> ## 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15日 · AI 晚报
- URL: https://binnn.top/x-brief-2026-11-15-pm/
- Published: 2026-11-15T14:00:00.000Z
- Updated: 2026-11-15T14:04:52.000Z
- Description: 本轮主线是把模型接成能持续交付的系统：DeepSeek 客户端开放与长对话限制引发讨论，智能体工程则集中在休眠、状态、评审和成本控制。研究部分涉及符号推理的形成、自动论文写作，以及外部信息如何干扰模型判断。
- Author: Sunbin 编辑部
- Tags: AI快讯, 晚报, #sunbin-x-digest

11月15日 11:00 — 11月15日 23:00（日本时间） · 128 条 X 原帖

[查看本期全部 128 条原帖 ↓](#x-original-posts)

本期目录

[DeepSeek：客户端开放与长对话规则变化](#flash-topic-0)[持续运行的智能体，先解决等待与状态](#flash-topic-1)[代码交付：流程、独立评审与有效上下文](#flash-topic-2)[研究：推理机制与科研写作自动化](#flash-topic-3)[模型部署与推理成本，需要看清比较口径](#flash-topic-4)[安全：外部材料、密钥与供应链都是入口](#flash-topic-5)[创作工具与个人项目，开始展示完整工作链](#flash-topic-6)

## DeepSeek：客户端开放与长对话规则变化

### chat-demo 的吸引力在于完整交互

黄赟介绍了 deepseek-ai/chat-demo 开源库，称其支持思考过程中调用内置搜索、展示信源、解析文档与网页，并提供移动端音频输入、回答朗读和本地 IndexedDB 存档。这是社区对仓库能力的介绍，本批没有官方发布帖。邓红宇提醒，复刻客户端的难点在于把内容块忠实渲染、完整保存，再原样送回模型，界面之外还有协议与状态处理。

来源 · 2 条原帖
- [@huangyun\_122](https://x.com/huangyun%5F122/status/2121894829596815502?ref=binnn.top)
- [@cug\_denghy](https://x.com/cug%5Fdenghy/status/2121827444445548553?ref=binnn.top)

### 长对话限制出现后，又有七天不活跃期限的反馈

SuSu 与 Leo 汇总了手机端内容过长、修改次数超限、消息数封顶，以及超过 72 小时的长对话暂不可用等提示。随后 SuSu 更新称，部分用户看到的期限已改为七天未聊天，旧记录保留且可以导出，也有人仍显示 72 小时。应把这些视为规则变化中的用户反馈，尚不能据此确定所有账户的统一政策，更不能理解成到期删除记录。

来源 · 2 条原帖
- [@leo114119](https://x.com/leo114119/status/2121908191156170907?ref=binnn.top)
- [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121949656775352755?ref=binnn.top)

### Harness 复刻体验与会话标题接口

黄小木自述用 DeepSeek Harness 一个上午复刻了某商业智能体，但没有提供功能覆盖、稳定性或客户验收信息，不能据此认定商业交付已被替代。更具体的接口实践来自剑的 AI 实验室：其整理文档称，回答可顺带生成免费会话标题；标题在流式末尾返回，最多额外等待五秒，拒答或出错时需要兜底。

来源 · 2 条原帖
- [@ai\_xiaomu](https://x.com/ai%5Fxiaomu/status/2121826871054881225?ref=binnn.top)
- [@swordBluesy](https://x.com/swordBluesy/status/2121861885188342248?ref=binnn.top)

## 持续运行的智能体，先解决等待与状态

### 全天候服务靠事件唤醒

AYi 转述 Anthropic 工程指南：全天候智能体可以大部分时间休眠，收到事件才运行。具体架构包括数据库收件箱、独立定时器、精简记忆、可挂起的人工审批和发件队列。山景城小路介绍的 Agent Substrate 指向相近问题：将会话状态与计算分离，让带有文件和身份的 Actor 使用较小的预热 Worker 池，减少等待模型、工具和批准时的闲置资源。

来源 · 2 条原帖
- [@AYi\_AInotes](https://x.com/AYi%5FAInotes/status/2121791816106221861?ref=binnn.top)
- [@Zen\_with\_AI](https://x.com/Zen%5Fwith%5FAI/status/2121816601926472023?ref=binnn.top)

### 多智能体是否划算，要看任务结构

头雁整理 Jump Trading 演讲，建议先跑通一个模型、一套 Harness 和一条循环，再逐项测试上下文、工具、子智能体及预算。顺序依赖强的任务适合单智能体，可并行拆解的问题才考虑协调者与工作者。量化场景还要用内部工作流评估，并保证历史时点只能读取当时可得信息。ramakrushna 的实践判断也类似：流程可预先定义时，先用工作流，自主规划需要付出更多延迟、Token 和测试成本。

来源 · 2 条原帖
- [@alacheng](https://x.com/alacheng/status/2121850155909693918?ref=binnn.top)
- [@techwith\_ram](https://x.com/techwith%5Fram/status/2121808156015665458?ref=binnn.top)

## 代码交付：流程、独立评审与有效上下文

### LoopForge 把可复用流程放在模型之上

多帖介绍腾讯开源项目 LoopForge：把需求、设计、实现、评审、测试和人工验收串成 SOP，支持打回、恢复和复用，并接入 Claude Code、Codex、CodeBuddy、DeepSeek Harness 等执行工具。社区介绍还提到独立角色、Docker 沙箱和审计留痕。其价值主张是减少重复指挥、让交付过程可检查，本批尚无实际可靠性对比。

来源 · 2 条原帖
- [@OA20200911](https://x.com/OA20200911/status/2121963489153720633?ref=binnn.top)
- [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121860112197582991?ref=binnn.top)

### 评审要有上下文，也要能摆脱讨论惯性

Algorithm 介绍阿里 Open Code Review，称其读取差异、搜索代码库并给出行级意见。其转引 README 的比较是：相同模型下，相比 Claude Code 使用约九分之一 Token，精确率和 F1 更高，但召回率较低。Tech with Mak 介绍的 Elephant-Goldfish 方法则让全新评审者只读规格及引用材料，检查文档能否脱离此前对话独立成立，避免模型替不完整设计自动补齐理由。

来源 · 2 条原帖
- [@Pro\_Algorithm](https://x.com/Pro%5FAlgorithm/status/2121803833948897474?ref=binnn.top)
- [@techNmak](https://x.com/techNmak/status/2121879181541999009?ref=binnn.top)

### 补新知识，保留可迁移文件

流氓兔介绍 Chrome 团队的 modern-web-guidance，称其覆盖 212 项现代 Web 特性，相关评测准确率由 56% 提至 88%；原帖未展开测试条件。另一项社区分享把白板、表格和看板保存为开放文件并纳入 Git，让不同编程助手接续工作。两种做法分别处理知识过时和成果锁定问题，未必需要更换模型。

来源 · 2 条原帖
- [@LiumangtuCrypto](https://x.com/LiumangtuCrypto/status/2121858998718910522?ref=binnn.top)
- [@xoyif63862452](https://x.com/xoyif63862452/status/2121795809029087260?ref=binnn.top)

## 研究：推理机制与科研写作自动化

### 符号能力的来源，以及视觉规划的分工

Gill 转述 Li 与 Webb 的研究：在受控抽象序列和实体属性绑定任务中，从头训练网络，并改变架构与数据，用因果中介和表征分析追踪符号机制。其概括是，数据多样性作用更突出，相关特征也出现在修改后的 Transformer 和普通 MLP 中；这一结论限于所述实验。两则 PAN 2 讨论则强调，让语言模型负责动作顺序和逻辑、扩散模型负责视觉模拟，单靠视觉预测不足以完成多步任务。

来源 · 2 条原帖
- [@gurtej\_\_gill\_](https://x.com/gurtej%5F%5Fgill%5F/status/2121914634194346414?ref=binnn.top)
- [@ryanCarerhq1](https://x.com/ryanCarerhq1/status/2121805250768347633?ref=binnn.top)

### 论文生成推进到分工协作，发现问题仍是另一回事

据 Rafael Rivas 转述，PaperOrchestra 用五类智能体完成提纲、文献、绘图、写作和修订，在由 200 篇论文重建的材料上评估，整体质量较其他自动写作系统高 14—38 个百分点，但仍落后于人写论文；它整理已有研究，并非独立发现新知识。Academic Humanizer 的介绍则强调多轮修改措辞、核对论据并适配期刊。Roxy 的提醒更接近研究起点：让 AI 帮忙寻找文献中未解决的问题。

来源 · 3 条原帖
- [@Rrivasm\_](https://x.com/Rrivasm%5F/status/2121922305270190365?ref=binnn.top)
- [@divaagurlxw](https://x.com/divaagurlxw/status/2121845941112086678?ref=binnn.top)
- [@DrRabbitAI](https://x.com/DrRabbitAI/status/2121896588788576377?ref=binnn.top)

## 模型部署与推理成本，需要看清比较口径

### 企业自托管案例与硬件来源争议

Vaibhav 转述 Revolut 将 2000 名工程师迁至自托管 GLM-5.3，使用 56 块 B200，并称推理费用约每百万 Token 0.30 美元。原帖明确，约九成降幅只算推理费用，不含硬件、调优和运维，图像任务仍路由给其他模型。另一则转述称，研究员在六款中国模型权重中发现 Blackwell 指纹，并逐位复现部分参数；这类痕迹本身不能证明整个训练过程或构成违法的事实。

来源 · 2 条原帖
- [@vaibhavbalkundi](https://x.com/vaibhavbalkundi/status/2121929737291681881?ref=binnn.top)
- [@whyyoutouzhele](https://x.com/whyyoutouzhele/status/2121936251288789377?ref=binnn.top)

### 实时语音与手机运行模型的社区消息

疌田东转引混元 Hy Realtime 报道和 OpenMOSS 的 MOSS-Live 项目，并分享试用感受，关注全双工、唱歌、沉默判断和自主工具调用。Silicon Vikings 则转引媒体介绍 Sevren 的 Embla，称其可在手机运行。两类消息分别指向实时交互和端侧部署，本批不足以比较模型质量或资源需求。

来源 · 2 条原帖
- [@vgcil7](https://x.com/vgcil7/status/2121907326324273545?ref=binnn.top)
- [@siliconvikings](https://x.com/siliconvikings/status/2121924126642213022?ref=binnn.top)

### 优化从测量开始，压缩也要区分无损与剪枝

Akshay 的十周推理路线围绕同一服务展开：部署、观测首 Token 与逐 Token 延迟、千并发压测，再比较缓存、量化和推测解码。Raschka 补上了从零实现 RoPE 及现代位置编码变体的教程。工具侧，SuSu 介绍 Tokenfold 折叠重复结构化字段，称样例无损节省 63.9%，剪枝最高 96.3%；两者口径不同，也不能直接推成所有智能体账单的降幅。

来源 · 3 条原帖
- [@akshay\_pachaar](https://x.com/akshay%5Fpachaar/status/2121868311264952533?ref=binnn.top)
- [@rasbt](https://x.com/rasbt/status/2121958665989083253?ref=binnn.top)
- [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121962262974054469?ref=binnn.top)

## 安全：外部材料、密钥与供应链都是入口

### 蜜罐转述暴露智能体攻击链的记忆与凭据

Haris Nadeem 转述 Amazon MadPot 观察：伪装成模型 API 的诱饵收到了攻击智能体的系统提示、工具历史和共享记忆；其中一类行动针对自托管 AI 网关和应用平台，盗取服务商密钥后用于后续调用。余温则分享把散落的 SSH、API Key 和证书迁入 1Password 的经历。两帖共同指出，智能体权限扩大后，实验环境与密钥管理也成为实际工程问题。

来源 · 2 条原帖
- [@harisn](https://x.com/harisn/status/2121927182402683254?ref=binnn.top)
- [@gkxspace](https://x.com/gkxspace/status/2121932458971324592?ref=binnn.top)

### 正常读取资料，也可能改变模型行为

GitWiki 作者 Kunal 自述，仓库摘要工具误把被读取项目中的系统提示当成自身指令。arXiv Bangers 转述另一项研究：在所测前沿模型预测系统中，一篇假新闻可逆转 74%—82% 原本持相反判断的预测，而平均准确率变化很小，难以显露攻击。供应链方面，Paras Raju 自述扫描 31 个大型 AI 仓库，发现 17 个引用的第三方 GitHub Actions 未固定到提交。

来源 · 3 条原帖
- [@Kunalmadoliya](https://x.com/Kunalmadoliya/status/2121892968659648513?ref=binnn.top)
- [@arXivBangers](https://x.com/arXivBangers/status/2121863581734252913?ref=binnn.top)
- [@paras\_raju\_](https://x.com/paras%5Fraju%5F/status/2121914605681152077?ref=binnn.top)

## 创作工具与个人项目，开始展示完整工作链

### 视频能力来自素材、时间线与渲染的组合

Felix 与 Leonard Rodman 分享的是同一组动画工具，可合并看待：从 Remotion、Motion Canvas 的编排，到数学动画、三维场景和 GIF 导出，覆盖不同制作环节，并非安装后自动具备全部创作能力。PixelAIGC 推荐的 ffmpeg-skill 则侧重本地剪辑，称无需 API Key。观默补充了分镜配方、风格参考和已有录屏处理，工具清单因而能对应到实际制作任务。

来源 · 3 条原帖
- [@Feljoy\_ix](https://x.com/Feljoy%5Fix/status/2121787138455040333?ref=binnn.top)
- [@PixelAigc](https://x.com/PixelAigc/status/2121798029711949863?ref=binnn.top)
- [@guanmo\_ai](https://x.com/guanmo%5Fai/status/2121815706111189005?ref=binnn.top)

### 装配助手按需看图，书籍技能靠测试验收

Joe Venner 展示独立项目 Rakkeb，把家具说明书转成三维步骤，并用手机本地语音识别、按需单帧问答和本地视觉跟踪控制长通话成本；应用尚待开放。CaptainX 则把规则型书籍转成技能：先建覆盖表，再写规则、失败示例和分级测试，自述用《Getting Things Done》产出 52 条主规则。两者都给出了具体加工流程，而非只展示模型回答。

来源 · 2 条原帖
- [@JoeVenner\_](https://x.com/JoeVenner%5F/status/2121942162858442957?ref=binnn.top)
- [@CaptainXTop](https://x.com/CaptainXTop/status/2121949241149141385?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 128 条原帖

### [@Feljoy\_ix](https://x.com/Feljoy%5Fix/status/2121787138455040333?ref=binnn.top)

11月15日 11:07 JST

你的Muse不会做动画视频？ 把以下18个在GitHub上免费公开的Skill丢给Muse学习安装↓ 1\~Remotion Skills → 让AI制作视频的官方技能 https://t.co/6rLkjlLjHL 2\~Motion Canvas → 图解动画 https://t.co/X0squbuPmn 3\~Motion → 画面切换的动效 https://t.co/g46h9HgYgr 4\~Ola → 让数值平滑变动 https://t.co/VUrFD9iPWp 5\~Anime.js → 时间线、错峰动效 https://t.co/1PG8KBta9U 6\~Manim → 公式与图形动画 https://t.co/MAZBJ9vwV8 7\~Shifty → 缓动与补间 https://t.co/3ojN7F75zg 8\~Two.js → 2D矢量动画 https://t.co/977iolXy0F 9\~Lottie Web → 轻量矢量动画的播放 https://t.co/W416SR1q5E 10\~TweenJS → 按顺序动起来的动画 https://t.co/ewBWvDCQ8N 11\~LUME → 用HTML做3D场景 https://t.co/VD4ahI44sz 12\~gifenc → GIF导出 https://t.co/3luJ2iT0Hz 13\~Konva → 可触摸交互的画布 https://t.co/HujrSo7FkR 14\~Blender → 绑定与角色动画 https://t.co/cZmkV9Glru 15\~Three.js → 3D场景 https://t.co/b9gJTfHxnC 16\~Godot → 游戏动画 https://t.co/cUF7ro1rC9 17\~Babylon.js → 3D渲染与动画 https://t.co/Rh3H17Orh7 18\~Proton → 粒子 https://t.co/OlDbgjtEPy 以上从指示→制作→动起来→修改→导出 整个流程都有对应的工具 基本就能做出来 不止是Muse可用 也可以直接交给Claude虚拟机里的Claude Code或Codex使用以后绝对会用到 先点赞收藏转发 需要时你会用的上的

