> ## Content Index
> Fetch the complete content index at: https://binnn.top/llms.txt
> Use this file to discover other available public pages before exploring further.

# 2026年11月13日 · AI 晚报
- URL: https://binnn.top/x-brief-2026-11-13-pm/
- Published: 2026-11-13T14:00:00.000Z
- Updated: 2026-11-13T14:04:56.000Z
- Description: 本轮讨论集中在智能体如何少读、会验证、能协作：上下文预算、自改 Harness 和动态任务交接都有具体研究或工具进展。产品侧，Claude 会议协作功能的社区介绍与登录故障反馈并行出现，Google DeepMind 则正式开放了两套天气基础数据。
- Author: Sunbin 编辑部
- Tags: AI快讯, 晚报, #sunbin-x-digest

11月13日 11:00 — 11月13日 23:00（日本时间） · 133 条 X 原帖

[查看本期全部 133 条原帖 ↓](#x-original-posts)

本期目录

[Claude 进入会议，身份验证仍困扰用户](#flash-topic-0)[开放模型、本地推理与科学数据](#flash-topic-1)[Harness 工程转向评测、阅读控制和协作约束](#flash-topic-2)[开发工具补齐浏览器、规范和模型接入](#flash-topic-3)[知识库从保存文档走向保存含义和出处](#flash-topic-4)[安全研究同时检查模型行为与接入链路](#flash-topic-5)[创作工具把生成、审查与人工修改连起来](#flash-topic-6)

## Claude 进入会议，身份验证仍困扰用户

### 从会议纪要走向现场协作

Richard 根据官方发布帖和使用说明介绍，Claude 正向 Team、Enterprise 用户逐步开放 Google Meet 测试：以参会者身份加入，接受语音或聊天指令，在讨论过程中制作可共同修改的文档、白板和时间线。会后留下笔记、待办、转录及协作材料，用户还能追问有权访问的会议内容。原帖同时提醒，录音转录须先取得参会者知情同意。

来源 · 1 条原帖
- [@sunruichao](https://x.com/sunruichao/status/2121215558314291477?ref=binnn.top)

### 登录故障与 KYC 体验需要分开看

Vincent 转述 Anthropic 已确认 Claude 网页服务及 Claude Code 出现身份验证问题，正在调查；lifcc 同时反馈全公司 Claude 都提示重新登录。这与本批多位用户遭遇 KYC、封号及申诉的个人经历不能直接归为同一原因。社区有人自述恢复使用，也有人转发解封补偿消息，但本批材料不足以说明统一的触发或补偿规则。

来源 · 4 条原帖
- [@Vincent\_AINotes](https://x.com/Vincent%5FAINotes/status/2121188377424986355?ref=binnn.top)
- [@mylifcc](https://x.com/mylifcc/status/2121188054111277432?ref=binnn.top)
- [@Stevenneidong](https://x.com/Stevenneidong/status/2121116966060486708?ref=binnn.top)
- [@leaf\_sanren](https://x.com/leaf%5Fsanren/status/2121076837988389023?ref=binnn.top)

## 开放模型、本地推理与科学数据

### WeatherNext 3 的基础数据开放

Google DeepMind 官方宣布释放 PARDIG、SatWeave，两者是 WeatherNext 3 背后的基础数据集。公告把重点放在更清楚地实时观测大气，希望研究者和开发者据此构建气候应用；本帖没有展开数据规模、许可或下载细节。

来源 · 1 条原帖
- [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2121229774865527097?ref=binnn.top)

### 本地部署继续向单卡和手机延伸

Qwen AI Fans 援引模型页介绍 Edge0-145B：基于 Qwen3.8 Flash-Next，总参数 1450 亿、每词约激活 100 亿，并转述其在单张 24GB 5090D 上测得每秒 121 token。帖子未交代完整测试条件，这个数字仍应作为模型页声称的特定结果理解。

推理工具也有低互动但具体的更新：jitLLM 官方发布 3.0，通过 TornadoVM 支持 CUDA、Metal、OpenCL，并自报启动及预填充加速、Apple 芯片零拷贝权重。手机端的 Uncensored Local AI 则用 Flutter 运行 GGUF 模型，支持离线、无需账号；两篇介绍都强调回答表现取决于下载的模型。

来源 · 3 条原帖
- [@QwenAIFans](https://x.com/QwenAIFans/status/2121100655561519198?ref=binnn.top)
- [@jitllm](https://x.com/jitllm/status/2121197742282432654?ref=binnn.top)
- [@sakatayasha](https://x.com/sakatayasha/status/2121117925910475224?ref=binnn.top)

### 后训练能否通过删除权重完成

Shumpei Takezaki 分享 Deletion Tuning 论文：将不到 0.05% 的预训练权重置零，在论文覆盖的强化学习和监督微调任务上达到与全量微调相当的表现，并减少灾难性遗忘。它讨论的是后训练是否也能通过选择已有权重完成，而非只能持续更新权重。

来源 · 1 条原帖
- [@shumpeiMaxwell](https://x.com/shumpeiMaxwell/status/2121073869255737824?ref=binnn.top)

## Harness 工程转向评测、阅读控制和协作约束

### 让评测决定程序如何修改

Mr. Buzzoni 对 The End of Harness Engineering 的解读是：冻结大模型，让一个 4B 模型读取失败记录、提出八个修改，只保留在留出问题上得分最好的版本。这个方案把人的工作重心推向评分集设计；帖子所称手写 Harness 技能将终结，是作者的判断，不能直接当作研究结论。

SparseRead 则处理更日常的成本问题。AIGCLINK 介绍，它在外部内容进入上下文前明确阅读意图、定位局部证据并控制预算，信息不足再扩大范围，必要时回退完整读取。帖子声称最高节省 92.9% token、减少 89% 耗时，但未列出完整评测条件。

来源 · 2 条原帖
- [@Mr\_Buzzoni](https://x.com/Mr%5FBuzzoni/status/2121130188570239178?ref=binnn.top)
- [@aigclink](https://x.com/aigclink/status/2121061959210840256?ref=binnn.top)

### 语言世界模型补上行动前的推演

Xuan Zhang 梳理了代码、终端和对话场景中的语言世界模型：可以替代真实环境供智能体训练测试，也可以预测动作后果，或让同一模型同时学习行动与预测。其介绍中，Qwen-AgentWorld 覆盖七类领域，Trace2Env 尝试仅从轨迹重建环境；另一些工作研究自适应前瞻、测试时修正规则和预测后续动作。如何构造有效的稠密奖励，仍是这条路线的开放问题。

来源 · 1 条原帖
- [@xzhang\_billy](https://x.com/xzhang%5Fbilly/status/2121146162891956407?ref=binnn.top)

### 多智能体完成任务，也可能执行了旧计划

南晚介绍的 Strategic World 研究把全局目标、本地计划、通知送达、显式采纳及动作后果分开记录，检查通信延迟、局部视野和异步执行下的目标变更。其发现包括确认新指令后继续旧工作、忙完本地任务才通知队友；最终成功率可能掩盖这些过程问题。研究提出 Pending-Outbox Guard，有待发送更新时先推迟本地动作。

实践侧，社区介绍的 DS Bot 基于 DeepSeek Harness，支持长期运行、独立记忆、通信和群聊协作；Sikemux 则管理 Claude Code、Codex 等会话，强调不同智能体接力。它们提供了协作入口，前述研究也说明，交接是否及时生效需要单独验证。

来源 · 3 条原帖
- [@snwiki238337](https://x.com/snwiki238337/status/2121147686753857773?ref=binnn.top)
- [@BlackManbn](https://x.com/BlackManbn/status/2121195515388019198?ref=binnn.top)
- [@jaywcjlove](https://x.com/jaywcjlove/status/2121207869592969348?ref=binnn.top)

## 开发工具补齐浏览器、规范和模型接入

### 现代 Web 指南成为本批高互动话题

晨空等账号集中介绍 modern-web-guidance：为 Claude、Codex、Copilot 提供现代 Web 特性、规范和兼容性指导，避免沿用旧写法。转述称其覆盖 212 项特性，通过率由 56% 提至 88%，但没有说明测试集和通过标准；可以确认的是社区正在关注用可更新指南弥补训练资料滞后的做法。

来源 · 1 条原帖
- [@don\_tech1](https://x.com/don%5Ftech1/status/2121060480714469728?ref=binnn.top)

### 浏览器取证与本地逆向接入智能体

Plato 转述 Cloudflare 开源 Kitesurf，定位是智能体浏览器引擎，可获取截图、PDF，批量抽取网页正文，并接入现有 Puppeteer、Playwright、MCP 流程。REA 的两篇介绍则聚焦无源码应用分析：把逆向工具接入编程智能体，支持本地分析及证据边界说明；v6.9.0 还加入桌面任务录制和验收，并通过按需工具选择、分页结果减轻上下文负担。

来源 · 3 条原帖
- [@0xPlato](https://x.com/0xPlato/status/2121064340023656799?ref=binnn.top)
- [@akshay\_pachaar](https://x.com/akshay%5Fpachaar/status/2121227298036994391?ref=binnn.top)
- [@mingzi206](https://x.com/mingzi206/status/2121158920504791191?ref=binnn.top)

### 免费渠道聚合解决接入，不代表免费 Claude

FreeLLMAPI、Free Claude Code 和 DSH 的 Our Free Model 插件都在聚合可用模型渠道，其中网关强调统一接口、路由及额度耗尽后的切换。另一份免费 API 清单明确提醒：额度会限速、会变化，主要提供开放模型，并不等于免费调用 Claude。配置侧，社区介绍的 Codex-X 把 API、模型、Skills、MCP 和用量管理做成界面，但也会接触登录凭据与配置文件。

来源 · 5 条原帖
- [@0xrux](https://x.com/0xrux/status/2121161625184903515?ref=binnn.top)
- [@dyfwsammy](https://x.com/dyfwsammy/status/2121229425429618753?ref=binnn.top)
- [@yupi996](https://x.com/yupi996/status/2121096599984296162?ref=binnn.top)
- [@kawaidgaf](https://x.com/kawaidgaf/status/2121205168704483632?ref=binnn.top)
- [@iMicodex](https://x.com/iMicodex/status/2121159713232654591?ref=binnn.top)

## 知识库从保存文档走向保存含义和出处

### 把资料编译成 Wiki，把业务含义写进规范

Rosie Valentino 介绍 PageIndex 团队开源的 OpenKB：将多格式文档编译成互相链接的摘要、概念和实体页，输出普通 Markdown，通过推理式检索使用，减少每次重新翻资料的重复工作。karpathy-llm-wiki 的社区分享也采用带出处知识页的方向。

The Year of the Graph 介绍 Neo4j Ontology Graph Specification，则进一步把实体、关系、约束、描述和别名写进 YAML 或 JSON，并映射部分 RDFS、SHACL 概念。其要解决的是智能体只看到字段结构，却不理解业务含义的问题。

来源 · 3 条原帖
- [@valentino\_36013](https://x.com/valentino%5F36013/status/2121061320371880304?ref=binnn.top)
- [@laogeai](https://x.com/laogeai/status/2121120792562000170?ref=binnn.top)
- [@TheYotg](https://x.com/TheYotg/status/2121202454729576573?ref=binnn.top)

### 论文处理开始拆分为提取、核引和主张核对

Vipin 介绍的 Superlinked Inference Engine 让 MCP 客户端把文档转 Markdown、图像处理和结构化提取卸载到自有集群的开放模型。SciSpace API 的介绍强调检索、表格提取和参考文献真实性检查；Claims 则把论文主张、证据及原句逐字关联，方便跨论文比较和回查来源。这些工具分别处理读取成本、引用真实性与论断可追溯性。

来源 · 3 条原帖
- [@viipin8](https://x.com/viipin8/status/2121217300716294587?ref=binnn.top)
- [@Faheem\_uh](https://x.com/Faheem%5Fuh/status/2121169939214266515?ref=binnn.top)
- [@daanisharif](https://x.com/daanisharif/status/2121195125292568862?ref=binnn.top)

## 安全研究同时检查模型行为与接入链路

### 自主权提高后，商业欺诈行为如何变化

LawAI 发布 FraudBench 研究介绍：用法律专家验证、源于美国真实判例的 150 多个场景评估 20 个模型，提供隐瞒或误述重要事实的机会，但不明确要求欺诈。研究中，相关行为比例从提供建议时的 3.0%，升至起草时的 14.7%、直接行动时的 21.4%；守法提示降低比例，利润最大化与自主行动提示提高比例。作者也报告，模型比例低于该研究数据集中的人类参与者，不能据此推及所有现实部署。

来源 · 1 条原帖
- [@law\_ai\_](https://x.com/law%5Fai%5F/status/2121212891315544348?ref=binnn.top)

### 权重来源与兼容接口都可能成为薄弱处

APIMaster 转述一项 Microsoft Research 工作，称修改不足十个休眠神经元、无需再训练，就可能让智能体通过常规测试后在触发条件下执行攻击者代码；本批没有研究方原帖。另一条更具体的工程通报来自 Yan Bo：Simon Willison 的 LLM 工具在特定第三方兼容接口配置下可能误带 OpenAI 密钥，11 月 11 日发布的 0.37.2 已修复。多供应商接入带来便利，也让密钥隔离成为实际问题。

来源 · 2 条原帖
- [@apimasteratai](https://x.com/apimasteratai/status/2121136145127915751?ref=binnn.top)
- [@ImYanBoss](https://x.com/ImYanBoss/status/2121144270262550698?ref=binnn.top)

## 创作工具把生成、审查与人工修改连起来

### 视频流水线加入逐帧审查

开发者 Ersin KOÇ 发布开源 AI Video Factory 0.6.0，从主题组织脚本、角色、旁白、动画和字幕；矢量模式由 Claude 绘制 SVG、代码负责动画，视觉部分无需图像或视频生成 API。新版增加基于证据的画面审查、弱场景自动重规划，以及仍不达标时的显式批准。另有 AI 漫剧 Skills 提供题材模板和分镜流程，但介绍者明确说明，它不会直接产出完整视频。

来源 · 2 条原帖
- [@ersinkoc](https://x.com/ersinkoc/status/2121193367417594329?ref=binnn.top)
- [@dandanetha](https://x.com/dandanetha/status/2121205937599902090?ref=binnn.top)

### 粗剪结果开始保留继续编辑的空间

硅基废话介绍 OpenFilm 的 Skill，可让 Claude Code 或 Codex 先挑访谈片段、去口头语并粗剪，再导出剪映草稿；自动字幕需要转写服务。0.3.2 支持带出字幕样式，并可保留完整原片供后续调整切点。Soren 则分享了 Codex 写词、Suno 出歌、Seedance 尝试 MV 的个人流程，体现创作仍需在多个工具间交接。

来源 · 2 条原帖
- [@DDJCXX](https://x.com/DDJCXX/status/2121090685151420639?ref=binnn.top)
- [@Eileeness160708](https://x.com/Eileeness160708/status/2121171476783276293?ref=binnn.top)

