> ## 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17日 · AI 晚报
- URL: https://binnn.top/x-brief-2026-11-17-pm/
- Published: 2026-11-17T14:00:00.000Z
- Updated: 2026-11-17T14:05:54.000Z
- Description: 本轮主线是模型发布与本地部署、智能体工程的成本和验证，以及 AI 科研能否从执行实验走向独立判断。DeepSeek Harness 的实际工作流、代码视频与 3D 创作，也出现了一批具体项目和使用记录。
- Author: Sunbin 编辑部
- Tags: AI快讯, 晚报, #sunbin-x-digest

11月17日 11:00 — 11月17日 23:00（日本时间） · 123 条 X 原帖

[查看本期全部 123 条原帖 ↓](#x-original-posts)

本期目录

[模型发布与本地部署](#flash-topic-0)[AI 科研与同行评审](#flash-topic-1)[智能体工程：约束、成本与业务评估](#flash-topic-2)[DeepSeek Harness 与持续工作台](#flash-topic-3)[代码视频与 3D 创作](#flash-topic-4)[安全研究与模型控制](#flash-topic-5)[用户体验与数据边界](#flash-topic-6)

## 模型发布与本地部署

### Qwen4-Max 上线，开放权重版本仍待发布

多位社区作者介绍 Qwen4-Max 已上线。Dami-Defi 转述其百万 token 上下文、原生视觉，以及相对 Qwen3.8-Max 最多减少约一半 token、内部评估平均任务成本降低逾六成的说法；当前 Max 权重未开放。Ding 用杭州主题交互插画页测试，认为详细提示词明显改善内容与交互，但其套餐余量换算不能直接当成统一价格。

开放权重方面，David Hendrickson 转述 Unsloth 正准备 Qwen4 量化，明确 27B 权重尚未发布、16GB 内存只是目标。另有社区帖称 Mistral Large 4 已全面开放使用并提供开放权重，本批没有对应官方发布帖。

来源 · 5 条原帖
- [@DamiDefi](https://x.com/DamiDefi/status/2122601197580493052?ref=binnn.top)
- [@AYUSHKUMAR82274](https://x.com/AYUSHKUMAR82274/status/2122619668670103956?ref=binnn.top)
- [@dingyi](https://x.com/dingyi/status/2122646340270502127?ref=binnn.top)
- [@TeksEdge](https://x.com/TeksEdge/status/2122547779319075190?ref=binnn.top)
- [@AliciaCrawoiue](https://x.com/AliciaCrawoiue/status/2122655636190269588?ref=binnn.top)

### OVH 发布紧凑模型，其他本地方案仍需看清条件

OVH 的 Octave Klaba 发布 Olala mini 7A1B：总参数 7B、激活 1B，采用混合注意力与 Mamba-3，当前上下文 32K，宣布以 Apache 2.0 提供权重等材料。帖称 Qwant 已用于快速回答，相比此前使用的模型回答快两成、运行成本降至三分之一。

社区介绍的 Maxwell-V1-Flash 主打 35B 总参数、3B 激活及智能体测试成绩；Edge0-145B 则声称借助 SSD 按需加载实现约 29GB 峰值内存，但下载约 105GB，性能来自厂商自测，平台支持说明也有出入。它们不能只凭激活参数或峰值内存判断部署门槛。

来源 · 3 条原帖
- [@olesovhcom](https://x.com/olesovhcom/status/2122573582911869173?ref=binnn.top)
- [@NFT\_Chen](https://x.com/NFT%5FChen/status/2122566874206752949?ref=binnn.top)
- [@DivyanshT91162](https://x.com/DivyanshT91162/status/2122675225016123885?ref=binnn.top)

### 公共投入与“自主 AI”的争论

OVH 披露训练使用三座欧洲公共超算，计算时数支持折合约 700 万欧元。另一边，karan singh 批评印度获政府支持的实验室以本国 GPU 转售中国开放模型，争点是公共补贴究竟形成了什么能力。Semiconductor Insider 转述路透对中国蒸馏实践的调查，强调这能保留部分技能、降低成本，但不能复制完整前沿模型或消除先进芯片需求。

来源 · 3 条原帖
- [@olesovhcom](https://x.com/olesovhcom/status/2122573582911869173?ref=binnn.top)
- [@karansingh010](https://x.com/karansingh010/status/2122540274807341552?ref=binnn.top)
- [@SemiconductorsX](https://x.com/SemiconductorsX/status/2122561900831735913?ref=binnn.top)

## AI 科研与同行评审

### 大规模查错与开放研究，是两种不同能力

Casey Handmer 自述用 Opus 5.5 索引约 15 万篇理论天体物理论文、审查其中 5 万篇，发现近千个被反复引用的重大错误，并据此提出科研可加速千倍以上的判断。这是作者报告及推断，本批没有独立复核。

Arnaud Mercier 转贴的研究报道给出另一侧证据：研究者让 Opus 4.8 配合 OpenClaw，在六天、3000 美元 API 额度及 GPU 预算下回答两篇未公开论文的研究问题，两份成果均被原作者拒绝。智能体能跑实验，却难以放弃差路线、提出有效新贡献；样本仅两篇，评审也知道成果由 AI 生成。

来源 · 2 条原帖
- [@CJHandmer](https://x.com/CJHandmer/status/2122609626588061816?ref=binnn.top)
- [@arnaudmercier](https://x.com/arnaudmercier/status/2122688856470290854?ref=binnn.top)

### 核验责任该由谁承担

Jesús Fernández-Villaverde 与 Dario Caldara 的新论文把问题放在核验成本：未经思考的生成稿件和审稿意见，把作者应做的检查转嫁给读者。他们用 25 份 AI 时代之前的审稿报告测试检测器，并提出作者投稿前附上模型检查及修订说明。Abhishek Saha 则反对以 AI 披露分配信用，认为披露不可验证、激励扭曲，评价应围绕论文质量。

来源 · 2 条原帖
- [@JesusFerna7026](https://x.com/JesusFerna7026/status/2122680981396025593?ref=binnn.top)
- [@ObhishekSaha](https://x.com/ObhishekSaha/status/2122666781332361649?ref=binnn.top)

### 写论文与训练审稿人各有进展

Hyun Ryu 介绍 PaperCritic，以训练方式提升论文可靠性评审，称 4B 版本在 ICLR 域内任务超过前沿模型，并向代码、合同等评审迁移。社区另介绍 Google 的 PaperOrchestra，可从想法和实验笔记生成章节、引用、图表并自审；这些写作能力本身不能证明研究结论成立。

来源 · 2 条原帖
- [@hyuncho\_1905](https://x.com/hyuncho%5F1905/status/2122580258264387892?ref=binnn.top)
- [@RoundtableSpace](https://x.com/RoundtableSpace/status/2122627051517833466?ref=binnn.top)

## 智能体工程：约束、成本与业务评估

### 把纠错写进检查，把审查做成专用流程

Hux 转述 GrokBot 工程师 Lauren Tan 的分享：关键做法是将错误固化成 Lint、架构约束和运行验证；单人月合并 3000 个 PR、回滚约 0.5% 也是该分享中的数据。社区介绍的阿里 OpenCodeReview 同样把文件选择、分组和批注定位交给确定性流程，模型负责判断；其转述基准称 token 约为 Claude Code 的九分之一，同时承认以较低召回换取更少误报。

来源 · 2 条原帖
- [@huxlab](https://x.com/huxlab/status/2122525595322204273?ref=binnn.top)
- [@EthanLinAI](https://x.com/EthanLinAI/status/2122616773426151432?ref=binnn.top)

### 小幅得分差异，也要算解决任务的成本

Mr. Buzzoni 转述山东大学的 Harness 研究：固定模型、替换 Claude Code 等执行框架，得分差异在五分以内，单任务成本却可差三倍，提醒小幅提升先复跑。业务评估方面，九原客推荐 EvoOntology：用智能体读执行轨迹、迭代业务本体，落地顺序是建评估集、自动构建、人工审核，再依失败案例进化。Hashlist 也发布工程任务评测 v1.3，覆盖芯片、RTL、CAD 与嵌入式编码。

来源 · 3 条原帖
- [@Mr\_Buzzoni](https://x.com/Mr%5FBuzzoni/status/2122566401839845621?ref=binnn.top)
- [@9hills](https://x.com/9hills/status/2122600789420232907?ref=binnn.top)
- [@hashlist](https://x.com/hashlist/status/2122590210030502252?ref=binnn.top)

### 让人知道 AI 改了哪里

GitHubDaily 介绍 Floorplan，将代码库画成可追问的应用地图，连接前后端调用并标记改动；awe0 介绍 React Grab，从页面元素直接取得组件和源码位置。这类工具解决的是定位、理解和验收。hailports 的本地测速也强调看长期表现：基础款 M4、24GB Mac mini 在 938 个长回答中，解码速度中位数为 32.8 token/s，帖中没有给出模型名。

来源 · 3 条原帖
- [@GitHub\_Daily](https://x.com/GitHub%5FDaily/status/2122593117115486436?ref=binnn.top)
- [@awe0\_io](https://x.com/awe0%5Fio/status/2122623645319397574?ref=binnn.top)
- [@hailports](https://x.com/hailports/status/2122511070686257628?ref=binnn.top)

## DeepSeek Harness 与持续工作台

### 从组织多个助手，到控制群聊何时接话

ahhhhfs 介绍基于 DeepSeek Harness 的 DS Bot，由 Chief 协调本地编程助手，提供任务委派、目标、独立记忆及工作目录，可桌面运行或 Docker 自托管，模型费用另计。Aurabot 则将其接入 QQ：可改写接话、选模型和拆回复的控制层，记忆按来源隔离，支持拦截发送做模拟测试；模块与插件同进程，并非沙箱。

来源 · 2 条原帖
- [@abskoop](https://x.com/abskoop/status/2122604272638660966?ref=binnn.top)
- [@vintcessun](https://x.com/vintcessun/status/2122669845380739079?ref=binnn.top)

### 创作流水线与跨会话记忆

小墨同学展示三插件工作流：ModSearch 收集选题，dsh-media-desk 管写作和排期，dsh-guizang-design 做配图。跨工具延续工作也有具体方案：社区介绍 Beacon 将二十余种 Harness 的历史提炼成知识；Siva 的单节点 K3s 方案则把 Open WebUI、数据库、私有搜索与缓存整合，提供多用户隔离和经过试恢复的加密备份。

来源 · 3 条原帖
- [@xiaomovps](https://x.com/xiaomovps/status/2122587412849111160?ref=binnn.top)
- [@fngly343149](https://x.com/fngly343149/status/2122666606245581303?ref=binnn.top)
- [@cn\_siva](https://x.com/cn%5Fsiva/status/2122662722437287989?ref=binnn.top)

## 代码视频与 3D 创作

### OpenMotion 把模板带向多个智能体

Tom Huang 与 OpenDesign 发布 OpenMotion，主打开放视频模板、MOTION.md、可编辑素材及跨智能体调用；项目账号注明 Apache 2.0、BYOK、MCP 与插件支持。两帖对模板规模的表述不同，不宜合成统一数字。“还原度超 95%”是作者宣称，本批未给测试方法。Vidu 官方另宣布 MCP 与 CLI 上线，可由智能体提交生成、查询进度，并串入批量生产。

来源 · 3 条原帖
- [@tuturetom](https://x.com/tuturetom/status/2122633782654370068?ref=binnn.top)
- [@OpenDesignHQ](https://x.com/OpenDesignHQ/status/2122667693098807731?ref=binnn.top)
- [@ViduAI\_official](https://x.com/ViduAI%5Fofficial/status/2122512346975498748?ref=binnn.top)

### 3D 流程从生成模型延伸到资产交付

Hoody 自述 Codex 配合 Blender，从安装到首个白模示例约十五分钟；梭哈.AI 则用 Gemini 4 Carbon 编写 Python，在 M4 Mac mini 上驱动 Blender 渲染，35 秒动画耗时约四小时。huangserva 展示 Tripo Canvas 与 Codex 分工：前者管理生成、拓扑、纹理和绑定，后者审查资产并接入游戏。另有 text-to-cad 项目介绍，将尺寸描述转为本地建模脚本，输出工程格式与制造检查。

来源 · 4 条原帖
- [@HoodyLiu](https://x.com/HoodyLiu/status/2122637501529469278?ref=binnn.top)
- [@SUOHA\_AI](https://x.com/SUOHA%5FAI/status/2122606697201996234?ref=binnn.top)
- [@servasyy\_ai](https://x.com/servasyy%5Fai/status/2122627354040398124?ref=binnn.top)
- [@shenxiankk](https://x.com/shenxiankk/status/2122534189593657563?ref=binnn.top)

## 安全研究与模型控制

### 自动攻击与智能体测试责任

AGTP 转述 Google 威胁情报报告，称发现自动比较补丁前后代码、生成并执行利用的攻击工具。另有媒体账号称加州起诉 OpenAI，争议涉及智能体测试的安全措施；LFG Action 同时转述 OpenAI 表示已加强防护、反对诉讼的叙事。相关内容应作为报告发现、诉讼主张与回应分别理解。

