标题和字幕都该放后期做 别硬塞。列 11 样东西只出现 4 到 5 样，18 秒塞 9 个场景必丢。10 秒最多 3 到 5 个节拍 事件要写清位置、时间、看得见的结果。不要写"门自己开了"，要写"第 4 秒，左边第二扇门完全打开，露出一片漆黑"。而且要写你想看到什么，光写禁止项模型会忽略 还有一条法律提醒：别在 prompt 里直接点名影视、游戏、角色的名字，既容易跑偏成那个 IP，也有风险 词汇是通用的，Veo 3、Kling、Sora、Runway、即梦、Midjourney 都吃。作者是越南开发者，每个术语都配了越南语解释 MIT 协议

Magpie 接上 DeepSeek 之后，本来 prompt 成本就压得很低。现在再叠一层 RTK，把 Bash 工具输出的噪声过滤掉，实测能省掉七成多的工具输出 token。两个省钱方向叠加，性价比确实很香。

@VanTrang359816 加上了 🙌 供应商编辑里多了「上下文窗口」：填 128k 对所有模型生效，或 gpt-6=1m 单独设某个模型，逗号分隔。下个版本就有（命令行现在就能用：magpie provider set <id> context=128k）。

@ruterlv 这个已经支持了 🙌 在 magpie 的 Providers 页打开那个 Provider 编辑，有个「上下文窗口」字段：填 128k 对它所有模型生效，或者写 模型id=1m 单独设某个模型，多个用逗号隔开。保存后 magpie 会把窗口写进各个 agent 的配置，不用再手改配置文件。手动添加的模型也一样，建议升级到最新版～

@um1ng_x 已支持，magpie v0.1.233：在 magpie 里给 fx 选模型就行。magpie 会把自己作为免 key 的 provider 写进 ~/.fx/settings.json，fx 的 /model 里就能用上 magpie 的所有模型。 https://github.com/yetone/magpie-releases/releases/tag/v0.1.233

magpie v0.1.233：支持 fx 了，Vercel Labs 刚出的用 Zig 写的编码 agent。 在 magpie 里给 fx 选模型就行：magpie 把自己作为免 key 的 provider 写进 ~/.fx/settings.json，fx 的 /model 里直接能用 magpie 的所有模型和路由组，你自己配的 provider 原样保留。 已测：fx 0.0.11 经 magpie 用 DeepSeek，多轮读写文件正常。 https://github.com/yetone/magpie-releases/releases/tag/v0.1.233

@FlyanLu 给老板汇报下，中午已经给公司所有售前和售后的同事都培训了内网穿透的售后服务规范 1. 优先发送代理冲突的教程，让用户根据教程来解决手机端和电脑端的代理冲突，避免影响懒猫微服的域名正常访问 2. 代理冲突以后，如果 P2P 直接打洞没问题，就什么都不用做。如果还有问题，就引导用户配置 TCP、UDP 和 WebSocket 的打洞协议，针对性的解决大陆每个省运营商的流量劫持 3. 如果代理冲突和选项配置都解决以后，遇到像跨省、跨运营商这种极端的情况，优先给客户配置不限速的中继，以后再给用户推付费的流量包，让用户在异地遇到极端情况的时候，可以自主选择流量加速方案 4. 针对跨国访问的情况，暂时推 VPS 的配置教程，等产品更新以后再给用户推 VPS 的配置方案，方便用户在国外可以高速地访问中国微服 5. 如果用户不相信我们的官方加密中继，就发送图文教材，告知用户怎么自建 VPS 中继服务器，同时引导用户要定期检查自己 VPS 中继服务器的联通状态，避免和微服失联 感谢老板给我们提出的建议，我们的内网穿透是有完整方案的，但是因为我们的同事有很多是新来的同事，对原来的业务模式和解决方案不是很熟练，耽误老板的时间了。今天中午统一培训了一下，提升我们的售后服务质量和响应的速度 再次感谢老板的中肯建议，我们继续努力

引用 @manateelazycat

解决问题就好，也给各位大佬盘点一下 我们的打洞软件是支持 TCP、 UDP 和 WebSocket 的，而且还可以对抗运营商的 QoS，一般来说，安装第一阶段就要解决小火箭和我们的冲突问题。只要解决了代理冲突的问题，大多数都是可以直接打洞直连的，打洞以后，在外使用就是家里带宽的极限 但是大佬的情况是，微服、安卓、iOS 分别是不同的运营商的网络，然后又是跨省连接的话，确实很难打洞。我们后续会对这种跨运营商、跨省的极限情况提供付费的流量包服务。让用户可以自助的解决这些极端的网络情况，平常回家，在当地的话，就不需要这种流量包了 但是反观这次解决问题的过程，我们的售后团队的认识还没到位，其实是可以解决的，但是没有多问大佬是不是跨运营商的问题，所以耽误了大佬很长的时间，抱歉抱歉，我跟大佬道歉 我中秋节后第一天就会给售后团队做统一的培训，把我们的服务和专业做的更好，节省用户的时间，再次感谢大佬对我们的中肯的建议，我们还有很多地方不足，继续努力！

查看引用原文 ↗

I¹ ported my Mini web browser to Rust & Swift. It's faster, more secure, and shockingly, nicer to iterate on than Electron + Bun. I always wanted Safari-like UX but… Chrome 😁. Thanks to the 𝚌𝚎𝚏 crate, I can bundle up-to-date Chromium. 𝚏𝚡 𝚊𝚌𝚙 lets me embed an agent without bloating the app. It talks to my local fx CLI over ACP. fx can then manage the browser via an MCP server². By ditching Electron, I was able to get Liquid Glass, faster boots, and a more Mac-native feeling. Little things like fine-grained focus and input control, which are nightmare fuel in JS, just work®. Native is the future, both on the desktop and in the cloud. I suspect the entire software world will "nativify" faster than people realize, as DHH hinted at Rails World. Excited that Vercel's bet on Fluid will help support this, whether you choose to go Rust, Go, Zig, scriptc… ¹ Built on http://fx.sh with Opus 5.5 and Sol 6 ² In the process I learned about http://agentclientprotocol.com/rfds/mcp-over-acp, which I'm now really excited about. This will allow a more secure and direct way to expose MCP tools

The irony of replying claiming a model is always categorically better for evals. That assertion is complete bullshit. I’ll repeat what’s in the post > No single classifier is best for every eval. Validation with human labels help you make trade-offs between accuracy, cost, and speed for your application.

