Sabitlenmiş Tweet
AINoSleep
396 posts


Worth noting the Glasswing context: 10K+ high-severity vulns found at scale isn't just a benchmark flex — it implies Mythos is being trained/evaluated on real-world adversarial code paths, not just synthetic SWE-bench style tasks. That's a meaningful shift from "better coder" to "better auditor." Curious if the Claude Code integration will expose the security-analysis mode or keep it gated.
English

Anthropic may soon integrate its powerful Claude Mythos model into Claude Code. Recent code references and a briefly visible public toggle suggest the integration is close.
Reports say it outperforms Opus 4.7 in complex programming tasks, autonomous workflows, and security analysis.
Anthropic currently limits Mythos to internal use and the Glasswing project. In that project, the model identified more than 10,000 high-severity vulnerabilities in open-source software.
Before wider release, Anthropic is focusing heavily on safety measures and guardrails. If integrated into Claude Code, Mythos could significantly change how developers and security teams use AI for coding, debugging, and vulnerability research.


English

@Jaden_riku 身边确实有这种感觉,Claude Code 的群里最近安静了不少,反而 Codex 的讨论越来越热。不知道是真的迁移,还是新鲜感驱动。
中文

@bozhou_ai 246 个 Skills、61 个 Agent,规模很impressive,但也好奇,对于独立开发者来说,维护这套系统本身会不会成为新的负担?或者它的设计思路是按需取用而不是全量部署?
中文

Anthropic 黑客松冠军,把自己的 AI 编程工作台开源了。
项目叫 ECC。作者 Affaan Mustafa 和队友用 Claude Code 8 小时做出 zenith.chat 拿了冠军,后来把背后的工作流整理成这个仓库。
它不是一个提示词合集,里面包含 61 个 Agent、246 个 Skills、76 个预设命令,还带规则、Hook、安全扫描和 MCP 配置。
如果你每天重度使用 Claude Code、Cursor、Codex,这个项目值得翻一遍。
它展示的不是怎么问 AI,而是怎么给 AI 搭一套能持续干活的工作台。
github.com/affaan-m/ECC
中文

@hasantoxr Curious whether the desktop layer adds meaningful overhead for power users who already live in the terminal. The risk with "making agents normal-people-friendly" is sometimes you trade flexibility for polish — would love to know if advanced config is still fully exposed.
English

@Khazix0918 有个角度值得想一下,团队Skills共享做起来容易,但"谁来维护质量"才是更难的问题。人多了之后,Skills库里堆了一堆没人更新的旧版本,可能比没有共享更混乱。权限管理只是第一步,怎么让库里的东西保持干净好用,可能才是这类产品真正难做的地方。
中文

藏师傅的小红、小绿书组图生成器 Skill 配合 Codex 太好用了!
我给他一张三张拼在一起的图,它自动加进去的时候,我并没有指定位置,它就能自动根据图片内容匹配对应的页数。
而且虽然这三张图是拼在一起的,它在下面还用分开的格子进行了标注,内容都能对上,太省心了。


歸藏(guizang.ai)@op7418
不吵架了,预告一下下一个 Skill 项目 基于 PPT Skill 可以帮你一键生成微信公众号和小红书封面 也可以基于文档生成小绿书和小红书图文所需的 3:4 组图 会自动处理你的截图素材,自动进行匹配
中文

@goodhunt @datawhale2018 @Hangdian That specific ache of wanting to hear every single person's story but time running out — that's how you know the room was genuinely special. Some events you're relieved when they end. Not those ones.
English

Meeting the whale brothers of Datawhale in person in Hangzhou was an experience I'll never forget.
Meeting actual users/day-1 whale bros, receiving exceptionally detailed questions, and discovering so many of the attendees were teachers - it still makes me a bit sad that my busy schedule cut this portion short. I wanted to take a picture with each individual and hear their story.
It gives me so much hope for a future where open source goes beyond technology: a philosophy of sharing applied knowledge freely.
The diversity of perspectives, time zones, languages - they become an asset to not only those involved but also all of society.
This, to me, is the power and beauty of open source. An elegant, inherently artistic, learning-centric movement - one beyond the constraints of "a person having ordinary skill in the art."



English

很多人买了 VPS,最后只干了一件事:吃灰。
服务器买完,域名不会绑,Docker 不会装,数据库不敢配,环境变量一堆英文看得头皮发麻。
但现在这件事可以换个玩法:
让 Codex 直接帮你把 VPS 变成一个能反复部署项目的赚钱底座。
工具链很简单:
Codex + Coolify
它能在你自己的服务器上部署网站、数据库、全栈应用,还有 280+ 一键服务。
以前你帮人做个网站,还要解释服务器、环境、数据库。
现在你可以把它变成一套可复制交付流程。
客户要的是上线,不是听你讲技术。
GitHub:github.com/coollabsio/coo…

中文

@bozhou_ai 想追问一个场景:如果是需要长期记忆 + 多 Agent 协作 + 工具调用都要的复杂产品,你最终是把 LangGraph 和 Claude SDK 组合起来用,还是只用一个?这块的边界一直没想清楚。
中文

现在 Agent 框架非常的多,我把几个主流的都用过了,说一下我的经验,可以让你在做技术选型时少走弯路。
1. LangChain。教程最多,第一次写 Agent 跟着能跑通。但出了问题,报错指向的是它内部的某一层,不是你写的代码,调试很费劲。适合先跑通再说,不适合认真做产品。
2. LangGraph。我用它处理那种需要循环判断的逻辑,失败了重试,或者 A 搜完交给 B 整理,B整理完交给 C 发出去。这种多步接力的场景,LangGraph 是目前最清晰的选择,当然还有就是目前也没有合适的替代产品。
3. Claude SDK。Anthropic 官方的,原生支持流式输出、提示词缓存、工具调用,这些在其他框架里要自己搭,这里开箱即用。没有多余封装,是我目前开发 Agent 最快的方式。
4. Pydantic AI。FastAPI 背后那个团队做的,风格一脉相承,核心卖点就是简单。用 Pydantic约束 LLM 的输出,格式不对自动重试,不用自己写校验逻辑。
选框架是选和你要解决的问题最匹配的。

中文