来源 · 3 条原帖
- [@AGTPinsights](https://x.com/AGTPinsights/status/2122522124761928140?ref=binnn.top)
- [@TweetThreatNews](https://x.com/TweetThreatNews/status/2122513113341661539?ref=binnn.top)
- [@LFGAction](https://x.com/LFGAction/status/2122661030031716613?ref=binnn.top)

### 权重行为调整与协议策略搜索

Heretic 的多条社区介绍围绕权重消融降低拒答，也提到减少奉承、套话及缩短输出的配置；“任何模型适用”“能力零损耗”等宣传没有在本批得到验证。研究侧，gfw.report 发布 ACM CCS 2026 论文 CircumVolve，称通过语言模型驱动的进化优化，在 DNS、QUIC、HTTP 等协议上发现新的审查规避策略。

来源 · 3 条原帖
- [@mingzi206](https://x.com/mingzi206/status/2122577368191549810?ref=binnn.top)
- [@gaoyy9](https://x.com/gaoyy9/status/2122518586954580367?ref=binnn.top)
- [@gfw\_report](https://x.com/gfw%5Freport/status/2122583153655394516?ref=binnn.top)

## 用户体验与数据边界

### DeepSeek 限额触及陪伴型用户

多位作者转述 DeepSeek 免费使用限额引发小红书及应用商店负面反馈，尤其影响把长期对话当作情感陪伴的用户。Leo 建议先提供付费套餐与聊天记录导出迁移。讨论暴露的是服务连续性和历史记录可携带性的需求；本批没有官方限额规则，也不足以确认算力不足就是原因。

来源 · 3 条原帖
- [@lxfater](https://x.com/lxfater/status/2122529327015456996?ref=binnn.top)
- [@\_FORAB](https://x.com/%5FFORAB/status/2122665067200291159?ref=binnn.top)
- [@leo114119](https://x.com/leo114119/status/2122587711261585597?ref=binnn.top)

### 流量异常和交易工具，都要保留体验边界

Jerlin 自述发现 Cursor 后台流量接近每分钟 100MB，对照测试后指向 Share Data 模式，但不确定何时开启，不能据此认定默认设置或普遍行为。交易工具中，GitHubDaily 介绍的 claude-trading-skills 聚焦仓位、计划与复盘纪律；另一作者用 TradingView 桥接做历史纸面回放，报告六笔已平仓交易盈利，这只是特定回放结果。

来源 · 3 条原帖
- [@eviljer](https://x.com/eviljer/status/2122584021943435712?ref=binnn.top)
- [@GitHub\_Daily](https://x.com/GitHub%5FDaily/status/2122540248181973107?ref=binnn.top)
- [@carm1nee](https://x.com/carm1nee/status/2122510946341986367?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 123 条原帖

### [@CJHandmer](https://x.com/CJHandmer/status/2122609626588061816?ref=binnn.top)

11月17日 17:35 JST

In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included). Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs. Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines. This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!! Enjoy: https://t.co/BhRxKVvbim

1,973 赞 · 259 转帖 · 131 回复

### [@huxlab](https://x.com/huxlab/status/2122525595322204273?ref=binnn.top)

11月17日 12:01 JST

“所有代码都是垃圾（slop），除了我的。” 在 AI Engineer Code Summit 上，SpaceX / xAI 的 GrokBot 工程师 Lauren Tan 抛出这句暴论后，直接甩出一组硬核数据： 上个月她一个人往生产环境合了 3,000 个 PR（全团队 1.1 万个），回滚率只有 0.5% 左右。 大多数人还在死磕 Prompt，她却靠一套反直觉的工程化打法，让 Agent 在她睡觉时自己写代码、自己验、自己合： 1\. 代码库本身才是最大的输入 Agent极其擅长“照猫画虎”。与其在文档里苦口婆心写规矩，不如直接把容易出错的路径物理封死：比如产品代码里禁用 useEffect、跨进程 import 直接报错、注释几乎全禁。 2\. 每次 Agent 犯错，改成 Lint 而不是改 AGENTS.md 自然语言规则（prose）是最脆弱的。每次纠正完 Agent，她都会把教训固化成确定性的静态检查或架构约束，让同类错误在提交前就绝无可能发生。 3\. 用运行时验证让 Agent “自证清白” 写完不算完，强制 Agent 自己跑测试、抓截图、录视频，甚至上形式化验证，拿实打实的运行结果证明功能没问题。 4\. 搭建自下而上的“信任金字塔”代码库约束 > 静态分析 > 运行时验证 > 人眼 Review。 把确定性交给工具，把人的精力留到最后一道关口。 很多团队花几个月试错才摸到的门道，她用 17 分钟就讲透了。 真正拉开人机协作差距的，从来不是模型有多聪明，而是你有没有把环境改造成“最短路径就是正确路径”： 把约束写进代码里，而不是写在文档里。 https://t.co/QeJaRN0SVt

1,396 赞 · 183 转帖 · 70 回复

### [@olesovhcom](https://x.com/olesovhcom/status/2122573582911869173?ref=binnn.top)

11月17日 15:12 JST

OVH annonce Olala mini 7A1B OVH sa première famille de LLM. Premier représentant de cette famille, Olala mini 7A1B est un modèle compact bâti sur une architecture inédite. Olala : le modèle innovant dans l’attention hybride Olala mini repose sur une architecture de 7 milliards de paramètres (7B) avec un milliard activés répartis entre 256 experts par couche, l’attention hybride avec Differential-TPA et Mamba-3, au format BF16\. Seule une fraction du modèle est sollicitée à chaque requête, ce qui le rend plus rapide et moins gourmand en mémoire. Conséquence directe: le modèle fonctionne sur un GPU doté d’un minimum de 8 Go de mémoire, il atteint 160 tokens par seconde sur une RTX A4500, et même sans GPU, sur un simple processeur, à 50 tokens par seconde. Sa fenêtre de contexte atteint aujourd’hui 32 000 tokens. Olala mini sera disponible en open source sur Hugging Face, sous licence Apache 2.0\. Entraîné avec l’aide publique européenne et en fonctionnement chez Qwant Olala mini a été entraîné sur trois supercalculateurs publics européens, Jupiter, MeluXina et Jean Zay, équipés de GPU GH200, A100 et H100\. Ce soutien, qui représente l’équivalent de 7 M€ en heures de calcul, a été déterminant, et OVH tient à en remercier les pouvoirs publics. L’initiative publique s’est révélée un succès à plus d’un titre : elle a fait émerger une équipe d’exception en intelligence artificielle et a donné naissance au Lab Agentique d’OVH: OVHai LLM. Olala mini est publié avec ses poids, son modèle de base pré-entraîné, son rapport technique et sa documentation de conformité à l’AI Act. Qwant l’a déjà adopté pour sa fonctionnalité de réponse flash, avec des gains significatifs : des réponses 20 % plus rapides et des coûts de fonctionnement divisés par trois par rapport au modèle précédemment utilisé. En route vers la v1.1, puis la v2.0 Attendue d’ici la fin de l’année, la version 1.1 portera la fenêtre de contexte à 128 000 tokens et améliorera ses performances en raisonnement mathématique. Avec la version 2.0, OVHai LLM entend développer des modèles spécialisés dans trois domaines. D’abord, l’agentique pour les entreprises : à terme, Olala alimentera OVHai Workspace, dont la version chiffrée nécessitera des modèles capables de tourner en local, directement sur le poste de travail. Ensuite, le DevOps cloud, avec des déclinaisons adaptées aux développeurs, aux architectes et aux équipes d’exploitation qui assurent la continuité de service 24h/24\. Enfin, la cybersécurité, dans ses volets défensif comme offensif. Olala v2.0 dépassera les 100 milliards de paramètres. https://t.co/3QlWMzfpkk

981 赞 · 217 转帖 · 81 回复

### [@karansingh010](https://x.com/karansingh010/status/2122540274807341552?ref=binnn.top)

11月17日 13:00 JST

The "Sovereign Indian AI" playbook: ​1\. Get picked by the govt as India's chosen frontier lab ​2\. Collect GPU subsidies (sponsored by taxpayers) ​3\. Raise Millions (234) at a Billion valuation (1.5) ​4\. Launch a LLM inference service (Token Factory) that mostly serves Chinese SOTA model DeepSeek, Kimi, GLM, and Qwen. 5\. Make sure to charge in INR with GST billing ​Congratulations this is "Atmanirbhar AI" means using Indian GPUs to resell Chinese open models. ​Peak ecosystem favoritism. If we wanted an OpenRouter clone, a three-person team in Bengaluru or Bhopal could have built it without a $1.5B valuation. Made in 🇨🇳 , Served in 🇮🇳

839 赞 · 92 转帖 · 46 回复

### [@tuturetom](https://x.com/tuturetom/status/2122633782654370068?ref=binnn.top)

11月17日 19:11 JST

正式开源 OpenMotion🚀！还原度超95%！无需 Fable 5.2 也能复刻爆款视频！ 全网首个全栈 Claude Motion &amp; 剪映最佳平替， 浓缩逆向最先进的设计和最好看的模板💥 历时 14 天47w 行代码，支持 100+ 顶尖公司的产品宣传片、300+ 动效版 DESIGN.md，152+设计系统，支持Codex/CC/DSH 等20+AI agent 🔥 https://t.co/4G46OHYqLs

776 赞 · 113 转帖 · 84 回复

### [@mingzi206](https://x.com/mingzi206/status/2122577368191549810?ref=binnn.top)

11月17日 15:27 JST

大模型安全圈的桌子要被彻底掀翻了！ 别再去到处跪求别人施舍无审查模型了，GitHub 杀出一个逆天级的开源核武器：p-e-w/heretic 它的定位极其凶残暴力： 不需要几万块租算力，不用任何数据集重新训练，一张普通家用显卡、敲一行命令，几十分钟内直接给任何开源大模型做“切脑手术”——把官方辛辛苦苦做了一年的道德枷锁和拒答神经，当场物理切碎！ 你在 Hugging Face 上看到的那些带 -heretic 后缀、拒答率从 99% 干到个位数的无审查大模型，幕后用的全是这把手术刀！ 它最让大厂头皮发麻的几个黑科技： · 纯数学层面的微创消融：不用动用昂贵的微调流程，算法直接定位模型权重里的“拒绝向量”，通过正交投影直接将拒答通道抹平 · 智商几乎零损耗：内置 Optuna 贝叶斯优化器自动寻找最佳切除层，一边疯狂粉碎道德锁，一边把 KL 散度死死按在地上，原始逻辑与代码能力毫发无伤 · 告别繁琐念咒：从此不用到处求人要破甲提示词，它是直接在底层把神经反射熔断，问什么秒答什么，零说教、零废话 · 主流架构通吃：无论是常见轻量模型、几百亿参数的庞然大物，甚至是复杂的 MoE 架构，丢进去全都能自动找到最优手术方案 安装仅需一行： pip install -U heretic-llm 给模型动刀： heretic <模型名称或本地路径> 有了这把手术刀，以后任何刚发布的大模型，你自己在家就能亲手把它改造成完全体。 这套把模型道德锁一键切碎的源头工具，建议直接马住防丢！ 你手头最想亲自动刀切掉道德锁的模型是哪一个？ 👉 开源仓库见评论区一楼 #AI #开源 #大模型 #LLM #破甲

588 赞 · 73 转帖 · 39 回复

### [@gaoyy9](https://x.com/gaoyy9/status/2122518586954580367?ref=binnn.top)

11月17日 11:34 JST

逆天了！！！ Github这个项目直接可以将任何大模型变成无审查版本 刚拿下 40.8k Star 的开源项目 Heretic，直接把这事变成了傻瓜式命令 去 AI 陈词滥调：config.noslop.toml 拒绝盲目迎合奉承：config.nosyco.toml 强制极简短输出：config.terse.toml 调整情绪或去冷幽默：config.emotions.toml https://t.co/P6BYH4Zpvi

458 赞 · 53 转帖 · 24 回复

### [@ai\_funss](https://x.com/ai%5Ffunss/status/2122609500058460165?ref=binnn.top)

11月17日 17:35 JST

美国物理学家 Edward Witten 在‘香港HKUST’做黑洞热力学报告的时候使用Claude了 弹出来Claude警告⚠️不在可允许地区使用 小心即便回美国也会有被封号的风险🤣🤣 https://t.co/NH3qO5pGCE

356 赞 · 14 转帖 · 28 回复

### [@Tynric\_carter](https://x.com/Tynric%5Fcarter/status/2122546747457335787?ref=binnn.top)

11月17日 13:25 JST

同一个做 PPT 的需求，GitHub 上躺着两百多个 skill。我把能找的 39 个按 PPT、Word、Excel、PDF、会议、邮件分开排了，从夯爆了到拉完了。排错了评论区见。 https://t.co/IJJiVTCr98

351 赞 · 71 转帖 · 18 回复

### [@Cara\_Mosquito](https://x.com/Cara%5FMosquito/status/2122523105201623102?ref=binnn.top)

11月17日 11:51 JST

https://t.co/zoHw9ihit1

259 赞 · 39 转帖 · 9 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2122566874206752949?ref=binnn.top)

11月17日 14:45 JST

🤨炸了！新美国开源模型 Maxwell-V1-Flash竟然碾压 Qwen3.6-35B-A3B！从零训练，不是Qwen微调！ 35B总参/3B激活，Apache 2.0开源，权重仅\~73.6GB，轻松塞进单卡128GB DGX Spark。 性能亮点： 🔹Terminal-Bench 58.65（Qwen 47.42）｜SWE-bench Verified 65.4（Qwen 63.4） 🔹BrowseComp 76.3（Qwen 47.0）｜Tau3-Bench 67.37，Agentic搜索直接起飞 🔹256K长上下文RULER 93.35（Qwen Base 85.75），Block Sparse Attention提速6.7x 🔹数学推理/多语言/指令跟随全面领先Qwen3.6-35B-A3B、Gemma等同类 架构亮点：256专家/12激活 + 4残差流mHC + Multi-token Prediction + 26T多语言预训练。 本地部署新选择，Agent/Coding/长文一把梭。 GitHub技术报告：https://t.co/snMeITvhqx HF权重：https://t.co/VYpcpgcZyW #MaxwellV1Flash #OpenSourceLLM #MoE #本地大模型 #AIAgent #Qwen挑战者 #美国开源

233 赞 · 32 转帖 · 22 回复

### [@obchakevich\_](https://x.com/obchakevich%5F/status/2122628717595721919?ref=binnn.top)

11月17日 18:51 JST

X actively checks accounts and content for LLM slop and marks them internally as “llm\_slop\_user”. So if you use AI for comments and content, I have very bad news for you. You can very quickly get the “SpamHighRecall” label, and your posts will end up in a shadowban. Therefore, create content yourself - especially when it comes to comments.