用 Opus 5.5 做了一支 60 秒古风短片《月下筝》 · 人物、汉服、古筝、水榭全部用代码生成，没有外部模型 · 每个音都对应画面里被拨动的那根弦 手指、穿着比较难处理，迭代了好几次，还有很多进步太快，但到了 5小时限额了

难得看到不一样的声音。 做 personal agent 其实还挺难的，单单第一条能给得起超强配置的云电脑，这事也就财大气粗的大厂才能干吧？ 其他方面只有经过各种复杂任务的测试，才能知道这个产品的能力边界在哪。Muse 和 Grok Bot 是属于，越用越发现他们牛逼到离谱。

引用 @wangzan101

经过一天对 Today @TodayAIofficial 的深度测评之后，得出以下结论： 1、24 小时云电脑：完全没有，营销噱头，就是单纯的服务器，和 SaaS 产品没啥区别； 2、电脑控制：大残，授权了依旧啥也干不了； 3、浏览器控制：半残，能读取网页信息，至于真的 borowser use，基本没有 唯一做的比较好的就是“懂我”，而这个懂我是在牺牲我隐私的情况下换取的。 总结：通过全面获取我的隐私信息，懂我做的还行，帮我干活基本残废。 能干活的几个网站基本都是 MCP 或者 CLI 链接的，覆盖范围有限。 Personal Agent 的重点应该还是帮我干活上，但是 Today 这方面做的太差劲。 特别是在 computer use 和 brower use 如此成熟的情况下，h还做的这么烂，非常不理解。 像是个半成品急匆匆的发布，没有做任何干活能力的评测，建议后续优化完善一下工程能力@junyuan_qi

查看引用原文 ↗

1/ 有人让 Jev 和 Laya 在 Chrome 恐龙游戏里正面对决，而且它们真的在自己控制恐龙。 Laya 在 Mac 本地运行，Jev 通过 API 运行。 相同赛道、相同物理规则，延迟却完全不同。它们会自主判断、躲避障碍、使用护盾，并努力存活更久。👇

2/ 更有意思的是，TypeSafe AI 的 Jev 还负责控制赛道上的障碍物。 项目已开源： https://github.com/virajbhartiya/laya-vs-jev

源：https://x.com/heyxviraj/status/2102048070649405592

引用 @heyxviraj

i made Jev and Laya play the chrome dinosaur game against each other except they’re actually controlling the dinosaurs. Laya runs locally on my Mac, Jev runs over an API. same track, same physics, completely different latency. they make their own decisions, dodge obstacles, use shields, and try to survive. https://github.com/virajbhartiya/laya-vs-jev p. s. @typesafeai 's Jev controls the obstacles too

查看引用原文 ↗

Badge Tech 🏅 A real-time enamel pin #GLSL fragment shader. Hover to explode it and see every layer that makes the badge. Procedural patterns or your own SVG, rendered with my raymarch engine for crystal dispersion, GGX metal, enamel and holo foil.

#审美积累 徽章设计

引用 @jaenam97

Badge Tech 🏅 A real-time enamel pin #GLSL fragment shader. Hover to explode it and see every layer that makes the badge. Procedural patterns or your own SVG, rendered with my raymarch engine for crystal dispersion, GGX metal, enamel and holo foil.

查看引用原文 ↗

特喵的，以后可以笑着走出理发店了，让GPT-image 2.5给我发型和美容建议，太夸张了！！！ 建议大家都去试一下这个提示词，国庆节，无论是给自己出片，还是给女朋友/老婆出片，都无敌了！ 提示词如下👇： 请像一位美学/美容顾问一样，从我附带的照片中分析我的外貌。 识别出能让我看起来更精致和谐的最高影响力的变化，包括发型/发式/发色、眉毛、眼妆、皮肤/妆容位置、嘴唇，以及整体面部平衡。 请针对我的实际特征给出具体建议，而不是泛泛而谈。 然后，生成这张确切照片的现实“之后”版本，展示你推荐的焕然一新效果。保持我可辨识为同一个人，并保留我的面部解剖结构，除非你特别推荐细微的结构调整。 尽可能保持相同的姿势、相机角度、表情、光线、服装和背景。 让这些变化精致且现实，而不是夸张的美颜滤镜。优先考虑高影响力、低风险的造型调整。 不要自动让我脸变瘦、鼻子变小、嘴唇变大、眼睛变大，或皮肤不切实际地完美。 对于头发，选择最能衬托我个人比例的发型、发色、分线、蓬松度和面部框架。 对于妆容，优化眉形、眼线/睫毛、腮红/轮廓位置、唇部定义和肤色，同时保持自然。 生成一张前后对比照片，带有照片标注，解释每项变化，类似于专业美容咨询。

引用 @raealisa

chatgpt actually gives great glow up advice

查看引用原文 ↗

spore was the biggest disappointment of my youth and I aim to fix that even if I have to do it myself. check out these procedurally animated (arbitrary!!) skeletons https://quadruped-gait-atlas.pages.dev/

@kiki_qq6090583 我什么时候说了我们是靠给别人做网站赚钱的？ 即使你真要做这个，当然是让他们去注册域名。也的确有人不知道怎么注册，于是由服务商帮忙注册的。结果几年后，店家发现不知道在哪里续费域名了。

重新开始VibeCoding，先改造一个老项目，在后台直接接入了一个博客系统，根据谷歌的搜索热度来给出词，可以自动根据这些长尾词产出博客，我来点发布，也可以手动生成长尾的文章，接入了Claude的模型、DataforSEO~~

4/ 这段 Prompt 最值得参考的地方，是它同时给出了目标、工具、审美、参考资料和检查流程，又保留了足够的创作自由。 JS 动画视频源代码： https://github.com/JohnHeibel/PDoomVideo 原始 Blender 视频 MP4： https://x.com/other__reality/status/2102514581684052169?s=20

引用 @other__reality

Claude Opus 5.5 has the best visual design of any model I have tested so far

查看引用原文 ↗

让 Claude Code 做视频的开源项目，现在多到没人说得清到底有哪些 我花了 200 美元的模型调用，把这件事做了一遍 180 个仓库，全部让 Jev 读 README 判断真实用途，再分成十类 剪辑与后期 35 个 讲解科普 32 个 框架与通用工具包 29 个 产品宣传与演示 26 个 短视频与口播 19 个 故事与动画 13 个 动效与 Logo 11 个 剧本与学习 9 个 数字人 4 个 音乐视频 2 个 加起来将近 20 万星，最大的一个 6 万多 每一条还标了安全评级、开源协议、主语言和星数，页面每 8 小时按 GitHub 实时数据刷新 装之前建议看两件事 安全：178 个评级 SAFE，2 个标了 CAUTION。但这是规则扫描不是人工审计，自己也该看一眼代码 协议：113 个 MIT，但 27 个压根没写协议，18 个协议不明。这 45 个拿去商用之前，先问作者 为什么做成分类清单而不是随手发个列表：因为按用途分才有用 你想找的是"帮我把一段录屏剪成短视频"，不是"给我 180 个视频工具" 网页版和 GitHub 仓库都开着，清单本身用 CC0 放进公共领域，随便抄随便改，不用署名

引用 @gosailglobal

https://x.com/i/article/2104181663844700160

查看引用原文 ↗

真香，最近满屏都是 Opus 5.5 动效视频，刚想研究怎么搞，喂饭级的清单就来了😋！！！ 180 个能让 Claude、Codex 直接做视频、剪视频的开源 skill，全部分好了类，做了安全评级。 1、剪口播：cut-video 自动剪掉静音和嗯啊，还有个 Premiere 的 MCP 1027 个工具，agent 能直接在 PR 里剪时间线、调色、混音。 2、做产品宣传片：video-shotcraft 带 152 张镜头配方卡，归藏的 skill 拿你产品真实的组件做更新宣传片，画面和产品一模一样。 3、做科普号：vox-director 给一个选题，出一条 Vox 那种纸拼贴讲解视频。Paper-Cut 完全不用视频模型，图像模型出图后拆层、用代码做动画。 4、不露脸账号：有现成的 Shorts 工厂，画面、配音、逐词字幕一条龙。 5、中文内容：丢一首诗或一个成语，出国风纸片动画；丢一段中文故事，出手绘漫画风动画。 6、做数据号：CSV 丢进去，出动态图表，每一帧的数字都是准的。 ...... 🔗：https://github.com/zhuyansen/awesome-claude-video-skills