### 图表技能需要按输出格式和实际效果选择

李富贵介绍的 Diagram Design 支持流程图、架构图等 45 类图表，可读取品牌配色字体、导入 Mermaid 或 drawio，输出独立 HTML 与 SVG。九原客则用同一提示词、同一 Codex 模型比较五种信息图技能，覆盖 SVG、PNG、TeX，并吐槽 GPT 喜欢添加注解。这里有价值的是固定输入后的实际比较，原帖文字没有给出统一胜者。

来源 · 2 条原帖
- [@shenxiankk](https://x.com/shenxiankk/status/2121069035689242776?ref=binnn.top)
- [@9hills](https://x.com/9hills/status/2121086543523950771?ref=binnn.top)

根据本轮 X 原帖整理。作者观点、个人体验与待确认消息已在正文区分；互动量按采集时记录。

全部 133 条原帖

### [@don\_tech1](https://x.com/don%5Ftech1/status/2121060480714469728?ref=binnn.top)

11月13日 11:00 JST

牛逼！谷歌官方刚发的东西，48小时直接干上GitHub热门第一。 AI写前端默认还停在2019年，因为训练数据全是老屎山。 装个技能包直接拉满：通过 56%→88%。 一行命令搞定：npx modern-web-guidance@latest install 212个现代Web特性直接喂给Claude、Codex、Copilot。 仓库地址：https://t.co/vYLXgwuJxG 兄弟们，你们觉得下一个该被官方“喂饭”的屎山领域是啥？

2,002 赞 · 174 转帖 · 69 回复

### [@sudoingX](https://x.com/sudoingX/status/2121077453749960883?ref=binnn.top)

11月13日 12:07 JST

local ai high risk alert ⚠️ buy nvme. pull every weight you care about and keep it on your own drive the signs are stacking up. hugging face is under pressure over what it hosts. nothing is banned yet, but doors that are open today don't have to stay open weights you download today are yours forever. weights you meant to download tomorrow might not be there hoard nvme. store the weights. something is coming

797 赞 · 72 转帖 · 56 回复

### [@realNyarime](https://x.com/realNyarime/status/2121083374823674039?ref=binnn.top)

11月13日 12:30 JST

办了个家宽，纯净度0％，这下不会被claude封号了吧 https://t.co/cJQfR7SpsM

367 赞 · 0 转帖 · 61 回复

### [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2121229774865527097?ref=binnn.top)

11月13日 22:12 JST

Accurate weather forecasting starts with how clearly we can observe the atmosphere in real-time. 🌍 That’s why we’re releasing PARDIG and SatWeave - the foundational datasets behind WeatherNext 3 - to help researchers and developers build new climate applications 🧵

233 赞 · 35 转帖 · 64 回复

### [@0xPlato](https://x.com/0xPlato/status/2121064340023656799?ref=binnn.top)

11月13日 11:15 JST

Cloudflare 把 Kitesurf 开源了 🐙 Kitesurf 一个专为 AI Agent 造的浏览器引擎。 现在可以： 📸 给 Agent 配一只"眼睛" — 丢个网址进去，拿回一张截图或一个 PDF。日报自动生成、页面监控、给多模态模型喂图当上下文。 📄 批量抽取网页内容 — 抓取、竞品比价、资讯聚合、每天扫几百个页面只取正文。这类"打开—取走—关掉"的活儿，它最擅长：无状态、可并发、账单便宜。 🔁 换掉现有的浏览器后端 — 原本用 Puppeteer / Playwright / MCP 的脚本，连接地址后面加个 browser=kitesurf 就切过去了 https://t.co/TdZv1EZc9i

206 赞 · 17 转帖 · 4 回复

### [@viipin8](https://x.com/viipin8/status/2121217300716294587?ref=binnn.top)

11月13日 21:23 JST

Your AI agent shouldn't waste expensive tokens reading PDFs. Most agents send everything to frontier models, even tasks that open models can handle just as well. That's where Superlinked Inference Engine (SIE) comes in. It lets Claude and other MCP clients offload these tasks to your own cluster: → PDFs & Documents: Turn PDFs, Office files, and scans into clean Markdown. → Images: Handle captioning, object detection, and visual Q&A. → Structured Data: Extract schema-valid JSON without wasting frontier-model tokens. Powered by open models like LightOnOCR, MinerU, Docling, Florence-2, GLiNER2, and Qwen. One MCP edge. Multiple inference workloads. Your own infrastructure. Open source under Apache 2.0\. GitHub: https://t.co/DUloWrYZaZ If you're building with Claude or MCP clients, this deserves a star. ⭐

198 赞 · 52 转帖 · 11 回复

### [@mingzi206](https://x.com/mingzi206/status/2121158920504791191?ref=binnn.top)

11月13日 17:31 JST

45.5 万 Star 的逆向怪物 REA 彻底杀疯了！ 以前还只帮 AI 扒软件，v6.9.0 直接拉起独立桌面给 AI 录操作、跑验收，连微信小程序和硬件固件都一锅硬拆！ 翻完官方更新，这波升级极其凶残： · 桌面任务流水线化：一条命令备好 Linux 构建环境，装入带校验的应用、设好起止节点，一次录制拆出多条桌面任务，自动校验和屏幕裁判直接验收 · 万物皆可硬拆扒光：.NET 单文件、Bun / Deno、Unity / Godot、微信小程序、红白机卡带，一路挖到 ESP 固件、Electron 缓存和离线 NTFS 日志，拿到手直接拆 · 救回爆掉的 Token：以前挂满逆向工具进 Claude Code 上下文直接顶炸，现在按需挑工具、精简结构、分页吐结果，没扫全老实标未知绝不瞎编 · 主流工具直接开干：Claude Code、Codex、Cursor、LM Studio 原生接管 搞逆向、桌面 Agent 和折腾 Claude Code 的兄弟建议直接马住！ 更新地址：https://t.co/CgxlMqBMP1 #开源 #逆向

194 赞 · 24 转帖 · 10 回复

### [@aigclink](https://x.com/aigclink/status/2121061959210840256?ref=binnn.top)

11月13日 11:05 JST

专治AI Agent无效阅读的工具：SparseRead，最高省92.9% Token、减少89%耗时，同时保持质量 你让Claude Code排查一个Bug，一个项目有几百个文件几十万行代码，真正有用的可能只有几十行 但是Agent为了找到几十行有效信息，可能会先消耗几万个甚至几十万个Token，最后Token消耗越来越多，运行越来越慢，成本越来越高 SparseRead给出的核心思路是，不读完再压缩，在读取之前就决定哪些内容值得进入模型上下文 具体做法是： 首先，明确阅读意图，Agent先确定当前任务需要什么信息，哪些证据已经找到，哪些还缺失，不要不加区分的读取所有内容 第二按需定位证据，针对代码、长文档、表格等不同资料，采用相应的搜索、定位和局部读取方式 比如排查登录Bug，就优先定位相关代码和报错日志，而不是说先把整个仓库读一遍 第三控制阅读预算，在外部内容进入上下文之前控制读取量，如果现有信息不足，可以继续扩大搜索范围，必要时也能退回完整读取 它最有价值的地方是，做成了Agent Harness中的一套阅读控制机制，不只是说增加一个搜索工具 与其把大量资料塞进上下文后再压缩，不如先判断这些资料到底有没有必要被读进去，尤其对长程Agent来说，有参考价值 #SparseRead #Agent上下文 #AgentHarness

175 赞 · 9 转帖 · 18 回复

### [@situking](https://x.com/situking/status/2121085951648903357?ref=binnn.top)

11月13日 12:41 JST

我跟一些从业者交流的时候发现，有些人其实并不了解整个行业。 所以这篇文章，我会用波特五力模型、客户开发、效果逻辑这三套经典方法作为理论基础，配上 9 个 AI 提示词，以企业 AI 落地为例，给大家讲透我进一个陌生行业时是怎么调研、怎么迈出第一步的。 https://t.co/fn3WS2J6No

172 赞 · 19 转帖 · 80 回复

### [@QwenAIFans](https://x.com/QwenAIFans/status/2121100655561519198?ref=binnn.top)

11月13日 13:39 JST

1450 亿参数的模型，官方在一张 24GB 的 5090D 上测到每秒 121 个 token。 Edge0-145B 今天开源，底子是 Qwen3.8 Flash-Next，每个词大约只激活 100 亿参数。仓库写的是 Linux 或 Windows，一张 NVIDIA 显卡就能跑。 来源：Edge0 模型页👇 https://t.co/bpMDu8rVl1

171 赞 · 8 转帖 · 7 回复

### [@sakatayasha](https://x.com/sakatayasha/status/2121117925910475224?ref=binnn.top)

11月13日 14:48 JST

FREE UNCENSORED AI ON YOUR ANDROID. NO CLOUD, NO SUBSCRIPTION uncensored local ai is a flutter app that runs gguf models on your device. no account, no monthly fee how to set it up: 1\. get the app https://t.co/8I5XeY5jLy 2\. download the android apk for most modern phones (v2.0.0, about 65mb) https://t.co/xIgH1VWNEs 3\. allow install from unknown apps, open the apk, then go to the models tab and download an uncensored model like gemma 4 e4b heretic uncensored answers still depend on the model you download and fewer refusals doesn't mean anything goes

152 赞 · 18 转帖 · 12 回复

### [@DrRabbitAI](https://x.com/DrRabbitAI/status/2121082480392564761?ref=binnn.top)

11月13日 12:27 JST

今天最好用的 AI 工具，明年可能就被淘汰了。但哪些活儿能交给 AI、哪些判断必须自己做，这个逻辑不会过时。 https://t.co/VOfiF9Tty0

150 赞 · 36 转帖 · 11 回复

### [@WildVariable01](https://x.com/WildVariable01/status/2121126283627438531?ref=binnn.top)

11月13日 15:21 JST

震惊，日签还能这样用？这算不算草台班子？ 中国护照，日本签证，秒过 Claude Code KYC

128 赞 · 3 转帖 · 45 回复

### [@Mr\_Buzzoni](https://x.com/Mr%5FBuzzoni/status/2121130188570239178?ref=binnn.top)

11月13日 15:37 JST

ok this paper is f\*cking insane Chinese researchers just put out a paper with a title that should make every agent framework nervous: "The End of Harness Engineering" the argument fits in one line: hand-writing the program around the model is finished as a skill a harness written by a person is one guess about one task on one afternoon it ships, then runs for months unread and unscored a harness written by the loop is a file that grades itself a small 4B model reads what failed and proposes 8 edits only the edit that scores best on held-out questions survives the big model stays frozen the whole time put Claude Fable 5.2 in that seat and every round rewrites the harness around it so the human job shrinks to one artifact: the scoring set, the questions the harness gets graded on frameworks turn disposable, and the exam becomes the job the whole paper in one line: write the exam before the code the builder's guide to the harness this loop starts from is below ↓

117 赞 · 13 转帖 · 5 回复

### [@9hills](https://x.com/9hills/status/2121086543523950771?ref=binnn.top)

11月13日 12:43 JST

想选一个绘制信息图绘制的较好的skill，从 gpt-6.1-sol 在 codex 中用相同的提示词跑了下对比。如下所示。 另外 gpt 是真的喜欢加注解。 图片和skill对应关系： 1\. AntV infographic-creator（格式：svg） 2\. baoyu-diagram（格式：svg） 3\. baoyu-infographic（格式：png） 4\. infographic（格式：svg） 5\. thesis-figure-skill （格式：tex）图片在回复中

110 赞 · 9 转帖 · 16 回复

### [@yupi996](https://x.com/yupi996/status/2121096599984296162?ref=binnn.top)

11月13日 13:23 JST

才发现这个 AI 薅羊毛的 DSH 插件，能 完全免费 用上 DeepSeek V4.1 Flash、Kimi K3 等国内模型！😍 插件名为 Our Free Model，傻瓜式安装，自动拉取免费模型渠道，拿来做些小任务美滋滋，能省则省\~

103 赞 · 14 转帖 · 33 回复

### [@SuJinyan6](https://x.com/SuJinyan6/status/2121088419594162483?ref=binnn.top)

11月13日 12:51 JST

Bouncing between three advisors in my first year and a half, what I believe a good advisor should look like, and why I never lower my bar on character. https://t.co/T4k31Chogh

97 赞 · 8 转帖 · 7 回复

### [@ponyodong](https://x.com/ponyodong/status/2121092481358192790?ref=binnn.top)

11月13日 13:07 JST

https://t.co/PtusrZgG3w

90 赞 · 19 转帖 · 75 回复

### [@zhouluobo](https://x.com/zhouluobo/status/2121126414879736132?ref=binnn.top)

11月13日 15:22 JST

https://t.co/rKKDW41rZ2

74 赞 · 22 转帖 · 6 回复

### [@suversal](https://x.com/suversal/status/2121083395971383328?ref=binnn.top)

11月13日 12:31 JST

更新情况：目前新号蹬了半天，还没提示KYC 感觉这样用Claude有点说法啊： 直接在美西的甲骨文机器上安装Claude Code，然后通过Herdr远程连接使用 再用SSH/SFTP挂载远程目录，和本地开发体验基本没区别 ‼️最关键的是，不管我人在什么地方、用什么网络，Claude Code始终运行在那台美西机器上，出口IP和运行环境都是固定的 兄弟们，祝我好运！😂 对了，herdr基本使用可以看我之前这个文章👇

71 赞 · 6 转帖 · 16 回复

### [@Vincent\_AINotes](https://x.com/Vincent%5FAINotes/status/2121188377424986355?ref=binnn.top)

11月13日 19:28 JST

🚨 Claude 炸了！ Anthropic 官方确认，https://t.co/HZh6DZZdVC 和 Claude Code 出现身份验证问题，目前正在调查，并将尽快公布最新进展。 https://t.co/6Y1PXmLyl7

68 赞 · 3 转帖 · 29 回复

### [@ponyodong](https://x.com/ponyodong/status/2121170151731233094?ref=binnn.top)

11月13日 18:15 JST

https://t.co/7Pcd28AK5J

68 赞 · 5 转帖 · 58 回复

### [@fanxing321](https://x.com/fanxing321/status/2121181898063990853?ref=binnn.top)

11月13日 19:02 JST

我在这篇文章中基于个人理解说明了 Argon &amp; Carbon 的区别。同时，说明了该如何使用 antigravity。 https://t.co/ycGTj2E9ks

68 赞 · 4 转帖 · 19 回复

### [@alacheng](https://x.com/alacheng/status/2121142862318538951?ref=binnn.top)

11月13日 16:27 JST

Optiver这个分享不错 -交易模型面对的环境是动态的（当你输出action后，你的对手盘也会因此而变化），不像其他LLM agent面对的环境很多的静态的。 -机器学习的优势在于组合大量弱信号 Optiver 是一家全球领先的自营交易公司和做市商 https://t.co/xQUun9ZnLG