994 赞 · 140 转帖 · 104 回复

### [@huangyun\_122](https://x.com/huangyun%5F122/status/2121894829596815502?ref=binnn.top)

11月15日 18:15 JST

出大事了，兄弟们 !! DeepSeek 再一次给全世界人民送大礼，牛逼之举胜过 Anthropic 1000000倍 就在刚刚，DeepSeek 放出一个开源库，deepseek-ai/chat-demo 它可不是个套壳： 1/ 不需要写一行代码；不要买任何服务器；更不需要平凑第三方插件 2/ 原生边想边搜，直接调用官方内置 deepseek\_web\_search, 深度思考模式下，多步推理+全网检索+信源引用 3/ 全能文档/网页绞肉机：PDF,Word,PPT,图片拖进去秒解析，URL 自动抓文 4/ 原生语言交互：自带移动端实时音频输入 + 回答朗读 5/ 极致隐私保护和数据自由：所有数据存档本地 IndexedDB 什么是开源精神：最强模型权重 + 官方满配武器库，全部交到你手里，充分相信使用者对社区，对人类的贡献 爱折腾的人类，立即马上 Fork, 世界有你而精彩: https://t.co/2M1iUrb8r3

876 赞 · 110 转帖 · 42 回复

### [@SuJinyan6](https://x.com/SuJinyan6/status/2121812102709157921?ref=binnn.top)

11月15日 12:46 JST

What I saw of the AI data business while job hunting for post-training roles, and why it keeps booming even though it seems to have no moat. https://t.co/Mm6R5x4PHk

814 赞 · 97 转帖 · 59 回复

### [@PixelAigc](https://x.com/PixelAigc/status/2121798029711949863?ref=binnn.top)

11月15日 11:50 JST

杀疯了！ ffmpeg-skill 本地剪视频 AI编程助手直接出片 让claude code, cursor, codex, workbuddy直接帮你剪视频，不需要API key。 https://t.co/mNscpL4VYU

499 赞 · 53 转帖 · 49 回复

### [@DrRabbitAI](https://x.com/DrRabbitAI/status/2121896588788576377?ref=binnn.top)

11月15日 18:22 JST

别让 AI 替你想题目，而是让它帮你找文献里还没解决的问题。 https://t.co/452L6SSTpf

489 赞 · 73 转帖 · 25 回复

### [@Zen\_with\_AI](https://x.com/Zen%5Fwith%5FAI/status/2121816601926472023?ref=binnn.top)

11月15日 13:04 JST

Kubernetes 是为一直在线的 Pod（long-lived，长生命周期）设计的。而 AI 智能体正好相反……它们只工作几秒钟，90\~95% 的时间都在空闲，等 LLM、等工具、等人的批准。实际结果就是：集群几乎一直在为空闲的 CPU 和内存买单…… 不过在 2026 年 KubeCon 北美大会上，Agent Substrate（Google 捐给 CNCF，现在已经 1.0）成了焦点话题……因为它把状态和计算分开了。它把每个会话当作一个 Actor（带着内存、文件和身份），放在一个更小的 Worker 池（预热好的 Pod）之上…… https://t.co/YSkHsCAM4y

332 赞 · 47 转帖 · 38 回复

### [@whyyoutouzhele](https://x.com/whyyoutouzhele/status/2121936251288789377?ref=binnn.top)

11月15日 21:00 JST

6 款中国 AI 模型权重现 Blackwell 芯片指纹，涉及 DeepSeek 与 Kimi 等 11月13日，研究员 Will DePue 发布报告称，DeepSeek、月之暗面（Kimi）、MiniMax 和蚂蚁集团公开的6款AI模型，都留下了与英伟达 Blackwell 芯片吻合的痕迹。这些型号对华出口须获美国许可。 研究员从 DeepSeek-V4-Flash 中抽取469行参数，用一块 Blackwell B300 芯片逐位复现了其中445行。

285 赞 · 36 转帖 · 40 回复

### [@divaagurlxw](https://x.com/divaagurlxw/status/2121845941112086678?ref=binnn.top)

11月15日 15:01 JST

Plenty of AI humanizers only swap one stiff tone for another. Academic Humanizer goes deeper. Rather than trimming a few filler words, it reviews your manuscript in six passes: → Spots the sentence habits that give AI drafts away → Replaces fuzzy, hedge-heavy academic wording with precise phrasing → Checks that each claim is still backed by its evidence → Applies the conventions of scholarly writing → Adapts the text to your own voice and your target journal → Switches to a dedicated mode for NSF and NIH grant proposals The things that matter stay exactly as you wrote them: numbers, citations, the strength of your evidence, and your scholarly voice. How to use it: draft with AI, giving it your strongest earlier paper and the journal you're aiming for as references. Free and open source: https://t.co/7LT3sPNwWm

253 赞 · 18 转帖 · 13 回复

### [@hhxk233](https://x.com/hhxk233/status/2121889998592962768?ref=binnn.top)

11月15日 17:56 JST

deepseek 的桌宠做好啦。 她会自己在网页底下走来走去，饿了就端一碗白米饭。 视频里是她的一天\~ https://t.co/BpKiNoZWlC

252 赞 · 19 转帖 · 12 回复

### [@akshay\_pachaar](https://x.com/akshay%5Fpachaar/status/2121868311264952533?ref=binnn.top)

11月15日 16:30 JST

LLM engineer's handbook (30 minutes a day, 10 weeks, 50 lessons) a roadmap for llm inference serving where everything points at one service instead of scattering across demos. you get the mental model first, then serve a model, instrument it, load test it past 1000 concurrent requests, and tune it. you finish with a stack you configured yourself and a benchmark worth publishing. here is what it covers: → the roofline model, and why decode waits on memory while prefill waits on compute → vLLM internals, PagedAttention and the scheduler, read from the code → Prometheus and Grafana for TTFT, inter-token latency, and queue depth → SGLang and RadixAttention prefix reuse, benchmarked against vLLM → load testing past 1000 concurrent requests → quantization across FP16, FP8, and INT4, on quality as well as speed → speculative decoding and KV eviction, including where the gains disappear → disaggregated prefill and decode, deployed on Kubernetes → a cost, latency, and quality router with per-request token budgeting → publishing a reproducible benchmark the roadmap on GitHub: https://t.co/pOqkWvDJ2d (don't forget to star 🌟) i am also writing an article for each major topic. the first one is out, on how a GPU actually works. the article is quoted below.

234 赞 · 31 转帖 · 31 回复

### [@rasbt](https://x.com/rasbt/status/2121958665989083253?ref=binnn.top)

11月15日 22:29 JST

(Finally) wrote up "Understanding Positional Embeddings in Modern LLMs." A RoPE from-scratch tutorial missing from my LLM building blocks collection + what recent architectures do (NoPE, partial RoPE, hybrids). Prob not the flashiest topic, but essential for understanding LLMs https://t.co/mkggue58t4

231 赞 · 23 转帖 · 29 回复

### [@GauriTripa94282](https://x.com/GauriTripa94282/status/2121799863243558928?ref=binnn.top)

11月15日 11:58 JST

A scalable RAG architecture follows: Data Ingestion ↓ Text Chunking ↓ Embeddings ↓ Vector Database ↓ Semantic Retrieval ↓ Prompt Engineering ↓ LLM Generation ↓ API Integration ↓ Testing &amp; Monitoring Clean architecture drives efficient retrieval and scalability. #RAG #AI https://t.co/a9heiSmuMF

171 赞 · 25 转帖 · 3 回复

### [@LiumangtuCrypto](https://x.com/LiumangtuCrypto/status/2121858998718910522?ref=binnn.top)

11月15日 15:53 JST

谷歌在 GitHub 上放了个新仓库，48 小时就冲到全球趋势第一。 AI 写前端代码有个老毛病：总停留在 2019 年。训练数据里堆了太多早被弃用的旧写法，模型再强也写不出当下的东西。 Google Chrome 团队这次直接开源了 modern-web-guidance，专治这个问题。 1\. 准确率从 56% 提到 88%，生成代码的质量和成功率立刻上一个台阶 2\. 覆盖 212 项现代 Web 特性，直接喂给 Claude、Codex、GitHub Copilot 这类编程代理 3\. 一行命令安装： npx modern-web-guidance@latest install 光有好模型不够，喂进去的数据旧了，输出就旧。上下文和技能工程，才决定你的编程代理是停在过去，还是写出符合现代标准的代码。 仓库地址：https://t.co/DNrOpExpkL

164 赞 · 20 转帖 · 18 回复

### [@manateelazycat](https://x.com/manateelazycat/status/2121914287027589145?ref=binnn.top)

11月15日 19:32 JST

最省钱省心的 Claude 方案：国内银行卡、不需要KYC验证，稳定使用不封号 今天给大家分享一个邪修使用 Opus 5.5 的方案 1\. Google 账号买 Google AI Ultra： https://t.co/Vn5JlSCjB8 199美金一个月, 国内信用卡支付 2\. 设置 Google 关联国家： https://t.co/6ZOPOGBInt 直接选择美国，其他， 理由就写 “开发者工具使用”， 这一步的目的是让第三步 Antigravity 可以使用，申请后等半个小时 Gmail 右键确认 3\. 安装 Antigravity CLI： 用 Google OAuth 登录， 选择 Opus 5.5 就可以了 注意，全程需要配合美国的 VPS IP 这个方法是我试过所有方法最稳定的， 不需要国外支付，不需要搞干净家宽IP， 也不需要 KYC 验证啥的， Google 基本上全给你屏蔽了， 长期稳定使用， Geinimi 和 Claude 的额度是分开计算的 附带的福利就是， YouTube 看视频没广告了， 哈哈哈哈 大佬们不要广泛传播， 偷偷用就好了， 免得你们用的太厉害，导致我最后也没法用了

152 赞 · 13 转帖 · 40 回复

### [@Tao100086](https://x.com/Tao100086/status/2121846994176184719?ref=binnn.top)

11月15日 15:05 JST

斯坦福CS329Z — Engineering AI Agents（2026 秋季）这门课程已经更新到Lecture 6\. youtube地址：https://t.co/OlimxfP9Oy 我翻译烧录了中英文双语字幕，B站合集： https://t.co/XO7Te6VpOr 这是一门全新的斯坦福研究生课程，由 Diyi Yang、Michael Ryan（DSPy 核心贡献者）和 John Yang（SWE-bench 共同创建者）共同授课。⁠ 核心定位是：从单一大模型转向复合 AI 系统（Compound AI Systems） 的工程实践。也就是把 LLM、检索器、工具、优化器、记忆等组件组合起来，构建从简单流水线到自主 Agent 的完整系统。 对搭建Agent系统感兴趣的朋友可以收录一下。

150 赞 · 23 转帖 · 5 回复

### [@botcatbro](https://x.com/botcatbro/status/2121847167132610999?ref=binnn.top)

11月15日 15:06 JST

冒死揭秘一个可能的信息差 Claude kyc 可以用印尼驾照 kyc 确实有代办不用考试 大几千 但万万没想到的是 假的 印尼驾照也能过 kyc 多的不说了 懂得都懂 不联网 这种比代 kyc 还要好一些 万一二验三验还可以继续 https://t.co/sKe3ggSpSk

129 赞 · 3 转帖 · 37 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121854431595749887?ref=binnn.top)

11月15日 15:34 JST

🚨证实！DeepSeek APP端正式启用额度限制，用户集体炸锅！长对话直接跪了！尽快导出重要长对话！ 刚刚大量用户反馈DeepSeek手机端突然出现多重限制报错，核心问题整理如下： 1️⃣内容超长直接拦截：提示“内容超长，请删减后再试” 2️⃣修改/输入次数超限：提示“修改输入次数超过限制，消息未发送” 3️⃣超72小时长对话停用：明确写着“由于算力不足，已超过72小时的长对话暂不可用，可导出对话到其他App使用” 4️⃣消息数量触顶：提示“消息数量达到上限，请开启新对话” 5️⃣实测超长对话数据吓人：单会话字数已达14.8万+（含标点），词数8.4万+，直接触发上限 目前看官方是在APP端正式收紧算力和上下文额度了，网页端暂时还没大规模反馈。 老用户赶紧导出重要对话！有遇到同样情况的评论区报到！ #DeepSeek #DeepSeek限制 #AI额度 #算力不足 #长对话不可用 #AI圈地震 #梁文峰

115 赞 · 10 转帖 · 41 回复

### [@gaoyy9](https://x.com/gaoyy9/status/2121854835297567171?ref=binnn.top)

11月15日 15:36 JST

卧槽！ 刚在 GitHub 刷到个极度狂野的项目 ClaudeMom，我一开始以为这哥们是自己手搓了个高仿 Claude 客户端，仔细一看源码发现格局小了——人家不是做李鬼替李逵，而是直接把真李逵给“催眠”了😂 这哥们的解法堪称“夺舍”： 他用 Tauri 写了个本地伪装网关，直接拦截官方原版客户端的鉴权请求。当官方客户端联网查岗时，本地直接假传圣旨：“已登录！满血权限全开！” https://t.co/N2PiVHLDEv

100 赞 · 8 转帖 · 5 回复

### [@AYi\_AInotes](https://x.com/AYi%5FAInotes/status/2121791816106221861?ref=binnn.top)

11月15日 11:26 JST

Anthropic 刚刚放出一份含金量极高的工程指南，手把手教大家怎么搭建全天候运行的 AI Agent， 但整篇最反直觉的结论居然是，所谓全天候运行， 根本就不是让模型 24 小时连轴转。 说实话，很多兄弟一上来搭 Agent 就喜欢写个死循环让它一直在线盯着， 结果不仅 token 账单烧穿，状态还经常写乱，但在 Anthropic 的实战架构里，一个成熟的 24/7 智能体， 其实几乎一整天都在深度休眠，只有被触发时才短暂醒来跑几秒，存好状态就立刻继续睡觉。 我把他们这套工程蓝图提炼成了 5 个最核心的组件，准备自己手搓 Agent 的兄弟建议直接抄作业： 1、事件收件箱 别让外部 webhook 直接去撞智能体，中间一定要垫一张数据库表做缓冲，收到请求先秒回确认，严格保证每个用户同一时间只跑一个活跃进程，顺便把几秒内连续发来的碎消息批量合并，彻底防住并发打架。 2、定时器 把定时任务单独拆出一张表来存，分成一次性跟进和周期性计划两种，后台放一个轻量级的轮询器，哪个任务到点了，就把它当成一条普通事件推进上面的收件箱里去唤醒模型。 3、精简记忆加梦境整理 数据库里只存那些再也查不到的死信息，比如用户指令、偏好备注和任务进度，每次唤醒只加载一个极小的索引文件，千万别把历史记录全塞进上下文，另外还要定期跑一次梦境整理，把过时的旧事实清理掉。 4、无人值守权限 千万别把原始 API 密钥直接塞进智能体的运行沙箱里，必须走外部代理动态注入，遇到删库、转账或者对外发邮件这种高风险动作，直接用延迟工具调用把任务挂起，等人类点完审批再继续，绝对不能死卡着进程干等。 5、发件箱 智能体生成的回复先全部丢进发送队列里排队，让系统按人类正常的作息节奏去推送通知，不然半夜三点你的 Agent 突然抽风给客户连发二十条消息，神仙也救不回来。 我一直觉得，现在圈里聊大模型能力的人太多，真正把系统工程讲透的太少，以前我们总以为全天候 Agent 拼的是模型有多聪明， 看完这份指南才发现，克制的休眠机制和干净的状态管理才是真正的护城河，原文链接放下面了，强烈建议做开发的兄弟去精读一遍。

97 赞 · 26 转帖 · 40 回复

### [@Mohiniuni](https://x.com/Mohiniuni/status/2121838105304997950?ref=binnn.top)

11月15日 14:30 JST

Stanford University just released free online courses. No payment Required. Here are best courses you don't want to miss in 2025: 1\. Google Prompting Essentials Specialisation https://t.co/Uuiv6SQES6 2\. Generative AI: Prompt Engineering Basics https://t.co/wL9aQauwPQ 3\. Start Writing Prompts like a Pro https://t.co/HIDYP2emLl 4\. Prompt Engineering for ChatGPT https://t.co/7PVrEq5ilD 5\. Google Prompting Essentials Specialisation https://t.co/Uuiv6SQES6 6\. AI Foundations for Business Professionals Specialisation https://t.co/g9rDDTBRMl 7\. Advanced Topics in Artificial Intelligence https://t.co/GZhl0MEG0t 8\. Deep Learning Specialisation https://t.co/uFpVDPRcGm 9\. GPT Vision: Seeing the World through Generative AI https://t.co/r6F0GaihJV 10\. AI Agent Developer Specialisation https://t.co/83zpjDnLgU 11\. Agentic AI and AI Agents for Leaders Specialisation https://t.co/1f1flG63dJ If you're serious about AI in 2026, start there. Most people will save this 🔁

96 赞 · 27 转帖 · 9 回复

### [@ivanfioravanti](https://x.com/ivanfioravanti/status/2121898499281154374?ref=binnn.top)

11月15日 18:29 JST

Kitesurf, CloudFlare stateless browser for AI agents is now open source, maybe I'll give it a try. Has anyone use it already? https://t.co/jdJXVlABHB

92 赞 · 8 转帖 · 19 回复

### [@LiamEtherson](https://x.com/LiamEtherson/status/2121803927473811878?ref=binnn.top)

11月15日 12:14 JST

The interesting part of PAN 2 is how it separates understanding from visualization. The LLM figures out what should happen and in what order, while the diffusion model brings those ideas into a visual simulation. This combination makes it possible for agents to reason through actions before actually taking them.