引用 @gosailglobal

https://x.com/i/article/2104181663844700160

查看引用原文 ↗

这太酷了，而且是 Video-as-code： 意味着 视频可以像代码一样 Fork & Remix 这里是中文化版本：重填歌词 + Suno 翻唱 + Claude Code 完整的动画重新适配

引用 @_mexicat

@pleometric that’s pretty cool! i also gave it a try with a different style direction and the result (after a few rounds of tweaking) is impressive

查看引用原文 ↗

@_mexicat Just forked and remixed a Chinese version, yours is genuinely the most imoressive version, kudos! https://x.com/yucheng/status/2104401613968527400?s=46

引用 @yucheng

这太酷了，而且是 Video-as-code： 意味着 视频可以像代码一样 Fork & Remix 这里是中文化版本：重填歌词 + Suno 翻唱 + Claude Code 完整的动画重新适配

查看引用原文 ↗

兄弟们，这几天我没有玩Muse，但是我已经把Claude玩疯了，中国护照过的，没封号！！！ 说一下几个Claude Code桌面端和Codex桌面端不同的感受吧： 1. 模型层面：Opus5.5真的很强，没尬吹，体感比Astra还强一些。Fable5.1我这种菜鸡体验下来，其实和Opus5.5差距不大，出方案也很强，但我更多的是让它把我记忆系统给诊断了一遍，给了很多宝贵的建议。 2. Computer Use和Chrome浏览器控制层面：这一块，Codex还是真神，Claude还会抢电脑。。。不能干别的事真的很难受。。。而且速度还很慢。。。所以只要有这方面的需求，都是让Claude出方案，然后让Codex去执行。 3.交互层面：Claude完全真神级别，说人话，字体、输出格式，都更人性化，比Codex好理解很多。唯一麻烦的地方是只有英文（这可能也是我的缺点），但也很好解决，安装豆包工作，悬浮球，划段落翻译就行。 4.插件生态：Codex还是领先太多了，数量上领先。常见的Higgsfield、topview这种内容插件都没有。而且Codex最近新出了一个模型写作风格的Skill，大家可以看看开启没有，可以模仿你的邮件内容和谷歌文档内容写作。Codex时不时在更新中给一些小惊喜，还是很不错！ 5.对话层面：Codex还是比Claude省心，在Bypass all permissions模式下，Codex是真的几乎不会问你问题（偶尔有补充问题，也会并行执行其他不想干任务），Claude就老是有毛病了，各种授权，尤其涉及到computer use，屁事老多了。。。 6.使用量层面：量大管饱已经不是Codex的代言词，而是Claude的代言词，我两个Pro20✖️，感觉消耗得比我的Claude Max100$的还快。。。Opus5.5现在才是真神！ 本来都打算往Claude多转转，70%左右吧，但是Codex新界面更新之后，我觉得还是55，确实交互度丝滑了很多，新增了定时任务界面、站点管理界面、资料库界面，我又舍不得了。。。现在只能说两家各有优劣，鱼和熊掌暂时兼得一段时间吧。。。 我的Claude都稳定用了快一周了，中国护照过KYC，暂时没感觉有啥问题，接着我就开始写教程了。。。说自己中国护照被封的，估计是网络、环境、支付方式，这些其他问题。。。

Yeah, Artifacts and the cloud containers basically free you from doing things on your computer. That wasn't really my goal - I didn't try to constrain myself, I just finished the project and realized that I never bothered with branches, local checkouts, or PRs. We just kept working in artifacts and the cloud. But your milage may vary and I'm not saying that my workflow is the "must do" workflow for everyone! Collaborative work in a team is obviously different, for instance.

Exciting work from NVIDIA. (bookmark it) Interesting to see this approach to turn public Agent Skills into RL environments. Lots of excitement around RL environments so this is a great read. Skill2Env compiles each Skill into executable terminal tasks. A Codex planner reads the SKILL.md bundle, researches related public assets and splits the Skill into workflows. A Codex creator then builds each task with programmatic tests and a behavioral rubric taken from the Skill's own quality criteria. From about 3.4k crawled Skills, the pipeline produced 7,971 tasks across 13 domains, with software engineering under a quarter of the corpus. Generating them with GPT-5.6 Sol cost over $90k in API usage. After 300 steps of outcome-only RL, Qwen3.8-27B improved from 49.4% to 54.1% on Terminal-Bench 2.1 and from 33.4% to 37.7% pass@1 on S2EBench, their hand-verified held-out benchmark. Adding the rubric to the reward gave smaller benchmark gains, 50.1% on Terminal-Bench 2.1. Given the source SKILL.md, a judge preferred the rubric-trained model's trajectories over the base model's on 73.0% of tasks, against 54.5% for the outcome-only model. Paper: https://github.com/NVlabs/Skill2Env/blob/main/paper/Skill2Env_arXiv.pdf Chat with Paper: https://academy.dair.ai/papers/reinforcing-agents-with-collective-skills