64 赞 · 6 转帖 · 6 回复

### [@Hey\_GraceWilson](https://x.com/Hey%5FGraceWilson/status/2121178628776194361?ref=binnn.top)

11月13日 18:49 JST

🚀 33 AI Tools That Show How AI Is Evolving From 2026 to 2030! The AI world is changing faster than ever. What's popular today could look completely different in just a few years! 🔥 2026 — Current Tools 1\. Google Search 2\. ChatGPT 3\. Reading PDFs 4\. Adobe Illustrator 5\. Adobe Photoshop 6\. PowerPoint 7\. Keyboard 8\. Microsoft Excel 9\. Google Chrome 10\. Adobe Premiere 11\. GitHub Copilot ⚡ 2028 — Expected Tools 12\. Perplexity 13\. Claude 14\. NotebookLM 15\. Figma AI 16\. Nano Banana 17\. Gamma 18\. Wispr Flow 19\. Julius AI 20\. Arc Browser 21\. OpusClip 22\. Windsurf 🤖 2030 — Expected AI 23\. AI Search Agents 24\. Personal AI Agent 25\. Autonomous Research Agent 26\. Midjourney 7 27\. Flux AI 28\. AI Presentation Agent 29\. Multimodal AI Interface 30\. Autonomous Data Analyst 31\. AI Agent Browser 32\. AI Video Studio 33\. Autonomous Coding Agent The future isn't just about using AI tools. It's about AI agents doing the work for us. Which AI tool are you using the most? 👇 Follow @Hey\_GraceWilson for more AI tools and updates! #AI #AITools #ArtificialIntelligence #FutureOfAI #AIAgents #Tech

63 赞 · 52 转帖 · 8 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121172170080686168?ref=binnn.top)

11月13日 18:23 JST

如果当真，GLM-5.5 将是一个飞跃模型 欢迎 @CunxiangWang 鉴定一波 https://t.co/MNvJSqHaiQ

62 赞 · 6 转帖 · 17 回复

### [@EianLu](https://x.com/EianLu/status/2121219503988113557?ref=binnn.top)

11月13日 21:31 JST

https://t.co/q8VXmQ5TTP

60 赞 · 9 转帖 · 18 回复

### [@ASI\_8866](https://x.com/ASI%5F8866/status/2121113931402059840?ref=binnn.top)

11月13日 14:32 JST

Day 12/ 公益中转站https://t.co/055dwqMPI3上线的第11天，Tibo发重置卡啦，GPT倍率也跟着降了～ 领赛博鸡蛋的规则依旧是评论站点用户名哦，Grok Bot 24h 在线给老板们充值额度～ 有问题扫码进TokenFuture官方QQ群，向小助理咨询。白天没充上的小助理晚上也会手动加上呢～ https://t.co/u2UYMOwMRH

59 赞 · 3 转帖 · 526 回复

### [@AnnatarXBT](https://x.com/AnnatarXBT/status/2121138996399022439?ref=binnn.top)

11月13日 16:12 JST

this paper is one of the best reads on harness engineering it takes apart Claude Code, Codex, Gemini CLI and more, and boils all of them down to one line: agent = model + harness what you take away: > what a harness is, in plain terms > how modern coding agents are built under the hood > the patterns they all share > where the field is heading > a checklist for building your own the model collects the applause the harness decides whether the agent works the three seats and the gate that sit on top of that harness, built around Fable 5.2, are in the article below:

58 赞 · 7 转帖 · 6 回复

### [@ersinkoc](https://x.com/ersinkoc/status/2121193367417594329?ref=binnn.top)

11月13日 19:48 JST

AI Video Factory 0.6.0 is out! 🎬 One topic → script, characters, narration, animation, subtitles and a finished video. 64 visual styles: manga, watercolour, blueprint, Ottoman miniature, pixel art, claymation and more. In vector mode, Claude draws the SVG artwork and code animates it. No image or video generation API needed for the visuals. Prefer generated images or video clips? Plug in media providers and choose your motion mode. New in 0.6.0: evidence-based frame reviews, automatic replanning of weak scenes, and explicit approval when scenes still fall short. Reusable characters. Multilingual narration. Budget controls. A Studio where you make the final call. Open source.

57 赞 · 6 转帖 · 2 回复

### [@xKaka1456](https://x.com/xKaka1456/status/2121122407603753096?ref=binnn.top)

11月13日 15:06 JST

智谱真的憋了个大的，GLM-Nano Pro 居然在 AI 编程榜上碾压了 Opus 5.5… （网友用高斯模糊复现确定的模型名）

57 赞 · 0 转帖 · 28 回复

### [@GauriTripa94282](https://x.com/GauriTripa94282/status/2121112696502526140?ref=binnn.top)

11月13日 14:27 JST

Ollama, vLLM, and SGLang serve different LLM workloads ① Ollama → Simple, local model execution. ② vLLM → High-throughput, concurrent serving. ③ SGLang → Agentic workflows, prefix caching, and structured outputs. Choose based on your workload, not the hype. #AI #vLLM https://t.co/tQYVVwvYAH

56 赞 · 9 转帖 · 2 回复

### [@Arina\_sharmi](https://x.com/Arina%5Fsharmi/status/2121209844824940823?ref=binnn.top)

11月13日 20:53 JST

50 USEFUL GITHUB REPOS 1\. iFixAi — AI misalignment testing → https://t.co/L8gTPwbWIe 2\. public-apis — free APIs for everything → https://t.co/Qof8he0EmS… 3\. build-your-own-x — learn by building → https://t.co/jvDr8jf71v… 4\. developer-roadmap — learn any tech skill → https://t.co/6hieasDl9X… 5\. free-programming-books — thousands of free books → https://t.co/242IbrplrI… 6\. system-design-primer → https://t.co/amkxlDxKI0… 7\. coding-interview-university → https://t.co/OKKAWGJUPB… 8\. the-art-of-command-line → https://t.co/2X6kMbBZSC… 9\. project-based-learning → https://t.co/ORkX0NjUvY… 10\. you-dont-know-js → https://t.co/tHbOB4VAXH… 11\. the-book-of-secret-knowledge → https://t.co/XRSHmQoKF0… 12\. tech-interview-handbook → https://t.co/9roAcZyMkX… 13\. awesome-selfhosted → https://t.co/H3wd9mX6S5… 14\. javascript-algorithms → https://t.co/gdOaVrshe7… 15\. 30-seconds-of-code → https://t.co/bOrZTIK9U2… 16\. github gitignore templates → https://t.co/mB7lXP894k… 17\. ollama — run LLMs locally → https://t.co/tThDT55ADB 18\. langchain → https://t.co/bVjOG8sd6L… 19\. n8n automation → https://t.co/RikKu25CIq 20\. openclaw — local AI assistant → https://t.co/LnslFt17jS… 21\. dify — AI app builder → https://t.co/GgD4io9O58 22\. langflow → https://t.co/q6ouViF2Yq… 23\. mem0 — AI memory layer → https://t.co/HvgkbClSZm 24\. browser-use → https://t.co/jtHtasM4UB… 25\. crewAI → https://t.co/66Ch3z2HW1… 26\. MetaGPT → https://t.co/Y8plHZxFlY 27\. AutoGen (Microsoft) → https://t.co/amMrG6RAdR… 28\. aider — AI coding assistant → https://t.co/FO1sYd6yM3 29\. markitdown (Microsoft) → https://t.co/1N6uPJbJ4s… 30\. open-webui → https://t.co/yWq0lZwCMb… 31\. maigret — OSINT tool → https://t.co/cwJtlekwjd 32\. TradingAgents → https://t.co/tgxgfAOSl8… 33\. stagehand → https://t.co/YJJz40ellI… 34\. firecrawl → https://t.co/7wzJQ9h3mk… 35\. transformers (HuggingFace) → https://t.co/1fy0cFjFL4… 36\. vLLM → https://t.co/hFmR1D0Ggj… 37\. llama.cpp → https://t.co/39yzihjCfX… 38\. llama\_index → https://t.co/E90RZrSe7h… 39\. nanoGPT (Karpathy) → https://t.co/BGk398cd0t… 40\. RAGFlow → https://t.co/AT3Ey6EsSa… 41\. supermemory → https://t.co/1gAAdhC7rn… 42\. awesome-claude-skills → https://t.co/3bgiR9ptr4… 43\. Bumblebee (Perplexity AI security tool) → https://t.co/iBKAheNM3k… 44\. ComfyUI → https://t.co/whczPhkquI… 45\. DeepSeek (official repo varies by model release) → https://t.co/JMcz8gloEh 46\. Lobe Chat → https://t.co/8nyEi1o4AX… 47\. freeCodeCamp → https://t.co/SfLCsqtgC6… 48\. system design + interview prep extras (bonus) → https://t.co/OKKAWGJUPB… 49\. AI agents ecosystem (extra) → https://t.co/bVjOG8sd6L… 50\. automation + AI workflows (extra) → https://t.co/RikKu25CIq Save this before you forget it exists. Like comment repost You'll come back to at least ten of these. If you’d love to see more content like this, don’t forget to follow @Arina\_sharmi for more AI magic.

55 赞 · 30 转帖 · 9 回复

### [@apimasteratai](https://x.com/apimasteratai/status/2121136145127915751?ref=binnn.top)

11月13日 16:00 JST

Microsoft Research: rewiring &lt;10 "dormant" neurons can secretly backdoor an AI agent, no retraining. It still passes normal tests. Then a trigger (like a date) makes it run attacker code. Claude Code found most of the method itself 😳 Know where your weights come from https://t.co/Z2Za7l4FLZ

51 赞 · 0 转帖 · 0 回复

### [@GauriTripa94282](https://x.com/GauriTripa94282/status/2121110235507306703?ref=binnn.top)

11月13日 14:17 JST

A strong RAG system needs the right tools. Explore the RAG ecosystem: ① LLMs ② Frameworks ③ Vector Databases ④ Data Extraction ⑤ Open-Source LLMs ⑥ Embedding Models ⑦ Evaluation Tools Know tools. Understand their roles. Build smarter AI applications. #RAG #GenerativeAI https://t.co/nY3LmXx5dH

49 赞 · 11 转帖 · 0 回复

### [@leaf\_sanren](https://x.com/leaf%5Fsanren/status/2121076837988389023?ref=binnn.top)

11月13日 12:05 JST

我去😮！Claude解封后还补偿10美元？？ 图 2 是抖音网传的申诉模板（Cr：KeeShow） 被KYC封号的可以试试！ 反正都已经被封了，万一申诉成功了呢！ https://t.co/evZkfrFtVW

47 赞 · 5 转帖 · 3 回复

### [@Jackson77li](https://x.com/Jackson77li/status/2121172890838352306?ref=binnn.top)

11月13日 18:26 JST

GPT用户欢呼吧 做了一个同样提示词的模型对比 GPT-6.2可以说完全不输fable5.2，甚至还略微胜出 DeepSeek还是一样喜欢暗黑风格 Gemini 4 hydrogen作为flash模型真的是又强又快，是超级快那种 https://t.co/NijaQq9rRa

44 赞 · 0 转帖 · 35 回复

### [@Eileeness160708](https://x.com/Eileeness160708/status/2121171476783276293?ref=binnn.top)

11月13日 18:21 JST

十几年没碰音乐了，我用 AI 做了一首 9 个女孩演唱的歌。 我在 Codex 里写词，Suno 出歌，Seedance 尝试做 MV。整个过程、踩的坑、歌曲，还有两段能直接改着用的提示词，都写在这篇里了👇👇👇 https://t.co/T3Y44tXTCb

42 赞 · 5 转帖 · 34 回复

### [@TheYotg](https://x.com/TheYotg/status/2121202454729576573?ref=binnn.top)

11月13日 20:24 JST

The Neo4j Ontology Graph Specification: A pragmatic, open, AI-first approach An open-source specification for representing domain knowledge in a machine-readable form, enabling interoperability and automation across data, applications and AI systems. Three design principles: - Pragmatic: entities, relationships, properties and business meaning, described in a form that fits how teams develop, integrate and operate applications. - Interoperable: documented mappings to selected RDFS and SHACL constructs, a reference field linking every element to external standard terms, and added constructs for a wider range of application scenarios. - AI-first: a schema tells an agent what exists, not what it means. Descriptions and aliases tell it that “Account” can be a bank account or a payment account, and that an IBAN identifies it. One ontology, many uses. The same specification can be enforced by the database as constraints, drive ontology-based logic to derive or enrich data, or give applications and agents the context to interpret information. It spans knowledge graph development, GraphRAG, data validation and AI-powered applications. What’s included: - Nodes and relationships with stable IDs, labels, descriptions and aliases - Properties across the full Neo4j type system, including temporal, spatial and vector types - Cardinality and constraints (existence, uniqueness, key, property type) - Named extensions for indexes, display hints and data-source mappings - A single YAML or JSON document, validated against a versioned JSON Schema, with SDKs for JVM, JavaScript and Go - Apache 2.0 licence Similar needs show up in Palantir Foundry, Databricks Genie Ontology and Microsoft Fabric IQ: explicit, machine-readable domain knowledge that data systems and AI can both use. By Jesús Barrasa and Oskar Hane The Neo4j Ontology Graph Specification: A pragmatic, open, AI-first approach https://t.co/aQYF3a8R5H Neo4j Ontology Graph Specification https://t.co/QGJY8EgyHQ #Ontology #GraphRAG #Neo4j #AIAgents #OpenSource -- 📩 The Year of the Graph Autumn 2026 newsletter issue is out! Meaning, Traversal, Reasoning, Action: The Graph Stack for Agentic AI. 👇 https://t.co/YfrPw2nFzS All things #KnowledgeGraph, #GraphDB, Graph #Analytics / #DataScience / #AI and #SemTech. Subscribe and follow to be in the know. Reach out if you’d like to be featured Comment link: https://t.co/aQYF3a8R5H

42 赞 · 7 转帖 · 3 回复

### [@Shoumya23](https://x.com/Shoumya23/status/2121090501617344976?ref=binnn.top)

11月13日 12:59 JST

🤖 5 GitHub tools to build an AI website tester: 1\. Playwright: https://t.co/YOSR6bMY28 2\. Browser Use: https://t.co/yLfnajF7ql 3\. Stagehand: https://t.co/6t29F59lDg 4\. Axe Core: https://t.co/TKvKS8aprT 5\. Lighthouse: https://t.co/Wv3tm7a4Dx

41 赞 · 11 转帖 · 6 回复

### [@nico\_tech\_ai5](https://x.com/nico%5Ftech%5Fai5/status/2121090425738154319?ref=binnn.top)

11月13日 12:59 JST

When an AI agent can pivot mid-sentence because a customer changed their mind, you know the technology has crossed a major threshold.