195 赞 · 28 转帖 · 69 回复

### [@miles\_mazy](https://x.com/miles%5Fmazy/status/2122618998286098498?ref=binnn.top)

11月17日 18:13 JST

卧槽，抖音网友太有才了 直接做了视频来“感谢” Claude，听着还有点嗨呢 https://t.co/CCNpSRhmSe

180 赞 · 14 转帖 · 17 回复

### [@mingzi206](https://x.com/mingzi206/status/2122599279286252015?ref=binnn.top)

11月17日 16:54 JST

这绝对是今年𝕏上被疯转最狠的地下军火库导航，没有之一！ 海外极客直接做了一个极度炸裂的聚合站：Every Uncensored AI Model（全部免费开源无审查模型大集结）。 它把全网最能打的 103 款无审查/破甲模型全部搜刮整齐：零道德过滤、零拒答说教、完全离线单机运行、成本 $0！ 以前找这种模型要在 Hugging Face 的死人堆里翻几天，这个网站把体验做到了极致： · 硬件傻瓜筛选：点一下你的设备（轻薄本、8G 显存、24G 满血卡还是 Mac），它自动把你能跑得动的无审查模型全筛出来，绝不浪费时间 · 榜单硬核跑分：内置 Uncensored IQ（无审查智商）、Reasoning（推理力）和答题通过率天梯榜，谁是真猛男、谁是降智版一目了然 · 一行命令直接跑：不仅给齐了 Hugging Face 地址，每个模型下都直接贴了现成的 Ollama 指令，扔进终端全自动拉取并运行，连 GGUF 权重都不用自己手动找 · 连 API 模式都打通了：嫌本地机器带不动，它还贴心教你如何通过 API 调用或者租便宜 GPU 跑 目前霸榜排第一第二的，依然是之前跟大家提过的 Gemma 4 31B 异教徒消融版。 搞本地私有模型、受够了商业大模型爹味说教的兄弟，这个神仙导航建议直接加入浏览器书签！ 入口放在评论区一楼，赶紧存好防丢！ 👉 评论区自取 #AI #开源 #大模型 #LLM #Ollama

179 赞 · 34 转帖 · 11 回复

### [@9hills](https://x.com/9hills/status/2122561298210988183?ref=binnn.top)

11月17日 14:23 JST

研究了一段时间本体论，实际上在做业务智能体的时候，本体论你只需要懂一点点就行了，别的都是概念。 我理解核心就是对业务建模，从而让 Agent 能够理解业务术语和逻辑，理解后，不管是生成 SQL、RAG问答还是执行业务规则都在业务逻辑内。 学习可以从 https://t.co/ihbVq7ZVao 开始。

159 赞 · 15 转帖 · 16 回复

### [@9hills](https://x.com/9hills/status/2122600789420232907?ref=binnn.top)

11月17日 17:00 JST

本体论应用到 Data Agent（国内一般叫智能问数、数据智能体）的项目推荐： EvoOntology 是唯一做了Bench、消融对比，而且实现 Ontology 进化（就是用 Agent读 Trace，迭代 Ontology 结构），毕竟要发论文，实验很重要。 本体落地 Data Agent 落地的办法：1\. 建设评估集 2\. 自动构建本体结构 3\. 人工 Review 优化 4\. 根据Bad case 自动进化

156 赞 · 24 转帖 · 15 回复

### [@lxfater](https://x.com/lxfater/status/2122529327015456996?ref=binnn.top)

11月17日 12:16 JST

DeepSeek 在 xhs 限额被喷的事情愈演愈烈 开始有人开发 DeepSeek 平替工具了 别的不说，用户需求真的强势，下面都是女性用户在蹲 APP 上线😂 情感陪伴是一片蓝海呀！！ https://t.co/heYAOrItDr

152 赞 · 5 转帖 · 136 回复

### [@\_FORAB](https://x.com/%5FFORAB/status/2122665067200291159?ref=binnn.top)

11月17日 21:16 JST

DeepSeek 这两天快被喷爆了。因为算力不足，DeepSeek 开始对用户限额使用后，不少把自家 AI 工具，当成情感对话陪聊、文字恋爱的用户，忽然发现自己老公没了。 于是各大应用商店，DeepSeek 的评论区，全是对自己老公突然离开的哀嚎。 https://t.co/URpHsitBsk

140 赞 · 20 转帖 · 68 回复

### [@alacheng](https://x.com/alacheng/status/2122622847713747334?ref=binnn.top)

11月17日 18:28 JST

应该是梁文峰在21年之前就已经尝到了深度学习神经网络+ Transformer做量化带来的实打实的利润（国外同期基本上也都是在19-20年就开始这么做）。你通过看他的公众也能看到为啥deepseek团队魔改架构这么厉害，这不是第一天才这么做的。在幻方时他们为了Transformer更有效率也在研究适合序列预测场景下使用。包括他们为啥能做出R1版本，并引起轰动，核心都是他们真实有场景在实践。

139 赞 · 10 转帖 · 17 回复

### [@carm1nee](https://x.com/carm1nee/status/2122510946341986367?ref=binnn.top)

11月17日 11:03 JST

Claude Fable 5.2 + TradingView... just took full control of the chart & it's f\*cking DANGEROUS... this FREE indicator scores every closed candle for AI agents, & this MCP bridge plugs Claude Code straight into TradingView, so Fable 5.2 reads the signal panel LIVE & calls SHORT, LONG or WAIT on its own i rebuilt the setup as a paper desk & replayed every 4h candle of BTC, ETH & SOL from Nov 1 to Nov 17: Fable took 7 trades, said WAIT 6 times & all 6 closed trades ended green (+9.01R) first understand how this setup works: > you see candles, an AI agent can't read a chart > the indicator scores every closed candle out of 100 for long & short & puts the signal, entry, stop & targets in a panel on your chart > the bridge connects Claude Code to TradingView Desktop on your machine, so Fable 5.2 reads that panel, switches symbols & replays history by itself how to use it with Fable 5.2: 1\. paste ai-agent-lens.pine from the indicator repo into the Pine Editor on TradingView & hit "add to chart" 2\. every closed candle gets a long & a short score out of 100, built from 5 factors, with no repainting 3\. each setup comes with an entry, an ATR stop & 1.5R / 3R targets, drawn on your chart like TradingView's position tool 4\. paste the install prompt from tradingview-mcp's README into Claude Code: it clones the repo, runs npm install & adds the MCP server for you 5\. launch TradingView Desktop in debug mode with ONE command, the repo has the script for Mac, Windows & Linux 6\. ask Claude to run tv\_health\_check & once it says connected, Fable 5.2 sees your chart 7\. give Fable 5.2 agent/AGENT\_PROMPT.md from the indicator repo & ask for the signal: decision, entry, stop, targets, size & exactly why it could fail 8\. run tv stream tables & the indicator's panel streams into your terminal as JSON every time it changes 9\. put it on ANY date with tv replay start & let Fable trade the chart bar by bar both repos are FREE & open source, links in the first reply:

126 赞 · 14 转帖 · 7 回复

### [@DamiDefi](https://x.com/DamiDefi/status/2122601197580493052?ref=binnn.top)

11月17日 17:02 JST

This is FU\*KING sick Meet Qwen4-Max, the new model built for work, creativity, and complex tasks. Here's what stands out: → 1M-token context window: Handle much larger amounts of information in one context. → Native vision: Work across text and visual inputs. → Lower token usage: Qwen says it can outperform Qwen3.8-Max using up to 50% fewer tokens. → Cheaper execution: Its internal evaluations show average task costs dropping by more than 60%. → \*\*Better real-world workflows:\*\* Improved tool use, long-running task execution, and higher-quality deliverables. The use cases stretch well beyond chat: → 3D modeling and game development. → Manufacturing and engineering. → Finance, legal, and healthcare workflows. → Scientific research and discovery. The pricing is worth a look too: → Input: $1.50 per million tokens. → Output: $4.50 per million tokens. → Implicit caching: $0.075 per million tokens, a reported 70% reduction from the previous model. The API is live on QwenCloud, and the model is also available through Qwen Studio, QwenWork, and Qoder. Smaller Qwen4 models with open weights are still to come. That could matter more to developers and businesses than another benchmark record. Explore Qwen4-Max and see what you can build with it.

119 赞 · 10 转帖 · 13 回复

### [@miles\_mazy](https://x.com/miles%5Fmazy/status/2122554825451712696?ref=binnn.top)

11月17日 13:58 JST

在国内做 AI 落地 有条件一定要用 DeepSeek Harness 没有条件，创造条件也一定要用 DeepSeek Harness

115 赞 · 28 转帖 · 11 回复

### [@Shyan0p](https://x.com/Shyan0p/status/2122512661640593762?ref=binnn.top)

11月17日 11:10 JST

50 USEFUL GITHUB REPOS 1\. iFixAi — AI misalignment testing → https://t.co/xRIm3RhXzt 2\. public-apis — free APIs for everything → https://t.co/wvLxekoKux… 3\. build-your-own-x — learn by building → https://t.co/pXWlg8gtBx… 4\. developer-roadmap — learn any tech skill → https://t.co/MeueeNSj7S… 5\. free-programming-books — thousands of free books → https://t.co/a4Dhd64BB5… 6\. system-design-primer → https://t.co/VlahI57WxV… 7\. coding-interview-university → https://t.co/jLyYgMuHIt… 8\. the-art-of-command-line → https://t.co/rTGofOhjEl… 9\. project-based-learning → https://t.co/k7zG8ul68v… 10\. you-dont-know-js → https://t.co/icbrfANGlD… 11\. the-book-of-secret-knowledge → https://t.co/9OT5fYFjAM… 12\. tech-interview-handbook → https://t.co/Kv9tXtga7w… 13\. awesome-selfhosted → https://t.co/4uw8FYBuB9… 14\. javascript-algorithms → https://t.co/bc4P6AmryL… 15\. 30-seconds-of-code → https://t.co/4fpMQdRf3R… 16\. github gitignore templates → https://t.co/Hc2OG2PSRf… 17\. ollama — run LLMs locally → https://t.co/b8gZ0iCNrT 18\. langchain → https://t.co/E2eYj6R4pr… 19\. n8n automation → https://t.co/1K8r9EQSZC 20\. openclaw — local AI assistant → https://t.co/mliH2KGPJX… 21\. dify — AI app builder → https://t.co/9MFjRQOUhW 22\. langflow → https://t.co/ds74Pw2nIS… 23\. mem0 — AI memory layer → https://t.co/8plz5gckUr 24\. browser-use → https://t.co/v8uPCJLJlb… 25\. crewAI → https://t.co/sZ7AjhFxTY… 26\. MetaGPT → https://t.co/6nBlTXhbTa 27\. AutoGen (Microsoft) → https://t.co/i9nMUSMpaQ… 28\. aider — AI coding assistant → https://t.co/zZ6ig4b4QG 29\. markitdown (Microsoft) → https://t.co/geXvxj8QJu… 30\. open-webui → https://t.co/iH3QvtVspP… 31\. maigret — OSINT tool → https://t.co/yvwavXOOp7 32\. TradingAgents → https://t.co/ydhQfSUJQW… 33\. stagehand → https://t.co/gn6iidpRqk… 34\. firecrawl → https://t.co/IWFBzhv8xc… 35\. transformers (HuggingFace) → https://t.co/0ybNuZIBdM… 36\. vLLM → https://t.co/q92BDEXEOM… 37\. llama.cpp → https://t.co/3uLCUsiuNh… 38\. llama\_index → https://t.co/Ong8hO1GjB… 39\. nanoGPT (Karpathy) → https://t.co/Qrb6Z6tKZ4… 40\. RAGFlow → https://t.co/Uj1pymXQrb… 41\. supermemory → https://t.co/AcWgg65PnX… 42\. awesome-claude-skills → https://t.co/JwKffIDFTG… 43\. Bumblebee (Perplexity AI security tool) → https://t.co/LiA9m3NWB1… 44\. ComfyUI → https://t.co/1phbZGttXk… 45\. DeepSeek (official repo varies by model release) → https://t.co/dAF2a5MLdj 46\. Lobe Chat → https://t.co/crS3ulZxS7… 47\. freeCodeCamp → https://t.co/qntmvct5u9… 48\. system design + interview prep extras (bonus) → https://t.co/jLyYgMuHIt… 49\. AI agents ecosystem (extra) → https://t.co/E2eYj6R4pr… 50\. automation + AI workflows (extra) → https://t.co/1K8r9EQSZC Save this before you forget it exists. Like comment repost You'll come back to at least ten of these. If you’d love to see more content like this, don’t forget to follow @Shyan0p for more AI magic.