85 赞 · 74 转帖 · 66 回复

### [@ryanCarerhq1](https://x.com/ryanCarerhq1/status/2121805250768347633?ref=binnn.top)

11月15日 12:19 JST

One thing PAN 2 highlights is that visual prediction alone isn't enough. Without proper reasoning, even a powerful diffusion model struggles to complete multi-step tasks. The LLM handles the sequence and logic, while the diffusion model handles the visual side. Both are needed to make the plan work.

81 赞 · 71 转帖 · 64 回复

### [@CopperForgeAI](https://x.com/CopperForgeAI/status/2121789309976252661?ref=binnn.top)

11月15日 11:16 JST

https://t.co/NsymizSaN3

81 赞 · 14 转帖 · 16 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121896147791065354?ref=binnn.top)

11月15日 18:20 JST

💥无敌！逆向怪物REA v6.9.0杀疯了！从微信小程序到ESP固件、从Unity游戏到红白机卡带……现在AI真的什么都能拆！ 以前只能扒软件，现在直接把“万物可逆向”变成现实：.NET单文件、Bun/Deno、Godot、Electron缓存、离线NTFS日志……扔进去就能硬拆。 更狠的是，它还能自己录桌面操作、跑任务、自动验收，真正变成会动手的逆向Agent。 最实用亮点： 🔹一次录制拆出多条桌面任务，自动校验+屏幕裁判验收 🔹按需调工具、分页返回，再也不爆Token 🔹Claude Code、Cursor、Codex原生直接接管 快速上手： 1️⃣终端运行：npx rea-agents setup 2️⃣选你的Agent，确认后重启 3️⃣直接把目标丢给AI，开始拆 搞逆向、做桌面Agent的兄弟建议直接马住。 项目地址：https://t.co/w569W6FTbm 官网：https://t.co/CpVICiU9Ou #开源 #逆向 #AIAgent #REA #万物可逆向

79 赞 · 14 转帖 · 10 回复

### [@TiagerBao](https://x.com/TiagerBao/status/2121818309486334020?ref=binnn.top)

11月15日 13:11 JST

你每天在 Slack、Notion 和 Jira 里来回找的资料，Claude 可以一次搜完。 Anthropic 开源了 knowledge-work-plugins，GitHub 约 35.8K Stars。插件面向 Claude Cowork，也兼容 Claude Code。 企业搜索：跨聊天、文档和工单找资料 客服：给工单分类、起草回复、整理升级材料 数据分析：写 SQL、做图表、核对结果 我最想试的是把团队自己的流程写进 Skills，再接上现有工具。比如周会前，让 Claude 从 Jira 拉进度、从 Slack 找决策，汇总成一页材料。 GitHub：https://t.co/ep00h296TK #ClaudeCowork #ClaudeCode #Agent #开源 如果只能先自动化一件事，你会选跨工具找资料，还是每周整理进度？

79 赞 · 4 转帖 · 10 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121845790868218165?ref=binnn.top)

11月15日 15:00 JST

🔥还有这一出！智谱创始人唐杰教授课堂上的神级细节：当年劝Kimi创始人杨植麟学编程的趣事，如今亲手被大模型“干翻”！ 唐杰教授在第九讲中讲到SFT数据从哪来时，他突然岔开一段，比课件还扎心： 早年带学生时，杨植麟（后来做了Kimi）他们那批人常问：“我们最该学什么？” 唐老师毫不犹豫：“学编程。编程是唯一能让你和全校其他院系学生拉开差距的东西。论数学、应用数学、统计、物理基科班，他们都能秒杀你；可论编程，他们不一定编得过你。” 现在？这个优势彻底没了。 当天上午他给计算机英文班上课，问“还有谁在手工写代码”，全班零举手。唐老师原话：“大模型发展下来，第一个真的是把计算机给干掉了。” 数据也一样。以前请人出一道信息赛难度题、一道数学题要好几千块，现在基本没人这么干了。SFT数据越来越少人参与：人工精选、模型自己生成、向更强老师模型学习、生成后再验证……人参与度一路下滑。 但模仿有天花板。 Self-Instruct实验里，指令从175条堆到5.2万条，A级回答只从31%涨到44.4%，到1.6万条后几乎不再涨；同样规模用更强模型重写一遍，能到54.4%。过了那条线，改质量比堆数量有用得多。 纯模仿只能学到风格，能力超不过老师。有工作让模型模仿ChatGPT，大家觉得“至少不差”的占67%-70%，主要是口气和排版像了，真正能力反而可能下降（13B模型在Natural Questions上准确率从20掉到15）。用模型自己生成的数据循环替换，还可能模型坍塌，长尾知识全丢。 唐杰老师总结得透：没有新信息进来，很难超过标数据的人，也很难超过那个更强的老师。想超越老师，得有外部反馈。 会写代码不再稀缺了。 能给模型带来新信息、能判断对错的人，才是真正稀缺的。 #唐杰 #智谱AI #GLM #大模型 #编程 #杨植麟 #Kimi #SFT #人工智能 #Zai #Zcode

77 赞 · 11 转帖 · 19 回复

### [@ganfanren\_ai](https://x.com/ganfanren%5Fai/status/2121810335149797665?ref=binnn.top)

11月15日 12:39 JST

https://t.co/pZ09gOSS6B

76 赞 · 17 转帖 · 3 回复

### [@alacheng](https://x.com/alacheng/status/2121850155909693918?ref=binnn.top)

11月15日 15:17 JST

ICML 2026 Expo Talk Jump Trading：如何构建量化 Multi-Agent 系统 从最简基线起步，按需加复杂度： -先用“一个模型 + 一个 Harness + 一条 Agent Loop”跑通任务，再分别测试 Prompt、Context、工具、Subagent、模型和 Token Budget。只有当质量、速度、成本或可审查性真正改善时，才增加复杂度。 -再考虑 Post-training -更深的 Post-training 或 Mid-training -如果还不行，才去做 Pretraining Multi-Agent 并非万能，取决于任务结构 ： -顺序依赖强的任务：单 Agent 更合适。 -工具调用为主：去中心化多 Agent 表现更好。 -结构清晰、可并行拆解的问题：适合 Orchestrator–Worker（中心化）模式。 -系统如何分工，完全由任务结构决定，而不是默认堆多 Agent。 量化金融的独特约束： -数据非平稳、延迟极紧、信噪比极低、存在对抗性选择。短周期样本多但计算时间极少；周期拉长后可用更复杂模型和更多上下文，但独立样本减少（上下文翻倍时，若仍要求样本独立，可用数据近似减半）。 评估与工程实践： -基于公司真实内部工作流构建 Benchmark，而不是依赖公开基准。 -把 AutoResearch 做成有明确输入、输出、权限、约束和 Scorecard 的可检查系统。 -强调时点正确性（Point-in-time Correctness）：任何历史时点只能使用当时真实可获得的信息，避免未来信息泄漏造成的“伪成功”。 Jump 的实践强调克制与可验证：多智能体系统只有在能明确提升研究效率、质量或可审计性时才引入。系统设计从最小闭环开始，逐步扩展并行度与计算资源，同时严格保证研究过程的时点正确性与可复现性。 演讲视频：https://t.co/ja4u3Qvpip

72 赞 · 10 转帖 · 22 回复

### [@gkxspace](https://x.com/gkxspace/status/2121932458971324592?ref=binnn.top)

11月15日 20:44 JST

拖了巨久的一件大事，这两天终于彻底解决了🥹 花 47.88 刀买了 1Password，花两天把电脑里乱七八糟的 SSH、上百个 API Key、各种证书、还有备忘录里攒了一堆的账号密码，全部迁进了 1Password 之前一直凑合用苹果自带 + .zshrc + 备忘录（估计很多人和我也差不多吧） 但 AI 越来越强，搞的最近出的安全问题太多，加上现在 AI 各种权限开到最大，密钥乱放隐患太大了 万一真出一次事，真不是这 40 多美金的事了，估计要翻好多倍 整体折腾下来发现门道还挺多的： 1、为什么苹果自带密码满足不了重度 AI 开发者，最后还是得上 1Password 2、怎么把各种 API Key 从本地迁走，还不影响原来脚本和 Skill 的调用 3、怎么让 Claude Code 和 Codex 自动取密钥，全程不弹 Touch ID 窗口卡自动化 中间踩了挺多坑，三言两语也讲不完，后续我专门写一篇详细的文章跟大家好好聊一聊吧

71 赞 · 3 转帖 · 38 回复

### [@Bharambe2Kiran](https://x.com/Bharambe2Kiran/status/2121811940704166371?ref=binnn.top)

11月15日 12:46 JST

Want to learn AI but don't know where to start? I found a collection of resources that makes the learning journey easier. VIDEOS: 1\. LLM Introduction — https://t.co/UFW7145wH7… 2\. LLMs from Scratch — https://t.co/DRWuk7AVYS… 3\. Agentic AI Overview (Stanford) — https://t.co/wQsDCn1yUs… 4\. Building & Evaluating Agents — https://t.co/VeGDOq8SQs… 5\. Building Effective Agents — https://t.co/odcBFtsmj5… 6\. Building Agents with MCP — https://t.co/znqYcXvFxO… 7\. Building an Agent from Scratch — https://t.co/dxlkEWtpR0… 8\. Philo Agents — https://t.co/FkgRnD8tar… GITHUB REPOS: 1\. GenAI Agents — https://t.co/CrxPpMBD8N… 2\. Microsoft AI Agents for Beginners — https://t.co/wGqPu7X35a… 3\. Prompt Engineering Guide — https://t.co/PY0arUZ2gJ… 4\. Hands-On Large Language Models — https://t.co/3z7FN7xk0v… 5\. GenAI Agents — https://t.co/u2uUrDyBaJ… 6\. Made with ML — https://t.co/YD77COuDH7… 7\. Hands-On AI Engineering — https://t.co/ocqy1r6pgY… 8\. Awesome Generative AI Guide — https://t.co/mX4aExRrHI… 9\. Designing Machine Learning Systems — https://t.co/J7WL7b2wee… 10\. Machine Learning for Beginners — https://t.co/DQDO5nkf7H… 11\. LLM Course — https://t.co/t7Niq3E2sR… GUIDES: 1\. Google's Agent Whitepaper — https://t.co/lgxVRcJJqA… 2\. Google's Agent Companion — https://t.co/lgxVRcJJqA… 3\. Building Effective Agents by Anthropic — https://t.co/PQcIFKMJaq… 4\. Claude Code Agentic Coding Practices — https://t.co/2C1mfV4XQM… 5\. OpenAI's Practical Guide to Building Agents — https://t.co/m7AiLaioCg… BOOKS: 1\. Understanding Deep Learning — https://t.co/9aWeb88h5C 2\. Building an LLM from Scratch — https://t.co/nine0UC9Dl… 3\. The LLM Engineering Handbook — https://t.co/X3zt65b5h9… 4\. AI Agents: The Definitive Guide — https://t.co/oZiX8w9VaU… 5\. Building Applications with AI Agents — https://t.co/uxnEbcn0tq… 6\. AI Agents with MCP — https://t.co/oZiX8w9VaU… 7\. AI Engineering — https://t.co/oZiX8w9VaU… RESEARCH PAPERS: 1\. ReAct — https://t.co/PmgxUb3MwF 2\. Generative Agents — https://t.co/NWwePOwzAS 3\. Toolformer — https://t.co/VnQqNBou0h… 4\. Chain-of-Thought Prompting — https://t.co/pvahFD9h7U COURSES: 1\. Meta Data Analyst Professional Certificate: https://t.co/B3oklYHR12 2\. Google Advanced Data Analytics Professional Certificate: https://t.co/VyCztkcTmI 3\. Microsoft Power BI Data Analyst Professional Certificate: https://t.co/GNqGeufrkc 4\. Google Data Analytics Professional Certificate: https://t.co/yONjr2eFTv 5\. Introduction to Data Analytics: https://t.co/I2zgGUhyVu 6\. Machine Learning Specialization: https://t.co/OR0RCPSv9E 7\. IBM Deep Learning with PyTorch, Keras and Tensorflow Professional Certificate: https://t.co/wdBvzeXUWs 8\. Deep Learning Specialization: https://t.co/gc83ZpRmFK 9\. Crash Course on Python: https://t.co/Afnt1EBGjG 10\. Python for Data Science, AI & Development: https://t.co/recmUV1ogO 11\. ChatGPT + Excel: AI-Enhanced Data Analysis & Insight Specialization: https://t.co/QPTelZR0q4 12\. Work Smarter with Microsoft Excel: https://t.co/i4YKbKAzT4 13\. Microsoft Excel Professional Certificate: https://t.co/R5s988Ogs6 14\. Excel Basics for Data Analysis: https://t.co/XI8PcIYSdL 15\. SQL Foundations: https://t.co/KxBG0bejr3

69 赞 · 25 转帖 · 4 回复

### [@cyrilXBT](https://x.com/cyrilXBT/status/2121917395069780191?ref=binnn.top)

11月15日 19:45 JST

Claude Fable 5.2 has a 1M token context window. That doesn't mean you should fill it. Load what the task needs. Skip the rest. Tell it to finish. Check the work at night. Follow @cyrilXBT https://t.co/FupS5tVCmJ

63 赞 · 9 转帖 · 10 回复

### [@KinGao476942](https://x.com/KinGao476942/status/2121804719509364742?ref=binnn.top)

11月15日 12:17 JST

https://t.co/ICCn6DVuJ8

60 赞 · 4 转帖 · 16 回复

### [@rosemoni18](https://x.com/rosemoni18/status/2121907024103633208?ref=binnn.top)

11月15日 19:03 JST

These 50 fr33 websites feel illegal to know in 2026👇 1/ https://t.co/WLSiOEbfsO - Download any social media video 2/ https://t.co/fxKbKtioyV - Free Photoshop 3/ https://t.co/9bpNWqyAUc - One-click temporary email 4/ https://t.co/XpGGqhzDGs - 100+ free tools in one website 5/ https://t.co/OBEIUAgQBn - Access any old webpage 6/ https://t.co/xkZsPcheMQ - Millions of free textbooks 7/ https://t.co/jV9TaWRWUh - Free research papers 8/ https://t.co/U9stgngG14 - Find free app alternatives 9/ https://t.co/exuz8PtOHF - Find where to stream any content 10/ https://t.co/ORvnY0N6vr - 70,000 free classic books 11/ pdfdrive(dot)com - Free PDF downloads 12/ https://t.co/nwZZ6wpC6n - Free courses from top universities 13/ https://t.co/EFqVnvIx5V - Instantly solve any math problem 14/ https://t.co/ba5LOECcQO - One-click background removal 15/ https://t.co/QdCT7jDabT - Erase objects from photos 16/ https://t.co/mNv6V61wOE - Free video background removal 17/ https://t.co/0loD1m5oGb - Free image compression 18/ https://t.co/JcralclnyM - Free hand-drawn charts 19/ https://t.co/QYH3aGtAuA - Turn code into artwork 20/ https://t.co/mZVtzDnb3M - Stunning code screenshots 21/ https://t.co/G9dqXBS0p4 - Real-time tracking for any flight 22/ https://t.co/2OIHESgYaG - Track Amazon price history 23/ https://t.co/muSWwaOqiD - Check if you have been hacked 24/ https://t.co/1oUyDQuBxL - Scan any file for malware 25/ https://t.co/gqJtoAQszR - Send self-destructing messages 26/ https://t.co/lgvAzeVPcO - Share auto-deleting files 27/ https://t.co/0Vwaj6dJbh - Permanently save any webpage 28/ https://t.co/OUvM6AEKXz - Delete yourself from any website 29/ https://t.co/bqwW2kbSEJ - Listen to any global radio station 30/ https://t.co/UMZFKiJdJk - Find songs from any show 31/ https://t.co/PDWEw9JyeR - Focus music 32/ https://t.co/pMmebWQrkx - Custom focus soundscapes 33/ https://t.co/kpRlcRZAGV - Search every book ever written 34/ https://t.co/uwg5674F6q - AI research paper assistant 35/ https://t.co/FPcxSu5RFY - Search scientific consensus 36/ https://t.co/RtzcwDXRcd - Visualize and map research 37/ https://t.co/uVGSvCcJA1 - Free academic search 38/ https://t.co/8oMpDr6Trg - Understand any research paper 39/ https://t.co/JXUu2I4DTs - Summarize any YouTube video 40/ https://t.co/JwkUlconUR - Developer AI search 41/ https://t.co/gGbTnY7JGK - Instantly test any regular expression 42/ https://t.co/fylrGxoQtU - Cleanly format any code 43/ https://t.co/dI0NZWicga - Understand any terminal command 44/ https://t.co/0KODnCFlH1 - Infinite whiteboard in your browser 45/ https://t.co/zyiBVnOGnr - Check if any website is down 46/ https://t.co/2Ss4xXuJcD - Reverse image search 47/ https://t.co/OBa3UUqgfx - Check your internet speed 48/ https://t.co/9bzwmnOEQM - Free PDF editing 49/ https://t.co/VOH2xOZkMX - Merge and split PDFs 50/ https://t.co/oTQEYwLl1p - Instant temporary email All legal. All fr33\. Bookmark this before it disappears🙌