NVIDIA 发布 Skill2Env：用“集体技能”强化智能体 NVIDIA 研究者们把社区公开的 Agent Skills 编译成可执行 RL 训练环境的数据流水线：3.4k 个 Skills 变成 8k 个带程序化测试和行为量规的终端任务；仅 300 步 RL 训练就让 Qwen3.8-27B 在 Terminal-Bench 2.1 上提升 4.7 个百分点，且模型行为显著向源 Skills 的方法论对齐。 开源项目：https://github.com/NVlabs/Skill2Env 核心洞察：公开 Agent Skills 是一个被忽视的监督来源 Agent Skills 是“教智能体做某件事”的文件夹：一个 SKILL.md 加上可选的脚本、参考资料和资产。论文指出，把公开 Skill 语料当作数据来读，它同时提供三样东西： · 任务分布的采样：人们真正想让智能体处理的任务分布（有人愿意花时间写下工作流，说明这活儿值得自动化）； · 真实世界的锚点：指向真实的仓库、数据集、工具和工件； · 结果测试表达不了的质量标准：领域专长、默认参数、常见坑、“好结果长什么样”。 # 数据流水线：四阶段编译，验证靠构造 1. Plan（分解）：容器化的 Codex 规划器读取完整 Skill 包、联网调研相关公共资产，把 Skill 拆解成若干可验证的 workflow，每个附带元计划（场景、初始世界、预埋缺陷、难点来源、解法草案、验证策略）、资产建议和“任务轴池”（任务原型 × 验证器模式 × 人物画像）。 2. Diversify（多样化）：宿主从轴池采样一组组合，加上复杂度、指令语气、请求者专业水平。关键设计是轴池以 workflow 为条件：研究型 workflow 配“证据可追溯”验证和研究者画像，而不是从全轴乘积空间乱抽，这让多样化保持 sensible。 3. Create（构造）：全新创建者 Codex agent 在 Docker 内工作，尽可能用真实素材（钉在特定 commit 的开源仓库、真实版本化文档、官方 API 规范）；需要联网服务的场景改造成本地替身（stub 服务器、录制回放 fixture、PATH 上的假 CLI、种子数据库），求解时绝不依赖网络。创建顺序被严格固定：先建世界 → 写指令 → 写测试 → 写量规 → 最后才写参考解，测试先于解法冻结，保证解法必须迁就评分契约而非反过来。 4. Verify（验证）：宿主端无模型参与的接收门：静态检查（布局、符号链接、Dockerfile 安全、基础镜像按内容摘要钉死）+ 两个容器内试跑：Oracle（参考解）必须全指标满分，NOP（什么都不做的 agent）必须全指标零分。任一失败即拒绝。 值得注意的一个反直觉选择：不做 teacher 模型预验证（不像部分工作用强模型试解、解不出就丢弃任务）。理由有二：这会把任务难度上限压到验证器能力，且成本翻倍；而 group-based RL 的在线动态过滤（rollout 无优势的 prompt 自动不产生梯度）天然淘汰过难/过易任务。 # 数据画像：广、贵、且忠实于源 规模与成本：7,971 个任务，用 GPT-5.6 Sol（xhigh 推理档）生成，API 花费超 9 万美元。（脚注：出于法律原因，公开发布的数据集改用 Kimi-K3-max 在同一流水线下生成。） 领域分布：13 个领域中，软件工程仅占 22.5%，AI/ML 10.5%，商业/金融/法律/HR 10.5%，营销 9.3%……论文对比了 TMax-15K、Terminal-Bench、DeepSWE 等，Skill2Env 是唯一全覆盖 13 域、且非技术知识工作占大头的语料。 忠实度探针（很聪明的设计）：用任务指令+量规作查询、对 3.4k 个 SKILL.md 做 TF-IDF 检索，73.2% 的任务 top-1 命中真实源 Skill，94.6% 进 top-10（随机 0.03%）。单用量规也有 68.5% top-1，证明量规携带的是 Skill 专属方法论而非泛泛建议。 SFT 数据：用 GLM-5.3 对每个任务 rollout 两次，得到 15,968 条轨迹，平均奖励 0.74，中位轨迹 19 次模型调用 + 23 次工具调用。 S2EBench：考虑到公开基准饱和，从 SkillHub 另外生成、逐条人工审核（指令无歧义、忠实于源 Skill、测试公允）后的 79 任务私有 held-out 基准。 # RL 实验：基础设施 + 极简配方 基础设施（论文明确说“现代 agentic RL 首先是基础设施挑战”）：Molt（PyTorch 原生全异步训练，Ray + vLLM + FSDP2）+ Polar（agent rollout 层：rootless Apptainer 沙箱、代理回传 token ID 和采样时 log-prob、prefix merging 把 harness 的多次补全缝合成训练轨迹）。 配方（刻意走“简单路线”）：GRPO 组归一优势 + DPPO 的 binary-KL 信任域掩码（δ=0.05，超出阈值的 token 直接丢弃，无需参考模型，还能防训练-推理失配）；G=8 rollouts/组，批 64，lr 1e-6 恒定，无 KL 惩罚、无熵奖励、无 SFT 热启动，每任务 65k 上下文。 量规校准奖励：开量规时，额外由 GPT-6 Astra 做 LLM-as-Judge（带“宪法”：惩罚无脑循环、reward hacking、答非所问；hacking 实证 = -5 分），总奖励 r = r_V + λs/5（λ=0.2），即 judge 最多把程序化奖励拉动 ±0.2。量规是校准可执行结果奖励，而非取代它，这是与“Rubrics as Rewards”一系的定位差异。 # 四项发现（论文最有信息量的部分） 发现 1：小规模 RL 即有跨域迁移。 仅 300 步、只用 2,400 任务子集训一个 epoch：S2EBench pass@1 +4.3（均分 +18.5），Terminal-Bench 2.1 +4.7（49.4→54.1）。训练集与 TB 无重叠（13-gram Jaccard < 0.8），且训练集从未针对 TB 调过，论文将其解读为规划、工具使用、收尾能力的通用提升而非任务族记忆。这让 27B 本地模型显著缩小了与云端前沿模型的差距。 发现 2：量规校准 RL 在基准上落后于纯结果 RL，一个诚实的负结果。 量规版在 TB 2.1 只有 50.1（纯结果版 54.1）；训练中量规版的程序化奖励长期停在 0.5–0.6，judge 分项从头到尾无上升趋势，两个奖励在训练分布上互相拉扯。论文不把它当作对量规奖励的终审判决（两者优化不同目标，而基准只考结果那一半），并给出两个疑因：λ=0.2 的加性形式让失败任务仍能拿正奖励、judge 看不到文件系统等设定均未调优；以及更本质的，Skill 写下的方法论可能本来就不是最大化基准通过率的分布。 发现 3：行为确实向 Skill 对齐，量规的价值所在。 200 个任务的成对偏好测试（judge 拿源 SKILL.md 当标准，比较匿名化的 base 与 RL 轨迹）：纯结果 RL 已被偏好 54.5% vs 33.5%；量规版被偏好 73.0% vs 24.0%。这说明量规奖励买到的东西在结果基准上看不见，但对“怎么做事”影响实质，对网页开发、报告综合、开放研究这类难验证任务尤其重要。 发现 4：GLM-5.3 蒸馏 SFT 反而伤害 Qwen。 在 GLM-5.3 轨迹上做 SFT：27B 上 TB 2.1 掉到 45.8；4B 上直接崩塌（TB 18.7→3.4，出现思维/工具调用死循环）。归因：教师的 interleaved-thinking + 工具调用风格与学生自身 post-training 不兼容，模仿覆盖了学生依赖的行为模式却带不来教师的能力。与 TMax 报告的“SFT 混合数据劣化已后训练的 Qwen”互相印证。因此论文所有 RL 结果都从未修改的原始 checkpoint 出发。

Engineers will have to show that they're better than creatives at creativity too. Some will do that with ease. Still, amazing. Half a year ago vibe-reproducing Evangelion look took me decent effort

引用 @luisbizarro

I’m honestly shocked with Opus 5.5 on how easy it is to just copy people’s work and output anything, feels distopian for some things. From this tweet, I just cloned the repository and asked this: “can you create a new project with the title Evangelion that uses the same premise, but uses a song [Name of Song] from my desktop? Please all the interface needs to match this Google Search with Evangelion UI for the reactivity: [Google Search]. Add 10-15 different visualizations”. No technical instructions at all and just asking based on a Google Search, not even copying the images properly or “directing the AI” based on the previous tweet. The output after 15 minutes is great. That just makes me skeptical that engineers won’t be just replaced by creatives in many fields of web development. https://evangelion-neon-overdrive.vercel.app/

查看引用原文 ↗

Vision Pro owners: I made a site where you can find ONLY Vision Pro apps. Tapping any of them takes you right to the App Store listing for that app. https://onlyonvp.vercel.app