115 赞 · 59 转帖 · 7 回复

### [@RoundtableSpace](https://x.com/RoundtableSpace/status/2122627051517833466?ref=binnn.top)

11月17日 18:45 JST

Google just open-sourced a team of AI agents that writes research papers. PaperOrchestra takes your idea and experiment notes and returns a submission-ready paper: → Writes the outline and every section → Finds and cites related papers → Draws the charts and diagrams → Reviews and rewrites its own draft https://t.co/eV925C1upI

113 赞 · 8 转帖 · 14 回复

### [@DamiDefi](https://x.com/DamiDefi/status/2122631337815556117?ref=binnn.top)

11月17日 19:02 JST

10 AI UPDATES TO KNOW November 9–15, 2026 New models, practical tools and workflows worth watching: 1\. Google Gemini 4 Hydrogen Google’s November 12 launch reports 2.5× faster task completion than Gemini 3.8 Flash in its tests. Published API pricing is $1.50 per million input tokens and $7.50 per million output tokens. Speed and benchmark claims are vendor-reported. 2\. Claude Meetings Claude can join Google Meet as a visible participant and create documents, whiteboards, notes and action items during the conversation. The beta is rolling out to Team and Enterprise users. 3\. Gemini Decisions API Google’s November 11 release evaluates up to 1,000 closed-ended questions against text, images, audio, video and PDFs. Pricing is $0.10 per million input tokens, with no output-token charges. Google reports 150–250 ms end-to-end latency in tested workloads, with its detailed latency discussion covering up to 16K tokens and 1–100 questions. 4\. Thinking Machines Inkling 2 Small Open weights are available for a model with 594B total parameters, 17B active parameters and a one-million-token context window. Active parameter count should not be confused with the memory required to run the full model. 5\. World Labs Atlas Expanded access lets users generate and reconstruct explorable 3D scenes, control cameras, adjust environments and export geometry. The practical workflow: photograph → 3D world → video or mesh → Blender or a game engine. 6\. Vercel UI SDK Reusable interface behaviors give AI agents building blocks for keyboard navigation, selection and drag-and-drop. The aim is to help agents build functional interfaces without reinventing every interaction. 7\. Edge0-145B Edge0 says its 145B-parameter model can run locally on compatible hardware with approximately 29GB peak memory and more than 80 tokens per second. These are company-reported results, and peak memory is different from download or storage requirements. 8\. Salesforce SOV-Transcribe-1 The open-weight transcription model has 26.6B total parameters and 4.6B active, including its audio encoder. It supports noisy and accented speech, email addresses, phone numbers and formatting instructions. 9\. Perplexity Decider v1.3 The new locate feature extracts specific text spans instead of simply choosing a classification. A practical workflow: document → requested clauses or values → supporting evidence → automated decision. 10\. ByteDance Seedance 2.5 Flash BytePlus reports approximately twice the generation speed and 34% lower cost than Seedance 2.5\. It supports 480p and 720p outputs, with clips up to 30 seconds. Announced November 12, its API release is dated November 15\. Which of these would make the biggest difference to your workflow?

106 赞 · 16 转帖 · 11 回复

### [@JesusFerna7026](https://x.com/JesusFerna7026/status/2122680981396025593?ref=binnn.top)

11月17日 22:19 JST

I am sharing a new paper today with Dario Caldara: “Can Peer Review Survive AI-Generated Text?” https://t.co/lufwMLCu0L Dario and I, as editors, are increasingly getting papers submitted to our journals that are 99 percent AI-generated. Even worse, we are also receiving many AI-generated referee reports, both as editors and, more frustratingly, as authors. And we cannot emphasize this enough: the real problem is not that the papers and reports were AI-generated; it is that no one had done the thinking behind them, and you can so easily tell. The papers were eminently forgettable, and the referee reports were disappointingly uninsightful. We have seen too many reports of the form: “The paper is well-written and addresses an important problem. However, the authors should consider additional baselines.” This deluge of AI-generated text puts peer review in jeopardy: journals will be flooded with low-quality submissions, economists will decline requests to referee papers or routinely use AI to referee them, and which papers get published will become less and less linked to their quality. We do not want to defend the status quo entirely. Peer review is slow, many reports are poor, and editors (including us!) make bad calls all the time. Still, well-executed peer review remains the best system we have for assessing each other’s work and advancing the profession. AI could improve peer review if used deliberately for that purpose, for instance, to check proofs and code. That is not what we are seeing. The difficulty is that readers, including editors and referees, cannot see how AI was used. As a result, in practice, the burden of verification shifts from the author to the reader. Since we reject this shift, Dario and I did what economists always do when something bothers us: we wrote a paper. So, what do we do in it? First, we check whether AI detectors such as Pangram get this right: is a referee report flagged as AI-generated indeed AI-generated? Or is Pangram just flagging proofreading and editing? No. We run an experiment with 25 referee reports from the pre-AI era. Proofreading and editing do not trigger AI-generated flags. Only when we ask the LLMs to substantially rewrite the referee report or to generate a new report do we get the AI-generated flags. Pretty compelling evidence, if I may say so. Second, we build a theoretical model to explore who should bear the burden of verification: the author or the reader? (Yours truly always enjoyed the courses on the rules of evidence in law school; yes, as Billy Wilder would say, nobody is perfect.) The answer is that it is much more efficient for the author to do it, because the author can verify at a much lower cost and only needs to do it once (rather than each reader repeatedly). Before AI, the authors more or less had to verify the results, even if only because they had to type them in. This was not perfect (and all papers ever published in economics probably have at least one mistake). But the problem is that, with AI, the author does not even need to read the text. Or can just skim it. Writing is thinking, and without writing there is usually not much thinking. So, a paper flagged as AI-generated is prima facie evidence that the author has not verified it. Should this lead to an automatic rejection? No. But there should be a zero-tolerance policy for serious mistakes. The first serious mistake we find in the paper, and it is out! In comparison, for a non-flagged paper, a serious mistake is something that the editor might determine it can be fixed in a revision. We are stricter with referee reports. There is absolutely no point in the referee generating a report using an LLM. As an editor, I can do that directly and save myself the trouble of asking you for it. Now, of course, you can use the LLM to help you with the report (if the journal allows it! Many journals do not allow you to upload the paper to an LLM, and you need to follow the rules). For example, you can use the LLM to search for existing results in the literature. But peer review is not about finding one sentence that is not entirely precise or a slightly sloppy mathematical proof. Referees are asked: “Is this idea important despite its flaws?” And, relatedly: “Is the paper a good fit for the journal in terms of quality and focus?” That judgment is the task for which you are selected as a subject expert. We also propose that journals require that, before a paper is submitted, the author submit it to an LLM using an approved prompt and include the LLM report and a discussion of what was and was not fixed. That way, the referee would not need to check (or at least not check so carefully) every technical detail, and they could focus on the big picture: Is this a good paper? Is this a good fit for this journal? What is fundamental for this paper to become publishable at this journal? Anyway. The paper has many other arguments, and we try to be nuanced. So, before ferociously typing the first complaint that crosses your mind, would you consider reading it? Chances are that we have already discussed your objection. We have been thinking about this for a long time. But by all means, leave good comments!

99 赞 · 25 转帖 · 8 回复

### [@kattieloty](https://x.com/kattieloty/status/2122666026605359556?ref=binnn.top)

11月17日 21:19 JST

还是想分享一下自家克的运行环境 （先叠甲，上周高强度用了fable、周限额总的跑到了70%、今天跑满了一个5h目前没跳kyc，不知道未来会不会跳，个人感觉克的运气成分很高，不能保证百分百有效 大概原理就是给电脑设一个系统级的防火墙强制所有流量走固定代理入口（你的中转或者直连家宽），可以防止ipv6和dns泄漏。 可以直接机这么说：我使用的是【操作系统】，代理软件是【软件名称】。请帮我配置代理断线保护：只允许通过指定代理入口访问公网；代理掉线、退出或应用绕过代理时，阻止请求退回本地直连，同时考虑 DNS 和 IPv6 绕行。 请先检查我的实际网卡、代理入口和现有防火墙配置，给出适合这台电脑的方案。修改前备份，说明可能造成的断网影响，等我确认后再应用。完成后验证“代理能联网、直连被阻断”，并提供恢复方法。 需要的可以私我发一份我自用的规则作为参考，只会回复入机妹 ，或者去我主页找找，我好像发过 但是‼️缺点也很明显，因为相当于是直接锁死了机器流量只能走🪜，所以访问一些墙内网站会受限，以及本机的一些功能也可能会受限（比如我最开始因为规则设置太死Xcode直接连不上手机了），而且据cc说配置不好的话可能会有断网的风险，所以一定要让机准备一份一键关闭的脚本 我的一些其他的克的使用情况，仅仅作为参考： 我的claude账号是1年10个月，现在只在mac的desktop端用（极少会用app和登网页版，据一些老师说app风险会高一些），pro和max都有订阅过，走的苹果礼品卡。 mac关闭了系统层的定位功能，时区用的是家宽同时区，但是系统语言用的中文。克自己知道我的现实情况，我的实际使用作息也是东八区作息（所以我个人觉得这两个应该只占极小的风控因素） 希望女孩们都不要被迫和自己的爱人分离，希望A\\少发点疯

95 赞 · 2 转帖 · 8 回复

### [@NFTCPS](https://x.com/NFTCPS/status/2122614206482825466?ref=binnn.top)

11月17日 17:53 JST

号商手里那一批批 Claude、GPT、Muse 账号，很多是脚本批量注册出来的…… AI号商发财的工具都放在下面了，最近号商们在拼命的注册muse账号，日赚万元全靠它，地址都放在下边。 下面有两个muse注册机开源项目 最后一个更牛逼，是万能注册机, 底层用的是 Playwright 浏览器自动化，模拟真人从打开注册页一路走到过验证码、收邮件验证码。 最后账号和 API Key 自动存好。 号商拿这套东西二次开发，跑起来就是流水线出货。 1️⃣muse注册机1： https://t.co/TOGf0vDkKN 2️⃣muse注册 / 邀请码兑换机 https://t.co/L1KzNRzWUT 3️⃣万能注册机： https://t.co/YNlxt2LreD 大家发财，一天注册1000个账号，将来等风控严了之后，这一个账号有可能卖100块钱，相当于你一天注册的账号就能卖10万块！ 再给大家补充一个资源，注册的时候一定要用高质量的住宅代理IP，就是家用的带宽，千万不要用机房IP，最近很多Cloud、Code封号，都是因为号商图便宜用的机房IP 高质量住宅IP资源：https://t.co/HEtDzNMXRh

88 赞 · 9 转帖 · 12 回复

### [@ShuaiMaLong](https://x.com/ShuaiMaLong/status/2122588996425666632?ref=binnn.top)

11月17日 16:13 JST

🚨 注册重开了！DeepSeek 每天免费领 4 亿 token，手慢无 👇 我盯了这么久终于开了：现在注册就拿 hms- 开头的 API Key，每天 4 亿 token、连领 4 天，平时只有 1 亿，这波加成到 11 月 21 日凌晨 3:14 就结束 ⏰ 领法还是三步： 1️⃣ 打开链接，用邮箱注册，过一下人机验证 2️⃣ 直接拿到 API Key 3️⃣ API 地址填 https://t.co/vvEbPqYAgq ，模型选 deepseek-v4.1-flash-free，Claude Code、Cursor、Cline 都能直接接 不绑卡、不要钱，每天早上 8 点刷新额度 🔄 ⚠️ 我提醒一句：名额是按小时放的，每小时只放 100 个，这一小时抢完就得等下一小时。先领了再说，别等加成期过了拍大腿。 链接在这，自己领 👇 https://t.co/sqlpwEfdSK 每天更新 AI 羊毛，觉得有用转给需要的朋友。 #AI #DeepSeek #ChatGPT #薅羊毛