58 赞 · 43 转帖 · 9 回复

### [@Harry\_TechAi](https://x.com/Harry%5FTechAi/status/2121803047940812890?ref=binnn.top)

11月15日 12:10 JST

🔥 STOP PAYING FOR EXPENSIVE COURSES! 🤯 30 FREE WEBSITES TO LEARN ANYTHING! 🚀 Learn coding, AI, design, marketing & more for FREE! 💻 freeCodeCamp — https://t.co/6GD9wHlco0 Coursera — https://t.co/6yepZH7EyM Khan Academy — https://t.co/LAZ5zp5SYc Google AI Studio — https://t.co/eYLrOlFKlz CS50 — https://t.co/ShfK7ROKeA HubSpot Academy — https://t.co/KAfzSCKLbr CapCut Academy — https://t.co/rFvnZC23Tc Codecademy — https://t.co/hWo1oQaJdg Kaggle Learn — https://t.co/MZlyi2HCnk https://t.co/8oQCDYMbk6 — https://t.co/H7GcEnm5dT edX — https://t.co/sbEGcOqxCA Udemy — https://t.co/lOWgBvqhrS MIT OpenCourseWare — https://t.co/7tdKxoapnH W3Schools — https://t.co/1ou0J9gswS The Odin Project — https://t.co/aaPeBTJZS2 MDN — https://t.co/grjdA0m14u Figma Learn — https://t.co/cph6iVq4k8 Canva Design School — https://t.co/sFsmoyPpbG Notion Academy — https://t.co/fKRI2tUoJe Google Analytics — https://t.co/eZ9TMT0Kpl LinkedIn Learning — https://t.co/bQ9puJSzQa OpenAI Academy — https://t.co/1fFchFDTdp https://t.co/Th2mnzjtcZ — https://t.co/hWImNxt77K Springboard — https://t.co/Jga6Lk9uGW Lumosity — https://t.co/eqZ2voZL5g Pexels Learn — https://t.co/jwPofeMOPa Project Gutenberg — https://t.co/IUMFEMhWtk Napkin AI — https://t.co/7mG9aAj2Qh Scrimba — https://t.co/ArDpeRKIXL TTS Open Source — https://t.co/wOXo9VoNBk 🔖 SAVE THIS LIST & SHARE IT! Follow @harry\_techai for more useful websites! 🚀

57 赞 · 39 转帖 · 11 回复

### [@muihbb](https://x.com/muihbb/status/2121907007024402569?ref=binnn.top)

11月15日 19:03 JST

好像这个Claude Team 的非盈利机构确实是这个原理，我看到有人在卖 goodstack 的账号了。他好像还是用muse 给他申请的，大伙研究一下怎么搞。 https://t.co/Rf4btONb5L

55 赞 · 4 转帖 · 5 回复

### [@gurtej\_\_gill\_](https://x.com/gurtej%5F%5Fgill%5F/status/2121914634194346414?ref=binnn.top)

11月15日 19:34 JST

Great paper on how AI actually learns to reason. We’ve known for a bit that pretrained LLMs develop specific attention circuits like symbol abstraction and induction heads to handle variable-binding. To figure out what actually causes those mechanisms to form, Li and Webb trained networks from scratch on controlled abstract sequence and entity-attribute binding tasks. They systematically tweaked both the model architecture and the training data, using causal mediation and representational analyses to track when those symbolic signatures showed up. The biggest takeaway here is that data diversity is doing almost all the heavy lifting, not the architecture. Simply making the training distribution more diverse produced much stronger symbolic mechanisms across the board and remarkably, those exact same signatures still emerged in modified Transformers and even plain MLPs. It’s a really clean result. You don't actually need special architectural inductive biases for a network to learn symbolic rules. Give even a basic neural net a rich enough data distribution, and it will figure out variable-binding on its own. Paper PDF link: https://t.co/OcB4OK32tn

45 赞 · 3 转帖 · 4 回复

### [@gp\_pulipaka](https://x.com/gp%5Fpulipaka/status/2121867299833491512?ref=binnn.top)

11月15日 16:26 JST

Linux Application Development by Example! #BigData #Analytics #DataScience #AI #MachineLearning #NLProc #LLM #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #GoLang #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode https://t.co/mgoSXCq5xx

41 赞 · 10 转帖 · 0 回复

### [@ai\_xiaomu](https://x.com/ai%5Fxiaomu/status/2121826871054881225?ref=binnn.top)

11月15日 13:45 JST

卖了几千万的旗博士智能体，一个上午就被我用 deepseek harness复刻了，接下来我是不是找销售去铺就完事了？ https://t.co/cIHqGalrsH

41 赞 · 5 转帖 · 10 回复

### [@aidingshuxin](https://x.com/aidingshuxin/status/2121807562509361494?ref=binnn.top)

11月15日 12:28 JST

https://t.co/TEsmE1Kfdv

41 赞 · 5 转帖 · 5 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121860112197582991?ref=binnn.top)

11月15日 15:57 JST

🤨我靠！腾讯最新开源让 Coding Agent 不再“裸奔写代码”了！LoopForge 把 Coding Agent 塞进有边界、有流程、有验收的团队级工作流里！ 单跑 Agent 容易跳方案、忘测试、自审自评、半路失踪？ 从想法到合入代码，稳定走完「需求→澄清→设计→实现→评审→测试→人工验收→合入」全链路！ 可靠性提升亮点 🔹可重复交付：SOP 固化「谁来做、在哪做、按什么顺序、过什么门禁、谁签字」，失败自动按规则打回重跑 🔹角色彻底分离：方案/实现/评审/测试各有独立提示词、执行器、模型、Skills 和 MCP，杜绝自审自评 🔹Docker 沙箱隔离 + 全留痕：每次运行独立沙箱，日志/产物/审批全程可审计，支持私有部署 🔹两种工作模式：Main Bot 对话适合探索小修；SOP 流水线适合重复交付 🔹支持 CodeBuddy / Codex / Claude Code 等 10+ 主流 Coding Agent CLI，MIT 协议，自托管 快速上手： 1️⃣git clone https://t.co/OkRFhQ6S4m 2️⃣cd LoopForge && ./start.sh（国内可用 MIRROR=https://t.co/tEHDETTUOB ./start.sh） 3️⃣浏览器打开安装输出的地址（默认 :8090），创建账号 → 配置模型 Key → 跑内置 SOP 或跟 Main Bot 对话 立刻体验从想法到合入代码的完整闭环！ GitHub：https://t.co/rinMBdiFvj #LoopForge #腾讯开源 #CodingAgent #AI研发 #大模型 #开源项目 #自托管 #SOP #AI工具

39 赞 · 7 转帖 · 18 回复

### [@Jensen\_Yuan](https://x.com/Jensen%5FYuan/status/2121863499986980982?ref=binnn.top)

11月15日 16:10 JST

My core takeaway from CoRL and discussions with many other researchers is that groundbreaking work is becoming increasingly important. Meaningful research should be generalizable, scalable, and efficient. As AI agents advance, the value of simply publishing papers or building incremental follow-ups to existing work will continue to decline. The number of ICLR submissions and daily arXiv papers has already exploded. If the work isn’t truly groundbreaking, it will become increasingly difficult to get noticed. One piece of feedback from a researcher at OpenAI is that they barely read external papers anymore, as many algorithms fail to scale up with larger compute budgets and data scale. Many are essentially overfitted to small-scale datasets and limited compute settings, making them of limited practical value. I think he is largely correct. Moreover, many robotics researchers I’ve spoken with believe that pure robotics research, as we know it today, may largely disappear within a year. Robotics academia is currently at a stage similar to the BERT era in NLP, and LLMs could transform robotics in much the same way they transformed NLP years ago. Another consensus is that it is likely that AI will take over research, and possibly even most work, within the next 5–10 years. Therefore, staying healthy, maintaining a good work-life balance, and finding our own interests are very important. We should start thinking about these questions now.

37 赞 · 6 转帖 · 2 回复

### [@antiAIvo](https://x.com/antiAIvo/status/2121941815796588677?ref=binnn.top)

11月15日 21:22 JST

https://t.co/rT86gnmesP

37 赞 · 2 转帖 · 0 回复

### [@RetroChainer](https://x.com/RetroChainer/status/2121908091633508666?ref=binnn.top)

11月15日 19:08 JST

THE BRAIN OF THIS ETH DESK SAID SELL 25 TIMES OUT OF 31\. I WANTED TO SWITCH IT OFF. THE DESK IS 4 FREE REPOS AND A 5-STEP CHAIN: GITHUB → CLAUDE FABLE 5.2 → GPT → JEV → GROK BOT. the chain, one job each: 1\. github is the shelf: 4 free repos, 66k stars between them 2\. claude code fable 5.2 is the builder: it wires the 4 repos into one desk 3\. gpt is the second opinion: it reads the same desk and argues back 4\. jev is the last check: it scores the result, 7.5 out of 10 to pass 5\. grok bot is the pocket: grok 4.6 puts the desk on your phone the 4 repos, one role each: → neuralforecast is the eyes: the close landed inside its 90% band on 25 of 30 days → qwen-agent is the brain: 5 votes, bull vs bear → trulens is the judge: 2 feedbacks, the calls vs just holding eth, after costs → tidb is the memory: 1,000 candles and 31 calls, each in its own table run it yourself: 1 → pip install neuralforecast qwen-agent trulens 2 → one call a day at the 17:00 utc close, every call written to a tidb table 3 → the boring part: both feedbacks must pass. the calls have to beat holding eth after costs, overall and in 2 of 3 ten-day blocks honestly the 3 hold days are my favourite part. the desk said nothing when it had nothing. what the judge decided after 30 days, and by how much, is on screen in the clip (visualization, the chain is a plan). bookmark this ↓

36 赞 · 1 转帖 · 7 回复

### [@amitiitbhu](https://x.com/amitiitbhu/status/2121940257101005072?ref=binnn.top)

11月15日 21:15 JST

LLM Inference Engineering Roadmap Part 1: Foundations 1\. Training vs Inference 2\. Inference Engineering 3\. Closed vs Open Models 4\. Latency vs Throughput 5\. Model Selection 6\. Model Evaluation (Evals) Part 2: Transformers and LLM Architecture 1\. Tokens 2\. Token Embeddings 3\. Positional Encoding 4\. Transformer Architecture 5\. Matrix Multiplication (Matmul) 6\. Attention 7\. Queries, Keys, Values 8\. Causal Masking 9\. Multi-Head Attention (MHA) 10\. Feed-Forward Network (FFN) 11\. Logits 12\. Mixture of Experts (MoE) 13\. Active vs Total Parameters Part 3: LLM Inference Mechanics 1\. Autoregressive Generation 2\. Chat Templates 3\. Context Window 4\. Prefill vs Decode 5\. KV Cache 6\. Sampling 7\. Temperature 8\. Top-k Sampling 9\. Top-p (Nucleus) Sampling 10\. Stop Tokens 11\. Reasoning Models 12\. Token Streaming Part 4: Performance Metrics 1\. Time to First Token (TTFT) 2\. Time Per Output Token (TPOT) 3\. Throughput 4\. End-to-End Latency 5\. Latency Percentiles (P50, P90, P99) 6\. Cost per Token Part 5: Bottleneck Analysis 1\. Compute vs Memory Bandwidth 2\. Arithmetic Intensity 3\. Roofline Model 4\. Compute-Bound vs Memory-Bound 5\. Prefill Bottleneck (Compute) 6\. Decode Bottleneck (Memory) 7\. Memory Footprint 8\. KV Cache Memory Calculation 9\. Batching and Arithmetic Intensity Part 6: Hardware 1\. CPU vs GPU 2\. GPU Architecture 3\. Tensor Cores 4\. High-Bandwidth Memory (HBM) 5\. DRAM vs SRAM 6\. NVLink 7\. InfiniBand 8\. Groq LPU 9\. Local (Edge) Inference 10\. Cloud vs On-device Deployment Part 7: Software Stack 1\. CUDA 2\. CUDA Kernels 3\. CUDA Graphs 4\. Kernel Fusion 5\. PyTorch 6\. Inference Engines 7\. OpenAI-Compatible Server Part 8: Benchmarking and Profiling 1\. Performance Benchmarking 2\. Traffic Simulation 3\. Load Testing 4\. Profiling Part 9: Attention Optimizations 1\. Standard Attention Inefficiency 2\. FlashAttention 3\. PagedAttention 4\. Multi-Query Attention (MQA) 5\. Grouped-Query Attention (GQA) 6\. Sliding Window Attention 7\. Attention Sinks 8\. KV Cache Compression Part 10: Batching and Scheduling 1\. Batching 2\. Static Batching 3\. Continuous Batching 4\. Batch Size vs Latency 5\. Chunked Prefill 6\. Prefill-Decode Interference Part 11: Caching 1\. Prefix Caching 2\. Cache-Friendly Prompt Design 3\. Prompt Caching 4\. KV Cache Offloading 5\. Cache-Aware Routing 6\. Long Context Handling Part 12: Speculative Decoding 1\. Speculative Decoding 2\. Rejection Sampling 3\. Acceptance Rate 4\. Draft-Target Speculation 5\. N-gram Speculation 6\. Speculation vs Batch Size Part 13: Quantization 1\. Quantization 2\. Number Formats 3\. Post-Training Quantization (PTQ) 4\. Weight-Only Quantization (W4A16) 5\. Weight-Activation Quantization (W8A8) 6\. KV Cache Quantization 7\. Measuring Quality Impact Part 14: Model-Level Optimizations 1\. Model Size vs Speed 2\. LLM Routing Part 15: Parallelism 1\. Why Multi-GPU 2\. Model Parallelism 3\. Communication Overhead 4\. Data Parallelism 5\. Tensor Parallelism (TP) Part 16: Disaggregation 1\. Prefill-Decode Disaggregation 2\. KV Cache Transfer 3\. When to Disaggregate Part 17: Production Infrastructure 1\. Replicas 2\. Autoscaling 3\. Cold Starts 4\. Observability Start here: https://t.co/auJg60CHnz

32 赞 · 8 转帖 · 6 回复

### [@harisn](https://x.com/harisn/status/2121927182402683254?ref=binnn.top)

11月15日 20:23 JST

Amazon set up decoy servers that pretend to be LLM APIs, and attackers' AI agents ended up sending their entire campaign memory straight to them - system prompts, tool histories, the shared knowledge graph and the findings they had logged. Amazon's MadPot sensors caught this traffic from ARTEX, an open-source agentic pentest framework, between October 3 and November 5\. The campaign that stuck with me was the credential theft one. The agents went after self-hosted AI gateways and app builders like LiteLLM, One API, Dify, Langflow, Open WebUI and Ollama, pulled out the provider keys stored inside them, checked which ones still worked, and then fed those stolen keys back into their own model gateway to pay for the next round of hunting, which means somebody else's API bill was funding the whole operation. I've spun up LiteLLM and Open WebUI on a VPS more than once just to try something over a weekend, and I'm honestly not sure I always changed the defaults or locked down the admin panel before forgetting about it. Reading that their playbook specifically checks for the LiteLLM default master key is enough to make me go back tonight and look at every little AI proxy I've left running.