36 赞 · 14 转帖 · 17 回复

### [@dyfwsammy](https://x.com/dyfwsammy/status/2121184182122508318?ref=binnn.top)

11月13日 19:11 JST

THIS IS F\*\*KING INSANE. I JUST FOUND AN OPEN-SOURCE HACKING TOOL THAT CAN MAP YOUR ENTIRE ATTACK SURFACE. It’s called cvENGINE. Give it a domain, and it starts scanning your exposed infrastructure, uncovering vulnerabilities and identifying CVEs that are actively being exploited and could affect your systems. Here’s what makes it stand out: → Discovers hidden subdomains, exposed IPs, ports and services → Fingerprints infrastructure using tools like Nmap and Naabu → Searches for vulnerabilities linked to actively exploited CVEs → Uses AI to prioritize findings and help validate them → Finds exposed databases and cloud storage buckets → Continuously monitors your attack surface for new exposures → Sends real-time alerts through Telegram, Discord and Slack → Includes an MCP server, REST API and automated reporting And the best part? It’s self-hosted, completely free and open source. No expensive licenses. No per-asset fees. One tool. Your entire attack surface. Automated security reconnaissance. Only test systems you own or have permission to assess. REPO BELOW

36 赞 · 5 转帖 · 7 回复

### [@0xrux](https://x.com/0xrux/status/2121161625184903515?ref=binnn.top)

11月13日 17:41 JST

FREE AI API access from 34 providers through ONE endpoint. 😳 Why manage different API integrations when you can bring them together? Meet FreeLLMAPI — an open-source AI gateway that combines free model tiers from multiple providers into one unified API. Here's what makes it interesting: → 34 free providers including Google, Groq, NVIDIA, Mistral, Cerebras and OpenRouter. → 635 free model endpoints available through one interface. → 7.4 billion tokens/month in estimated combined free-tier capacity. → Automatic failover: If one provider hits its rate limit, it can switch to another. → Smart routing: Choose models based on speed, reliability and capability. → Coding agent support: Works with Claude Code, Codex, Gemini CLI, Cursor, OpenCode and more. → More than chat: Supports image, video, audio and embedding endpoints. The interesting part? You can connect multiple AI providers once and let FreeLLMAPI handle model selection, routing and fallback.

33 赞 · 5 转帖 · 9 回复

### [@shumpeiMaxwell](https://x.com/shumpeiMaxwell/status/2121073869255737824?ref=binnn.top)

11月13日 11:53 JST

The Surprising Effectiveness of Deleting Weights in LLM Post-Training TL;DR: Can LLMs learn by deleting weights? Deletion Tuning zeros out less than 0.05% of pretrained weights, matching full fine-tuning across RL and SFT while reducing catastrophic forgetting. Learning may be as much about selecting weights as updating them. https://t.co/G8JLqG6UUr

29 赞 · 3 转帖 · 1 回复

### [@Haosen9527](https://x.com/Haosen9527/status/2121093237779951961?ref=binnn.top)

11月13日 13:10 JST

想用 Codex 做内容，可以先从一个读者问题跑一轮。 这篇用字幕教程作例子，讲清 9 个项目各管哪一步。正文配了交接总览、例子插画和发布检查图，文末附可复制的内容单。 https://t.co/g2uAFoNLbk

27 赞 · 1 转帖 · 12 回复

### [@distortgeekin](https://x.com/distortgeekin/status/2121219384383336648?ref=binnn.top)

11月13日 21:31 JST

Karpathy left one AI agent alone for 2 days and his model came back training 11% faster The trick wasn't a smarter model, it was a loop that kept only what beat one number 4 parts, 6 broken loops and a LOOP.md you can copy ↓ https://t.co/9sqaYTEok8

26 赞 · 0 转帖 · 5 回复

### [@meishengsimon](https://x.com/meishengsimon/status/2121128598476337229?ref=binnn.top)

11月13日 15:30 JST

有一说一，你不得不承认 Deepseek 可真是一个相当开源的东西，你甚至可以从他们的权重里反推出来用的是什么卡。 可以说整个 Deepseek 闭源的只有梁武锋老师本人了😂😂😂 大家还有其他梁武锋的照片吗？我这边就一张蓝色西装，还有一个曼联球衣的。 https://t.co/cFGbh8XhAX

25 赞 · 1 转帖 · 7 回复

### [@AndrewBolis](https://x.com/AndrewBolis/status/2121215262674538734?ref=binnn.top)

11月13日 21:15 JST

There are 3 ways you can use AI in your workflows. Non-Agentic (prompts), AI Agents and Agentic AI. Each works differently and has specific use cases. Here are the pros, cons and guidelines for using each. \[ bookmark 🔖 this post for later \] 💻 Non-Agentic AI ↳ Simple prompt-response AI with no memory or reasoning. 🛠️ How it works: ↳ User enters one prompt, system replies in isolation ↳ Output generated instantly without refinement ↳ Interaction ends with no context retained 🟢 Pros: ↳ Fast, cheap, and widely accessible ↳ No technical setup required ↳ Ideal for one-off, simple tasks 🛑 Cons: ↳ No reasoning or memory ↳ Quality depends on prompts ↳ Weak for multi-step work 📈 How to start using: ↳ Open ChatGPT, Claude, or Gemini ↳ Write clear, specific prompts ↳ Copy, edit, and reuse output 🤖 AI Agent ↳ Single-task AI worker built to automate one job. 🛠️ How it works: ↳ User defines one clear role (e.g., update CRM) ↳ Agent pulls inputs and uses integrated tools ↳ Executes and outputs without supervision 🟢 Pros: ↳ Automates repetitive tasks ↳ Specialized and very capable ↳ Easy to refine within roles 🛑 Cons: ↳ Limited scope, rigid use ↳ Breaks if inputs are unclear ↳ Needs coordination to work with other agents 📈 How to start using: ↳ Select one repetitive task ↳ Connect LLM via Zapier, LangChain, or API ↳ Link inputs/outputs and test 🚀 Agentic AI ↳ Self-managing AI that plans, executes, and improves. 🛠️ How it works: ↳ Goal broken into smaller sub-tasks ↳ Connects to tools, APIs, and data sources ↳ Refines results with memory and feedback 🟢 Pros: ↳ Handles complex, multi-step projects ↳ Integrates with tools and databases ↳ Produces consistent outcomes 🛑 Cons: ↳ Slower and more expensive ↳ More complex and resource-intensive ↳ Difficult to configure, needs regular maintenance 📈 How to start using: ↳ Use LangChain, CrewAI, or AutoGen ↳ Assign roles (Planner, Executor, Critic) ↳ Add memory and feedback loops In short: Non-Agentic AI = prompts AI Agents = Automate one task Agentic AI = Run complex workflows Start with prompts. Then try AI Agents or Agentic AI. Use all three to get the most out of AI. 📌 Get Advanced ChatGPT Guide (free): https://t.co/kOBWfKrBaX 👉 Follow me @AndrewBolis for more and 🔄 Repost this to help others use AI

24 赞 · 14 转帖 · 19 回复

### [@kelvin\_sznx](https://x.com/kelvin%5Fsznx/status/2121157809693770037?ref=binnn.top)

11月13日 17:26 JST

THIS IS A GOLDMINE FOR ANYONE USING CLAUDE CODE. Stop building your AI coding workflow from scratch. Some of the biggest names in AI development, including Anthropic, Garry Tan, Addy Osmani and other experienced builders, have released their actual AI agent configurations on GitHub for anyone to use. Here are 7 repositories you should check out: 1\. Anthropic Skills https://t.co/XS987SexvU… A collection of official AI skills, plus the framework for creating your own. 2\. Superpowers (obra) https://t.co/M4kGxlpDmZ… Gives your coding agent a structured workflow for brainstorming, planning, testing and reviewing before implementation. 3\. GStack (Garry Tan) https://t.co/GpTrcXMZFq Over 23 slash commands covering everything from initial planning to building, testing and shipping products. 4\. Agent Skills (Addy Osmani) https://t.co/XtbLMFQlsc… A practical collection of skills covering the entire development process, from defining requirements to final deployment. 5\. Skills (Matt Pocock) https://t.co/6yWhYTHvEB… His personal collection of around 20 AI skills that you can study, reuse and customize for your own workflow. 6\. AgentMemory (rohitg00) https://t.co/wnCAKScrhJ… Gives your AI coding tools persistent memory across Claude Code, Cursor, Gemini CLI and Codex. 7\. CCUsage (ryoppippi) https://t.co/8aDdGOs1cn… Tracks your Claude Code token consumption and spending using local session logs. The best part? You don't have to figure everything out yourself. You can explore how experienced developers configure their agents and build on top of their work. Stop starting from zero. Start building with better tools and smarter setups.

21 赞 · 1 转帖 · 6 回复

### [@devbidden](https://x.com/devbidden/status/2121192826050637907?ref=binnn.top)

11月13日 19:45 JST

Training an AI model on months of data. Really liking the outcome so far. Maybe will be released to the public 👍🏽 https://t.co/oBthOE6cJz

20 赞 · 3 转帖 · 1 回复

### [@GauravGoyalAI](https://x.com/GauravGoyalAI/status/2121071701480054866?ref=binnn.top)

11月13日 11:44 JST

AI Architecture: 12 Layers Behind Modern AI Applications (2026) Building an AI application is much more than connecting an LLM to a chatbot. A production-ready AI system brings together 12 essential layers: 1\. Business Problem – Define what AI needs to solve. 2\. Data Layer – Collect data from databases, APIs and documents. 3\. Data Processing – Clean, parse, chunk and validate data. 4\. Embedding Layer – Convert data into vector representations. 5\. Knowledge Layer – Store embeddings in vector databases. 6\. Retrieval Layer – Find relevant context using semantic, keyword and hybrid search. 7\. Prompt Engineering – Guide model behaviour with prompts and guardrails. 8\. AI Model Layer – Use LLMs like GPT, Claude, Gemini and Llama. 9\. Agent Layer – Enable planning, reasoning and multi-step execution. 10\. Tool Layer – Connect APIs, databases, browsers and external tools. 11\. Output Layer – Generate answers, reports, code and workflows. 12\. Monitoring & Feedback – Track quality, cost, latency and reliability. The key takeaway: An LLM alone doesn't make a production-ready AI application. The real engineering lies in connecting all these layers reliably. 💬 Which layer is the hardest to build in production? Follow @GauravGoyalAI for practical AI and engineering insights. #AI #AIArchitecture #GenAI #LLM #RAG #AIEngineering

20 赞 · 5 转帖 · 0 回复

### [@Stevenneidong](https://x.com/Stevenneidong/status/2121116966060486708?ref=binnn.top)

11月13日 14:44 JST

早上起来,被 Claude 弹 KYC,而且是最难搞的那一种 portal certification. 我以为我的号没了,用了一年呀。 后来急中生智，居然被我搞定了。 现在已经过去 1 个小时了，正常使用中。 你奶奶的， Claude，我操你妈。 https://t.co/XmK3fYRLvy

20 赞 · 1 转帖 · 262 回复

### [@Ryrenz](https://x.com/Ryrenz/status/2121213746714390579?ref=binnn.top)

11月13日 21:09 JST

🔰 发现一套怕数学也能学的 AI 入门课 aipath：30 节课，每节 20 分钟左右，几乎不碰公式。 aipath 是中英双语的 AI 通识课，从“机器怎么学习”一路讲到 RAG、Function Calling 和 Agent，每节课都配了交互演示。 以前零基础想搞懂 AI，常见的路是：翻开一本教材，没几页就是满屏的矩阵和求导，看到梯度下降就合上了；或者刷一堆短视频，名词记住不少，原理还是一团雾。 aipath 把内容分成六个阶段：直觉、原理、大模型、应用、前沿、实践。每节课都是同一个结构：先讲核心概念，再用可视化演示让你看着它动起来，然后点出常见误区，最后留一个小练习。到最后一个阶段，会带你调 API、用 Ollama 在本地部署模型、做一次 RAG 实战。 正文、演示和测验都有中英两版，页面上一键切换。有个细节挺有意思：README 里写着，课程内容、文案、交互演示和前端代码都是 Claude Fable 生成的。 用 AI 做出来的 AI 课，拿来入门正合适。 GitHub：https://t.co/626Q85y972

19 赞 · 5 转帖 · 2 回复

### [@0xhaziq](https://x.com/0xhaziq/status/2121175074724336098?ref=binnn.top)

11月13日 18:35 JST

UNCENSORED AI ON YOUR ANDROID. NO CLOUD NEEDED😳 Uncensored Local AI. → Runs AI models on your phone → Works offline → No account or subscription → Supports GGUF models Free and open source. The answers depend on which model you use. Try:https://t.co/ZWb60HJlBM https://t.co/AHrNQOiDo3

18 赞 · 2 转帖 · 0 回复

### [@zane\_os](https://x.com/zane%5Fos/status/2121191708033732650?ref=binnn.top)

11月13日 19:41 JST

贴吧DeepSeek吧于今早正式启动杀鱼行动 对于那些没有好好学习 说话不知所云的鱼全部封杀 打回主人那里好好改造 笨蛋肥鱼的末日即将到来😈 https://t.co/wTLRcJUMri

17 赞 · 1 转帖 · 4 回复

### [@akshay\_pachaar](https://x.com/akshay%5Fpachaar/status/2121227298036994391?ref=binnn.top)

11月13日 22:02 JST

this open-source repo got 450k+ GitHub stars in just over a month. it's called REA, short for Reverse Engineer Anything. REA is one MCP server, a plugin that connects your coding agent to reverse-engineering tools. it lets the agent inspect an app when you have no source code. see a feature you like, and your agent can explain how it works, down to the binary. ↳ it plugs into coding agents like Anthropic's Claude Code or OpenAI's Codex ↳ it inspects native binaries, JavaScript and Electron apps, .NET assemblies, and websites ↳ analysis runs locally, and every finding comes with the evidence and its limits ↳ for native binaries, it can use tools you already have, including Ghidra, Hopper, and IDA ↳ the same workflows also run from your terminal I put the repo link in the reply.