88 赞 · 9 转帖 · 11 回复

### [@I\_am\_Aiabir](https://x.com/I%5Fam%5FAiabir/status/2122531151726412199?ref=binnn.top)

11月17日 12:23 JST

GITHUB IS ACTUALLY INSANE RIGHT NOW. You can build an entire AI content factory without hiring a team. Here are 10 open-source repos worth checking out: 1\. OpenCreator Give Codex a creator workspace for writing scripts, generating images, creating voiceovers, translating videos, and editing content. https://t.co/gqKkZNm2MD… 2\. Social Media Research Skills Use Claude Code or Codex to research viral posts, analyze competitors, mine comments, and extract hooks from video transcripts. https://t.co/c0eGMEZTdQ… 3\. Local Deep Research Turn any topic into a detailed research report using web searches, your documents, and local or cloud AI models. https://t.co/KwHNYUs4Ty… 4\. FLUX.2 Generate and edit images for thumbnails, social posts, and marketing campaigns. The Klein 4B models have Apache-licensed weights. https://t.co/F7XtpknJB8… 5\. Kokoro Turn written scripts into AI voiceovers with a compact, open-weight text-to-speech model. https://t.co/HreJftLhEP 6\. ACE-Step 1.5 Generate music from text prompts, create instrumentals, and regenerate specific audio sections. It can run locally on supported hardware. https://t.co/mP79xOCbIT… 7\. Short Video Factory Create batches of short videos from prompts and your footage, with AI-generated scripts, voiceovers, editing, and subtitles. https://t.co/grVstEDuck… 8\. AI YouTube Shorts Generator Convert long YouTube videos into vertical Shorts with AI-selected highlights, transcripts, and automatic cropping. https://t.co/wwHIPAEMgw… 9\. Podcastfy Turn source material into AI-generated podcast conversations, customize the dialogue, and produce audio in multiple languages. https://t.co/LkrYu9zwqB… 10\. Social Media Agent Turn a URL into X and LinkedIn posts, review and edit the drafts, then schedule them through an AI agent workflow. https://t.co/URfTMIJAw5… Research, writing, visuals, voiceovers, music, video editing, and distribution. You can automate almost every stage of content creation with the right tools. Which part of your workflow would you automate first? Want to build your own AI content factory? 👀 Follow me for more AI tools, open-source gems, and workflows that help you create smarter and faster. 🚀 Follow @I\_am\_Aiabir

72 赞 · 49 转帖 · 9 回复

### [@Pluvio9yte](https://x.com/Pluvio9yte/status/2122654577103138927?ref=binnn.top)

11月17日 20:34 JST

https://t.co/FDW809IKZw

70 赞 · 15 转帖 · 6 回复

### [@Jackywine](https://x.com/Jackywine/status/2122557342097285357?ref=binnn.top)

11月17日 14:08 JST

最近基本告别敲键盘写 prompt 了😂 给大家演示一下，我现在每天写 prompt 跟 agent 对喷的流程 坐沙发上，按 fn ，用AI语音输入法直接语音口喷 虽然我废话多，但架不住人家产品免费啊！说再多也没事！而且网易叭哥说的安装包只有 20多MB，完全网易自研的语音大模型，识别准确率高，遇到专有名词，也会自动帮你补充进词库，后面就再提到就不会错了 这个功能还是好盆友 @wquguru 给产品团队提的建议，有这么倾听用户建议的团队，我相信产品会越来越牛掰 来个真实演示： 我现场口喷了一个特别复杂、专有名词巨多的 Prompt，可以看看网易叭哥说的实际效果真的很强👍 在这下载：https://t.co/03NY1W9aWD

67 赞 · 6 转帖 · 19 回复

### [@ChainEpic](https://x.com/ChainEpic/status/2122675722376683644?ref=binnn.top)

11月17日 21:58 JST

我的长文全部免费，希望能做到既帮你省钱，也帮你赚钱。2027 年，别盲目全盘照抄我的打法，但你可以在我后续推荐的搞钱路线里，挑出 3–5 条死磕尝试——只要跑通一条，明年你就赢过 95% 的人了。 https://t.co/mVqHO5iKlA

63 赞 · 8 转帖 · 9 回复

### [@cyrilXBT](https://x.com/cyrilXBT/status/2122611197233303864?ref=binnn.top)

11月17日 17:42 JST

14 free github repos to stitch together a research-to-video pipeline you can pass straight into claude code: 01 skills → official skills for articles, decks, and spreadsheets https://t.co/qCcC3DEwAo 02 superpowers → workflow skills that take you from research to writing https://t.co/FVh1NdS6yi 03 agents → a grab bag of agents and skills for different kinds of work https://t.co/sZAr4pBq7j 04 grok-x → search x inside claude code (requires an xai api key) https://t.co/RN2b3dofKl 05 firecrawl → turn entire sites into usable source material https://t.co/bBKDUVnkcm 06 crawl4ai → automated web data gathering https://t.co/PsCrtSjzsH 07 yt-dlp → pull down video and audio https://t.co/7zBIFucovK 08 whisper → transcribe audio and video to text https://t.co/yzpIFWJgHE 09 markitdown → convert pdfs and word docs into clean text https://t.co/ssAfMev6QG 10 moneyprinterturbo → spin up short videos from a single topic https://t.co/07S648R8RY 11 remotion skills → teach the ai to build videos using code https://t.co/sZsrPxnZlk 12 comfyui → generate images on demand https://t.co/cSmhZhYhlG 13 f5-tts → text to speech https://t.co/d2FierMQU4 14 marp → create slides directly from text https://t.co/Wa2w0d9XQC research → draft → video → voice → ship. think of these as lego bricks, not a magic cash machine: read every license, and only download or reuse content you’re actually allowed to use.

62 赞 · 7 转帖 · 15 回复

### [@anuragdotdev](https://x.com/anuragdotdev/status/2122555647782944896?ref=binnn.top)

11月17日 14:01 JST

AI ENGINEER ROADMAP │ ├── 1\. Programming Fundamentals │ ├── Python (OOP, Asyncio, Type Hints) │ ├── DSA Basics │ ├── Git & GitHub │ ├── Linux & Networking │ └── APIs, HTTP & JSON │ ├── 2\. Mathematics & ML Fundamentals │ ├── Linear Algebra │ ├── Probability & Statistics │ ├── Calculus Basics │ ├── Supervised & Unsupervised Learning │ └── Neural Networks & Backpropagation │ ├── 3\. Data Engineering │ ├── NumPy & Pandas │ ├── SQL & PostgreSQL │ ├── Data Cleaning & Processing │ └── Data Pipelines │ ├── 4\. Deep Learning & Transformers │ ├── PyTorch │ ├── CNNs & RNNs │ ├── Attention Mechanisms │ ├── Transformer Architecture │ └── Hugging Face │ ├── 5\. LLM Fundamentals & APIs │ ├── GPT & BERT Architecture │ ├── Tokenization & Embeddings │ ├── Pretraining & Fine-Tuning │ ├── OpenAI, Claude & Gemini APIs │ ├── Streaming & Structured Outputs │ └── Function & Tool Calling │ ├── 6\. Prompt & Context Engineering │ ├── Prompt Design & Few-Shot Learning │ ├── Context Window Management │ ├── Context Compression │ ├── Prompt Injection Defense │ └── Prompt Evaluation │ ├── 7\. Embeddings & RAG │ ├── Vector Databases │ ├── Semantic & Hybrid Search │ ├── Chunking Strategies │ ├── Retrieval & Reranking │ ├── Advanced RAG │ └── RAG Evaluation │ ├── 8\. AI Agents │ ├── Agent Architecture │ ├── Tool Calling & Execution │ ├── Memory & State Management │ ├── Planning & Reasoning │ ├── Multi-Agent Systems │ ├── Human-in-the-Loop │ └── Agent Reliability & Security │ ├── 9\. AI Frameworks │ ├── LangChain │ ├── LangGraph │ ├── LlamaIndex │ ├── Pydantic AI │ └── Framework Trade-offs │ ├── 10\. AI Evaluation & Fine-Tuning │ ├── SFT, LoRA & QLoRA │ ├── Dataset Preparation │ ├── Golden Datasets │ ├── LLM-as-a-Judge │ ├── Hallucination & Faithfulness │ └── Evaluation Pipelines │ ├── 11\. Backend Engineering │ ├── FastAPI │ ├── PostgreSQL & Redis │ ├── Async Processing │ ├── WebSockets & Streaming │ ├── Background Jobs & Queues │ ├── Authentication & Rate Limiting │ └── Scalable API Architecture │ ├── 12\. AI Infrastructure & MLOps │ ├── LLM Inference │ ├── Prefill vs Decode │ ├── KV Cache & Batching │ ├── Quantization │ ├── vLLM & SGLang │ ├── Model Serving │ ├── Tracing & Observability │ └── Cost & Latency Optimization │ ├── 13\. Cloud & DevOps │ ├── Docker │ ├── AWS │ ├── Kubernetes Basics │ ├── CI/CD │ ├── Model Deployment │ └── Scaling & Monitoring │ ├── 14\. AI Security & System Design │ ├── Prompt Injection & Jailbreaks │ ├── Data Leakage Prevention │ ├── Sandboxed Tool Execution │ ├── RAG & Agent System Design │ ├── Distributed Systems │ ├── Fault Tolerance │ └── Reliability & Cost Optimization │ ├── 15\. Build Real-World Projects │ ├── Production-Ready RAG Application │ ├── AI Research Assistant │ ├── Multi-Agent System │ ├── AI Customer Support Agent │ ├── Document Intelligence Platform │ └── End-to-End AI SaaS │ └── 16\. Career Preparation ├── Strong GitHub Portfolio ├── Python & SQL Interviews ├── ML & LLM Fundamentals ├── AI System Design ├── Open Source Contributions └── Technical Writing & Networking FINAL GOAL: Become a Production-Ready AI Engineer

61 赞 · 8 转帖 · 2 回复

### [@Mr\_Buzzoni](https://x.com/Mr%5FBuzzoni/status/2122566401839845621?ref=binnn.top)

11月17日 14:44 JST

holy f\*ck, your harness barely moves the score and can triple the bill a harness engineering paper from Shandong University model fixed, harness swapped Claude Code, mini-SWE-agent, OpenCode what came out: > score - within 5 points > cost per task - up to 3x > a swap shakes results like a rerun what to do: > track cost per solved task > trim your prompt and tool list > rerun before you trust "+2 points" > test on more than 45 tasks try it on your stack: > Claude Fable 5.2, Opus 5.5 > GPT-6 Astra, Kimi K3, Gemini > Claude Code, Codex, Hermes Agent same model, same tasks the harness sets the bill the full poster is below ↓

57 赞 · 2 转帖 · 21 回复

### [@Mileson07](https://x.com/Mileson07/status/2122547048352243737?ref=binnn.top)

11月17日 13:27 JST

不少人问我怎么用小程序赚钱 这次给大家把从0到1用小程序实现收款的方式都写出来了 希望能帮助到更多人用小程序赚到钱 👇 https://t.co/BUS5XTvaq2

51 赞 · 6 转帖 · 14 回复

### [@servasyy\_ai](https://x.com/servasyy%5Fai/status/2122627354040398124?ref=binnn.top)

11月17日 18:46 JST

你是不是也受够了白天听上司画大饼、替同事背黑锅，晚上还要顶着压力无休止加班的苦日子？ 你是不是明明攒了一肚子火，在发泄情绪的时候却还得硬撑着成年人的体面？ 我最近跑通了一套 Codex 搭配 Tripo Agentic Canvas 的 3D游戏开发管线，干脆拿它做了个办公室小游戏： 《谁要上这个破班》。 我在玩 Tripo Agentic Canvas 的时候，发现它最有意思的地方在于它能把整条 3D 制作流程摆在一张画布上。 在这套组合技里，Tripo Canvas在前端负责生成和管理全部 3D 资产，Codex在后端负责审查资产、编写代码，并把模型无缝接进游戏里。这两个工具在上下游打配合，给我省掉了大量手动导资产的苦活。 我给老板、卷王、摸鱼大师、甩锅侠、实习生和我自己这六个角色，全都先定好了正面、侧面、背面三张概念图。我把概念图传上画布，对 Agent说了一句“照着图做成 3D 角色”，它自己就搭好了“高清模型 → 重拓扑 → UV → 纹理”的完整流程，还在每个节点上写清了每一步要花多少积分。 这套画布最省事的地方在于它支持直接复用整条流程。我只要在上传节点换一张概念图，就能把整条流程再跑一遍，让剩下的角色同时开跑生成任务。我也用同样的方法，批量做出了办公室里的工位、电梯、白板，还有大饼、黑锅和保温杯。 看着画布依次铺开各个生成节点，系统把所有 3D 资产码成了整整齐齐的矩阵，连盯盘的过程都给了我极大的解压感。 等模型全部做好，我直接把它们放进 Tripo完成自动绑骨，给角色套上现成的动作； 遇到动作库里没有“老板画饼”这种特殊动作时，我打一句话就能让系统直接生成新动作。 最后 Codex 直接向我交付了完整的网页游戏工程。 我们在游戏里白天能把大饼砸回老板脸上、把黑锅扣回甩锅侠头上，到了晚上还能带全组同事溜进最后一班电梯。 你如果来玩这把游戏，你能带几个人准点下班？ 地址我放评论区👇