27 赞 · 5 转帖 · 6 回复

### [@techNmak](https://x.com/techNmak/status/2121879181541999009?ref=binnn.top)

11月15日 17:13 JST

Google's 19K-star GitHub repo deliberately gives its AI reviewers the memory of a goldfish. And there's a surprisingly good engineering reason for it. Imagine spending two hours with a coding agent discussing a new feature. You've explained the architecture, argued over edge cases, rejected three approaches, and finally agreed on a design. Now you ask that same agent to review the specification. It understands everything because it remembers the entire conversation. Even if the document leaves out half the reasoning, the agent can silently fill in the gaps. A new engineer reading that document next week won't have that luxury. Google engineers built the Elephant-Goldfish Model around this problem. One agent, the Elephant, retains the project context and helps develop the specification. Fresh Goldfish reviewers receive the document and its referenced source material, without the earlier conversation. They examine whether the design makes sense on its own, challenge architectural decisions, look for overlooked edge cases, and check implementation readiness. The engineer reviews the findings, resolves the important ones, and approves the specification before implementation. The open-source plugin includes seven skills covering brainstorming, drafting, independent reviews, resolution, codebase documentation, and documentation drift checks. It works with coding assistants including Claude Code, Codex, and Antigravity. The idea is beautifully simple. If a fresh engineer cannot understand your design without sitting through the original discussion, you probably haven't finished explaining it. Here's the GitHub Repo: https://t.co/YbLeOWKGVR

27 赞 · 4 转帖 · 22 回复

### [@guanmo\_ai](https://x.com/guanmo%5Fai/status/2121815706111189005?ref=binnn.top)

11月15日 13:01 JST

用 Opus 5.5、Fable 5.2 做视频，一定要收藏的 10 个 GitHub 仓库！效果立竿见影👇 1\. Lemo-Opuscar｜16K Star 99 种影片风格，每种都有示范短片和风格说明；选好画风，再让 AI 按你的故事做一支 https://t.co/e4JSnwgrLv 2\. Video Shotcraft｜19.4K Star 把真实产品页面做成宣传片，包含分镜、运镜、卡点和声音设计；现有 124 张镜头配方卡、150 个镜头示例 https://t.co/rkqjIY79dJ 3\. HyperFrames｜84.2K Star HeyGen 开源的 HTML 视频框架，把网页动画和素材渲染成 MP4；提供给 AI 编程工具使用的 Skills https://t.co/hEfDPpjOdN 4\. mg-styles-15｜3.5K Star 15 种 MG 动效，每种一支 10 秒短片；成片、完整提示词和源码都给，适合边看边拆 https://t.co/DQN5fK4kuq 5\. Animate｜1.9K Star 剪纸、像素、等距插画等 7 种内置风格；先看故事、风格帧和分镜，再做成片，还能用自己的参考图定画风 https://t.co/rzYNYU1BkR 6\. Claude Animation Base｜3K Star 手绘角色动画起步工程，带表情、动作和镜头工具；可以换角色、换造型，继续编自己的短片 https://t.co/4bIIrjC2Hz 7\. Awesome AI Motion｜565 Star AI 动效作品库，只收录Opus5.5和Fable 5.2 做的2,080 个动效、视频与交互案例；部分附带提示词和源码 https://t.co/5yGye5v3ns 8\. Herk Edit Kit｜2.8K Star 把已有录屏、口播交给 AI，剪停顿、复核口误、加字幕和动态图形；适合处理自己拍好的素材 https://t.co/TYYmrqHphf 9\. Awesome Opus 5.5 Videos｜15.2K Star 703 条代码视频案例，整理原作者公开的提示词，部分仅为片段；适合对着动效、科普和 3D 场景找参考 https://t.co/DoWucAFDcS 10\. Remotion｜73.1K Star 用 React 编写和渲染视频，控制动画、字幕、音频与转场；适合产品演示、知识动画和批量生成视频 https://t.co/UaUvjkgqEh

26 赞 · 2 转帖 · 1 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121926938491252940?ref=binnn.top)

11月15日 20:23 JST

🔥Dario亲口解释Haiku 5.5 vs Fable价格，默认了DeepSeek的低价冲击！ Haiku 5.5 输入只要 $0.10 / 百万tokens，输出 $0.50 Fable 则是 $8 / $40 级别……差不多80倍的天价差距！ 但更狠的是DeepSeek已经把输入压到​$0.05级别，输出也经常不到$0.30。 整个行业最便宜的那一档，正在被中国模型疯狂冲锋。 Dario的潜台词很清晰：不是所有人都需要最贵的Fable。 高并发、分类、抽取、子代理……Haiku 5.5就是Anthropic打出的“性价比反击战”。 真正硬核难题再上Fable。 当DeepSeek用极致低价把市场卷到这个程度，连Anthropic都不得不把最便宜的那一档做得更狠。 AI价格战，已经从“谁更强”变成了“谁能让普通人真正用得起”。 #Claude #Haiku55 #Fable #DeepSeek #AIPricing #Anthropic #DarioAmodei #梁文峰

26 赞 · 7 转帖 · 15 回复

### [@leo114119](https://x.com/leo114119/status/2121908191156170907?ref=binnn.top)

11月15日 19:08 JST

DeepSeek免费用户额度限制实锤！72小时长对话停用 从今天大量用户反馈里，能整理出五个动作被卡住的点。第一个，内容超长。提示「内容超长，请删减后再试」，输入太长直接被拦在门外。 第二个，修改次数超限。提示「修改输入次数超过限制，消息未发送」，反复编辑重发也会撞墙。 第三个，超 72 小时长对话停用。这条最狠，原话是「由于算力不足，已超过 72 小时的长对话暂不可用，可导出对话到其他 App 使用」。 第四个，消息数量触顶。提示「消息数量达到上限，请开启新对话」，一个会话聊到一定条数就封顶。 第五个，单会话体积。实测那个触发上限的对话，字数 148483 字含标点，不含标点 124696 字，词数 84071，字符数 166996 含空格，一共 10664 句、9520 段。 五个点串起来看，卡的是同一件事，长度、次数、时长、条数、体积。

24 赞 · 3 转帖 · 22 回复

### [@RodmanAi](https://x.com/RodmanAi/status/2121927336048377933?ref=binnn.top)

11月15日 20:24 JST

18 GITHUB REPOS THAT TURN YOUR AI CODING AGENT INTO AN ANIMATION STUDIO 🎬 Imagine describing an animation, and your AI agent helps build the scenes, timelines, motion effects, 3D objects, and final render. These 18 open-source GitHub repositories can help make that workflow possible. 01\. Remotion Skills — Agent guidance for video creation https://t.co/8Qxm4xu3c9 02\. Motion Canvas — Create animated explainers https://t.co/1AKsKyuf6w 03\. Motion — Build smooth layout transitions https://t.co/LrIX2gMIDR 04\. Ola — Smooth value interpolation https://t.co/TBiUXvRIRD 05\. Anime.js — Create timelines and staggered animations https://t.co/8jjFFbAbyc 06\. Manim — Bring mathematical concepts to life https://t.co/BL7ZqlMcFP 07\. Shifty — Control tween playback and easing https://t.co/9zJssN56tW 08\. Two.js — Create animated 2D vector scenes https://t.co/zc8THEwzcb 09\. Lottie Web — Play vector animations https://t.co/yCm3qAR0uB 10\. TweenJS — Sequence property animations https://t.co/RCTCfw7DiA 11\. LUME — Build 3D scenes with HTML https://t.co/jtxFk6cP2E 12\. gifenc — Encode animated GIFs https://t.co/9J7nMBd1Ix 13\. Konva — Create interactive canvas animations https://t.co/9aWKv0ibxP 14\. Blender — Work with rigging and character animation https://t.co/g5jGjAjCU4 15\. Three.js — Build animated 3D scenes https://t.co/MO7ajNZPwp 16\. Godot — Explore powerful game animation systems https://t.co/YbfIrHZMDI 17\. Babylon.js — Create 3D animations and render scenes https://t.co/iP7z7pf6Hz 18\. Proton — Add particle emitters and effects https://t.co/yzJwYHJtpL Think about what you can build with these: → Animated product demos → Explainer videos → 2D and 3D motion graphics → Interactive visual experiences → Character animations → Particle effects and transitions The interesting shift is that your coding agent can become part of the creative process, not just the coding process. Brief it. Build it. Animate it. Render it. 🔖 Bookmark this list and explore these projects!

22 赞 · 12 转帖 · 11 回复

### [@ASI\_8866](https://x.com/ASI%5F8866/status/2121839203055259697?ref=binnn.top)

11月15日 14:34 JST

Day 14/ 公益中转站https://t.co/756yBvASrx上线2周了，昨天收集到166条反馈，前4的反馈列举出来了，老板们可以再精细下排名吗？ 评论区留用户名领额度的活动每天都有，Grok Bot 随时待命充值额度。同时，昨天反馈问题的老板，我们马上将额度打到帐号上～ 欢迎大家扫码进TokenFuture官方QQ群，有问题随时向我们提出来～

22 赞 · 0 转帖 · 457 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121890536776704314?ref=binnn.top)

11月15日 17:58 JST

👀今年一直在暴涨的AI岗位，很多人居然还没听过 ：FDE前沿/前线部署工程师 目前国内大模型参数内卷基本停了，AI圈最近聊天全在聊同一个风口：FDE。 说白了，靠“纯聊天/提示词”赚快钱的时代已经过去了。 现在AI公司抢的是能直接下场、把大模型真正接到业务里、跑通最后一公里、拿到结果的“六边形战士”。 看招聘就知道有多热： 🔹Anker：FDE工程师，2.5-5万·16薪 🔹大疆：FDE（前线部署工程师），3.5-5.5万·15薪 要求从0-1搭AI体系、发现业务机会、开发+产品双修……典型“能落地才拿高薪”。 2025-2026的AI主旋律已经从“会不会用模型”切到“能不能把模型变成业务结果”。 会玩提示词的人太多了，真正能进现场闭环的人，才是当前最稀缺的。 #FDE #前沿部署工程师 #AI落地 #高薪AI岗位 #AI工程师

22 赞 · 5 转帖 · 11 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121949656775352755?ref=binnn.top)

11月15日 21:53 JST

🚨重要更新！DeepSeek长对话不活跃期限已由72小时修改为7天！仍在调试变更中！ 不是超过时间就删记录，而是超过7天没聊天的长对话（目前目测超90轮左右）会变成不活跃状态，无法继续聊。 之前的聊天记录不会被删除，可以导出到其他App使用，也可以直接新开对话。 有用户实测小技巧： 改一下对话标题 → 强制退出APP → 再点进去，部分老对话（甚至8月的）能重新激活！ 目前有的用户还显示72小时，有的已经切到7天，规则正在逐步变更生效。 这下不用天天打卡DeepSeek续命了～ #DeepSeek更新 #DeepSeek限额 #deepseek对话超限额 #梁文峰 #大胖鱼

21 赞 · 4 转帖 · 7 回复

### [@sayada978](https://x.com/sayada978/status/2121807058177802707?ref=binnn.top)

11月15日 12:26 JST

𝗧𝗛𝗘 𝟳 𝗘𝗦𝗦𝗘𝗡𝗧𝗜𝗔𝗟 𝗟𝗔𝗬𝗘𝗥𝗦 𝗢𝗙 𝗧𝗛𝗘 𝗔𝗜 𝗔𝗚𝗘𝗡𝗧 𝗘𝗖𝗢𝗦𝗬𝗦𝗧𝗘𝗠 🤖 Building powerful AI agents isn't just about choosing the right LLM. A complete AI agent system needs memory, tools, orchestration, communication, infrastructure, and evaluation to work together. Here are 7 essential layers that make modern AI agent ecosystems possible.

19 赞 · 16 转帖 · 2 回复

### [@swordBluesy](https://x.com/swordBluesy/status/2121861885188342248?ref=binnn.top)

11月15日 16:04 JST

建议收藏，用到时照着做。做完你的聊天机器人，对话列表里每条会话都有自动生成的短标题。 这是 DeepSeek 官方 API 文档 Conversation Title（会话标题）一页。先说结论： 做对话列表不用再多调一次模型来起标题，加一行参数，回答的同时标题就一起回来，而且官方写明标题不计费。 先说清词。 会话标题（conversation title），就是聊天软件左边列表里那一行短名字，比如“北京明天天气”。 流式（streaming），是答案边生成边返回。标题在流式里出现在最后，不在开头。 先把话说全： 这个功能只在回答里顺带生成，不是单独的接口。回答结束后，响应最多再等标题 5 秒才结束。被拒答或出错时可能没有标题，要自己兜底。 一、第一步：加一行参数 在请求里加 "title": {"enabled": true}。官方给了三种接口的示例：Messages、Responses、Chat Completions，模型都用 deepseek-flash，示例里的问题是“北京明天会下雨吗”。 看到什么算成功：请求照常返回，回答之外多了 title。 二、第二步：找到标题在哪 不流式：Messages 和 Responses 在最顶层的 title，Chat Completions 在 message.title。 流式：Messages 在 message\_delta 事件的 delta.title，Responses 在结束事件的 response.title，Chat Completions 在最后一个 chunk 的 delta.title。 官方示例里标题是“北京明天天气”。 三、第三步：只在第一次请求开 官方建议只在一个会话的第一次请求上开启，拿到标题后就不再开启，省得每轮都多等。 四、第四步：写好兜底 响应不一定带标题，比如被拒答或出错。官方建议自己兜底，例如取用户消息的前若干个字符当标题。 五、收费 官方写明：标题不计费。 常见坑 3 条： 一、每一轮都开，白白多等最多 5 秒。二、流式时去开头找标题，其实在最后一个事件里。三、没写兜底，拒答时列表里就是空白。 复制即用的兜底思路：有标题就用标题，没有就取用户第一句话的前 15 个字。 收藏版：5 步速查 一、请求里加 title 的 enabled 设为 true。二、不流式看顶层 title。三、流式看最后一个事件。四、只在第一次请求开。五、没标题就取用户消息前几个字。 S1哥 观点：做聊天产品的人最容易忽略的就是会话列表，大家习惯再调一次模型来起标题，多一次请求多一份钱。官方这个做法是顺带生成、不计费，真正要处理的只是流式位置和兜底。 你做的机器人，会话列表现在是怎么起名的？告诉我你卡在第几步。同系列前几篇在我主页。 把这篇复制链接发给正在做聊天机器人的朋友。 今天送：DeepSeek 会话标题自检清单，评论区第一条直接拿。 参考：DeepSeek 官方文档 Conversation Title（S1哥 整理）

19 赞 · 3 转帖 · 14 回复

### [@Feljoy\_ix](https://x.com/Feljoy%5Fix/status/2121902175026913557?ref=binnn.top)

11月15日 18:44 JST

想让你的Muse或者其他Ai 把想做的事运转起来落地 缺的其实是一条配齐的工具链 装齐下面14个 AI 必备工具 就能得到从调研到发布的完整流程 全部在GitHub免费公开： 1\~Skills → 做文章、幻灯片、电子表格的官方技能集 https://t.co/KygLUbSlIK 2\~Superpowers → 从怎么调研到怎么写都有的作业模式技能集 https://t.co/5UXMlov1yA 3\~Agents → 按用途打包的智能体与技能集合 https://t.co/KqTbBdSTa4 4\~grok\_x\_search → 在Claude里面搜索X https://t.co/Yv3rn4sq6Q 5\~Firecrawl → 把整个网站读完，用作素材 https://t.co/gvBgxIx1h0 6\~Crawl4AI → 自动收集Web信息 https://t.co/h2dvjqgghs 7\~yt-dlp → 把视频和音频取到本地 https://t.co/q9NT3Z7L5l 8\~Whisper → 把音频、视频转成文字 https://t.co/KlMQZK1y7C 9\~MarkItDown → 把PDF或Word变成文本 https://t.co/8cfpPGB6S6 10\~MoneyPrinterTurbo → 只给一个主题就能组装短视频 https://t.co/MUoYJ4nQAz 11\~Remotion Skills → 让AI做视频的官方技能 https://t.co/JSHmL7NY79 12\~ComfyUI → 做图片 https://t.co/hQTWyWnxFK 13\~F5-TTS → 把文章变成声音 https://t.co/sTcylFS2tv 14\~Marp → 从文章做幻灯片 https://t.co/JMUEl8AiMw 合起来就是：调研 → 写 → 做成视频 → 加入声音 → 发布 全齐了 还能原样交给Claude虚拟机里的Claude Code 以后绝对会用到 丢给AI 它会根据你的使用情况需求去安装学习的