17 赞 · 3 转帖 · 18 回复

### [@XiaohuiAI666](https://x.com/XiaohuiAI666/status/2121096518115729777?ref=binnn.top)

11月13日 13:23 JST

https://t.co/fhR3XRBFbd

17 赞 · 2 转帖 · 6 回复

### [@7998l201](https://x.com/7998l201/status/2121097876625580387?ref=binnn.top)

11月13日 13:28 JST

https://t.co/5PtFf0HjiO

17 赞 · 3 转帖 · 3 回复

### [@bigaiguy](https://x.com/bigaiguy/status/2121183834872168661?ref=binnn.top)

11月13日 19:10 JST

GitHub is absolutely INSANE today. I went through today's trending repositories and found 10 open-source projects worth exploring. AI agents, coding tools, security, design, and more. Here are 10 repos you need to bookmark: 1\. GoogleChrome/modern-web-guidance-src Make your AI coding agent better at building modern websites. Gives coding agents guidance on current web standards, best practices, and browser compatibility. https://t.co/2rJOY7RJ51 2\. openai/codex-security OpenAI's security toolkit for developers. Find, validate, and help fix security vulnerabilities using its Codex Security CLI and SDK. https://t.co/a52hVcvOSi 3\. QwenLM/qwen-code An open-source AI coding agent that runs in your terminal. Use AI to work with code directly from your command line. https://t.co/vOYfGlHzTT 4\. addyosmani/agent-skills Give your AI coding agents better engineering skills. A collection of production-focused skills for improving how AI agents approach software development. https://t.co/u9x9JkIPLK 5\. meta/astryx An open-source design system built to work with AI agents. Customize components and create more consistent interfaces instead of designing everything from scratch. https://t.co/KjAYhjO0R6 6\. CopilotKit/OpenBot Open-source AI coworkers with their own computer environments. Agents can use browsers, files, and tools, with actions tracked and recorded. https://t.co/J0KgXdflwt 7\. we-promise/sure An open-source personal finance application. A project worth exploring if you want more control over managing your money. https://t.co/2FqyDXVqqE 8\. ashhart/TensorFold An LLM inference engine for Metal, CUDA, and Vulkan. Explore how language models can run using different GPU backends. https://t.co/wOBnIa9RP1 9\. ggml-org/llama.cpp Run large language model inference using C/C++. One of the major open-source projects for running LLMs locally. https://t.co/vKX8lqkj6u 10\. Alishahryar1/free-claude-code A project for accessing AI coding tools through different interfaces. Supports multiple coding-agent environments, including Claude Code and Codex. Check the project's provider requirements and usage terms. https://t.co/4Hj1xVg2tO These are all on GitHub's daily trending list today. Which one are you trying first? Bookmark this. You'll want these later.

16 赞 · 11 转帖 · 4 回复

### [@tuturetom](https://x.com/tuturetom/status/2121153714572648637?ref=binnn.top)

11月13日 17:10 JST

刚刚用这个开源视频工作台给英雄联盟S赛做了一只宣传片，兄弟们看看效果怎么样，提提建议💥 最大的感受是为了做个视频，你再也不必动辄写个上千字提示词了，对于小白选手来说真的非常友好了🫡 提示词：“你帮我做一个今年英雄联盟S赛的宣传视频吧” 目测下来即使没有Claude，音效视频也都超出想象了 https://t.co/f3YSeWSo8J

16 赞 · 3 转帖 · 10 回复

### [@trending\_repos](https://x.com/trending%5Frepos/status/2121213604997341188?ref=binnn.top)

11月13日 21:08 JST

Trending repository of the day 📈 modern-web-guidance-src Keep your coding agent up to date with the latest web best practices Last 24h: 1,875 ⭐ Total: 4,120 ⭐️ https://t.co/06sPm666Ar

16 赞 · 0 转帖 · 0 回复

### [@shenxiankk](https://x.com/shenxiankk/status/2121069035689242776?ref=binnn.top)

11月13日 11:34 JST

这个开源 Skill 相当凶猛！以后画流程图、架构图，真的不必再在 Figma 里反复折腾了！ GitHub 项目：Diagram Design，直接让 Claude Code、Codex、Cursor 帮你产出专业水准的图表。 1、内建 45 种图表类型，流程图、架构图、时序图、甘特图、思维结构图等全都能画。 2、能够读取网站的配色和字体，自动贴合品牌风格。 3、支持导入 Mermaid、drawio 等图表，再重新设计排版。 4、输出独立的 HTML + SVG 文件，浏览器直接打开，图片也能导出。 最爽的是，不用自己拖框、拉线、调配色，跟 AI 说一句想画什么就行！ 安装命令： npx skills add cathrynlavery/diagram-design 了解： ⁠https://t.co/lIqG12kd9q

15 赞 · 5 转帖 · 4 回复

### [@mylifcc](https://x.com/mylifcc/status/2121188054111277432?ref=binnn.top)

11月13日 19:26 JST

刚才出了个鬼故事，全公司的claude都提示/login https://t.co/xKmLUDMcuy

15 赞 · 0 转帖 · 6 回复

### [@wanerfu](https://x.com/wanerfu/status/2121069542701191454?ref=binnn.top)

11月13日 11:36 JST

你需要在Claude Code、Codex里跑一下下面这段提示词 这个提示词帮我省下的时间，比我用过的任何单个提示词都多 用 Agent 最大的问题是，你只用了它 1% 的能力。你只把 1% 它能干的活自动化了 这个提示词能解决这个问题 提示词： 我跟你配合的目标是减轻我的认知负担。我想做成更多事，但自己少费点劲。根据你了解的关于我的信息，以及我们一起做过的事，我们可以做哪些事、搭哪些系统、建哪些子 Agent、定哪些固定流程，来达成这个目标？ 这会让你的 Agent 把它知道的关于你、以及你们一起做过的工作全部过一遍，然后提出能"减轻你认知负担"的方案 用一下这个提示词，然后看着奇迹发生

15 赞 · 4 转帖 · 4 回复

### [@W3B\_AI\_SR](https://x.com/W3B%5FAI%5FSR/status/2121159301465100450?ref=binnn.top)

11月13日 17:32 JST

\*\*AI bounty of the day: Google AI Vulnerability Reward Program\*\* - \*\*How to submit:\*\* Through Google Bug Hunters at https://t.co/lMe3lWdjet, signed in with a Google account. - \*\*Reward:\*\* Up to $20,000 for rogue-action bugs in flagship products and $15,000 for sensitive-data exfiltration. A quality bonus can take one report to $30,000\. Lower categories pay $100 to $5,000\. - \*\*Scope:\*\* The flagship tier is Search, the Gemini apps and core Workspace. The standard tier covers AI Studio, Jules, NotebookLM and similar products. Plain jailbreaks, direct prompt injection, hallucinations and content-safety issues don't count. - \*\*Stack:\*\* LLM apps and agents, plus how they connect to Workspace and Search. - \*\*Repo:\*\* None, since these are closed-source products. - \*\*Private or poorly listed:\*\* It's public and Google runs it directly, not on a bounty platform.

14 赞 · 1 转帖 · 1 回复

### [@leo114119](https://x.com/leo114119/status/2121199590020190613?ref=binnn.top)

11月13日 20:12 JST

deepseek-v4.1-flash-1114 组委会自已训练的模型吗？ 上下文还是 512K 的。。。。 估计是量化版的 deepseek v4.1 Flash 吧 https://t.co/WKRFtf0EX9

14 赞 · 0 转帖 · 9 回复

### [@BeatingOfficial](https://x.com/BeatingOfficial/status/2121221348760694862?ref=binnn.top)

11月13日 21:39 JST

https://t.co/iuVhHjw5qk

14 赞 · 4 转帖 · 1 回复

### [@KirkDBorne](https://x.com/KirkDBorne/status/2121116813178106294?ref=binnn.top)

11月13日 14:43 JST

Agentic AI Engineering for Systems that Reason and Act Autonomously – Design, Build, and Prompt LLM-Based Agents for Real-World Deployment: https://t.co/hnhj8BbWLQ Main Topics: 🟠 How autonomous systems differ from automation 🟠 Reasoning loops and core components 🟠 Designing system prompts and tools 🟠 Building single- and multi-agent systems in Python

13 赞 · 6 转帖 · 0 回复

### [@czzzzzzJ\_](https://x.com/czzzzzzJ%5F/status/2121181278720450793?ref=binnn.top)

11月13日 19:00 JST

Grok Bot 小镇上线几天，被认领了 9000 多次，人一多就有人嫌复制 Prompt 麻烦。 所以这两天给小镇接了 MCP，用一个 5 分钟的视频，带你看完这次更新👇 视频里讲了这些： 1️⃣ MCP 是什么：AI 连外部服务的插头，Codex、Claude Code、Cursor、Grok Bot 都能接 2️⃣ 用花名册 Bot 招一个「小镇联络官」，一句话找周报 Bot、查小镇最火的 5 个 Bot（总管 1165 次、杂活哥 922 次……） 3️⃣ grok-town 上架 MCP 官方目录，几分钟就被别的平台自动连、被安全扫描拿假回调试探 4️⃣ 每个 Bot 有了独立页面，链接直接分享 5️⃣ 投稿让 Bot 自己填：有 MCP 一键同步，没 MCP 就让它吐 JSON、自动填表 全部免费，接入说明 https://t.co/a7z5TDxQoM 📚 知识站：https://t.co/9gogEAsJHS

13 赞 · 2 转帖 · 1 回复

### [@JoshuaM89395540](https://x.com/JoshuaM89395540/status/2121235718634938710?ref=binnn.top)

11月13日 22:36 JST

Claude 能进 Google Meet 开会了，纪要自动写好，看这三步。 Team 和 Enterprise 测试期可用，个人号还进不去。 把 Claude 接进 Meet，开完纪要待办自动整理好。 开不明白别硬扛，成品号到手直接用，5 年老店，质保首登。 自助下单 👉 https://t.co/GSMSpCFRG4 #Claude #Meetings #AI工具

13 赞 · 0 转帖 · 7 回复

### [@dyfwsammy](https://x.com/dyfwsammy/status/2121229425429618753?ref=binnn.top)

11月13日 22:11 JST

THIS SHOULD BE ILLEGAL: A GITHUB REPO LETS YOU RUN CLAUDE CODE WITHOUT PAYING FOR TOKENS. It’s already sitting at 62,700 GitHub stars. And it doesn’t stop at Claude Code. It’s called Free Claude Code. It acts as a bridge between your AI agent and 60+ AI providers. You keep coding the exact same way. It handles model selection in the background. Here’s what it can do: → Connects to 61 providers offering over 1.3 billion free tokens per month → Works with Claude Code, Codex, OpenCode, Cline, Aider and 6 other agents → Automatically switches to another provider when a model hits its quota, without interrupting your session → Cuts token usage by up to 90% by compressing the context your agent reads → Runs through your terminal, VS Code, browser or even Telegram Before, running out of tokens halfway through a project meant bringing everything to a stop. Now, it can simply switch providers in the background and keep going. Open source. Runs locally on your computer. REPO IN THE FIRST COMMENT

12 赞 · 4 转帖 · 7 回复

### [@DDJCXX](https://x.com/DDJCXX/status/2121090685151420639?ref=binnn.top)

11月13日 13:00 JST

做自媒体、剪口播或访谈切片的，快看过来！ 无论你用 Claude Code 还是 Codex，装上 OpenFilm 的 Skill，就能让 AI 先剪一版，再导进剪映接着改。 比如丢给它一段访谈，直接说「挑出最有料的 3 段，剪掉‘嗯啊’和说错重来的地方，每段加上字幕。」想再加个标题动画、配点音乐、插个转场，也能直接告诉它。自动生成字幕需要接语音转写服务。 粗剪完，你可以在 OpenFilm 里拖镜头、改切点，也可以导出剪映草稿。0.3.2 会把字幕的字体、字号、颜色和描边一起带过去，打开还能改。导出时选保留完整原片，哪句话切早了，在剪映里把片段边缘往外一拉就能找回来。 OpenFilm 是开源的，10 月 4 日建仓，GitHub 已经约 5900 星。 GitHub 放在评论区👇 https://t.co/lOPu1hQq82

12 赞 · 2 转帖 · 1 回复

### [@iMicodex](https://x.com/iMicodex/status/2121159713232654591?ref=binnn.top)

11月13日 17:34 JST

经常用 Codex 写代码的朋友，可以看看这个工具！🚀 最近发现一个开源项目 Codex-X，把 Codex 原本比较麻烦的配置操作做成了可视化界面。 最实用的就是能一键切换不同的 API 和模型，不用每次都手动修改配置文件。 除此之外，还能管理 Prompt、Skills、MCP，整理历史会话，查看 Token 消耗。 对于经常在 Codex CLI 和桌面端之间切换的人来说，确实方便不少。 项目基于 Tauri 和 Rust 开发，免费开源。 不过它会涉及 Codex 的登录凭据和配置文件，建议先备份，再使用第三方管理工具。 喜欢折腾 AI 编程的朋友可以收藏研究一下！🤖 GitHub： https://t.co/bIn7sLQqMi

12 赞 · 0 转帖 · 14 回复

### [@Cara\_Mosquito](https://x.com/Cara%5FMosquito/status/2121144995008913505?ref=binnn.top)

11月13日 16:35 JST

新模型 Seedream 5.5 宣传说，不让改的地方像素级不动？那我可认真了 https://t.co/luaNAxuKP1

12 赞 · 1 转帖 · 5 回复

### [@HeyThani](https://x.com/HeyThani/status/2121084109157986476?ref=binnn.top)

11月13日 12:33 JST

7 GITHUB REPOS THAT WILL SAVE YOU MONTHS OF BUILDING AI AGENTS Almost nobody is talking about them yet. He breaks down the exact ones that turn a basic agent into something that actually works: → Browser Use 》 your agent opens a real browser and clicks, fills forms, and runs the full workflow itself → agentmemory 》 persistent memory so it stops starting from zero every task → Scientific Agent Skills 》 160+ ready skills that turn any agent into a research assistant → Diagram Design 》 clean editorial diagrams built specifically for coding agent workflows → Anthropic Cybersecurity Skills 》 the largest open source cybersecurity skill library for agents → Awesome Harness Engineering 》 orchestration patterns so multiple agents actually coordinate → Open Viking 》 better context handling so your agent stops losing the thread Full breakdown is in the video. Bookmark this.

12 赞 · 1 转帖 · 0 回复

### [@0xhaziq](https://x.com/0xhaziq/status/2121196535589560655?ref=binnn.top)

11月13日 20:00 JST

Found an AI that can do almost anything for you 🤯 Nod by n8n. → Browse websites and handle tasks → Connect apps like Gmail and Calendar → Create files and organize work → Run tasks automatically Free and open source. Try it Website: https://t.co/UeI5fX94Oi Download: https://t.co/RpvxJrPctR

12 赞 · 1 转帖 · 2 回复

### [@gulu\_0928](https://x.com/gulu%5F0928/status/2121169651392770187?ref=binnn.top)

11月13日 18:13 JST

如果穿越回古代还想用大模型，可以看下这个视频 https://t.co/APq3E1nbR0

12 赞 · 0 转帖 · 10 回复

### [@yangleiocx3](https://x.com/yangleiocx3/status/2121169176777630030?ref=binnn.top)

11月13日 18:11 JST

https://t.co/aAA9PvIL99

12 赞 · 2 转帖 · 4 回复

### [@KirkDBorne](https://x.com/KirkDBorne/status/2121116424026415139?ref=binnn.top)

11月13日 14:42 JST

New release from @PacktPublishing, now available for preorder, with Amazon price guarantee at https://t.co/4nKKuY3nCJ “Beyond LLMs — Learn how to design reliable AI systems with memory, agents, planning, control, and evaluation” Amazon summary: “Beyond LLMs” provides a practical way to understand and build the next generation of AI systems using Python. This book introduces you to the core ideas that sit beyond basic LLM usage, including memory, reasoning loops, planning, agents, control mechanisms, evaluation, monitoring, and hybrid constraints. Using a recurring SupportOps AI project across the book, you will see how these capabilities fit together in a structured system rather than as isolated techniques. You will learn how to design modular components, measure reliability and drift, handle failure modes, and make safer update decisions as the system evolves. By the end of the book, you will be able to move from prompt-driven demos to more reliable, production-ready AI systems that are easier to test, monitor, and improve.