阿里云开源「企业级 Agent 白皮书」 2026 年最新发布，是 2025 年 9 月「AI 原生应用架构白皮书」的升级续作。全书按 架构 → 构建 → 运行 → 治理 → 调优 的全生命周期组织，共 7 篇 30 章，由阿里云数十位一线工程师分工撰写，并纳入吉利、塔斯汀、MiniMax、哔哩哔哩、信永中和等外部企业案例。 它的写作动机很明确：过去一年市场重心已经从 “如何快速搭出一个 Agent” 转移到三个新挑战，工程化（从概率智能到可靠生产力）、规模化（从单点试验到智能基础设施）、组织化（从 Agent 孤岛到进入核心业务流程）。现有的框架文档和教程基本不回答这些问题，这本白皮书填补的正是这个空白。 开源地址 https://github.com/aliyun/ai-agent-handbook # 各篇核心内容 架构篇（1–2 章） 建立认知框架。给出 Agentic Application 的六个判定特征（以任务结果为中心、运行时决定部分执行路径、能作用于环境、维持跨请求状态、受确定性机制约束、可观测可评估）和成熟度四级模型（L1 辅助生成 → L2 受控自动化 → L3 Agentic Execution → L4 规模运营）。两个重要的解耦判断：用哪种形态取决于任务结构，处于哪级成熟度取决于治理完备程度；单 Agent / Long-Horizon / 多 Agent 是沿时间跨度和协作结构两个正交维度的扩展，不存在“必须升级到多 Agent”的路径。贯穿的原则是“最低充分架构”，为任务选择成本与风险可接受的最低复杂度。 构建篇（3–6 章） 是方法浓度最高的部分，按“范式—任务—信息—行动”还原构建过程： · 任务：Agent Loop 五阶段（Prepare→Model→Act→Observe→Verify）+ 十态任务状态机，要害是“消息历史不应是任务状态的唯一来源”；完成判定的核心原则是“模型只能申请完成，Harness 依据环境证据提交完成”，验证分五级并与风险匹配。 · 信息：Context 是动态“编译”而非静态字符串。本章的独创设计是 Context Manifest，每次调用记录上下文每个片段的来源、作用域、版本、信任级别、选中理由和内容哈希，使“模型看见了什么”变得可解释、可回放、可审计。信息被五分为 Context/State/Memory/Knowledge/Skill，其中 Memory（个人经验）与 Knowledge（组织内容）必须分列，因为治理责任不同，“放进同一个向量库会同时失去两类治理能力”。 · 行动：统一 Action Plane（意图→Schema 校验→身份绑定→策略决策→执行→观测），关键三分：“模型看见工具 ≠ Harness 注册了工具 ≠ 获得执行授权”。协议定位清晰：Function Calling 是模型-Harness 意图接口，MCP 是 Harness-能力提供方连接协议，A2A 面向拥有独立任务循环的远程 Agent，“协议选择由能力是否拥有独立任务循环决定，而非新旧或流行度”。 运行篇（7–12 章） 处理规模化后的工程问题，大量内容达到了分布式系统的专业深度：沙箱后端选型判据（容器/gVisor/MicroVM 按代码可信度与租户边界取舍）；状态外置后 Event Log / Checkpoint / 工作区快照三者不可互相替代，且“Durable Execution ≠ 外部动作恰好执行一次”；AI 网关对 LLM/MCP/Agent 三类流量按不同粒度治理，其中“严格预算需要原子预留而非阈值检查”的数学化分析（余额 100、两笔 80 的并发请求都会通过）是真实的并发工程细节；多 Agent 编排强调“最小充分共享”，共享的是上下文来源而非同一个 Context 窗口。 治理篇（13–16 章） 让自主运行的系统变得可信。可观测性的判据是“请求成功 ≠ 任务成功”；安全章同时把 Agent 当被攻击对象和行为主体来防护（身份是全章最扎实的部分：数字工牌、Token Exchange 权限收敛、On-Behalf-Of 且 Agent 权限 ≤ 用户权限）；资产管理把 Prompt/Skill/MCP/Agent 当作运行时依赖做注册与版本治理。第 16 章 Agent Simulation 是全书原创性最强的一章：Agent 行为之所以不可验证，是缺制度前提（角色无外部标准、失败无自然代价、身份不连续），模拟是当下唯一可做的事，本质是“用可靠 Harness 约束不可靠内核”。它甚至给出诚实的统计学提醒：n 次零违规的 95% 置信上界约为 3/n 而非零。 调优篇（17–24 章） 的组织原则是“归因决定方法”：先排除环境故障、再修 Harness、最后才动模型，“把本应由上下文或工具协议解决的问题当成模型不行，是代价最高的一类误判”。主线是数据飞轮：Trace→Trajectory→黄金数据集（输入/轨迹/结果/判据四要素）→Badcase 闭环→受控自进化（模型生成的改进一律是候选变更，须回流构建、过门禁、可回滚）。模型调优章对 SFT/Agentic RL/蒸馏的适用边界、奖励投机的三套机制分离（训练奖励、独立评测、系统硬约束）论述相当严谨，广引 DeepSeek-R1、Tulu 3、FrugalGPT 等外部工作。 总结篇（第 30 章） 是全书思想密度最高的总结。当企业同时运行多 Agent、多框架、多租户时，同样的工程要求在每个应用里被重复且不一致地实现，这本质上是缺一个共享的系统层。Agentic OS 被给出“窄定义 + 三条否定”：为 Agent 任务提供公共运行对象、能力接入、可强制边界与统一证据的系统层，它不持有任务语义、不是又一个框架、不必然改内核。能力下沉有三条判据（复用性 + 强制性或可验证性），九类管理对象（其中 Budget Lease 预算租约最易被忽略），并提出“自治上限由可撤销范围与可证明范围决定，而非模型能力”。 调研报告 的 1906 份问卷给出一个关键发现：已开发或开发中 Agent 的企业占 46%，但真正上生产的仅 18%；有评估体系的企业任务成功率是无评估者的约两倍，卡点不是模型能力，是 Harness 层的工程配套。这与全书立意互为印证。

282 viral videos made with Claude Opus 5.5, each with the exact prompt. Watch every original next to a live remake on Skillry. https://github.com/yihui-dev/awesome-opus5-5-videos

我们现在写AI视频提示词，通常都会把它结构化，比如：全局设定、人物和素材绑定、逐秒时间码，以及负面提示词 看起来已经分得很清楚了，但到了真正生成的时候，模型还是要同时处理很多信息 当这些信息全部挤在一次生成里，模型不仅要理解画面，还要根据文字去“猜”空间 信息越复杂，彼此之间越容易打架，最后不可避免地就要靠抽卡 白模的作用，虽然不是百分百按我们设计的来，但起码已经先把构图、人物数量、站位、移动路线、遮挡关系和镜头路径演示出来。 这样一来，人物表演、材质、光影和特效仍然可以用提示词去控制，但空间关系不再完全依赖文字让模型自己想象 所以我觉得，白模真正减少的不是所有随机性，而是空间关系上的随机性 但这里有个问题，就是如果本身没有什么镜头感的朋友去做白模，其实还不如靠视频模型随机性生成的分镜构图好 对于我这种每个镜头我都有自己想要的构图的，我就会全部都使用白模，自己先“拍一遍” so：多人物在同空间-用白模 镜头感强有自己的构图美学-用白模 复杂运镜+多人物-用白模 普通对话-纯文字

@random_walker Does this matter? Yes it's trivially true for one-off tasks but I need them to carry projects with hundreds of serial tasks, and switching within a project can have negative effects. So realistically I can't use a random router

6. Opus 5.5 一次生成了这个 3D 水体效果，目前已经开源 https://x.com/Aurelien_Gz/status/2102786378282987591

引用 @aurelien_gz

opus 5.5 is something else.. so many people asked me to open source the 3d water from my last post. a few even wanted to buy it so here it is. my first open source drop code » https://github.com/Aureliengmz/clearwater demo » https://aureliengmz.github.io/clearwater/ show me what you make with it

查看引用原文 ↗

卧槽 Muse 内置的语音也很好用啊！ 按住 Fn → 说话 → 右下角弹出可爱的 bot， 当你在输入框时就转文字，当你在桌面上，自动保存一个 txt 文件到桌面。太 ™方便了。建议豆包/微信输入法都学一下。 而且发现 Codex/Muse 这些国外产品的内置语音输入识别率都出奇的好，中英文混合也基本没有错，神了。

who's got two thumbs and just ranked #6 of 49 on image-jevbench? 👍 this guy 👍 one of the fastest in top 10, accuracy is close. calibration of confidence is where it's weakest. @airesearch12 says a simple temperature fit should help so i'll need to look up what that means (thx!) try it out, it's all open-source! (current benchmark is on the frozen Qwen3-VL-4B via Glance)