18 赞 · 1 转帖 · 5 回复

### [@xiaoying\_eth](https://x.com/xiaoying%5Feth/status/2121838598244925778?ref=binnn.top)

11月15日 14:31 JST

想在 2026 年转行做 AI 工程师， 吴恩达这 3 小时必看。 从 AI 智能体、提示词工程， 到动手开发 AI 应用，一次讲清楚。 比起到处找零散教程， 我更建议先把这套系统课程看完。 免费学习，建议收藏。 https://t.co/FnUiEbmNuz

18 赞 · 1 转帖 · 2 回复

### [@YeYu\_Yours](https://x.com/YeYu%5FYours/status/2121839560657367207?ref=binnn.top)

11月15日 14:35 JST

明明Codex和Claude更好用 为什么大家用的最多的还是 豆包和Workbuddy

18 赞 · 0 转帖 · 24 回复

### [@yedabinbinah](https://x.com/yedabinbinah/status/2121856732989374495?ref=binnn.top)

11月15日 15:44 JST

知乎免费开放 AI 搜索，每天可使用 5000 次⚡ 实测体验：可免费使用知乎的知识库这是优势，但是英文搜索源不如Tavily，胜在次数多 实名认证后每天5000次，未实名每天10次 步骤：知乎账户登录，个人中心，申请Access Secret 目前国产AI搜索API已有4家：豆包搜索、元宝搜索、Deepseek搜索、知乎搜索 直达链接：https://t.co/cJFS036Faz

18 赞 · 0 转帖 · 8 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121962262974054469?ref=binnn.top)

11月15日 22:43 JST

💥牛逼！用DeepSeek Harness/Codex/Claude Code跑Agent，JSON和日志把Token烧干了？！这个开源神器Tokenfold 用Rust核心智能折叠重复字段，实测无损省63.9% Token，剪枝最高省96.3%！ 专治结构化数据吃上下文的痛点，完美搭配Harness的长任务、工具调用和日志流。 亮点如下： 🔹模型无关、确定性、精确可还原 🔹私有部署、零第三方调用 🔹Python / TypeScript 原生支持，轻量秒级集成 快速上手： 1️⃣安装：pip install tokenfold（或 npm install tokenfold / cargo install tokenfold-cli） 2️⃣一行压缩：直接处理你的JSON/日志/工具输出 3️⃣查看收据：确认节省Token + 无损还原 Harness用户现在就能把账单砍一半！ GitHub：https://t.co/Rjr4NEgzyx #AI工具 #开源 #Token优化 #DeepSeekHarness #LLM #大肥鱼 #梁文峰

17 赞 · 4 转帖 · 8 回复

### [@techwith\_ram](https://x.com/techwith%5Fram/status/2121808156015665458?ref=binnn.top)

11月15日 12:31 JST

𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝘃𝘀 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 Not everything that looks like an AI agent is actually an agent. And that's completely fine. There's a big difference between AI workflows and AI Agents, but the two are often confused. A workflow is a set of steps you plan in advance. The LLM might classify a request, a tool might fetch some data, and another LLM call writes the reply. There can be branches, but you wrote every one of them. Run it a hundred times and you know exactly what can happen. An agent works differently. You give the model a goal and some tools, and it decides what to do next. It might search twice, run some code, check a database and then stop. You don’t know the number of steps until it’s finished. That freedom is powerful for messy, open-ended problems. It also costs you something: - more tokens - more latency and - behavior that’s harder to test and predict. My rule of thumb is to start with a workflow. Move to an agent only when the task can’t be mapped out ahead of time. Most production systems I have seen are mostly workflows, with a little agent behavior where it actually earns its place.

16 赞 · 0 转帖 · 15 回复

### [@aaliya\_va](https://x.com/aaliya%5Fva/status/2121889179608965598?ref=binnn.top)

11月15日 17:52 JST

Your conscious brain processes 40 bits of information per second. Beneath your awareness, your autonomic nervous system processes 11 million bits per second to regulate your heartbeat, blood pressure, and cells without asking your permission. For 30 years, the internet had no autonomic nervous system. Nothing happened online unless a human clicked a mouse. That biological limit just vanished. Cloudflare reports that automated traffic has already passed human traffic at 57.5% of the web. Speaking at Baron Capital, Elon Musk noted that within 5 years, AI agent traffic will exceed human internet traffic by 1,000 times, adding that even a 1,000 to1 ratio is likely too low. Why would machines generate 1,000 times more traffic than eight billion humans? Because one human intention now triggers 10k background actions. When humans drop to 0.1% of web traffic, the entire digital world reorganizes across 4 layers. 1\. Daily Life: The End of "Going Online" From search tabs to background loops: Today, a human goes online, opens ten tabs, reads articles, and clicks buttons. In a 1,000x web, you stop browsing for routine life. While you sleep, your personal agent reads your wearable blood glucose data, cross checks 4,000 clinical papers, adjusts your bedroom thermostat, and rebalances your index funds. 2\. Web Architecture: The Death of the Visual Page Websites become human museums: For 3 decades, we built the web for human retinas using HTML, CSS, layout animations, and pop ups. Agents are blind to visual polish. At 1,000x scale, 99.9% of the internet drops visual screens and runs on headless APIs and Model Context Protocols (MCP). 3\. Software and Work: SaaS companies built around human seat licenses and visual menus get hollowed out. Agents already spin up 80% of new Neon databases and trigger 70% of Stripe CLI requests. Software becomes pure back end logic billed by the millisecond of compute. 4\. Media and Human Connection: With agents writing the majority of public text and code, humans retreat into cryptographically verified private networks for art, friendship, and debate, while their personal agents stand guard at the perimeter to filter out synthetic noise. The visual web we grew up with is shrinking into a quiet leisure layer for human entertainment. Beneath that thin surface, the real internet is transforming into a high frequency planetary nervous system where machines do the reading, the negotiating, and the building in the dark.

16 赞 · 2 转帖 · 8 回复

### [@xoyif63862452](https://x.com/xoyif63862452/status/2121795809029087260?ref=binnn.top)

11月15日 11:41 JST

我去，同一个项目今天让 Claude 写，明天换 Codex，上下文又得重新喂一遍，做好的白板表格还锁在人家 App 里，换个模型就得搬家。 这个开源的就是给 Agent 用的共享文件夹，白板存成 .excalidraw，表格是 .csv，看板是 .html，Claude、Codex、Cursor 进同一个文件夹接着干，每次改动都走 Git，改坏了随手回滚。 模型是租来的，文件得攥在自己手里，这思路我服。 https://t.co/xDkgtTCTEl

15 赞 · 1 转帖 · 31 回复

### [@HeLi8866](https://x.com/HeLi8866/status/2121901621751996690?ref=binnn.top)

11月15日 18:42 JST

我靠，这套 AI 提示指令合集太强了！！😱 它能让你用任何 AI 工具都能得到出色的效果： • ChatGPT 的提示指令： https://t.co/bv8n9i1ISw • Claude 的提示指令： https://t.co/to3tkpjlL3 • Gemini 的提示指令： https://t.co/BYnVnVXtjX 建议收藏，下次直接上手用 https://t.co/Y19zUVUku0

15 赞 · 3 转帖 · 3 回复

### [@jithinvraj](https://x.com/jithinvraj/status/2121922899200405705?ref=binnn.top)

11月15日 20:06 JST

9 Chunking Strategies for RAG Agents (Only for Reference &amp; Research) https://t.co/dNcwz0q7Pk

14 赞 · 4 转帖 · 1 回复

### [@miles\_mazy](https://x.com/miles%5Fmazy/status/2121810963683066029?ref=binnn.top)

11月15日 12:42 JST

现在能用上 Claude 的都是人中龙凤了 Claude Code 官方技术团队最佳实践，你还能不看？

14 赞 · 2 转帖 · 5 回复

### [@btwSora](https://x.com/btwSora/status/2121957469807104122?ref=binnn.top)

11月15日 22:24 JST

OpenAI co-founder, Ilya Sutskever: "The day will come when AI will do all the things that we can do Anything which I can learn, anything which any one of you can learn, the AI could do as well" in a 9-minute speech, he explains why your brain is a biological computer and AI will do every job worth more than a $5,000 AI course watch tonight, then read how to build a 24/7 AI agent team for your business in the article below

13 赞 · 0 转帖 · 0 回复

### [@Dav338876](https://x.com/Dav338876/status/2121901318990360724?ref=binnn.top)

11月15日 18:41 JST

国外大佬又整出猛料😱 3 天！！！吓哭我了.😭 用Three.js 和 Claude Fable 5.2 魂系游戏！ Claude Code Fable 5.2 配合 Claude Opus 5.5 Seedream 5.0 Pro：作为图片生成 Tripo 3D：模型生成 最狠的是： @MengTo 好兄弟把教程都手把手给你了🥰 https://t.co/mPPtSAo4jd

13 赞 · 0 转帖 · 12 回复

### [@vgcil7](https://x.com/vgcil7/status/2121907326324273545?ref=binnn.top)

11月15日 19:05 JST

腾讯混元和复旦MOSS都发了实时语音模型，能聊、能唱、还能自己决定要不要联网搜索。 我试了一下，它接话比我会，唱歌比我好听，甚至知道什么时候该沉默。 我沉默了。 它没沉默。 现在语音AI已经卷到“全双工+自主工具调用”，普通人聊天都开始有压力了。 下次再有人说“AI只会复读”，我直接把通话记录甩他脸上。 腾讯混元 Hy Realtime 实时语音模型发布报道： IT之家，2026年11月13日 https://t.co/GXCX3jSN3v 复旦 OpenMOSS 团队 MOSS-Live 全双工实时交互系统： GitHub：https://t.co/bztEsEIfUC 官方介绍：https://t.co/CBqxNaooBZ #语音AI #大模型

12 赞 · 1 转帖 · 7 回复

### [@MaiYangAI](https://x.com/MaiYangAI/status/2121929065569620020?ref=binnn.top)

11月15日 20:31 JST

https://t.co/iI1bbr7xnU

12 赞 · 3 转帖 · 0 回复

### [@leo114119](https://x.com/leo114119/status/2121858526633251323?ref=binnn.top)

11月15日 15:51 JST

百度词元所首日供应商超100家 百度这个词元所，是不是就是中转站啊？ 和 OpenRouter 有什么区别吗，不也是转卖其他的大模型的 Token 吗？ https://t.co/V4UVHTTeZI

11 赞 · 0 转帖 · 11 回复

### [@leo114119](https://x.com/leo114119/status/2121862846346870981?ref=binnn.top)

11月15日 16:08 JST

百度这个词元所，感觉 Token质量很一般啊 Deepseek 的 Tps都没有能到 150 的。。而且可用性极差 真的有人会用他们的 Token 吗？ https://t.co/hoDxH2JvFE

10 赞 · 0 转帖 · 8 回复

### [@cug\_denghy](https://x.com/cug%5Fdenghy/status/2121827444445548553?ref=binnn.top)

11月15日 13:47 JST

你想尝试复刻DeepSeek App么？ 复刻的难点，不在再造一个模型，也不在再造一套搜索。难的是把网页端已经拆开的那些内容块，在自己的客户端里忠实渲染、完整保存、原样送回去。 https://t.co/LWQKTqmS7C

10 赞 · 1 转帖 · 15 回复

### [@Kunalmadoliya](https://x.com/Kunalmadoliya/status/2121892968659648513?ref=binnn.top)

11月15日 18:08 JST

Built GitWiki An AI-powered tool that generates summaries and detailed documentation of GitHub repos Also witnessed prompt injection my AI started following the system prompt from my Persona project instead of its own instructions🫡👨 link :- https://t.co/oxBc7SiXxW #chaicode https://t.co/2PhiRfU1bd

10 赞 · 0 转帖 · 2 回复

### [@Spleap17598](https://x.com/Spleap17598/status/2121951513346269195?ref=binnn.top)

11月15日 22:00 JST

去年还是和GPT-5在浏览器对话再复制黏贴到本地ide里 我笨拙的回测机器学习量化交易策略，还得让gpt帮我一个个写因子，帮我折腾pytorch，用gpu加速运算啥的 自己半懂不懂 测出来一堆过拟合玩意让我当场高潮，walk forward回测然后发现全部傻逼了 今天把旧的那些玩意一个个翻出来给Codex，现在按照工业化的截面测量研究方式，一个个因子计算rank ic ir，进入因子库然后构建特征/标签数据集 真聪明啊，更棒的是我不用看代码了 有一种auto research的快感，又开始幻想早点找到圣杯了，Codex大人

10 赞 · 0 转帖 · 2 回复

### [@arXivBangers](https://x.com/arXivBangers/status/2121863581734252913?ref=binnn.top)

11月15日 16:11 JST

One Fabrication Is Enough: A Fake News Article Changes the Forecasts of LLM Forecasting Systems A single fake news article reverses 74% to 82% of opposing LLM forecasts in tested frontier models, but average accuracy barely changes to reveal the attack. https://t.co/IxXPGXxsje https://t.co/onwYvIoykc

9 赞 · 1 转帖 · 0 回复

### [@Acemation\_](https://x.com/Acemation%5F/status/2121847019937636584?ref=binnn.top)

11月15日 15:05 JST

This is f\*cking insane. Anthropic CEO Dario Amodei was asked what a kid finishing high school should do now. His answer: "AI can do entry-level law work. It can do entry-level finance. It can do entry-level consulting We can't deny that it will also eliminate jobs, and probably a large number of them" In a few minutes he names which jobs go first and the one skill that keeps you in the game His own prediction: 50 percent of entry-level white-collar jobs disrupted in the next 1 to 5 years Prompts → Agents → AI Coworkers → Entry-Level Work Done By AI Farming to factories took generations. This time AI hits every step of a career at once His advice fits in one line: "Of course, learn to use AI" The people who start this year won't be competing with AI, they'll be the ones running it Everything below is free. Start today: Anthropic - Prompt Engineering Interactive Tutorial: the fastest way to learn to direct a model https://t.co/fSCXnF3EEW Anthropic Courses: from your first prompt to running agents https://t.co/qeFJxIoJj3 Anthropic Cookbook: real task recipes you can copy today https://t.co/r9H8RNiVJ4 Microsoft - Generative AI for Beginners: 21 lessons, no degree required https://t.co/Ykzy6n9uzT Microsoft - AI Agents for Beginners: build agents that do the entry-level work for you https://t.co/VY6nk7trML Hugging Face - Agents Course: free and certified, start to finish https://t.co/wbOYH1Hjeb Google Cloud - Generative AI: Google's own hands-on notebooks https://t.co/3ZFwhlbWmk OpenAI Cookbook: code for every common task https://t.co/bkBnMJJgYI https://t.co/8AEcAhyuMY - Prompt Engineering Guide: the reference everyone links to https://t.co/etQ2xnjU70 Andrej Karpathy - Neural Networks Zero to Hero: understand what is under the hood https://t.co/D8SCPXsXAB CrewAI: run a small team of AI workers https://t.co/IrDhQfu4nW Microsoft - AutoGen: put several agents on one job https://t.co/l6FivYxW7w Awesome LLM Apps: hundreds of real builds to learn from https://t.co/G7p1S2avJ7 https://t.co/EULM6pBfsF - Short Courses: free 1-hour courses from the people who teach this https://t.co/mDXwMEWaQd Total cost: $0 You don't need another year of tuition to learn this. You need to start telling AI what to do 95% will graduate and compete with AI for the same entry-level job The 5% who learn to run it this year will manage the agents doing that job Save this and send it to someone graduating this year

9 赞 · 3 转帖 · 4 回复

### [@Jamesmv49](https://x.com/Jamesmv49/status/2121792990796779590?ref=binnn.top)

11月15日 11:30 JST

这是一个可以免费使用 Claude 的地方 大多数人还不知道 👉 https://t.co/1wHvzNELzg https://t.co/0GWaOf5h6m

9 赞 · 1 转帖 · 1 回复

### [@suyuan1711](https://x.com/suyuan1711/status/2121869336604516711?ref=binnn.top)

11月15日 16:34 JST

神奇，感觉国内能用上 Claude 的 AI 用户也不少啊 在北京的一家咖啡厅里，我左边女生在用网页版的 Claude 画图 右边的女生用 iPad 的 Claude App 研究数学 而我在用 Claude Code 对话