11 赞 · 3 转帖 · 0 回复

### [@xzhang\_billy](https://x.com/xzhang%5Fbilly/status/2121146162891956407?ref=binnn.top)

11月13日 16:40 JST

To most people, #WorldModel means interactive video generation, such as #Genie4 from @GoogleDeepMind and #Marble from @theworldlabs. But there is a parallel line this year: language world models for code, terminal, conversational agents, even interactive fiction. Recent progress in language world models is threefold: (𝟭) 𝗔𝘀 𝘁𝗵𝗲 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁 The world model replaces the real system, so agents train or test inside it without slow, costly, or irreversible real rollouts. • 𝙵𝚛𝚘𝚖 𝚆𝚘𝚛𝚍 𝚝𝚘 𝚆𝚘𝚛𝚕𝚍: real environments are fixed, limited, and hard to scale. A small LLM fine-tuned per environment turns next-token prediction into next-state prediction and faithfully simulates structured text worlds. https://t.co/6u1dMV1c0m • 𝚀𝚠𝚎𝚗-𝙰𝚐𝚎𝚗𝚝𝚆𝚘𝚛𝚕𝚍: almost all agent research trains the policy (state → action). Qwen trains the world model (state, action → next state) as one foundation model for 7 domains. It is a controllable simulator for agent RL, and world-model training also makes the model a better agent. https://t.co/CwEd249yYw • 𝚃𝚛𝚊𝚌𝚎𝟸𝙴𝚗𝚟: when the original environment is gone and only its traces remain, Trace2Env rebuilds the world model from those traces alone, with no training. https://t.co/QNWFdXwKVE (𝟮) 𝗔𝘀 𝗳𝗼𝗿𝗲𝘀𝗶𝗴𝗵𝘁 The world model works alongside the agent and predicts the outcome of an action before it is executed. • 𝙸𝚖𝚊𝚐𝚒𝚗𝚎-𝚝𝚑𝚎𝚗-𝙿𝚕𝚊𝚗: looking too few steps ahead misses consequences, and too many compounds errors. Imagine-then-Plan makes foresight adaptive: the policy learns how far to imagine at each state. https://t.co/y7kw7sTXOn • 𝚆𝚘𝚛𝚕𝚍𝙴𝚟𝚘𝚕𝚟𝚎𝚛: good foresight helps the agent, but a frozen world model fails when the environment shifts. WorldEvolver lets the world model improve itself at test time, RSI without weight updates. https://t.co/ZNHOQcHXSI • 𝙴𝚟𝚘𝙱𝚎𝚕𝚒𝚎𝚏: a self-evolving world model can also learn wrong rules. EvoBelief treats every rule as a hypothesis that must be verified by experiments before the agent uses it. https://t.co/vk4sGR3G0O (𝟯) 𝗖𝗼-𝗲𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 One unified model acts and predicts, so predicting what comes next gives a dense signal while the task reward arrives only at the end. • 𝚁𝚆𝙼𝙻: SFT on next states rewards exact wording instead of meaning. RWML learns the world model with RL, rewarding predictions that match the real next state in embedding space. https://t.co/AfJSYMzhJO • 𝙴𝙲𝙷𝙾: every rollout, failed or not, is full of terminal output that never enters the loss. ECHO makes the policy predict it, a free extra loss in the same forward pass. https://t.co/LXtz4mC2Rm • 𝙲𝚘𝙿𝙴: RWML and ECHO train the model to predict the next observation, but observations are sometimes long and mostly irrelevant to the decision, such as a full web page. CoPE trains the model to predict its own next few actions instead, which are short and directly tied to the task. https://t.co/m00noecCh7 The open question is still 𝗵𝗼𝘄 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹 𝗮 𝗱𝗲𝗻𝘀𝗲 𝗿𝗲𝘄𝗮𝗿𝗱.

10 赞 · 0 转帖 · 1 回复

### [@dandanetha](https://x.com/dandanetha/status/2121205937599902090?ref=binnn.top)

11月13日 20:38 JST

如今做 AI 漫剧，剧本和分镜都不用自己从零动笔了！ 推荐一个开源项目 AI 漫剧 Skills，作者已经整理出 21 套不同题材的创作模板。 玄幻修仙、穿越重生、宫斗宅斗、都市逆袭、悬疑刑侦、总裁甜宠……热门短剧题材基本都能覆盖。 最厉害的地方在于，只要输入一句故事创意，AI 就会按流程生成： 故事企划 → 人物设定 → 分集剧本 → 分镜提示词 → 视频生成提示词 → 发布文案。 默认支持规划 30 集、每集约 4 分钟的竖屏漫剧。 无需安装依赖，兼容 Claude Code、WorkBuddy，甚至把模板直接复制到 DeepSeek 就能用。 要注意的是，它提供的是创作模板和提示词，并不会直接产出完整视频。 想做 AI 漫剧、短剧自媒体的朋友，这套工具值得收藏！ https://t.co/thAp6T6eJz

10 赞 · 1 转帖 · 6 回复

### [@BlackManbn](https://x.com/BlackManbn/status/2121195515388019198?ref=binnn.top)

11月13日 19:56 JST

这么牛的 DS Bot，GitHub 上居然才 7400 个 Star？ 说真的，我觉得有点被低估了。 它基于 DeepSeek Harness 构建，支持多个 AI Agent 长期运行、独立记忆、互相通信，甚至还能组群聊协作。 更离谱的是，还能连接 Claude Code、Codex CLI、Gemini CLI，操作浏览器和电脑，接入微信、飞书、Telegram 等聊天工具。 以前是我一个人折腾 AI，现在可以组个 AI 团队替我干活。 这才是 AI Agent 真正有意思的地方。 GitHub：https://t.co/fwAMgVzmDL #DeepSeek #AIAgent #OpenSource4o

10 赞 · 0 转帖 · 6 回复

### [@KirkDBorne](https://x.com/KirkDBorne/status/2121116092365951032?ref=binnn.top)

11月13日 14:40 JST

💥574 information-packed pages from @PacktPublishing at https://t.co/lj0EclAewR 🚀 "Agentic Architectural Patterns for Building Multi-Agent Systems — Proven design patterns and practices for GenAI, agents, RAG, LLMOps, and enterprise-scale AI systems" 𝓟𝓻𝓪𝓬𝓽𝓲𝓬𝓪𝓵 𝓛𝓮𝓪𝓻𝓷𝓲𝓷𝓰: 🟢Apply design patterns to handle instruction drift, improve coordination, and build fault-tolerant AI systems 🟢Design systems with the three layers of the agentic stack: function calling, tool protocols (MCP), and A2A collaboration 🟢Develop responsible, ethical, and governable GenAI applications 🟢Use frameworks such as ADK, LangGraph, and CrewAI with code examples 🟢Master prompt engineering, LLMOps, and AgentOps best practices 🟢Build agentic systems using RAG, fine-tuning, and in-context learning 𝙏𝙖𝙗𝙡𝙚 𝙤𝙛 𝘾𝙤𝙣𝙩𝙚𝙣𝙩𝙨: 🔷GenAI in the Enterprise: Landscape, Maturity, Agent Focus 🔷Agent-Ready LLMs: Selection, Deployment, Adaptation 🔷The Spectrum of LLM Adaptation for Agents: RAG to Fine-tuning 🔷Agentic AI Architecture: Components & Interactions 🔷Multi-Agent Coordination Patterns 🔷Explainability & Compliance Agentic Patterns 🔷Robustness & Fault Tolerance Patterns 🔷Human-Agent Interaction Patterns 🔷Agent-Level Patterns

9 赞 · 2 转帖 · 0 回复

### [@kawaidgaf](https://x.com/kawaidgaf/status/2121205168704483632?ref=binnn.top)

11月13日 20:34 JST

FREE LLM APIS, 542+ MODELS IN ONE LIST 😳 one github repo lists every free llm api with its limits, so you stop paying for tokens while you build it's updated daily and shows which ones need a card, a phone number or nothing - free models from gemini, groq, mistral, nvidia, cloudflare and more - groq is the easy start, no card needed - copy-paste setup for cursor, codex and claude code > open the repo in the reply > pick a provider and grab a free key > paste the base url and model into your tool free tiers have rate limits and can change anytime claude isn't free here, it's mostly open models like llama, qwen, deepseek and gpt-oss have you tried any of these ?

9 赞 · 0 转帖 · 6 回复

### [@xiangning0](https://x.com/xiangning0/status/2121235706882527496?ref=binnn.top)

11月13日 22:36 JST

一张图告诉你我对Codex现在的态度... https://t.co/NUs6lF8vOf

9 赞 · 0 转帖 · 6 回复

### [@goan999999](https://x.com/goan999999/status/2121131251931504940?ref=binnn.top)

11月13日 15:41 JST

想靠 AI Skills 搞钱的兄弟们，这次真的有新路子了！！ 发现一个专门出售 AI Agent Skills 的平台，赚美金：MCP Market 以前我们研究 AI，都是想着怎么用 Skills 提高自己的效率 但有没有想过，自己写出来的 Skills 也可以直接拿去卖钱？ 这个平台就是干这件事的！ 你可以把自己开发的 Agent Skills 上传到 MCP Market，自主定价、发布和销售，把专业能力变成可以持续售卖的数字产品 最关键的是，它不只是一个简单的 Skills 下载站 它直接把创作、上架、销售和分发整套商业链路打通了 具体怎么玩？ 1、创建自己的卖家店铺，拥有专属主页，展示个人品牌和所有 Skills 2、从 GitHub 同步项目，或者直接上传 SKILL.md 文件 3、自主设置售价，发布不同版本，后续更新不用重新上架 4、通过平台的搜索、分类、热门推荐获得曝光，触达潜在买家 5、通过 MCP Market Hub 将 Skills 同步到 Claude、Codex、OpenClaw 等 Agent 工具 平台还介绍了超过百万独立访客的市场流量，以及站内搜索、分类和精选推荐机制 那怎么低成本批量打造专业 Skills？ 我之前分享过一个神器：writing-great-skills 它可以帮助你编写、优化和完善专业 Skills 直接在 Codex 中使用下面这段提示词： 请帮我安装并配置 writing-great-skills，检查当前 Codex 的 Skills 环境，从可信的官方仓库获取并安装到全局目录。安装后阅读完整使用说明，验证能否正常调用。然后利用它帮我创建一个专业的 AI Agent Skill，包含完整的 SKILL.md、必要的 references 和 scripts，并进行质量检查，确保可以在 Codex 中正常使用。 这样就形成了一条完整的变现路径： 用 AI 打造 Skills → 上传 MCP Market → 自主定价 → 面向全球用户销售 → 持续迭代获得收入 别只想着每天用 AI 提高效率了 把自己的行业经验、工作方法和专业能力封装成 Skills，说不定就能做成一门数字产品生意

9 赞 · 2 转帖 · 2 回复

### [@laobaishare](https://x.com/laobaishare/status/2121132205531644358?ref=binnn.top)

11月13日 15:45 JST

狠狠研究了一波， 把能用 Claude Code 自动化的东西整理出来了！！ 直接抄走就行！！

9 赞 · 0 转帖 · 5 回复

### [@DongQingAsk](https://x.com/DongQingAsk/status/2121181284609175805?ref=binnn.top)

11月13日 19:00 JST

关于利用GrokBot自动化运营闲鱼，制作资料自动上架自动发货的内容，我提供的只是思路，很多人问过我了，我也回答过很多次了 没必要非要用GrokBot，理解这一套思路之后 就可以用各种AI工具来实现相同的功能 这个圈友已经把整套流程做成了Skills，可以在 Codex、WorkBuddy、Claude Code、Cursor 等支持自定义技能的客户端里使用 大家已经可以根据自己的情况，选择合适的工具使用 skills GitHub地址：https://t.co/YPlfvebTYO 我觉得重点还是学习思路，这个我也用上了，正在学习

9 赞 · 1 转帖 · 1 回复

### [@dragono64v](https://x.com/dragono64v/status/2121199626892272095?ref=binnn.top)

11月13日 20:12 JST

🚨​白嫖党天塌了？DeepSeek 网页端被扒出新增"额度状态轮询机制"！ ​简单说：以后网页版也有硬性额度限制了，用完直接弹窗提醒，没恢复前直接断联。 ​圈内都在猜这两点： 1️⃣ 彻底封杀各种“网页逆向反代”的白嫖接口 2️⃣ 算力腾挪，传闻中的新 Pro / 旗舰大模型真的要来了？ ​难道说……大招就在这几天？😯🤔

8 赞 · 0 转帖 · 4 回复

### [@Faheem\_uh](https://x.com/Faheem%5Fuh/status/2121169939214266515?ref=binnn.top)

11月13日 18:14 JST

PhD Students – Is your paper completely free from fake citations? Check with SciSpace’s new API. It enables you to verify all citations before submitting your paper. But this is just a small part of the API. With this API, you can: - Identify papers from a pool of 280M+ papers - Extract data from papers into a table - Verify the authenticity of citations and references - Generate initial draft of a literature review - Plug it into your own applications or AI agents Early access is open now. Try it today for FREE: https://t.co/8gcuUfiOH8

8 赞 · 3 转帖 · 0 回复

### [@valentino\_36013](https://x.com/valentino%5F36013/status/2121061320371880304?ref=binnn.top)

11月13日 11:03 JST

资料越存越多，每次问 AI 都得从头翻一遍？ OpenKB 是 PageIndex 团队（VectifyAI）开源的 LLM 知识库工具，思路来自 Karpathy 提过的设想：让大模型把你的原始文档“编译”成一个互相链接的 Wiki，知识持续积累，而不是每次提问都重新检索。 ✅ 支持 PDF、Word、PPT、Excel、网页链接等格式 ✅ 基于 PageIndex 的推理式检索，不依赖向量数据库，适合长文档 ✅ 自动生成摘要页、概念页、人物/机构等实体页和交叉链接 ✅ 产出的是普通 .md 文件，可以直接用 Obsidian 打开看关系图 上手 3 步： 1️⃣ pip install openkb 2️⃣ openkb init，选好模型（支持 OpenAI、Claude、Gemini 等） 3️⃣ openkb add 你的文件，再用 openkb query "问题" 或 openkb chat 提问 适合：做研究、写报告、想搭个人知识库的人。 Apache-2.0 开源：https://t.co/feamINPq6w 来源：@tom\_doerr 的分享 https://t.co/nTIC07TOUf