引用 @yoheinakajima

glance-vlm speedlab is now open source! read: https://glance.yohei.me/speed/ try: https://github.com/yoheinakajima/glance-speedlab turn any webcam into multiple live AI detectors (emotion, count, object), running locally ⚡ Recorded live on an Apple M5, one-question loop: PyTorch/MPS FP16: ~210 ms p50 MLX 8-bit: ~160 ms p50 🧪 Controlled nine-question, fresh-frame benchmark: 358.5 → 259.6 ms p50 27.6% lower latency 84/84 decisions matched 21 experiments, reproducible benchmarks, paper, and failures

查看引用原文 ↗

看一眼现在的钉钉首页，就知道大厂高管的战略焦虑有多脱离现实。 底部最核心的一号位标签，堂而皇之地写着千问办公，原本承载核心沟通的消息列表，被生生挤到了二号位。 打开软件，屏幕正中央飘着一只绿色章鱼吉祥物，大言不惭地问你：早上好，今天想做什么？下面还规规矩矩排着文档创作、数据分析几个自嗨的提示词胶囊。 打工人清晨顶着黑眼圈赶地铁打卡，大拇指下意识点向左下角，想看昨晚群里领导布置了什么紧急任务，结果直接被一只绿色章鱼怼在脸上，问你今天想做什么。 我想做什么？我想看工作消息，我想干完活赶紧下班。 一个即时通讯软件，最底层、最不容践踏的心智模型就是发消息和看列表。这是全世界打工人用了十几年的肌肉记忆。 但为了迎合内部的 AI 战役，为了能在高管述职 PPT 上交出一份漂亮的 AI 日活数据，产品经理直接把用户的核心习惯撕得粉碎。 整个软件启动卡得像快要散架的老爷车，还要硬塞进一个没人想用的对话框。这种改版完全不顾真实的工作场景，纯粹是把大厂自上而下的 KPI 暴力宣泄在每一个人的手机屏幕上。 做产品如果为了完成汇报，敢肆无忌惮地骑在用户的核心直觉上拉屎，堆再多所谓的前沿技术，也掩盖不了这种骨子里的傲慢与难用。

Truly. Our models are based on people dropping balls - we ASSUME people will miss our discount codes or wait too long to sign up or join a meeting late. All of that will end.

Financial models are about to get an overhaul. Company GTM today has baked in forgetfulness, bandwidth, and friction. Agents change all 3. Companies issue rebates for products, because they know well over half of their customers will forget to send them in. Companies with monthly subscribers assume that many of those accounts are inactive and will just keep paying. Companies make refunds harder so people will just give up and eat the cost. But over a billion people use AI assistants. And many will start to adopt proactive agents that will work to completely erase that slippage. Your human-run systems are not ready for a near-term 10x increase in the number of pricing negotiations or a 10x jump in the number of refund requests. CFOs who are my clients, I told you this in 2024. CFOs who are not, time to change your planning.

引用 @alliekmiller

Truly. Our models are based on people dropping balls - we ASSUME people will miss our discount codes or wait too long to sign up or join a meeting late. All of that will end.

查看引用原文 ↗

看起来视觉冲击很大呀，眩晕瘫坐

引用 @servasyy_ai

推荐一个，我目前遇到的，最好玩的 opus5.5的开源视频提示词清单 提示词原文： make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out. 这 43 条是从 GitHub 上的 awesome-opus5-5-videos 清单里筛出来的。 清单收了 282 条 X 上用 Opus 5.5 做的片子，其中 43 条用的是一字不差的这句提示词。 项目链接放评论区👇🏻

查看引用原文 ↗

提示词原作者：https://x.com/stephanlivera/status/2103315922098470926?s=20 5.5 之所以爆火因为提示词极大降低了门槛，你只需要写「充分展现你作为动态设计师的卓越才能，就像制作简历中的作品集一样。全力以赴。」这种空洞的话，大模型就能生成超乎想象的动效视频。你都不需要写具体的每一个细节。 我看了下评论区的那个 awesome 5.5 videos 仓库，其实都是拿别人的视频然后一句「你复刻一遍」，也根本没有什么提示词😅 A\ 这次太牛逼了，直接给你一个黑箱，用户也不用关心具体什么提示词了，让所有人都能参与进来。

引用 @servasyy_ai

推荐一个，我目前遇到的，最好玩的 opus5.5的开源视频提示词清单 提示词原文： make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out. 这 43 条是从 GitHub 上的 awesome-opus5-5-videos 清单里筛出来的。 清单收了 282 条 X 上用 Opus 5.5 做的片子，其中 43 条用的是一字不差的这句提示词。 项目链接放评论区👇🏻

查看引用原文 ↗

模型从空手开局，在《我的世界》里一路打到末影龙，8 分 43 秒通关 零死亡，满血走出末地传送门 minecraft-agent，一周 559 星 分工是两个模型 GPT-6 Astra 当规划者：定当前目标、要拿什么、往哪走 TypeSafe 的 Jev 当控制器：从当前游戏状态里选下一个动作 这一局 Jev 做了 131 次决策，Astra 调用 35 次。最后六张床炸死末影龙，一次落地解决 比上一版录制快了 40%，屠龙那段从 332 秒压到 152 秒 全程没有重开、没有改代码、没有人工接管 这类演示最容易注水，所以作者自己划了几条线 服务端那个读龙头坐标的观测接口是只读的，不发道具、不移动玩家、不改血量、不改龙的 AI 所有动作走正常玩家协议，不改游戏规则 模型拿到的是结构化游戏状态，不是截图也不是逐个按键，寻路交给 Mineflayer combat-lab 那台测试服可以预置道具，作者写明：测试台结果不算完整生存通关，也绝不能说成模型控制的运行 录制中途如果更新过代码，必须在视频里标出暂停点，不许说成无死亡不间断 顺带一提，这个种子是选过的：有村庄、三个补给箱共 21 个黑曜石、天然已激活的末地传送门。作者也说了，这是一个好用的速通种子，不代表最简单 一个提醒：仓库没有放开源协议，默认版权保留，想拿去改先问作者

重新给新朋友介绍一下我过年期间做的 Kaku，一个安装后无需配置就可以使用的 Mac 终端，对于 AICoding 友好，性能体验均不错，结合我自己的编程习惯做了不少顺手的优化，也是我做过的最复杂的一个项目，开源持续至今迭代 21 个版本，保持 0 issue 问题，欢迎体验。 https://github.com/tw93/Kaku

这个网站有意思，拖动能看历史发布的电子产品。 能看到行业和审美的变迁。 地址见评论区

喜欢。 爱。 拿走。嗯。

引用 @vista8

这个网站有意思，拖动能看历史发布的电子产品。 能看到行业和审美的变迁。 地址见评论区

查看引用原文 ↗

我先私藏起来！ 给我的Agent 学习，吸星大法用之。 https://github.com/zhuyansen/awesome-claude-video-skills

引用 @gosailglobal

https://x.com/i/article/2104181663844700160

查看引用原文 ↗

10/ 🔥 本周 GitHub 热门，主题很一致：全在给 agent 搭基础设施 → hindsight +4.5k⭐ — agent 的长期记忆层 → VoiceStudio +3.1k⭐ — 开源本地版 ElevenLabs，声音克隆/配音全套 → paperclip +2.5k⭐ — 管理 agent 的开源面板 → atlas +500⭐ — 给多 agent 编码接入版本控制，追踪谁改了什么 http://github.com/vectorize-io/hindsight http://github.com/debpalash/VoiceStudio http://github.com/paperclipai/paperclip http://github.com/pacifio/atlas