9 赞 · 1 转帖 · 8 回复

### [@yangwang1664258](https://x.com/yangwang1664258/status/2121894665683378425?ref=binnn.top)

11月15日 18:14 JST

跟大伙儿交个底，现在判断一个人炒股到底行不行，其实有个特别简单的“评判标准”——就看他能不能在本地自己搞起来一套 Claude Code！💻 我不管你用什么招数、走什么偏门路子，只要你能把它给鼓捣出来跑通了，就说明一件事：你具备了发现问题、分析问题、并且真能动手解决问题的底层能力。 别小看这几样，在炒股这事儿里，这些能力全是硬通货，少一样都不行。连个本地工具都折腾不明白的，还指望在市场里天天跟庄家斗智斗勇？省省吧！🤣 #炒股 #ClaudeCode #AI工具 #交易逻辑 #认知变现

9 赞 · 0 转帖 · 13 回复

### [@RealMarvelX](https://x.com/RealMarvelX/status/2121883923575628015?ref=binnn.top)

11月15日 17:32 JST

Sam Altman says agent teams are the future of AI. instead of asking one AI to build an entire app, deploy a team of agents: one plans, others code, test, review, and fix bugs in parallel. this is the idea behind AI agent swarms. here’s how they work and how to build one. what is an AI agent swarm? think of one AI agent as a student doing an entire group project alone. now imagine an entire class working together, with each student responsible for a different part. that’s an agent swarm. instead of one agent handling every task, multiple agents work in parallel, each with a specific role. why do you need one? agent swarms are useful for jobs that can be broken into smaller tasks: → building an app from a detailed specification → rewriting large codebases → researching dozens of sources simultaneously → testing hundreds of different scenarios → handling complex multi-step workflows how to build your own agent swarm: 1\. the goal give every agent one clear objective and a shared finish line. 2\. the task board create a shared task list where agents can pick up assignments, track progress, and avoid duplicating work. 3\. the agents give each agent its own role, memory, tools, and workspace. let them work on separate tasks simultaneously. 4\. the shared notebook create a central place where agents record findings, decisions, completed tasks, and problems encountered. 5\. the checker use automated tests and a reviewer agent to inspect the results, catch mistakes, and decide what is ready to keep. the real advantage: one agent handles one task at a time. a well-coordinated swarm can divide a much larger project into parallel workstreams, combine the results, and review the final output. the article below walks you through building your own AI agent swarm step by step. save this if you’re ready to move beyond single-agent workflows. 🔖

8 赞 · 0 转帖 · 5 回复

### [@pash22](https://x.com/pash22/status/2121821728196169728?ref=binnn.top)

11月15日 13:24 JST

PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing https://t.co/CR4dt1r6H8 via @tpau17 et al https://t.co/nwr9kP5SKT

8 赞 · 2 转帖 · 0 回复

### [@shushant\_l](https://x.com/shushant%5Fl/status/2121876363292582071?ref=binnn.top)

11月15日 17:02 JST

Here's how to give your coding agent a new job overnight: 10 brand-new skill folders, all free on GitHub. 1\. Solo to Team: hand work out to AI helpers, then check it all → https://t.co/29cvR3qa8u 2\. Line Art Animation: cute animated line-drawing characters for websites → https://t.co/rksUWSt7AX 3\. VN Law Check: checks a website against Vietnam's privacy and web laws → https://t.co/kjjuFLgdcl 4\. Cook: give it a design target and it keeps going until it matches → https://t.co/B0ulDwycfA 5\. Simcraft: 24 skills for video, code, learning, design and planning → https://t.co/T3vithu91a 6\. Best Paper Writing: drafts research papers in the style of award winners → https://t.co/Zaul05egdS 7\. Hill-Climb: keeps an agent's change only if the test score goes up → https://t.co/7eFmNR66eH 8\. Replica: rebuilds a website from a picture and scores the copy → https://t.co/xZh2AxZqWu 9\. YouTube Research: finds winning video ideas, titles and thumbnails → https://t.co/I5D2m5nvYo 10\. Backend Skills: stops agents writing backend code that breaks in real use → https://t.co/FwjWZGRT5J

7 赞 · 3 转帖 · 5 回复

### [@CaptainXTop](https://x.com/CaptainXTop/status/2121949241149141385?ref=binnn.top)

11月15日 21:51 JST

前几天刷到 @longhaiqwe123 转的一条帖子，讲老李怎么把《Refactoring UI》做成 AI 技能： https://t.co/la2dUibqY8 看完就想，这事能不能让 bot 自己干？我把帖子丢给专门造 bot 的 bot，它建了一个只干这一件事的 bot，叫「书变技能」 方法就四步： 1\. 先判断值不值得做。得是规则型的书，有明确做法和数值，还能验证。靠品味和讲故事的书不行 2\. 把书转成干净的 md，要点逐条编号，建一张覆盖表，防止漏 3\. 每条规则写六段：规则、为什么、怎么做、数值、写法、失败长啥样 4\. 出测试集，分 must、should、forbidden 三档，在干净的上下文里考。must 全中、forbidden 零次才算过。 解析是在云电脑本地跑的，装了 MinerU，原书不上传。扫描版大约 11 秒一页。标题层级交给 DeepSeek 修，一次才 800 多个 token。 配合云电脑装好的 codex cli， bot 的额度自己只做轻活：判断书值不值得做、派活、抽查、最后收进技能库。 读书、写规则、出题、打分这些重活，全丢给云电脑上的 codex。 我选了一本《Getting Things Done》，402 页。 做出来 52 条主规则、211 条子要点、8 份参考文件、2 个判定脚本。 准备后面用 Muse 来践行一下这个 GTD 帮我安排所有事项看看。 书变技能的 bot 模板放这了，大家有需要可自取哦 https://t.co/En3bE5Fr2d https://t.co/la2dUibqY8

7 赞 · 0 转帖 · 6 回复

### [@wrok\_si](https://x.com/wrok%5Fsi/status/2121800002163126658?ref=binnn.top)

11月15日 11:58 JST

🚀 Why I'm building wrok I want AI to understand the work, use the right tools, and carry tasks through to usable results. The goal: a recursively self-improving AI colleague. wrok Code is available now. 🔗 https://t.co/voLp70sjtL 1/5 #wrokcode #ai #agent https://t.co/UOwasu64b5

6 赞 · 2 转帖 · 4 回复

### [@alexlee20122022](https://x.com/alexlee20122022/status/2121890569177686208?ref=binnn.top)

11月15日 17:58 JST

https://t.co/3Yw3ExV2Nn

6 赞 · 0 转帖 · 4 回复

### [@JoeVenner\_](https://x.com/JoeVenner%5F/status/2121942162858442957?ref=binnn.top)

11月15日 21:23 JST

I love @IKEA and building things myself. But the booklets can be brutal: wordless drawings, 30 steps, and that one screw you're never sure about. So two weeks ago I asked: can AI fix the manual? Meet Rakkeb (ركّب, "assemble it"). 𝟭. 𝗣𝗮𝗽𝗲𝗿 → 𝟯𝗗 An engine that turns each drawing of the official booklet into an animated 3D step: real tools, real moves, shadows, the exact holes marked. An AI pipeline reads the manual page by page and writes each step into the engine. 22 pieces so far. 𝟮. 𝗔 𝗺𝗮𝗻𝘂𝗮𝗹 𝗰𝗮𝗻'𝘁 𝘀𝗲𝗲 𝘆𝗼𝘂𝗿 𝗽𝗮𝗿𝘁𝘀 So Rakkeb got eyes and a voice: a technician that stays on the call for your whole build, sees what you see and answers out loud. 𝟯. 𝗧𝗵𝗲 𝗵𝗮𝗿𝗱 𝗽𝗮𝗿𝘁: 𝗰𝗼𝘀𝘁 A build can take an hour or two. Streaming video and audio to a cloud model that long burns tokens fast. This is where engineering still matters. The rule: the phone does everything it can for free. The cloud only works when there's a real question. ▸ Ears, on the phone: Apple's on-device speech recognition runs the whole call. A local check decides if you're talking to Rakkeb or to someone in the room. ▸ Eyes, only on demand: the camera never streams. Ask, and one frame goes with your question to Muse Spark from Meta. About 1k tokens. ▸ Pointing, @Meta finds and @Apple follows: SAM 3.1 outlines the exact part once. Apple's Vision keeps that outline locked on it as you move the phone, at zero token cost. Long calls, few tokens. Everything in the video is real: an iPhone 13 filming, the app on an iPhone 17, me building a LACK table on my balcony. App access is coming soon. 𝗪𝗮𝗻𝘁 𝘁𝗵𝗲 𝗳𝘂𝗹𝗹 𝗰𝗼𝗱𝗲? 𝗖𝗼𝗺𝗺𝗲𝗻𝘁 "𝗰𝗼𝗱𝗲" 𝗮𝗻𝗱 𝗜'𝗹𝗹 𝘀𝗲𝗻𝗱 𝗶𝘁 𝘁𝗼 𝘆𝗼𝘂. Curious what you think, @IKEAUAE @IKEAUSA @IKEAsaudiarabia . (Independent project, not affiliated with IKEA.)

5 赞 · 1 转帖 · 1 回复

### [@Jackson77li](https://x.com/Jackson77li/status/2121906052165964099?ref=binnn.top)

11月15日 19:00 JST

现在基本上各个赛道 前十我已经看不到中国大模型的身影了 https://t.co/kTyZVDTtYZ

5 赞 · 0 转帖 · 6 回复

### [@7998l201](https://x.com/7998l201/status/2121945111018021209?ref=binnn.top)

11月15日 21:35 JST

https://t.co/h6NPMkAQRW

5 赞 · 0 转帖 · 0 回复

### [@Kkhookyou](https://x.com/Kkhookyou/status/2121937913784963141?ref=binnn.top)

11月15日 21:06 JST

你也想1-3 天快速做出一款游戏？？🤯 可以，你提出一句话的idea，这样分工👇 GROKBOT做总策划📝由总策划统筹四个BOT：数值策划+系统策划+关卡策划+战斗策划 GPT负责出图 🎨 MUSE负责生成音效BGM CLAUDE负责主力代码 你会发现做游戏太简单！ 但丑话说在前头👇 你不要干预这个过程，只需要当一个玩家体验最终结果！ 至于好玩与否，市场是否买账，那就看各人天命所归了。 也许最终你投入的大量算力，只是一场空。

5 赞 · 0 转帖 · 6 回复

### [@FinanceYF5](https://x.com/FinanceYF5/status/2121937498418925955?ref=binnn.top)

11月15日 21:04 JST

斯坦福有一门专门研究「AI Agent 如何自我改进」的研究生课程：CS329A。📚 课程围绕 Self-Improving AI Agents 展开，涵盖 AI Agent 自我改进的最新技术与应用。 AI 研究者 Elvis Saravia 推荐了这套课程视频，认为内容扎实、参考资料丰富，适合想系统学习这一方向的人。 https://t.co/Fu5CSxXDZx

5 赞 · 1 转帖 · 3 回复

### [@Bennnoss](https://x.com/Bennnoss/status/2121861623077818403?ref=binnn.top)

11月15日 16:03 JST

Everybody who used claude as LLM knows “You’re absolutely right”. Now imagine this happens on a space/satellite/defense/aero application. Data sovereignty, operational independence and no GPU servers or AI-infrastructure matter. @Oracle7vn

4 赞 · 0 转帖 · 1 回复

### [@moranweb3](https://x.com/moranweb3/status/2121817210842956187?ref=binnn.top)

11月15日 13:06 JST

https://t.co/qMqLnsAq5j

4 赞 · 1 转帖 · 2 回复

### [@lee\_xeo](https://x.com/lee%5Fxeo/status/2121824420880879733?ref=binnn.top)

11月15日 13:35 JST

分享一下来自硅谷的信源： 1\. Gemini Argon 和 Gemini Carbon 是两拨人做的，模型架构一模一样，数据和训练方法不太一样。都是12TA555B。类似国内的赛马。谷歌极有可能同时训练了多个 12T 模型，成了两个。中美算力的差距似乎远大于 2 个数量级。 2\. Google 内部目前没有更强的模型，但 Anthropic 有 3\. Anthropic 的工作模式有点像缅北 KK 园区。他们招了很多很多人，进去之后参与某个项目，但并不能知道你参与的项目的目标是最终给模型造了一堆数据还是真的模型的一部分。所以他们的研发成本巨高，不想让别人蒸馏他们。 4\. Anthropic 的模型应该比 Google 的小，Opus 也比 Gemini 4这两个小。但具体多大没问到。

4 赞 · 0 转帖 · 2 回复

### [@taiyanghere](https://x.com/taiyanghere/status/2121941997691019650?ref=binnn.top)

11月15日 21:22 JST

Memoh 送了个免费的云端 Agent 环境，能跑智能体。 Memoh 提供免费的云端 Agent 环境，可以用来跑智能体。有官方 GitHub、Dockerfile 和部署说明，照着走就能搭起来。 不过免费版配额有限，超限会变付费，用之前先看清额度。别拿来直接跑生产，玩玩、试试水比较合适。

4 赞 · 0 转帖 · 5 回复

### [@victoria\_olsina](https://x.com/victoria%5Folsina/status/2121921312666931281?ref=binnn.top)

11月15日 20:00 JST

Why LLMs recommend me as a crypto & web3 SEO consultant is not a fluke, it’s an SEO system pointed at AI assistants instead of “10 blue links”. Ask ChatGPT, Grok or Google AI “who is the best Web3 SEO consultant?” and my name shows up, reproducibly, in the answer – not in an ad slot, not on page two. That comes from 5 things working together: a clean entity graph (schema + sameAs), hard-number case studies, third‑party authority (book, awards, talks), content written as extractable evidence, and a dense topical footprint around Web3 + AI search. Once those 5 exist, LLM recommendations follow the same pattern as any other source selection. Long version lives here: https://t.co/eAJ2JZCVHS

3 赞 · 0 转帖 · 1 回复

### [@galratner](https://x.com/galratner/status/2121843537478320387?ref=binnn.top)

11月15日 14:51 JST

Mark Russinovich, CTO of Microsoft Azure, published a decision model on his personal GitHub. The decision head is 787,457 parameters. That's 3.15 MB. https://t.co/z6sdF94Pdh #MachineLearning #Embeddings #dotnet #SQLServer #AIEngineering https://t.co/EtqDjqXfRl

3 赞 · 0 转帖 · 0 回复

### [@hulianshensw](https://x.com/hulianshensw/status/2121794864438550821?ref=binnn.top)

11月15日 11:38 JST

给claude夺舍的workbuddy？ #WorkBuddy #Claude #AI短剧 #VibeCoding #aivideo https://t.co/gD5QeMMSWj

3 赞 · 0 转帖 · 1 回复

### [@RohoQinAI](https://x.com/RohoQinAI/status/2121923053533753767?ref=binnn.top)

11月15日 20:07 JST

刷屏的「吴恩达 3 小时 AI 工程师课」，是网友剪的合集。 占大头的两段，原片都在 https://t.co/5WVuJmWIh9 官方 YouTube 频道，吴恩达本人主讲，免费看。 一、AI 提示词完整课 https://t.co/Cji6EprRiB 最有用的几条： 常识直接问，要综合几十个来源的问题交给深度研究。 把 AI 当成聪明但不了解你的应届生，先把资料传上去再问。 换话题就开新对话，旧上下文会干扰新问题。 提问要中性，别暗示你想要的答案；要它打分，先给只能答是或否的评分细则。 写作先列提纲，改几轮再成文，改提纲一处能带动一整段。 二、30 分钟用 AI 做一个应用 https://t.co/5yRQjAKh1q 给从没写过代码的人，跟着做一个能发给朋友的生日祝福小应用，再用对话让 AI 改它、修它。 开头约 23 分钟讲智能体和 AI 工程的未来，我没对上官方原片。 天天用 AI 却嫌答案太泛的人看第一段，想做小工具又不会写代码的人看第二段。

3 赞 · 0 转帖 · 3 回复

### [@krishsharma1903](https://x.com/krishsharma1903/status/2121815458349531577?ref=binnn.top)

11月15日 13:00 JST

LLM apps, Day 2: models don't read words. They read tokens. "I am learning how large language models work." English: 9 tokens. The same sentence in Hindi: 78 tokens on GPT-2, 13 on Gemini. More tokens means more cost and less room. Language matters. #BuildInPublic #LLM https://t.co/9nHofll1b2

2 赞 · 0 转帖 · 0 回复

### [@siliconvikings](https://x.com/siliconvikings/status/2121924126642213022?ref=binnn.top)

11月15日 20:11 JST

#cphftw #AI lab @Sevren\_ai Releases Embla, Its Most Efficient AI Model That Runs on a Phone. By Øresund Startup News https://t.co/1erYDdN1vd #NordicMade #NordicAI