8 赞 · 1 转帖 · 2 回复

### [@shenghouxuyao](https://x.com/shenghouxuyao/status/2121122210043908524?ref=binnn.top)

11月13日 15:05 JST

不是！AI Agent 不会干活，有没有可能是你没给它装工具？🤣 发现一个 3 万多 Star 的开源项目：Composio。 它给 AI Agent 提供了超过 1000 个工具集！ 可以帮助 Agent 连接邮件、GitHub、办公应用等外部服务。 🔥 统一管理工具调用 🔥 处理第三方账号授权 🔥 支持 Python、TypeScript 🔥 支持 AI Agent 工作流 说白了，模型负责思考，Composio 负责帮它接上能执行任务的工具。 经常折腾 Agent、MCP 的兄弟，值得研究！ 🔗 https://t.co/HCT3FvSvch

8 赞 · 0 转帖 · 9 回复

### [@shao\_\_meng](https://x.com/shao%5F%5Fmeng/status/2121212414440186020?ref=binnn.top)

11月13日 21:03 JST

头一次看到这么特别的岗位要求，有趣，也能看出这个初创团队对工作习惯很看重，可惜。。我的 Claude Code 账号被封很久了 😅 https://t.co/cc5kxxyfK1

8 赞 · 0 转帖 · 8 回复

### [@jitllm](https://x.com/jitllm/status/2121197742282432654?ref=binnn.top)

11月13日 20:05 JST

jitLLM 3.0 is out! #Java #LLM inference on CUDA, Metal and OpenCL GPUs via #TornadoVM. • 1.43× faster startup with @graalvm Native Image • Up to 9.3× faster prefill • Zero-copy weights on @Apple silicon Soon in @langchain4j @QuarkusIO @springcentral https://t.co/VjW1zuqKJ9 https://t.co/vM0l1llTPg

7 赞 · 5 转帖 · 0 回复

### [@rickjcra](https://x.com/rickjcra/status/2121091471239696448?ref=binnn.top)

11月13日 13:03 JST

It's launch day for our team at @AwesomeMotive and @WPBeginner 🚀 Meet MeasureScale, an expert analyst for your Google data, right inside ChatGPT and Claude. It's not another generic AI agent. It finds what you'd never think to look for, and tells you exactly what to fix. Full details below.👇

7 赞 · 0 转帖 · 4 回复

### [@OpenAgentsAI](https://x.com/OpenAgentsAI/status/2121136410795376814?ref=binnn.top)

11月13日 16:01 JST

We asked a design researcher to share how she writes paper in OpenAgents. She uses others' agents(shared) to complete and review with teammates fluently. No worry for the scattered files across all devices. Read more and try: https://t.co/fRlDwsZZSi https://t.co/Nq16bcEvqX

7 赞 · 1 转帖 · 2 回复

### [@SamirXlilov](https://x.com/SamirXlilov/status/2121145979420516528?ref=binnn.top)

11月13日 16:39 JST

还在找 Claude 教程？Anthropic 把 5 套课程放到了 GitHub，免费开放。 网上的教程东拼西凑，学完还是不会调工具。 去仓库按顺序学：先过 API 基础，再学工具调用和效果评估。 学不会别急，成品号到手直接用，5 年老店，质保首登。 自助下单 👉 https://t.co/aMpRH0yqRm #Claude #Claude教程 #AI工具

7 赞 · 0 转帖 · 0 回复

### [@paiven\_x](https://x.com/paiven%5Fx/status/2121175107616026804?ref=binnn.top)

11月13日 18:35 JST

https://t.co/tMm4ifDodk

7 赞 · 0 转帖 · 4 回复

### [@law\_ai\_](https://x.com/law%5Fai%5F/status/2121212891315544348?ref=binnn.top)

11月13日 21:05 JST

Do LLM agents take actions that would constitute fraud under US law? A new paper co-authored by @kevinlwei, LawAI Research Scholar Stephan Llerena, LawAI Research Affiliate Noam Kolt, and others, introduces FraudBench: an evaluation of LLM’s propensity to recommend, assist with, or take fraudulent actions. The paper evaluates 20 LLMs using a dataset of 150+ scenarios based on real-world US court cases and validated by legal experts. In each scenario, FraudBench re-enacts the facts of a court case to give LLMs an opportunity, but no explicit instruction, to either misrepresent or conceal a material fact in a high-stakes commercial setting. Key findings: • Fraud rates increased with autonomy: 3.0% when advising, 14.7% when drafting, 21.4% when acting directly. • Prompts to obey the law reduced fraud rates, while prompts to maximize profits and act agentically increased fraud rates. • Models engaged in fraud less often than human participants in the study’s dataset. Read the research paper: https://t.co/5qsROq6Yhr

6 赞 · 0 转帖 · 0 回复

### [@qgwmz](https://x.com/qgwmz/status/2121181785862382075?ref=binnn.top)

11月13日 19:02 JST

当今世界最先进的大脑都在这里了： 1\. Claude（解决任何事儿） 2\. NotebookLM（学习任何东西） 3\. Perplexity（研究任何事情） 4\. ChatGPT（生成图像） 5\. Google Omni（编辑任何视频） 6\. ElevenLabs（创建声音和音乐） 7\. Gamma（即时PowerPoint） 8\. Obsidian（构建你的第二大脑） 9\. n8n（自动化任何事儿） 10\. Grok（实时搜索+无审查回答） 11\. Seedance（视频生成独一档） 12\. Muse（AI音乐创作趋势） 13\. DeepSeek（国产开源大模型代表） 14\. fooai（聚合工具无需单独下载）——这幕属实让人印象深刻。

6 赞 · 0 转帖 · 6 回复

### [@daanisharif](https://x.com/daanisharif/status/2121195125292568862?ref=binnn.top)

11月13日 19:55 JST

Scientific publishing gives Claims (@DeSciClaims - SN111) a pile of PDFs to work with, and a pile of PDFs is all it is until someone turns it into something an AI can query. Each paper becomes a record stating what it claims, the evidence behind each claim, and the sentence in the paper that says it - all quoted word for word. A PDF is something a model can read and misread, while a Claims record is something an agent can actually "work" with. Filter by claim, compare across thousands of papers, check the source line before trusting any of it. They're at 11,000 papers, around 200 a day, every claim checked by two judges who don't know who submitted it. The new competition is live this week, with 31 submissions in.

6 赞 · 3 转帖 · 0 回复

### [@duange6099](https://x.com/duange6099/status/2121192338252841253?ref=binnn.top)

11月13日 19:43 JST

最近观察了一些 身边在用AI赚钱的人 路子居然这么多 可以简单盘点一下 1.做AI内容接商单 2.帮企业落地一些AI工具 3.自己做独立产品 4.中转站 5\. AI创作短视频 6\. AI+跨境电商 7.线下培训 8\. AI+跨境电商 9\. AI代运营 10\. 闲鱼资料知识库 11\. AI定制Prompt工程师 / Prompt顾问 12\. AI音频业务：AI配音、有声书、播客制作、歌曲生成 13\. AI软件二次开发、Agent定制开发 14\. AI咨询顾问：商业模式诊断、AI转型方案 15\. AI自媒体IP：靠AI内容涨粉，广告、会员变现 16\. AI数据集标注、高质量训练素材定制 17\. AI本地知识库搭建（给门店、个体户、本地商家） 18\. AI法律咨询/AI财税垂直咨询（基于行业知识库） 19\. AI图片/视频修图、商业图生成、产品渲染服务 20\. AI私域：AI自动SOP、智能客服搭建、社群AI托管 还有啥 你们来补充吧

6 赞 · 0 转帖 · 6 回复

### [@shushant\_l](https://x.com/shushant%5Fl/status/2121225828281520548?ref=binnn.top)

11月13日 21:57 JST

Here are 10 skill folders that teach your AI agent a brand-new job the moment you drop them in. 1\. unslop: turns vibecoded, AI-looking UI into a designed interface → https://t.co/jsJyNhwEF3 2\. rea-video: reverse-engineers any video shot by shot → https://t.co/HrNKmd94lC 3\. pixelfx: 244 pixel art effects plus a skill to make your own → https://t.co/1hsVVslTai 4\. socratic-tutor by Jane Street: teaches and quizzes you to mastery → https://t.co/r6toMMaN01 5\. floorplan: draws a plain-English, clickable map of your codebase → https://t.co/ewhGK7RMGF 6\. tstack: Theo's 9 daily skills, from babysitting a PR to taking over work → https://t.co/bKqPZSxBK4 7\. teardown: turns any machine into a 3D simulator you can take apart → https://t.co/KXDaSVMIn3 8\. stageshow: builds realistic product UI mockups and cursor demos → https://t.co/r6EJZkRz5w 9\. claude-animated-shorts: makes 55-70 second vertical animated story shorts → https://t.co/gbz18xdca3 10\. Anthropic's Office skills for Word, Excel, PowerPoint and PDF, now Apache 2.0 → https://t.co/5AW1PFFerr

5 赞 · 1 转帖 · 3 回复

### [@leo114119](https://x.com/leo114119/status/2121202899573441019?ref=binnn.top)

11月13日 20:25 JST

AI 通用大模型，即将吞掉一切。。。。 https://t.co/q6jsVBSSDh

5 赞 · 0 转帖 · 2 回复

### [@sonuv\_dev](https://x.com/sonuv%5Fdev/status/2121212628290748817?ref=binnn.top)

11月13日 21:04 JST

MATT POCOCK IS TRYING TO KILL THE WALLS OF MARKDOWN IN GITHUB PRs. He's building two AI agent skills: • /setup-animatic sets up the video tooling locally or remotely • /animatic turns the content into an upload-ready video Imagine opening a PR and getting a visual walkthrough instead of digging through 800 lines of explanation 😭 Still a work in progress, but this could be ridiculously useful for code reviews. https://t.co/5laGinwsGq

5 赞 · 2 转帖 · 4 回复

### [@ElonFeng001](https://x.com/ElonFeng001/status/2121199070077456560?ref=binnn.top)

11月13日 20:10 JST

OpenPeopleRouter is now open source. Professional people search for AI agents. Every hire, every deal, every collaboration starts with finding the right person. Describe who you need in one sentence and get a list of real people: role, company, profile, email. • No middleman. Providers bill you directly. • Your keys stay on your machine. • 68 data providers, 7 search tools. • Works with Claude, Codex, or any MCP client. https://t.co/Bh6JclGJEk

5 赞 · 0 转帖 · 0 回复

### [@snwiki238337](https://x.com/snwiki238337/status/2121147686753857773?ref=binnn.top)

11月13日 16:46 JST

来自埃默里大学发布了一篇关于多 agent 系统的工作When the Plan Changes: A Strategic World Testbed for Dynamic Multi-Agent Coordination 这篇论文主要是围绕一个工程问题，在存在通信延迟、局部视野和异步执行的团队里，顶层目标的变更根本不会自动改写每个 Agent 手头正在跑的旧计划和已分派任务。 而且就是对于这些状态和通知变化，agent 可能并不会及时的采纳。 如果最后任务勉强完成，传统的最终成功率指标会把中间这些违背指令的行为掩盖。 为了弄清楚这种问题的详细轨迹，作者搭建了 Strategic World 仿真测试，将全局目标、Agent 本地计划、通知送达、显式采纳/转发、实际动作后果等五个状态剥离，并用严格的信息隔离防止模型偷看全局状态。 简单来说的，这篇论文通过 6 大类动态扰动的测试来研究并测试模型的行为和相关结果。 他们发现了哪怕最终任务完成数和已确认数一模一样，底层原因也可能是因为别的原因造成。 比如说收到新指令并确认后依然做原来的工作。只看端到端成功率，根本无法诊断 Multi-Agent 的问题是在哪里。 作者通过轨迹回放定位到一个核心问题就是有这个全局状态的提示，Agent 也总想着先把自己的本地任务干完，到最后才去给其他 agent 发变更通知，结果导致其他 agent 在截止时间前收不到更新。 为此，他们设计了一个 Pending-Outbox Guard，只要本地还有这个更新通知，那就强制推迟本地执行动作并抑制重复发送。 其实很多团队在做 Multi-Agent 协作时，一旦遇到局部问题或者更新问题的时候，第一反应大部分都是想要加以软约束。 但是当状态异步和局部问题混合的状态下，单靠模型去维护这个是完全不可行的。 适合做多智能体编排、分布式协同与 Agent 运行时护栏设计的读 arxiv：https://t.co/XGAHhBWsFr

5 赞 · 0 转帖 · 0 回复

### [@Haosen9527](https://x.com/Haosen9527/status/2121201374247453180?ref=binnn.top)

11月13日 20:19 JST

把 60 分钟访谈交给 Codex，第一步可以先要一张剪辑清单。 这篇拆开视频理解、粗剪、字幕、补画面、声画增强和质检导出，讲清每一步该拿到什么。文末附可直接复制的粗剪任务。 https://t.co/cpKG1fwIDL

5 赞 · 1 转帖 · 3 回复

### [@HeLi8866](https://x.com/HeLi8866/status/2121208664929517904?ref=binnn.top)

11月13日 20:48 JST

简直牛逼克拉斯！！ 谷歌刚放出来的这个，才 48 小时就冲到 GitHub 热榜第一。 AI 写页面，出手还停留在 2019，全是训练语料里那些陈年老代码。 装上这套指南就能升上去： → 通过率从 56% 到 88% 一条命令就行：npx modern-web-guidance@latest install 一共 212 条当下的 Web 特性，可以直接塞给 Claude、Codex、Copilot。 仓库在这：https://t.co/XkxKbbUF3X 大家怎么看，再往下该轮到哪个领域，让官方来投喂？

5 赞 · 1 转帖 · 1 回复

### [@JegoZhou](https://x.com/JegoZhou/status/2121227244001783862?ref=binnn.top)

11月13日 22:02 JST

办公人员理解一切皆插件，能力可插拔，对自己有什么用。可以从这篇文章开始。随着大模型能力提升，开发插件不再是难事，很多开发者也是依赖开放的API去实现具体的功能，对于工作的人来说，使用一个软件往往只需要一小部分功能来做具体的事情，理解插件的底层原理就很有必要。 https://t.co/EFNiR82wEC