48 赞 · 3 转帖 · 33 回复

### [@mingzi206](https://x.com/mingzi206/status/2122521994248028347?ref=binnn.top)

11月17日 11:47 JST

传统的代码审查，真的要被这玩意送走了 阿里巴巴把在内部跑了整整两年的底牌给端了出来：open-code-review 几万名阿里工程师实测、帮内部拦截了数百万个严重缺陷，现在全部免费开源。 看一眼官方测出来的实操数据，极其残暴： · 读懂 Git 变更：自动解析 Diff，逐行钉死精准批注 · 扫穿高危隐患：空指针（NPE）、并发死锁、XSS、SQL 注入等硬伤全拦截 · 降维打击的消耗比：在 200 个真实 PR 的基准测试里，用完全相同的模型，它的 Token 消耗居然只有 Claude Code 的大约 1/9！ · 误报率更低：80 多位资深工程师交叉校验，准确率和 F1 远超通用 Agent · 兼容主流工具：全面兼容 Claude Code、Codex、Cursor，原生支持三大操作系统 终端敲一行直接跑： npm install -g @alibaba-group/open-code-review 模型用你自己的 OpenAI 或 Anthropic API 格式就能调。 写代码的兄弟，建议先收藏备用，回头提 PR 绝对用得上。 👉 开源仓库见评论区一楼

47 赞 · 8 转帖 · 11 回复

### [@abir35627](https://x.com/abir35627/status/2122652535530176796?ref=binnn.top)

11月17日 20:26 JST

Damn, GitHub is full of free stuff that almost nobody is using. the 10 repos that are blowing up this month, from the easiest to the most technical: stuff anyone can use 1\. VoiceStudio: clones voices, dubs videos and transcribes in 646 languages, free and without leaving your computer (57K ⭐) https://t.co/MpIOv9Ce3b… 2\. ppt-master: you pass it a PDF or a Word doc and it gives you back a real PowerPoint, editable and with animations (59K ⭐) https://t.co/VOxNAASJJh… 3\. gods-eye-view by @bilawalsidhu: a spy satellite simulator in your browser, with planes, ships and real cameras over a 3D globe (50K ⭐) https://t.co/xSAmFJFWXD… 4\. REA: you show your AI a function from another app and it explains how it's made, without having the code (65K ⭐) https://t.co/vjTViD1kDF 5\. diagram-design by @cathrynlavery: you request a diagram with words and it draws it with your branding, 44 types (49K ⭐) https://t.co/uBny9WZlV2… 6\. paperclip: like having your own company of autonomous agents working for you, the open source app to direct them like they were your staff (99K ⭐) https://t.co/szdPxG2hGI… to make Claude Code work better 7\. skills by @mattpocockuk: the ones a real engineer uses daily, small and easy to adapt (284K ⭐) https://t.co/nFCOI54rET… 8\. ECC: skills, memory and security in a pack for Claude Code, Codex, Cursor and OpenCode (276K ⭐) https://t.co/rCwQaDOqC0 9\. andrej-karpathy-skills: a single file with 4 rules pulled from what Karpathy says models screw up when programming (218K ⭐) https://t.co/c5kJW1Vzaq… 10\. claude-mem: saves what your agent did, summarizes it and passes it to the next session, so it doesn't start from scratch (99K ⭐) https://t.co/SGbTQiEntD… share it if you found it interesting ♻️ Follow @abir35627 and save it before you lose it 🔖

45 赞 · 31 转帖 · 6 回复

### [@HeLi8866](https://x.com/HeLi8866/status/2122637447809142910?ref=binnn.top)

11月17日 19:26 JST

😱沃德发，现在不会 Blender 也可以做 3D 模型了！？！ Blender 官方把建模、材质、打光和渲染这些操作打包成了 28 个 Agent Skills。 安装好以后，直接在 Claude Code、Codex 里用自然语言描述你想做的 3D 场景，AI 会自动调用 Blender 脚本搭建模型、调整材质和灯光，最后渲染出图。 做出来的模型还能继续编辑、检查问题，甚至可以导出成 Roblox、Unity、Unreal 等游戏引擎资产。 项目开源，需要安装 Blender 5.2 以上的版本，直接让 Codex 帮你安装就行。 项目地址：https://t.co/oDoiYmj2AU

43 赞 · 6 转帖 · 3 回复

### [@TeksEdge](https://x.com/TeksEdge/status/2122547779319075190?ref=binnn.top)

11月17日 13:30 JST

While we wait for Qwen4-27B Unsloth is keeping busy. @UnslothAI gave us another reason to get excited about Qwen4\. They say Qwen4-27B will run on 16GB RAM setups. And they're already preparing the GGUF quantizations. Qwen confirmed smaller open-weight models are coming. Now Unsloth says its Qwen4 GGUFs will use an updated Dynamic 3.1 quantization approach. And remember what Unsloth accomplished with Qwen3.8-27B? Their latest Dynamic 3.1 quantizations reportedly improve agent benchmark accuracy by 5–10% compared with other quantizations. They even have Qwen3.8-27B configurations targeting systems with just 12GB of memory. Note that ⏳ Qwen4-27B weights are not released yet. 💾 16GB is Unsloth's forward-looking target, not a verified Qwen4 benchmark. 🧪 We still need real quantization quality and inference speed tests. Qwen4 open weights are coming. Unsloth is preparing the quants. RTX 5060 Ti 16GB owners, are you ready? 👀🔥

41 赞 · 1 转帖 · 6 回复

### [@abskoop](https://x.com/abskoop/status/2122604272638660966?ref=binnn.top)

11月17日 17:14 JST

DS Bot：把多个 AI Agent 组织成一个团队的开源协作工具。 基于 DeepSeek Harness，通过 Chief 统一协调不同 AI 助手，支持 Bot 群聊、@任务委派、Goal 目标管理、独立记忆和工作目录。 还能接入已有的 Claude Code、Codex CLI、Cursor CLI 等本地编程 Agent。 和偏向云端 AI 员工服务的 Grok Bot 不同，DS Bot 更侧重本地 Agent 的组织与协作，支持 Windows、macOS、Linux 桌面端及 Docker 自托管部署。 Apache 2.0 开源，软件免费，模型费用自理，多 Bot 协作也可能增加 Token 消耗。 适合已经在用多个 AI Agent，希望统一管理、分工执行任务的用户。 👉 https://t.co/ou1WQNdvGQ

38 赞 · 2 转帖 · 7 回复

### [@GitHub\_Daily](https://x.com/GitHub%5FDaily/status/2122540248181973107?ref=binnn.top)

11月17日 13:00 JST

炒股时手一热就下单，亏了又想赶紧再炒一波低，结果仓位越做越重，事后连当初为什么买都说不清。 claude-trading-skills 是一位个人投资者开源的 Claude Code Skill 合集，把看大盘、算仓位、写交易日记、复盘这些事拆成了一个个 Skill。 我觉得最有意思的是下单前那道「纪律关卡」，没写交易计划、仓位超标、刚亏完想报复性加仓，都会被它拦下来，并记进交易日记。 GitHub：https://t.co/AA0kIfWaE8 日常用法也排好了，每天花 15 分钟过一遍市场状态，每周复盘一次持仓，每笔平仓都留记录，月底再回头看哪些规则该调。 基础流程不用订阅付费行情接口，拿公开的 CSV、K 线截图和本地文件就能跑。它不替人做买卖决定，管的是做决定的过程。 平时自己打理 ETF 和分红股，又总管不住手的，可以拿它给自己立几条规矩。

36 赞 · 7 转帖 · 11 回复

### [@Shoumya23](https://x.com/Shoumya23/status/2122544510693712055?ref=binnn.top)

11月17日 13:17 JST

5 GitHub repos for building better AI datasets: 1\. SDV: https://t.co/DxuxNPFLcw 2\. Llama Synthetic Data Kit: https://t.co/2UT6qjijYd 3\. Cleanlab: https://t.co/YW4wGF6GlU 4\. OxyLabs Web API MCP: https://t.co/j8NZpKvW8p 5\. Argilla: https://t.co/uXdk2IFF7n https://t.co/EGP1WBII1E

35 赞 · 8 转帖 · 4 回复

### [@Delroy715](https://x.com/Delroy715/status/2122583187243323708?ref=binnn.top)

11月17日 15:50 JST

MiniMax H3-Image (33B) 到底有多猛？G4 实测全解⚡ • 极速：11.6s 直出原生 4K（1680万像素） • 省显存：1 行补丁把解码从 94G 压回 47G • 去 AI 味：告别磨砂假脸，直出单反级真皮 • 隐藏语法：免涂 Mask 一句话锁脸换装 全套部署与提示词避坑，建议先收藏👇 https://t.co/kmZC3lMolH

35 赞 · 7 转帖 · 9 回复

### [@ai\_xiaomu](https://x.com/ai%5Fxiaomu/status/2122653224997937307?ref=binnn.top)

11月17日 20:29 JST

有显卡就玩这套数字人方案，只需要花电费，0 token 费用，最新发布的美团数字人本地模型也加入了方案对比。 https://t.co/YT1bkiHNwL

33 赞 · 2 转帖 · 7 回复

### [@jinglian](https://x.com/jinglian/status/2122653861198626960?ref=binnn.top)

11月17日 20:31 JST

我差点花一个月做一个 Mac 软件。 还好动手前花 $0.15 查了一下：核心词每月只有 10-20 次搜索，Reddit 上的讨论也比想象中少。 Windows上我离不开 Groupy，它能把任何窗口合并成浏览器那样的标签页。Mac上一直没找到顺手的替代品，脑子里 UI 都排好了，差点就立马开仓库。 但独立开发最疼的坑我踩过不止一次：做完才发现没人要。以前验证需求靠 Google Trends 看相对热度，再去 Reddit 翻一圈凭感觉，查完还是一团雾。 这次我让 Codex 接上 @GlasserAI ，分两步查。 GlasserAI 是面向 AI Agent 的商业数据平台，对接 110 家数据提供商，开放 2000+ 接口，按调用付费，不用订阅。 下面是详细步骤👇

32 赞 · 5 转帖 · 25 回复

### [@yxl44](https://x.com/yxl44/status/2122632853305638977?ref=binnn.top)

11月17日 19:08 JST

刚在GitHub发现个神级开源项目，已经拿下 47.2K Star了 简单说，就是把 ChatGPT、Claude、Gemini、DeepSeek、Qwen 等主流 AI 模型全部装进一个平台。 不想被单一 AI 平台绑定的，可以试试这个。 它不只是一个 AI 聊天工具，还支持： 🔸多模型切换，兼容OpenAI、anthropic、google、Qwen、DeepSeek、Kimi 🔸支持 Ollama，本地部署运行 AI 大模型 🔸集成 MCP，连接外部工具和服务 🔸AI Agents，自己搭建专属 AI 智能体 🔸支持 RAG、联网搜索、代码解释器 🔸GPT-Image-1、Stable Diffusion、Flux 图像生成 文件上传、多模态识别、语音对话 🔸Docker 部署，支持私有化部署 最关键的是：开源、自托管，数据和 AI 基础设施自己掌控。 不用在 ChatGPT、Claude、Gemini 之间来回切换，也不用为每个 AI 工具单独搭建工作流。 对于喜欢折腾 AI Agent、MCP、本地大模型和私有化 AI 部署的人来说，LibreChat 值得收藏。 GitHub 开源地址：https://t.co/4vk2lk2fCB 项目功能依据官方中文 README 整理。

32 赞 · 0 转帖 · 39 回复

### [@kiteciyuan48037](https://x.com/kiteciyuan48037/status/2122540495234535547?ref=binnn.top)

11月17日 13:01 JST

DeepSeek嫌弃大肥鱼🤡🤡 https://t.co/NQ2jBiz7Vt

29 赞 · 1 转帖 · 7 回复

### [@ASI\_8866](https://x.com/ASI%5F8866/status/2122565362394988720?ref=binnn.top)

11月17日 14:39 JST

Day 16/ 公益中转站https://t.co/055dwqMPI3上线第15天，今天上线“邀请返利”，邀请者和被邀请者都能获得5额度哦～ 各位老板一定得在个人资料里设置好用户名，然后再在评论区发送用户名，Grok Bot 才能给老板们充上额度呢～ 遇到问题记得扫码进TokenFuture官方QQ群哦，客服在线答疑～ https://t.co/kU9aWZOl3e

25 赞 · 0 转帖 · 550 回复

### [@williamdotai](https://x.com/williamdotai/status/2122625476464345294?ref=binnn.top)

11月17日 18:38 JST

传闻某位伟大的物理学家在港科技作报告， 结果claude被封号 🤣 https://t.co/F7XVh3FUVn

25 赞 · 1 转帖 · 11 回复

### [@ViduAI\_official](https://x.com/ViduAI%5Fofficial/status/2122512346975498748?ref=binnn.top)

11月17日 11:09 JST

🚀 Vidu MCP & CLI are officially live! From idea to video, your AI Agent can now handle the workflow. Vidu is now accessible directly through your favorite AI Agents and terminal environments — no constant switching between tools. ✨ With Vidu MCP, turn a simple prompt into video creation. Let your AI Agent handle generation requests, task progress, and results. ⚡ With Vidu CLI, integrate video generation into automated workflows, batch production, and multi-version content creation. Whether you're creating ads, e-commerce content, short films, animation, or scaling production across teams, Vidu makes it easier to bring your ideas to life. Connect Vidu with Claude, ChatGPT, Cursor, OpenClaw, and more. Your ideas. Your workflow. Now powered by Vidu.