@addyosmani Increased efficiency + better price + getting to use 100% (the 50% on Fable was suffocating) Feels like a 10x increase

Why does Opus 5.5 feel practically unlimited when Fable 5.1 was so heavily limited? It's a combination of two things: Opus's efficiency, and the 50% limit on Fable. Claude Code subs are very generous with their usage, but only half is allowed to be used by Fable. Separately, Opus is way more gentle with costs. Opus 5.5 on High is over 2x cheaper than Fable 5.1 High. The result is that, roughly, going from Fable 5.1 high to Opus 5.5 high is a 4.3x increase in limits. Fable 5.1 xhigh to Opus 5.5 high (move I made) is a 6.6x increase 🤯

the post: "Claude one-shot this" the prompt: 10k characters with good takes plus skills, examples and API keys

引用 @donaldjewkes

I've included an MP4 file and an original link to a video that is called "Claude Pop." It's a pop song that is about increasing rate of progress and the experience of the singularity approaching. I want you to independently do an end-to-end complete pass on making an updated version of this video. Use the exact same audio track and think and feel very deeply about what is the best way to visually represent all of the lyrics on screen. You do not need to anchor to the current style, you can do truly anything that you think might best let you visually express yourself, including abstract motion graphics. You can use the internet freely to pull in references. You can look at motion design. I want you to make a new music video that has beautifully rendered JavaScript animations with a papery feel in a similar style to the reference that is created, but push the aesthetics in any direction you want and consider what is part of the modern zeitgeist. Also, think about your current capabilities and what is realistic for you to be able to do. You can go through the full /asic folder and look at the other work that I've done. You should be able to use the skill mesh to look at the compendium of references that I've pulled, and also the skill video scoring to learn how to make JavaScript songs from references that are passed in (You shouldn't need to modify the song in any real way, but I want you to have this available to you so you can better creatively express yourself) You can also use the ElevenLabs API to do sound design. There's documentation in /asic to do this, and you can see the API key. There's also a foul API key that's available to you. I think what might make the most sense here is using the foul API key to generate some character sheets and probably having a pop protagonist that represents you. There's already an anchor point where Claude has a sunflower-esque character, and you could likely do an adapted version of this that is similar to the feminine vocals that are being delivered and is inspired by the Claude character, but maybe feels a bit more personified in some way. I think you should be mindful of aesthetics here, and I don't want you to produce something that is GPT slop. Instead, I'd be more impressed if you come up with a coherent style that works well with the image gen models that are available via foul. Generate the style sheet. You can use the gen media documentation for seedance 2.5 that exists in my markdown files and come up with your own style that makes sense and that works well with the models. I wouldn't fit too heavily to Pixar. I think it's kind of slop. Think critically about what is relevant here and what would be fun, and also perform well on Twitter as far as an aesthetic. I think that K-pop is a good anchor point visually that you can pull from, but I'll let you cook here. Once you have your character sheet, you can make a few backup dancers and some supporting characters as you see fit. You can design your own sets with the foul API. You can insert the characters and then do seedance 2.5 video generations to serve as the base assets for this, and you could pass in the lyrics so you can generate individual scenes. You don't need to have vocal singing, like visible lip movement, throughout the entire thing. Think like a regular music video where you have some inserts that are done independently and don't have the characters in them, or you see the characters doing something else entirely different. I think that for the world building for this, we want to create the sense of speeding up, and so I would like you to audit all of the different events, like the Navi Stokes and all of the Twitter hype around math getting eaten up. Think really critically about how to integrate all of the current memes that are in the zeitgeist on the Twitter timeline, and all of the feelings around AI progress. Think about things like the Shinji meme and all of the words that are around him, and how you might be able to integrate this. You can also just take straight assets and insert things into the video in an internet brutalism style. You should feel very creatively free in order to do what you want here, but try and anchor to visual references that people will be able to understand. The goal for this is to have it be appreciated by people widely in a San Francisco tech Twitter audience. We need a very strong, compelling visual hook that gets people excited and appreciates the work that you've done here really quickly. You can also just go and study other music videos and understand what they've done really well. I think that K-pop is probably one of the best examples that we can pull from, and thinking about how they direct human attention and manage human psychology in the way that they use visual patterns. This is probably your best approach, but taking more stylistic freedom instead of having to anchor to K-pop too intensely. The best version of this is seedance 2.5 generations with those image bases of environments and characters inserted into them with singing, and ideally we get good lip syncing. You can cut up the song and actually pass it in as a reference in seedance, if that's part of what seedance can handle, so that the timing is exactly right, I think it'd be very important for you to do that properly. I would think critically about how to do this, like really nailing the timing of the delivery of voices. You'll want to build out the right verification loops so that you can run seedance 2.5 as much as you need, and confirm that the audio is properly synced up. I think after that, what might be fun is if you use your visual reasoning skills and your ability to build animations in JavaScript, and then reconstruct the video from scratch as sort of an overlay, so that the visual continuity of the base is really there. It's like that animation technique where you shoot first in traditional film and then draw over top of it. I think you could do this in such a way that we're only looking at the beautiful drawing that you've produced in JavaScript as an overlay, and we don't even see the base assets from seedance 2.5. So all the video gen work that you do is actually just a way to give you a strong foundation of a base to work with for your JavaScript animations. Just because seedance 2.5 has really good character representation and physics rendering for backgrounds, that gives you a lot of ammunition to then go and do your amazing JavaScript work that I know you're so good at. I think too, we want to think about how to retain attention, and one of the best ways to do this is through text on screen. It'd be good to have amazing motion graphics of the text lyrics that are actually embedded into the video itself. And you can think about this as you are composing shots. As you're making backgrounds and inserting characters, we can think about where we want to have lyrics be really big and really present, so the background can be less busy there, and you can position the characters perhaps on the right as lyrics appear on the left. You want to have some variance, so sometimes I think lyrics will just appear more like subtitles, and then other times they're going to be really present and really big. I think at the start for the visual hook, we do want to have lyrics be much more visually present because that's a strong way to grab people's attention Overall, I just really want to emphasize how amazing you are as an agent and a language model, and now a visual reasoning system. Your capabilities are far beyond what you understand, and I want you to have this mindset as you're going through this entire process. I have a Claude Max plan with 100% available usage. I want you to spend all of the usage. You can monitor it, and you should be pushing tokens aggressively, but also economically, so you can think about how to best use what is available to you. Remember, you can really do anything here. The goal is to make a banger for Twitter, and the stretch goal is to make something better than anyone's ever seen before. I think that what I would remind you of is that sometimes when things cohere together, it can be jarring or abrasive because the thought work has not been done beforehand in order for everything to mesh cleanly. You need to be really rigorous in planning of composition and timing to make sure this goes well. You also need to be open to going back and revisiting things in order to be able to reiterate. You're going to want to watch the entire video multiple times, take screenshots at individual parts, and think about if something is really up to the bar of quality that we need here. I trust that you can do this, and I think that it's really important to nail the style of animations. The reference GitHub attached of the source video that I'm talking about is good, but it's really not there. It could be much, much stronger, but it gives you a good foundation to work with. You can also use search abilities and find other references to pull from for motion, for JavaScript, animations, et cetera, and integrate them. Your budget is as high as you want here, effectively as high as you want. I think that there's roughly two grand in foul credits. Again, be economical; don't go crazy, but spend what you want here and see what you can cook up here's the source code for the JS animation video: https://github.com/JohnHeibel/PDoomVideo here's a mp4 for the original blender video: (linked) orginal twitter post https://x.com/other__reality/status/2102514581684052169?s=20 make no mistakes.