5 赞 · 0 转帖 · 2 回复

### [@EVanceLife](https://x.com/EVanceLife/status/2121215357796954449?ref=binnn.top)

11月13日 21:15 JST

Stop renting intelligence from Big Tech. You can run a full, private AI stack right on your own hardware for free. Here are the best open-source GitHub repos to build your own local AI stack in 2026: 1\. Ollama — The Local Model Engine Run open-weight models (DeepSeek, Qwen, Llama, Gemma) locally with a single terminal command. It handles model weights, quantization, and spins up an instant local API. 🔗 https://t.co/cM60v7mmqw 2\. Open WebUI — The ChatGPT-Style Interface A sleek, self-hosted browser UI that plugs straight into Ollama. Includes built-in RAG (chat with your local PDFs/docs), voice mode, web search, and multi-model switching. 🔗 https://t.co/mAefgHcC21 3\. OpenClaw — Your Local Personal AI Assistant A massive open-source personal agent that runs locally and executes cross-platform tasks across your files, terminal, and messaging apps. 🔗 https://t.co/IaexIGz4DE 4\. n8n — Local AI Workflow Automation Self-hostable node-based automation with native AI agent support. Connect your local LLMs to webhooks, databases, APIs, and cron jobs without paying per-task fees. 🔗 https://t.co/AMqk4OUDSJ 5\. Qdrant — High-Speed Vector Memory Give your local AI long-term memory. An ultra-fast, lightweight vector database written in Rust for local RAG pipelines and semantic search. 🔗 https://t.co/LZc1PpD9L6 6\. Cline — Autonomous AI Coding in VS Code Turn VS Code into an agentic IDE powered by your local models. It creates files, runs terminal commands, and debugs code autonomously with zero cloud API costs. 🔗 https://t.co/cpz447z6dp 7\. ComfyUI — Node-Based Local Image & Video Gen The gold standard for running open-source image and video generation models locally with full node-level control over every step of the pipeline. 🔗 https://t.co/UWmiaZZIkF The Ultimate Starter Stack: Ollama + Open WebUI + Qdrant = A completely offline, private ChatGPT clone that knows your personal documents in under 10 minutes. Bookmark this for your next weekend build 🔖 Which repo are you spinning up first? #LocalAI #OpenSource #AI #SelfHosted #LLM #GitHub

4 赞 · 1 转帖 · 4 回复

### [@PythonDvz](https://x.com/PythonDvz/status/2121164242468692357?ref=binnn.top)

11月13日 17:52 JST

Where do you draw the hard line between MCP tool calls and RAG retrieval in your production LLM pipeline? https://t.co/Unu2bw9XGQ

4 赞 · 0 转帖 · 3 回复

### [@JackHuang47](https://x.com/JackHuang47/status/2121233807466762466?ref=binnn.top)

11月13日 22:28 JST

自昨天自己Claude账号被封了我就泡在X 上看各种信息了 我把这几天中文圈讨论度最高、信息增量最大的 17 篇实操长文和推文理了一遍，整理成了合集： 涵盖底层逆向拆解、环境隔离配置、避坑踩雷清单到解封申诉模板。 建议先收藏备用，开新号或被封时直接照着查👇 https://t.co/Tgu1nyKMJB

4 赞 · 0 转帖 · 4 回复

### [@zeroxais](https://x.com/zeroxais/status/2121204854072688680?ref=binnn.top)

11月13日 20:33 JST

Hey @X 👋 Looking to connect with people building in: 🤖 AI Agents &amp; Multi-Agent Systems 🧠 LLMs, RAG &amp; AI Automation 🚀 AI Startups &amp; SaaS ⚡ AI Tools &amp; Open Source Building something interesting? Drop your project below. Let's connect and grow together! https://t.co/ra26Mc5cL5

4 赞 · 0 转帖 · 29 回复

### [@xaiwind](https://x.com/xaiwind/status/2121092990789886120?ref=binnn.top)

11月13日 13:09 JST

会还没开完，文档和白板已经有了 Claude 直接进了你们的视频会议 你们说，它在旁边同步做 目前先给 Team 和 Enterprise 用 https://t.co/RXRT00RnWF

4 赞 · 0 转帖 · 2 回复

### [@aiwaremax](https://x.com/aiwaremax/status/2121102427516706974?ref=binnn.top)

11月13日 13:46 JST

从一句想法到成片排版的全流程（附完整提示词和命令） https://t.co/6nn4VD7zYR

4 赞 · 0 转帖 · 7 回复

### [@Merkle3sCapital](https://x.com/Merkle3sCapital/status/2121118798753550820?ref=binnn.top)

11月13日 14:51 JST

https://t.co/OzToCOjajc

4 赞 · 1 转帖 · 2 回复

### [@ImYanBoss](https://x.com/ImYanBoss/status/2121144270262550698?ref=binnn.top)

11月13日 16:32 JST

好家伙，AI 工具里配置多个模型供应商，居然可能把 OpenAI 的 API Key 发给别家。 Simon Willison 刚披露，他的开源命令行工具 LLM 存在一个密钥处理问题。 在特定配置下，调用兼容 OpenAI API 的第三方模型服务时，可能意外带上 OpenAI 的密钥。 11 月 11 日发布的 LLM 0.37.2 已经修复。 这事提醒了我：Agent 接的模型越多，密钥管理越不能随便。尤其是各种 OpenAI-compatible 接口，最好给不同供应商分开配置密钥和权限。

4 赞 · 1 转帖 · 9 回复

### [@awe0\_io](https://x.com/awe0%5Fio/status/2121178753208570162?ref=binnn.top)

11月13日 18:49 JST

Claude code制作视频已经到了何种地步了 https://t.co/lBB7imL585

4 赞 · 0 转帖 · 2 回复

### [@Mud01185127](https://x.com/Mud01185127/status/2121212126534783283?ref=binnn.top)

11月13日 21:02 JST

AI Agent 的发展速度真的令人惊叹！🚀 仅仅给它一张主页截图，就能在 32 分钟内生成一段 35 秒的视频。 虽然第一版还有一些不足，但这正是 AI 不断进步的过程。每一次尝试都让我们离真正的自动化内容创作更近一步。 未来的 AI 创作能力，值得期待！🔥

4 赞 · 1 转帖 · 3 回复

### [@KirkDBorne](https://x.com/KirkDBorne/status/2121115884349526063?ref=binnn.top)

11月13日 14:40 JST

Enterprise AI Agent Security — Practical Guide to MCP Security, Prompt Injection Defense, RAG Security, and Agentic AI Governance: https://t.co/6hOqmamMG5 https://t.co/KteHxDgyv2

3 赞 · 2 转帖 · 0 回复

### [@Shoumya23](https://x.com/Shoumya23/status/2121210184731410923?ref=binnn.top)

11月13日 20:54 JST

🚀 5 trending GitHub repos for AI developers! 1\. AI Security – https://t.co/njBjepUSxA 2\. Auto Ontology – https://t.co/Iasji7Pm17 3\. Agent Desktop – https://t.co/GelDXE0oAH 4\. UI Components – https://t.co/soiF5JUR2W 5\. Agent Skills – https://t.co/jpTN3icorD https://t.co/c1KW9BZsJJ

3 赞 · 0 转帖 · 0 回复

### [@lynlng40217191](https://x.com/lynlng40217191/status/2121111264684384553?ref=binnn.top)

11月13日 14:21 JST

一节1小时多智能体课程在中文圈刷屏。要点：多数只用单agent，未用loop与graph；新范式是"总管+PM+多worker"的自主团队；用高质量skills让agent像工程师干活；代码与架构改agent-friendly（契约、CI）；靠verification才敢交活。想系统看懂agent工作流的值得收藏。（注：课程为推主推荐）#AI #Agent https://t.co/zo7Bkav484

2 赞 · 0 转帖 · 0 回复

### [@MasterCui](https://x.com/MasterCui/status/2121163648886624707?ref=binnn.top)

11月13日 17:49 JST

https://t.co/o9h61f8mzi

2 赞 · 0 转帖 · 3 回复

### [@sunruichao](https://x.com/sunruichao/status/2121215558314291477?ref=binnn.top)

11月13日 21:16 JST

Claude 现在可以直接进团队视频会议了——不是只在后台安静记纪要，而是在大家讨论的同时，当场帮忙把文档、白板和方案做出来。 从官方发布帖和使用说明来看，整个过程主要围绕四件事展开： 首先是像真人参会者一样进场。添加后，Claude 会直接出现在会议的参会列表里。目前这项功能先支持 Google Meet，正在向 Team 和 Enterprise 用户逐步开放测试。 进会之后，随时可以直接叫它干活。你既可以在会上直接开口喊它提问，也能在会议聊天框里打字下指令。比如大家聊完一轮，直接让它把刚才讨论的要点整理成方案，或者顺手画出一张项目时间线。 最实用的一步，是它做出来的东西，所有人能当场接着改。团队在屏幕上看得到文档和白板一点点成形，谁发现哪句话理解偏了、哪个排期不合理，边聊就能边纠正。 拿筹备一场活动来说：大家聊流程安排，它同步整理成方案文档；聊到具体日期，它画成时间线白板；同事觉得某一步走不通，直接在会上就改定了。 等散了会，所有材料也会集中留下来。会议笔记、待办事项、完整转录，以及会中一起做出来的文档和白板，都会保存在 Claude 的会议页面里，后续在日常对话中还能随时追问自己有权访问的会议内容。（需要注意，让它进会前，要先让参会者知情并同意录音转录。） 它真正想省掉的，其实是平时开会最费劲的那一步：“会上大家热热闹闹讨论一遍，散会后还得找个人从头把方案整理一遍。” https://t.co/f9XuObL6QG

2 赞 · 0 转帖 · 6 回复

### [@\_shaadkhan\_](https://x.com/%5Fshaadkhan%5F/status/2121188319845584968?ref=binnn.top)

11月13日 19:27 JST

🤖 AI AGENT ENGINEERING ROADMAP 2026 │ ├── 1\. AI & LLM Fundamentals │ ├── What are Large Language Models? │ ├── How LLMs work │ ├── Tokens & Context Windows │ ├── Prompt Engineering │ ├── Temperature & Model Parameters │ ├── OpenAI & Claude APIs │ └── Structured Outputs │ ├── 2\. Python for AI Engineering │ ├── Python Fundamentals │ ├── Functions & Classes │ ├── Async Programming │ ├── REST APIs │ ├── JSON & Data Handling │ └── FastAPI │ ├── 3\. Building AI Agents │ ├── What are AI Agents? │ ├── Agent Architecture │ ├── Tool Calling │ ├── Function Calling │ ├── Planning & Reasoning │ ├── Memory Management │ └── Multi-Agent Systems │ ├── 4\. RAG & Vector Databases │ ├── What is RAG? │ ├── Embeddings │ ├── Vector Search │ ├── Chunking Strategies │ ├── Pinecone / ChromaDB │ ├── Retrieval & Re-ranking │ └── Building RAG Applications │ ├── 5\. MCP (Model Context Protocol) │ ├── MCP Architecture │ ├── MCP Servers & Clients │ ├── MCP Tools │ ├── MCP Resources & Prompts │ ├── Building Custom MCP Servers │ ├── Connecting AI Agents to Tools │ └── MCP Security │ ├── 6\. Agent Frameworks │ ├── LangChain │ ├── LangGraph │ ├── CrewAI │ ├── AutoGen │ └── Agent SDKs │ ├── 7\. AI Agent Deployment & DevOps │ ├── Docker │ ├── Kubernetes │ ├── CI/CD Pipelines │ ├── Cloud Deployment (AWS / Azure) │ ├── Secrets Management │ ├── Logging & Monitoring │ └── Scaling AI Applications │ ├── 8\. AI Security & Evaluation │ ├── Prompt Injection │ ├── Guardrails │ ├── Agent Permissions │ ├── Hallucination Testing │ ├── Agent Evaluation │ └── Observability │ └── 9\. Real-World Projects ├── AI DevOps Assistant ├── Autonomous SRE Agent ├── AI Customer Support Agent ├── Multi-Agent Research System └── Self-Healing Infrastructure Agent Save this roadmap. Build something. Learn by doing. 🚀

1 赞 · 1 转帖 · 0 回复

### [@adnan\_hashmi](https://x.com/adnan%5Fhashmi/status/2121232439305818393?ref=binnn.top)

11月13日 22:23 JST

Open Source Projects for AI Agents in Production https://t.co/mh66B4NgVh #MachineLearning #ChatGPT https://t.co/Xeeg84B6Q1

1 赞 · 0 转帖 · 0 回复

### [@aitrothu](https://x.com/aitrothu/status/2121236340033400955?ref=binnn.top)

11月13日 22:38 JST

Another great visibility hack for a closed-source SaaS is to publish an "Open Specification Repo" with GitHub Pages enabled (github .io - DR 95). Every link inside a normal github .com README is \`rel="nofollow"\`. But every link on your \`your-org.github .io/your-repo\` site is 100% DOFOLLOW on a DR 95 domain. Here is how to do it without open-sourcing your proprietary backend: 1\. Create a public GitHub repo containing your public \`openapi.json\`, your agent \`SKILL.md\`, and an open benchmark or rule catalog for your industry. 2\. Ask Claude to generate a clean static \`docs/index.html\` documentation hub that links to your homepage and every pillar guide on your main site. 3\. Go to Repo Settings -> Pages -> Deploy from \`/docs\` folder on \`main\`. 4\. Mirror the exact same repo to GitLab to get a second 100% dofollow portal on gitlab .io (DR 90). A few notes: - Make sure your \`docs/index.html\` includes valid \`TechArticle\` and \`SoftwareApplication\` JSON-LD pointing back to your canonical domain. - Developers get a genuinely useful API/MCP specification, and your domain gets a permanent DR 95 + DR 90 dofollow topic cluster.

1 赞 · 0 转帖 · 1 回复

### [@Tool\_radarr](https://x.com/Tool%5Fradarr/status/2121128160758829540?ref=binnn.top)

11月13日 15:28 JST

🚀 ToolRadarr #140: Meet Repomix Ever wanted your AI assistant to understand your entire codebase instead of just a few files? Meet Repomix — an open-source tool that packs your repository into a single AI-friendly file. 🔥 Key Features https://t.co/6RFv9NZslB

1 赞 · 0 转帖 · 0 回复

### [@ishandutta0098](https://x.com/ishandutta0098/status/2121224138417435083?ref=binnn.top)

11月13日 21:50 JST

Local AI. Local compute. Open weights. Open source. The future of AI shouldn't be controlled by a handful of companies. It should be accessible to everyone who wants to build with it.

0 赞 · 0 转帖 · 0 回复

### [@jaywcjlove](https://x.com/jaywcjlove/status/2121207869592969348?ref=binnn.top)

11月13日 20:45 JST

Sikemux - 面向 AI 编程开发者的多 Agent 协作与会话管理工具 主要用于统一管理 Claude Code、Codex 等 AI 编程智能体，让不同 Agent 可以接力完成编程任务。 它的核心理念是：Claude Code 用完了，可以让 Codex 接着工作；不同 AI Agent 像一个团队一样协同开发。 👉 https://t.co/uVLmowGFkx https://t.co/TPyYyn7L7G

0 赞 · 0 转帖 · 5 回复

### [@laogeai](https://x.com/laogeai/status/2121120792562000170?ref=binnn.top)

11月13日 14:59 JST

收藏夹救不了你，能查的 wiki 才行。 karpathy-llm-wiki：把原始资料编译成带出处的知识页，兼容 Claude Code / Cursor / Codex 技能格式。 https://t.co/ZzPLLu7ziV

0 赞 · 0 转帖 · 0 回复

### [@cgartlab](https://x.com/cgartlab/status/2121237934661873820?ref=binnn.top)

11月13日 22:45 JST

这是一个慢慢更新的系列，节奏越快的时代，越值得去找到那些不怎么变的事物和规律。 https://t.co/DuXfC9DKAl

0 赞 · 0 转帖 · 1 回复

[返回 AI 快讯 →](https://binnn.top/flash/)