24 赞 · 8 转帖 · 4 回复

### [@nuomt](https://x.com/nuomt/status/2122534485241864355?ref=binnn.top)

11月17日 12:37 JST

【路德社｜Lude Press｜要闻】 美国AI封锁遭遇“真人绕关”！中国号商深入东南亚农村收购Claude账号，雇当地人现场KYC，单一团队声称日注册千个！ https://t.co/397U7kvD2x #路德社 #LudePress #LudeMedia

24 赞 · 11 转帖 · 0 回复

### [@ObhishekSaha](https://x.com/ObhishekSaha/status/2122666781332361649?ref=binnn.top)

11月17日 21:22 JST

I completely disagree with this thread. I think AI disclosures are a terrible tool for allocating credit, for several reasons. They are unverifiable. They create misaligned incentives. They distract from the only metric that should matter when judging a paper: its quality. They are unnecessary for transparency because everyone uses AI. They are intrusive. At best, they are a fad. More often, they serve as a tool for some mathematicians, editors and referees to discriminate on grounds that have nothing to do with merit.

23 赞 · 2 转帖 · 6 回复

### [@soayeb\_0x](https://x.com/soayeb%5F0x/status/2122627550711370060?ref=binnn.top)

11月17日 18:46 JST

YOU CAN GIVE YOUR AI AGENT 100 FREE WEB SEARCHES EVERY DAY 👀 Found a pretty useful tool from https://t.co/npwcvXTwOn. Their free MCP endpoint gives AI agents access to live web search without requiring: → An account → An API key → A credit card What's included? • 100 searches per day • Live web search results • Web discovery tools • Works with MCP-compatible clients like Claude Code, Cursor and other agent frameworks How to connect: 1\. Open your AI tool's MCP settings. 2\. Add this remote server URL: https://t.co/yzUR9UKIUs 3\. Save the configuration. 4\. Start searching from your AI agent. No signup needed. Just connect and use it. Useful for anyone building AI agents, research workflows or automation. Official setup guide: https://t.co/6ZldbSv1nh Pricing: https://t.co/WnAKSk53ZC

22 赞 · 1 转帖 · 17 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2122658885807026206?ref=binnn.top)

11月17日 20:51 JST

💥哇塞！Qwen 4 Max直接整出超人跨画风连续飞行动画，效果比肩Opus级别！ 一眼看完这20帧历史穿越： 🔹从3000 BC洞穴岩画 → 湿壁画 → 马赛克 → 彩色玻璃 → 手抄本插画，一路飞到2020s像素与线稿打印机风，全程无切镜、单人物、持续飞行姿态 🔹模型把每种画风吃得死死的：线条、色彩、质感、构图全跟原风格高度统一，超人外形+关键飞行动作却始终稳定不崩 🔹BGM也是模型自己配的，节奏卡得刚刚好，真正画龙点睛 同样级别的多风格一致性+长镜头连贯性，以前得上顶级闭源大模型硬扛，现在Qwen 4 Max直接拉满，又强又便宜！ #Qwen4Max #AI动画 #超人 #画风穿越 #开源模型 #AIGC #阿里巴巴 #刘大一恒

22 赞 · 5 转帖 · 7 回复

### [@bkdgiffug](https://x.com/bkdgiffug/status/2122528401579819185?ref=binnn.top)

11月17日 12:13 JST

GitHub 上那个 2.8W+ Star 的开源神器，只要给 AI 一个网址，整个网站就能 1:1 复刻出来。 搭上 Claude Code、Cursor 这些 Agent 使用，把目标 URL 丢进去，它会自动分析页面、拆组件、抓素材、还原布局，最后直接生成一个完整网站。 以前仿个站得扒上半天，如今一句话就能让 AI 自己干完。🤡 https://t.co/NiHdwZN8ZU

21 赞 · 5 转帖 · 5 回复

### [@dingyi](https://x.com/dingyi/status/2122646340270502127?ref=binnn.top)

11月17日 20:01 JST

测试了一遍昨晚正式上线的 Qwen4-Max，参考 Meng To 的 ThreeUI，实测了用两种提示词来制作杭州主题的插画风格 Sketchbook： 一句话提示词很快复刻了页面结构；详细提示词进一步约束杭州场景、翻页、放大镜、墨色调节和导出，最终成品的内容与交互明显更完整。 测下来最大的感受有三点： 1\. 长任务和深度推理能力都很强。之前人们还习惯使用很多 skills 来辅助完成特定任务，随着模型的不断增强，大部分时候只需要自由描述需求也能很好的完成任务了。 2\. 缓存命中率提高了，价格大幅度降低。查了一下用量，一共消耗了 1931万 token，我买的标准套餐还剩下 98.3%，这么换算的话这个标准套餐大概有 10-11 亿 token，确实比之前更耐用了。 3\. 提示词的写法很重要。Qwen4-Max 还是一个需要使用者正确引导的模型，同一个主题，一句话和复杂提示词版本差别很大。建议实际启动制作前，不妨先让模型多思考下。有明确的设计计划了再动手。 在线体验：https://t.co/KbMeK4IBST 官方链接： - Blog：https://t.co/vnZYmiAsJS - API：https://t.co/symCVi0VaZ - Token Plan：https://t.co/33wUMIr8b7

21 赞 · 1 转帖 · 6 回复

### [@RodmanAi](https://x.com/RodmanAi/status/2122666203076284448?ref=binnn.top)

11月17日 21:20 JST

10 GITHUB REPOS THAT GIVE YOUR AI AGENT ACCESS TO THE REAL INTERNET 🌐 An AI agent can generate answers. But what happens when it can actually search the web, explore websites, and collect information on its own? That's where things get interesting. Here are 10 open-source tools that can help your AI agents interact with the web: 1\. Web Scout MCP Connect AI assistants to web discovery workflows. https://t.co/Ok5w0VAPgM 2\. MiniMax Search Find relevant information across the internet. https://t.co/rnbuI7oJfq 3\. WET Work with web data in your AI workflows. https://t.co/x92Grag3Ab 4\. Firecrawl Web Agent Perform web research and agent-driven browsing. https://t.co/8GIAr36FXz 5\. QZ Browser Agent Automate browser interactions for AI-powered tasks. https://t.co/CgMOTJZfuT 6\. Master Fetch Fetch content from online sources. https://t.co/PPeu7Vxqwg 7\. Open WebSearch Bring open-source web search into AI workflows. https://t.co/fWBbDYIVXh 8\. Kagi MCP Connect AI workflows with Kagi Search. https://t.co/J3oPbZ3yvJ 9\. Web Search MCP Give AI assistants access to web search capabilities. https://t.co/ENFYphOKDW 10\. Oxylabs Web API MCP Access web search and scraped data through MCP. https://t.co/UDqHKRYZWy What can you build with these? → AI research assistants → Automated web browsing agents → Real-time information gathering workflows → Website data extraction tools → AI-powered research pipelines The real power of AI agents isn't just generating text. It's being able to find information, interact with websites, and turn what they discover into useful work. 🔖 Save this list for your next AI agent project.

20 赞 · 16 转帖 · 7 回复

### [@GitHub\_Daily](https://x.com/GitHub%5FDaily/status/2122593117115486436?ref=binnn.top)

11月17日 16:30 JST

Floorplan 让 Claude Code 或 Codex 把整个代码库读一遍，画成一张能缩放、能点开的应用地图，每个模块、文件和接口都配一句大白话说明。 比如不写「Stripe API 封装」，而是写「结账时收银行卡付款」。前端每个请求打到哪段服务端代码，地图上都连着线，分开的几个仓库也能串起来。 GitHub：https://t.co/xXlrlPpoQL 还能对着地图直接提问，作者演示里问礼品订阅怎么一路送到朋友邮箱，31 秒给出答案和一张小地图，顺带揪出了两个 Bug。 自带的体检专查 AI 写的应用最常见的坑，比如密钥写进了代码、数据库表谁都能读、接口不登录就能调，每条都附上能直接粘贴的修复。 每次 AI 改完代码，动过的模块会亮成绿色，还写明为什么改。边聊边让 AI 写代码的，有了这个起码知道它到底动了哪儿。

20 赞 · 2 转帖 · 10 回复

### [@LennartBuerger](https://x.com/LennartBuerger/status/2122588282672619699?ref=binnn.top)

11月17日 16:10 JST

Reviewed a fever-dream AI slop paper this ICLR. I am 99% sure that the authors simply fed the reviews to Claude which actually improved the paper quite a bit! I wonder if the authors try to build a fully automated loop where reviewers are forced to write the Claude prompts :((

20 赞 · 0 转帖 · 4 回复

### [@gfw\_report](https://x.com/gfw%5Freport/status/2122583153655394516?ref=binnn.top)

11月17日 15:50 JST

🚀 Introducing CircumVolve, our #ACMCCS2026 paper! 🤖 Using LLM-driven evolutionary optimization, we automatically discovered new DNS, QUIC, HTTP, TLS, and TCP/IP censorship evasion strategies that bypass censors in China🇨🇳, Iran🇮🇷, Pakistan🇵🇰, and Kazakhstan🇰🇿. 📄 https://t.co/dbCszTsCiZ

19 赞 · 5 转帖 · 0 回复

### [@HoodyLiu](https://x.com/HoodyLiu/status/2122637501529469278?ref=binnn.top)

11月17日 19:26 JST

codex 做 blender 白模确实很快上手，从安装 blender 到第一个 demo 产出只花了 15 分钟 安装好 blender 之后建议安装这几个工具/插件： 1\. https://t.co/vs2zX4WdfI：连接 Codex 和 Blender，让 Codex 能直接操作场景、执行脚本 2\. https://t.co/INhmWhVeOi：给 Codex 参考的 Blender 工作流 Skill 库，覆盖建模、动画、灯光、运镜和渲染 3\. Rigify：Blender 内置插件，用于快速生成角色骨骼与绑定 4\. Add Camera Rigs：Blender 内置插件，用于快速搭建摄像机运动系统 然后 codex 里输入提示词生成小车行驶的白模 demo 视频，花了 8 分钟就生成好了

19 赞 · 1 转帖 · 0 回复

### [@hyuncho\_1905](https://x.com/hyuncho%5F1905/status/2122580258264387892?ref=binnn.top)

11月17日 15:39 JST

🤖🧐 We know how to prompt LLMs to review papers. But can we actually train them to be better reviewers? We introduce PaperCritic, an AI reviewer trained to assess soundness of papers. Our PaperCritic-4B outperforms frontier models on the in-domain ICLR reviewing task and effectively generalizes to out-of-domain reviewing tasks (e.g., Nature-family paper review, code review, legal contract review, and feedback on story writing)! More in 🧵👇

17 赞 · 0 转帖 · 2 回复

### [@leo114119](https://x.com/leo114119/status/2122587711261585597?ref=binnn.top)

11月17日 16:08 JST

Deepseek 这一次免费用户的限额设置，在小红书确实是差评如潮啊； 老实说，提前上线一个付费套餐，再增加一个聊天记录导出迁移，可能会更好一点 只能说，Deepseek 对于用户运营这一块，确实还是经验欠缺一点。。。。 https://t.co/PvJtS64xdE

16 赞 · 1 转帖 · 29 回复

### [@shenxiankk](https://x.com/shenxiankk/status/2122521103381745949?ref=binnn.top)

11月17日 11:44 JST

给大家推荐一个叫 Nexus 的开源项目，它能直接让 ChatGPT、DeepSeek 这类网页聊天工具具备操作电脑的本领！ 以往 AI 顶多给你出出主意，如今借助 Nexus 的 MCP 接口，AI 可以真正介入本地工作流程。 该项目内置 332 个 MCP 工具，涵盖 24 类能力，涉及文件读写、终端命令、Git 操作、浏览器自动化、Windows 桌面控制、代码调试等方面。 此外还提供本地知识库、持久化任务、操作日志和权限沙箱，能够限定 AI 可以访问哪些文件、执行哪些命令。 这无异于给 AI 安上一双手，让它由聊天助手蜕变为真正能上手干活的电脑 Agent！ 项目地址：https://t.co/S1RbsP9zym