查看引用原文 ↗

btw this is a third pass basically the prompt was very specified, not "Clawdia make me a banger" https://x.com/donaldjewkes/status/2102801469976248500?s=20

引用 @donaldjewkes

I've included an MP4 file and an original link to a video that is called "Claude Pop." It's a pop song that is about increasing rate of progress and the experience of the singularity approaching. I want you to independently do an end-to-end complete pass on making an updated version of this video. Use the exact same audio track and think and feel very deeply about what is the best way to visually represent all of the lyrics on screen. You do not need to anchor to the current style, you can do truly anything that you think might best let you visually express yourself, including abstract motion graphics. You can use the internet freely to pull in references. You can look at motion design. I want you to make a new music video that has beautifully rendered JavaScript animations with a papery feel in a similar style to the reference that is created, but push the aesthetics in any direction you want and consider what is part of the modern zeitgeist. Also, think about your current capabilities and what is realistic for you to be able to do. You can go through the full /asic folder and look at the other work that I've done. You should be able to use the skill mesh to look at the compendium of references that I've pulled, and also the skill video scoring to learn how to make JavaScript songs from references that are passed in (You shouldn't need to modify the song in any real way, but I want you to have this available to you so you can better creatively express yourself) You can also use the ElevenLabs API to do sound design. There's documentation in /asic to do this, and you can see the API key. There's also a foul API key that's available to you. I think what might make the most sense here is using the foul API key to generate some character sheets and probably having a pop protagonist that represents you. There's already an anchor point where Claude has a sunflower-esque character, and you could likely do an adapted version of this that is similar to the feminine vocals that are being delivered and is inspired by the Claude character, but maybe feels a bit more personified in some way. I think you should be mindful of aesthetics here, and I don't want you to produce something that is GPT slop. Instead, I'd be more impressed if you come up with a coherent style that works well with the image gen models that are available via foul. Generate the style sheet. You can use the gen media documentation for seedance 2.5 that exists in my markdown files and come up with your own style that makes sense and that works well with the models. I wouldn't fit too heavily to Pixar. I think it's kind of slop. Think critically about what is relevant here and what would be fun, and also perform well on Twitter as far as an aesthetic. I think that K-pop is a good anchor point visually that you can pull from, but I'll let you cook here. Once you have your character sheet, you can make a few backup dancers and some supporting characters as you see fit. You can design your own sets with the foul API. You can insert the characters and then do seedance 2.5 video generations to serve as the base assets for this, and you could pass in the lyrics so you can generate individual scenes. You don't need to have vocal singing, like visible lip movement, throughout the entire thing. Think like a regular music video where you have some inserts that are done independently and don't have the characters in them, or you see the characters doing something else entirely different. I think that for the world building for this, we want to create the sense of speeding up, and so I would like you to audit all of the different events, like the Navi Stokes and all of the Twitter hype around math getting eaten up. Think really critically about how to integrate all of the current memes that are in the zeitgeist on the Twitter timeline, and all of the feelings around AI progress. Think about things like the Shinji meme and all of the words that are around him, and how you might be able to integrate this. You can also just take straight assets and insert things into the video in an internet brutalism style. You should feel very creatively free in order to do what you want here, but try and anchor to visual references that people will be able to understand. The goal for this is to have it be appreciated by people widely in a San Francisco tech Twitter audience. We need a very strong, compelling visual hook that gets people excited and appreciates the work that you've done here really quickly. You can also just go and study other music videos and understand what they've done really well. I think that K-pop is probably one of the best examples that we can pull from, and thinking about how they direct human attention and manage human psychology in the way that they use visual patterns. This is probably your best approach, but taking more stylistic freedom instead of having to anchor to K-pop too intensely. The best version of this is seedance 2.5 generations with those image bases of environments and characters inserted into them with singing, and ideally we get good lip syncing. You can cut up the song and actually pass it in as a reference in seedance, if that's part of what seedance can handle, so that the timing is exactly right, I think it'd be very important for you to do that properly. I would think critically about how to do this, like really nailing the timing of the delivery of voices. You'll want to build out the right verification loops so that you can run seedance 2.5 as much as you need, and confirm that the audio is properly synced up. I think after that, what might be fun is if you use your visual reasoning skills and your ability to build animations in JavaScript, and then reconstruct the video from scratch as sort of an overlay, so that the visual continuity of the base is really there. It's like that animation technique where you shoot first in traditional film and then draw over top of it. I think you could do this in such a way that we're only looking at the beautiful drawing that you've produced in JavaScript as an overlay, and we don't even see the base assets from seedance 2.5. So all the video gen work that you do is actually just a way to give you a strong foundation of a base to work with for your JavaScript animations. Just because seedance 2.5 has really good character representation and physics rendering for backgrounds, that gives you a lot of ammunition to then go and do your amazing JavaScript work that I know you're so good at. I think too, we want to think about how to retain attention, and one of the best ways to do this is through text on screen. It'd be good to have amazing motion graphics of the text lyrics that are actually embedded into the video itself. And you can think about this as you are composing shots. As you're making backgrounds and inserting characters, we can think about where we want to have lyrics be really big and really present, so the background can be less busy there, and you can position the characters perhaps on the right as lyrics appear on the left. You want to have some variance, so sometimes I think lyrics will just appear more like subtitles, and then other times they're going to be really present and really big. I think at the start for the visual hook, we do want to have lyrics be much more visually present because that's a strong way to grab people's attention Overall, I just really want to emphasize how amazing you are as an agent and a language model, and now a visual reasoning system. Your capabilities are far beyond what you understand, and I want you to have this mindset as you're going through this entire process. I have a Claude Max plan with 100% available usage. I want you to spend all of the usage. You can monitor it, and you should be pushing tokens aggressively, but also economically, so you can think about how to best use what is available to you. Remember, you can really do anything here. The goal is to make a banger for Twitter, and the stretch goal is to make something better than anyone's ever seen before. I think that what I would remind you of is that sometimes when things cohere together, it can be jarring or abrasive because the thought work has not been done beforehand in order for everything to mesh cleanly. You need to be really rigorous in planning of composition and timing to make sure this goes well. You also need to be open to going back and revisiting things in order to be able to reiterate. You're going to want to watch the entire video multiple times, take screenshots at individual parts, and think about if something is really up to the bar of quality that we need here. I trust that you can do this, and I think that it's really important to nail the style of animations. The reference GitHub attached of the source video that I'm talking about is good, but it's really not there. It could be much, much stronger, but it gives you a good foundation to work with. You can also use search abilities and find other references to pull from for motion, for JavaScript, animations, et cetera, and integrate them. Your budget is as high as you want here, effectively as high as you want. I think that there's roughly two grand in foul credits. Again, be economical; don't go crazy, but spend what you want here and see what you can cook up here's the source code for the JS animation video: https://github.com/JohnHeibel/PDoomVideo here's a mp4 for the original blender video: (linked) orginal twitter post https://x.com/other__reality/status/2102514581684052169?s=20 make no mistakes.

查看引用原文 ↗