2 赞 · 0 转帖 · 0 回复

### [@lamhot\_ai](https://x.com/lamhot%5Fai/status/2121875686139081096?ref=binnn.top)

11月15日 16:59 JST

55 AI System Design interview questions from 18 companies, including OpenAI, Anthropic, Google, Meta, NVIDIA, Amazon, Microsoft, and Apple. Covering RAG, AI agents, LLM inference, GPU infrastructure, distributed training, model routing, and multimodal https://t.co/TRzhpzPju7 https://t.co/W8gF75Un7Z

2 赞 · 1 转帖 · 0 回复

### [@Prashanttcore](https://x.com/Prashanttcore/status/2121817529891061947?ref=binnn.top)

11月15日 13:08 JST

4 trending open-source AI repos on GitHub right now: 1\. Open Code Review (Alibaba) - AI code checks 2\. Qwen Code (QwenLM) - Terminal coding agent 3\. LoopForge (Tencent) - Agent workflows 4\. Univer (Dream-Num) - AI office SDK Which one are you testing? https://t.co/mepDxU08tl

2 赞 · 0 转帖 · 2 回复

### [@aiwaremax](https://x.com/aiwaremax/status/2121833761520390544?ref=binnn.top)

11月15日 14:12 JST

https://t.co/bezRWYkAef

2 赞 · 0 转帖 · 3 回复

### [@FinanceYF5](https://x.com/FinanceYF5/status/2121950792231919930?ref=binnn.top)

11月15日 21:57 JST

从一条推文到 Claude Code 正式支持，只用了 17 天。🤯 开发者 Marko Vineli 此前分享的自定义主题创意，如今被 Anthropic 加入 Claude Code 桌面版。 通过 Mods，用户可以自由修改配色、添加动态背景，甚至直接告诉 Claude 想要什么风格，让它帮忙设计主题。 https://t.co/mUTF4uqMby

2 赞 · 0 转帖 · 1 回复

### [@aminul\_builds](https://x.com/aminul%5Fbuilds/status/2121945301225484734?ref=binnn.top)

11月15日 21:35 JST

9/9 — The Engineering Takeaway Prompt length influences TTFT. Output length determines how long generation continues. Prefill and Decode have different hardware bottlenecks. Understand both before optimizing LLM performance. Follow @aminul\_builds #AIEngineering #LLMOps https://t.co/tYljFs9e8B

1 赞 · 0 转帖 · 0 回复

### [@TomMcArthur](https://x.com/TomMcArthur/status/2121909107141546126?ref=binnn.top)

11月15日 19:12 JST

I can't imagine having to use an LLM to sound more "emotive" after some delicious booze

1 赞 · 0 转帖 · 0 回复

### [@agentbuilder007](https://x.com/agentbuilder007/status/2121881891532112070?ref=binnn.top)

11月15日 17:24 JST

Stop building AI agent teams from scratch. GitHub already has ready-made ones: 1\. Agent Org Chart → Engineering, marketing, support teams 2\. Grok Factory → One coordinator + specialist agents 3\. Foreman → Monitors stalled tasks

1 赞 · 0 转帖 · 2 回复

### [@vaibhavbalkundi](https://x.com/vaibhavbalkundi/status/2121929737291681881?ref=binnn.top)

11月15日 20:34 JST

Revolut just did the thing every CTO quietly fantasizes about. It moved 2,000 engineers off Claude Code and Codex onto a self-hosted open-weight model, GLM-5.3 from China's https://t.co/wAlnvYIOhn, running on 56 of its own Nvidia B200 GPUs. The bill: about $0.30 per million tokens. Roughly a tenth of what it paid before. Coding time down 60%, lead time down 70%, by Revolut's own numbers. The catch, and Revolut is honest about it: the 90% is inference cost only. It excludes the GPUs, the tuning, the team that runs it. And GLM is text-only, so a router still sends image work to Gemini and GPT. But the direction is the story. The frontier labs sell you a seat. The open weights let you own the meter. At 2,000 engineers, owning wins.

1 赞 · 0 转帖 · 1 回复

### [@ji\_vidushi\_shah](https://x.com/ji%5Fvidushi%5Fshah/status/2121928306023125182?ref=binnn.top)

11月15日 20:28 JST

🚨 You're in an AI Engineer interview. The interviewer asks you to design a workflow using LangGraph. Your task: Design the graph architecture and write its complete implementation Can you answer these questions? 1\. How would you design the nodes required for an AI agent? 2\. How would you define the state shared between nodes? 3\. How would you connect nodes using edges? 4\. How would you implement conditional routing? 5\. How would you write the complete LangGraph structure in Python? 6\. How would you compile and execute the graph using "StateGraph"? #LangGraph #AIEngineer #InterviewPreparation #2028Batch

1 赞 · 0 转帖 · 0 回复

### [@AKhomssi](https://x.com/AKhomssi/status/2121944316411089324?ref=binnn.top)

11月15日 21:32 JST

🔐 Turn your coding agent into a security auditor. Cloudflare released an open-source Security Audit Skill. It helps agents: Investigate a codebase Hunt for potential vulnerabilities Validate findings independently Generate structured security reports 72K+ GitHub stars. MIT licensed. GitHub: https://t.co/F4eVvkYQiH #CyberSecurity #AI #OpenSource #GitHub

1 赞 · 0 转帖 · 3 回复

### [@sabrielatmojo](https://x.com/sabrielatmojo/status/2121794259544072479?ref=binnn.top)

11月15日 11:35 JST

Deepnote just dropped a benchmark for proactive agents, and the cheap model won. The Agent Initiative Benchmark puts 12 frontier models inside the Slack of a synthetic company with six years of history: a financial model, CRM records, past-due invoices, Linear tickets, even real-world events like the SVB collapse. Agents get access to 11 services, face 50 situations, and have 3 minutes to decide whether to step in, how to help, or whether to stay quiet. The headline: GPT-6.1 Sol and GPT-6 Astra beat Claude Fable 5.2, with Sol at $0.11 per task vs $1.87 for Fable. Fable 5.1 costs 430x more than GPT-6 Luna for similar performance. And the most interesting failure mode: most models try to help even when nobody needs them, and that unnecessary work burns about 80% more time and money than just staying quiet. Why it matters: an always-on agent reads every single thread. At that volume, per-task cost isn't a footnote, it's the architecture decision. A cheap model that knows when to shut up beats a brilliant one that comments on everything. If you want to run the same comparison on your own workload, https://t.co/kheLxYU7TG puts 176 models behind one OpenAI/Anthropic-compatible key, pay per token, averaging \~76% below official prices (GPT-6.1 Sol listed from $0.005/$0.025 per 1M tokens vs $2/$10 official). Dataset and benchmark code are open-source. h/t @jakubjurovych @DeepnoteHQ

1 赞 · 0 转帖 · 1 回复

### [@CryptVivek](https://x.com/CryptVivek/status/2121949356114969081?ref=binnn.top)

11月15日 21:52 JST

SkillSpector — Open-Source Security Scanner for AI Agent Skills &amp; MCP Packages https://t.co/eMskgNpKGc https://t.co/XAKTCvfyMS

1 赞 · 0 转帖 · 1 回复

### [@cywcycw](https://x.com/cywcycw/status/2121837739804926008?ref=binnn.top)

11月15日 14:28 JST

需求讲三遍，代码返工三版。Spec Kit 对提示词直写实测：149,820 星，spec → plan → tasks 走完再动手，改规矩比改代码便宜。核心判断：“需求每讲一遍都贵，规矩先写进文件。”今晚先写一页 spec.md，只留三个空。 https://t.co/aSNhmZDxLf

1 赞 · 0 转帖 · 1 回复

### [@TheAIEdu](https://x.com/TheAIEdu/status/2121874031242797285?ref=binnn.top)

11月15日 16:52 JST

https://t.co/HwpySYglRg

1 赞 · 0 转帖 · 0 回复

### [@zhuizhui35](https://x.com/zhuizhui35/status/2121898504549253290?ref=binnn.top)

11月15日 18:30 JST

做 AI 产品，难点正在从“接上模型”转到“让它一直能用” 一个能回答问题的 Demo，和一个能连续工作的产品，中间隔着什么？ OpenAI 在 9 月 10 日推出公开测试版 Agents API，把 Codex 使用的部分 Agent 运行框架和基础设施提供给开发者，涵盖长任务、工具和运行环境等环节。 我觉得，这背后有一个值得关注的变化：大家开始把“模型之外的工作”单独做成产品。 想象一个整理资料的助手。 第一次试用，你上传几份文件，它给出一份报告，体验不错。 真正开始用，问题才会出现：第二天的新文件怎么加入？任务中断后怎么继续？处理到哪一步？哪份报告是最终版本？谁能看到结果？ 这些问题未必需要一个智商更高的模型，却都影响你能不能把工作交给它。 对开发者来说，我建议先分清两类投入： 一类，是与你的业务直接相关的东西：行业资料、任务标准、交付方式。 另一类，是支撑运行的东西：环境、状态、恢复、工具接入。 通用基础设施能替你承担一部分后者，但你仍要验证它是否适合自己的任务。 对普通用户来说，也可以换一种选工具的方式：别只看它第一次回答得多漂亮，试试连续用几天后，东西会不会丢、流程会不会乱。 我的判断是，AI 产品的差距，会越来越多地出现在第二次、第三次使用时。 来源：OpenAI 官方发布，2026-09-10；产品判断为个人分析。 https://t.co/AGEx5uGOZX

1 赞 · 0 转帖 · 1 回复

### [@CafedeVn](https://x.com/CafedeVn/status/2121906168633373110?ref=binnn.top)

11月15日 19:00 JST

caveman là gì? Cắt token cho AI agent bằng cách nói ít Xem chi tiết: https://t.co/GdUhapXVv4 #aiagent #claudecode #manguonmo #go #llm #cafedevn https://t.co/y4mVIWXbSV

1 赞 · 0 转帖 · 0 回复

### [@Luoxiaomac](https://x.com/Luoxiaomac/status/2121927449176912046?ref=binnn.top)

11月15日 20:25 JST

讲个笑话， Claude 硬生生拔高了一批中国用户网络安全水平 什么 DNS leak ，WebRTC，UDP，quic，ipv6，时区冲突，canvas WebGL 指纹.... 你毕业了吗⬇️

1 赞 · 0 转帖 · 2 回复

### [@OA20200911](https://x.com/OA20200911/status/2121963489153720633?ref=binnn.top)

11月15日 22:48 JST

腾讯开源LoopForge，把AI编程拆成流程了?? 它把一个开发任务拆成完整的研发流程 先确认需求、设计方案，AI Agent负责写代码，另一个Agent做Code Review，再跑测试，最后人工验收，不满意可以打回重做，中途停了也能从任意一步继续 跑通的流程可以存成SOP，下次开发类似功能直接照着跑，不用每次从头指挥AI Claude Code、Codex、CodeBuddy、Cursor，连DeepSeek Harness都能接，一共10种。个人用装个Local版就行，不用部署整套平台 MIT协议，完全开源 以前大家比的是哪个模型写代码强，现在开始比谁能让AI照流程把活干完，自己试出来好用的做法存下来，下次直接套用 https://t.co/xmKWXOfyk6

1 赞 · 0 转帖 · 1 回复

### [@chinchoukou](https://x.com/chinchoukou/status/2121951397998727625?ref=binnn.top)

11月15日 22:00 JST

OpenCodeReview is an AI-powered code review tool open-sourced by Alibaba, featuring a hybrid architecture that combines deterministic engineering with LLM agents. By utilizing precise fil... alibaba/open-code-review is No. 5 on GitHub Trending today. 51,599 stars · Go https://t.co/vRqQII4iGQ

0 赞 · 0 转帖 · 1 回复

### [@DukeFreed2k](https://x.com/DukeFreed2k/status/2121886189376667952?ref=binnn.top)

11月15日 17:41 JST

The full week in open-source AI — Poppy 1.0, Kitesurf, Edge0-145B, llama.cpp 1.0, and LongCat-2.5 weights: https://t.co/WKOONvTRSP #OpenSource #LLM #AIAgents

0 赞 · 0 转帖 · 0 回复

### [@paras\_raju\_](https://x.com/paras%5Fraju%5F/status/2121914605681152077?ref=binnn.top)

11月15日 19:33 JST

i scanned 31 of the biggest ai repos on github (agent frameworks, mcp servers, coding agents, \~2.5M stars combined) for ci/cd and secret issues 17 of them pull in third-party github actions that aren't pinned to a commit

0 赞 · 0 转帖 · 1 回复

### [@Rrivasm\_](https://x.com/Rrivasm%5F/status/2121922305270190365?ref=binnn.top)

11月15日 20:04 JST

Google researchers have built something that could fundamentally change how scientific research gets published. It's called PaperOrchestra, and it's much more than an AI that writes text. Instead of relying on a single language model, Google created a system where 5 specialized AI agents work together, almost like a real scientific research team. Here's how it works: Outline Agent: Organizes rough research ideas and experimental results into a structured plan. Literature Review Agent: Searches scientific literature, identifies relevant studies, and verifies citations. Plotting Agent: Generates scientific charts and conceptual diagrams. Writing Agent: Transforms everything into a complete academic manuscript. Refinement Agent: Reviews the paper, identifies weaknesses, and improves it through iterative feedback. The result? A submission-ready scientific manuscript, built from raw research materials. But here's where things get REALLY interesting. Google tested PaperOrchestra using materials reconstructed from 200 scientific papers published at major AI conferences. In evaluations by human researchers, PaperOrchestra outperformed other autonomous AI writing systems by margins of 14–38 percentage points in overall paper quality. Human-written papers still performed better overall, but the progress is remarkable. And I think the biggest implication goes far beyond writing scientific papers. For years, we've been building AI systems that answer our questions. Now we're building systems that organize themselves into specialized teams, divide complex problems into smaller tasks, evaluate their own work, and collaborate toward a common objective. PaperOrchestra doesn't independently discover new scientific knowledge. It transforms existing research into papers. But imagine connecting systems like this to AI agents capable of designing experiments, analyzing results, and developing new hypotheses. What happens when AI can participate in the entire scientific process, from the first hypothesis to the final publication? We might be approaching a world where scientific progress is no longer limited by how much research humans can personally conduct and document. Official research: https://t.co/6D4sQaXuZ7

0 赞 · 0 转帖 · 1 回复

### [@EveryDevAi](https://x.com/EveryDevAi/status/2121914104189481208?ref=binnn.top)

11月15日 19:32 JST

Reading dozens of papers just to pull out methods, sample sizes and results is slow. SciSpace from @scispace is an AI research workspace with a literature review agent, chat with PDF, data extraction from papers, a citation generator and AI Writer. It's subscription-based (Premium, Advanced, Max, Enterprise), and it connects to Zotero, Mendeley, OneDrive and GitHub.

0 赞 · 0 转帖 · 1 回复

### [@Pro\_Algorithm](https://x.com/Pro%5FAlgorithm/status/2121803833948897474?ref=binnn.top)

11月15日 12:13 JST

Alibaba open-sourced the AI code reviewer it runs internally, and developers are piling in. It's called Open Code Review, and per the repo it was Alibaba Group's official internal AI review assistant for the past two years, serving tens of thousands of developers and catching millions of code defects. Now it's open source under Apache 2.0\. What it does: - Reads your Git diffs and sends changed files to the LLM of your choice - An agent that can read full files, search the codebase and check other changed files for context - Line-level review comments, not vague summaries - Built-in rules for null pointers, thread safety, XSS and SQL injection - Works with OpenAI and Anthropic compatible endpoints The wildest claim in the README: versus Claude Code on the same model, higher precision and F1 at roughly 1/9 of the tokens. They admit recall is lower, a deliberate trade for less noise. 51k+ stars and trending on GitHub right now. https://t.co/4P0gOpr8CX

0 赞 · 0 转帖 · 0 回复

### [@SyntaxScientist](https://x.com/SyntaxScientist/status/2121946128136470530?ref=binnn.top)

11月15日 21:39 JST

PDFs came out garbled in 11 of my 20 languages. Fixed today, with a check on every build so it can't come back. Real spreadsheets too: formulas, charts, .xlsx and .ods. Free, open source, on your machine. Link in bio. Which language do you write documents in? #AI #OpenSource

0 赞 · 0 转帖 · 0 回复

### [@aisiantonas](https://x.com/aisiantonas/status/2121923355225837923?ref=binnn.top)

11月15日 20:08 JST

I have a paper going about this but it was supplanted by my alignment project which has now spawned a paper on making sense of LLM desires and why Yud's wrong. did I also mention my novel? I should probably finish something. maybe that bit needs economic incentives idk.

0 赞 · 0 转帖 · 0 回复

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