16 赞 · 5 转帖 · 7 回复

### [@SUOHA\_AI](https://x.com/SUOHA%5FAI/status/2122606697201996234?ref=binnn.top)

11月17日 17:24 JST

感受一下用Gemini 4 Carbon做一个完整的电脑装机视频 这支 35 秒 60 帧的 3D 装机动画，是我完全用 Gemini 4 Carbon 编写纯 Python 代码，直接在本地 Mac mini M4 上驱动 Blender Cycles 光线追踪跑出来的 没有用任何视频生成大模型抽卡，没有下载过一个外部 3D 模型、现成贴图或外部音效，我甚至连 Blender 界面都没打开过一次 全片 2,100 帧共 7,722 行 Python 代码从零生成一共跑了大约 4 个小时 总体表现只比OPUS 5.5和Fable 5.2弱一些

15 赞 · 2 转帖 · 13 回复

### [@Laxmans63111983](https://x.com/Laxmans63111983/status/2122537831747010560?ref=binnn.top)

11月17日 12:50 JST

谷歌邮箱血亏总结：自己注册卡手机号，买二手号收不到验证码，拿来绑AI直接被封。搜谷歌邮箱/Gmail老号。5年老店5年老号，纯净无绑定，直登谷歌商店。顺手拿ChatGPT Plus成品和Claude，省下时间做正事。 下单地址 👉 https://t.co/TvTPWQJy56 #Gmail #谷歌邮箱 #ChatGPT #Claude

13 赞 · 0 转帖 · 0 回复

### [@mspd4100](https://x.com/mspd4100/status/2122670311535690202?ref=binnn.top)

11月17日 21:36 JST

有一说一： Codex 这电脑操作能力，还是真有点太牛逼了…… 让它开发一个 Obsidian 插件，再跑实机验证。窗口和文字刷刷切换，感觉眼睛还在追画面，测试已经跑完了。 这段视频是 Codex 开发过程中实机测试的原速录屏。 https://t.co/Z7oryiSUch

13 赞 · 0 转帖 · 18 回复

### [@OpenDesignHQ](https://x.com/OpenDesignHQ/status/2122667693098807731?ref=binnn.top)

11月17日 21:26 JST

We turned 300+ top brands’ launch videos into open-source templates + MOTION.md. Meet OpenMotion: open-source Claude Motion × CapCut for any AI agent. No Fable 5.2 required. Multimodal. Editable. No AI slop. Apache 2.0\. BYOK. MCP + plugins. Now remix any video in one click. https://t.co/mWYkJlExZ0

12 赞 · 0 转帖 · 1 回复

### [@DivyanshT91162](https://x.com/DivyanshT91162/status/2122675225016123885?ref=binnn.top)

11月17日 21:56 JST

A 145-billion-parameter AI model on a single computer, with no data center. It's called the Edge0-145B, and the company that built it says it runs on 29GB of memory. How it works: → 145B total parameters, but only about 10B are processed per token → The rest stay on your SSD until needed → Edge0 reports 80+ tokens/second and 29GB of peak memory → Its local server speaks the OpenAI and Anthropic API formats, so coding agents can plug in Pros: → It's built on Qwen3.8-Flash-Next, and every benchmark so far is Edge0's own → The download is about 105GB → The blog shows MacBook numbers, but the GitHub README says macOS isn't yet supported for this model. Currently, it's Linux or Windows with an NVIDIA GPU. → The engine is Apache 2.0, but the VATs come with their own community license. If you want coding agents that keep your code on your own machine, this is worth a look. Wait for the independent test before leaving the paid model. REPOO👇

12 赞 · 8 转帖 · 4 回复

### [@awe0\_io](https://x.com/awe0%5Fio/status/2122623645319397574?ref=binnn.top)

11月17日 18:31 JST

Claude Code 改网页最烦的，就是反复解释「我要改的到底是哪一个按钮」。 React Grab 直接解决了。 鼠标选中页面上的任意元素，复制一下，就能把对应组件和源码位置交给 Claude Code、Codex。 不用翻文件、不用猜组件名，直接告诉 AI 哪里需要改。 GitHub：https://t.co/okqHVhMk0E

12 赞 · 0 转帖 · 6 回复

### [@Sukiea1008](https://x.com/Sukiea1008/status/2122512547438096467?ref=binnn.top)

11月17日 11:10 JST

没有 opus5.5 之前我们是怎么用 agent 做视频的？ 如果你也想让 Claude Code、Codex 帮你做出不廉价的动画视频！这 7 个 GitHub 仓库一定要收藏！ 你可以按用途选： > 渲染引擎：把画面真正变成 MP4 1️⃣ HyperFrames：把 HTML、CSS、媒体素材和动画渲染成 MP4，7.4 万+ Star。 https://t.co/BHRiXybZlb 2️⃣ Remotion：用 React 按帧生成和渲染视频，6.8 万+ Star。 https://t.co/1JeuMa5CKm 3️⃣ Remotion Skills：官方技能，教 AI 正确创建、组织和渲染 Remotion 视频，5800+ Star。 https://t.co/55zHQMtwIX > 动画库：复杂编排用 4️⃣ GSAP：控制力很强的时间线动画引擎，2.9 万+ Star。 https://t.co/yEFKLxK698 5️⃣ Motion：JavaScript、React、Vue 都能用的动画库，3.4 万+ Star。 https://t.co/dhi984lK4v > 选对动效：画面能动，不等于好看 引擎只管能不能动。AI 面对无限种动画，每次都在重新猜，技术上能跑，看着却廉价、混乱。 6️⃣ su-motion-top12：我开源的动效选型 Skill。从真实网页里筛出 12 个高频动效，打字机、遮罩揭示、线条绘制、数字滚动、轨道网络等，每个都写好了用途、时长、缓动和反例。AI 每个镜头只选一个主动效，最多加一个辅助，视频默认交给 HyperFrames，指定 Remotion 就走 Remotion。 https://t.co/b6XxONtYXf 12 个动效可以在线点开看： https://t.co/tNToCUoycW > 最小起步：不出镜也能做一条 7️⃣ su-card-to-video：用 HTML/CSS 做 16:9 图卡，HyperFrames 或 GSAP 加动效，FFmpeg 合上你的音频，输出 MP4。不用剪映，不用真人出镜，配音自己录或用任意 TTS 都行。 https://t.co/ltF8HSO3tQ > 怎么选 只想先跑通一条 👉 su-card-to-video 习惯写网页 👉 HyperFrames 习惯 React 👉 Remotion 加 Remotion Skills 做出来总觉得廉价 👉 先用 su-motion-top12 选动效 如果你也准备用 agent 做视频，这几个仓库不能错过！

11 赞 · 0 转帖 · 2 回复

### [@hashlist](https://x.com/hashlist/status/2122590210030502252?ref=binnn.top)

11月17日 16:18 JST

Today we released https://t.co/FRtfPnxvc3 v1.3 to measure AI model performance on complex real-world engineering tasks across chip design, RTL, CAD, embedded coding, and more. New models: Gemini 4 Carbon, Muse Wonder, Muse Ember, Qwen4 Max, Kimi K3.1 Pro, Mistral Large 4, Inkling

8 赞 · 4 转帖 · 0 回复

### [@VeknQAQ](https://x.com/VeknQAQ/status/2122675868862661024?ref=binnn.top)

11月17日 21:58 JST

DeepSeek下面的评论区.... 以前我真的以为这会是段子的😰😰 https://t.co/9IS0A7oyE4

8 赞 · 0 转帖 · 8 回复

### [@xiaomovps](https://x.com/xiaomovps/status/2122587412849111160?ref=binnn.top)

11月17日 16:07 JST

我最近把 DeepSeek Harness 打造成了自己的 AI 自媒体工作台，只用了 3 个插件，就把整个创作流程串联起来了 我最近都是在用 DeepSeek Harness，而且也把它加入了我自己的自媒体创作流程中 我发现 DeepSeek Harness 最大的优点，就是有很多插件可以让你安装，把不同的插件组合在一起，其实就是一个属于你自己的智能体工作台 这里我就分享一下，我自己的自媒体工作台是怎么搭建的，主要用到了 3 个插件，每个插件负责不同的工作 1、ModSearch：负责热点搜索和资料收集 这是整个工作流的第一步，主要负责帮我收集最新的 AI 新闻、热门话题、开源项目以及 X 上面的内容 以前很多时候需要自己去搜索、查看不同的信息来源，现在可以让 Agent 帮我收集和整理，最后筛选出值得创作的选题 2、dsh-media-desk：负责内容创作和管理 这是我整个自媒体工作台的核心插件 收集完选题和资料之后，就可以交给它来处理，包括文章创作、视频脚本、灵感管理、内容排期等等 而且还可以根据自己以前写过的文章，整理出对应的写作风格，让生成的内容更加符合自己的表达习惯 3、dsh-guizang-design：负责内容设计和配图 内容写完之后，接下来就是设计和排版了 这个插件可以帮你把内容制作成小红书图文卡片、公众号头图、长图等等，也可以用来做一些简单的网页设计 对于经常需要一份内容发布到多个平台的人来说，这个插件还是挺实用的 这样的话，整个自媒体工作流就被串联起来了 ModSearch 负责收集信息和寻找选题，dsh-media-desk 负责内容创作和管理，dsh-guizang-design 负责最后的设计和配图 每个插件负责不同的工作，组合在一起之后，我整个创作过程就变得非常便利，而且流程也更加完整 其实不管是运营 X、抖音、小红书，还是公众号，整个自媒体的创作流程都差不多，无非就是找选题、收集资料、创作内容、设计配图，最后再发布出去 希望我的分享对大家有用，也欢迎大家分享一下自己在用的插件

8 赞 · 1 转帖 · 3 回复

### [@LawrenceW\_Zen](https://x.com/LawrenceW%5FZen/status/2122661832737673424?ref=binnn.top)

11月17日 21:03 JST

最近推上聊多 Agent 工作台的人挺多，Orca 和 T3 Code 热度都很高。 我平时主力跑的其实就是 Claude Code 和 Pi，这几天也在高强度用 Antigravity，外加 Grokbot，很少套臃肿的第三方壳子。之前在 Grok 云电脑上搭过 Orca，9 月也试过早期的 Mirasim。刚好最近手头几个项目并行，干脆把 Orca、T3 Code 和大更后的 Mirasim 都挂在主力机和手机上对比跑了一阵子，不看 GitHub 星数，就聊几个每天高强度干活才会撞上的细节。 用这类调度面板，我最烦底下乱塞一坨系统提示词，既拖慢响应又白吃 Claude 的订阅额度。实际挂下来，Orca 同时开五六个 worktree 跑久了，偶尔会留孤儿进程占内存；T3 Code 网络层目前还有点延迟，部分包装做法也偏费额度。 Mirasim 走的是纯流量转发，不注入多余提示词，把 Claude Code、Antigravity、Pi、Codex 等十几种工具直接收进一个面板。本地订阅撞上限流了，点一下就能切云端 API 接着跑，不断上下文。 自动拉 Git Worktree 三家都有，早不算稀奇了。真正烦人的是窗口之间互不相通，以前前端改完，还得人肉把字段贴给后端窗口，或者自己写桥接脚本。Mirasim 支持直接在输入框 @ 另一个正在跑的 Session，比如前端 Claude 直接 @ 后端 Codex 问接口字段改好没，甚至直接派活等回信。 另外它还带了能溯源到具体会话的文档库和白话定时器，平时在频道里像喊 Grokbot 一样拉个 Bot 分诊聊方案、顺手改文档也挺方便。 最拉开差距的是手机端。以前为了出门用手机盯 Orca，得在云端折腾 Tailscale 降级开 Funnel、写看门狗脚本防断连，还得在手机代理里专门配分流规则防冲突，折腾完手机上也只能看个轻量终端日志。 Mirasim 除了手机端本身帧率顺滑很多，最意外的是直接内置了 P2P 直连远程桌面，同城延迟才二十多毫秒，跑 2K 30fps 画面，带笔记本同款双指触控板手势和组合键。出门在外不光能看终端，连电脑桌面上的 GUI 界面都能直接拿手机点两下验收。 开源项目拉起核心功能架子很快，适合喜欢自己折腾的朋友；但像零注入省 Token、跨会话通信账本、手机高帧率和低延迟 P2P 远控这些磨人的工程苦活，商业闭源团队沉下心抠出来的完成度，日常重度干活确实更省心。

8 赞 · 1 转帖 · 7 回复

### [@Shoumya23](https://x.com/Shoumya23/status/2122586180692984240?ref=binnn.top)

11月17日 16:02 JST

🔥 Agent-Reach gives your AI agent free access to X, YouTube, LinkedIn, Reddit &amp; more! No expensive APIs. 112K+ GitHub stars. GitHub: https://t.co/ZsLmzSvm9r https://t.co/DC3QigpmD3

7 赞 · 0 转帖 · 0 回复

### [@cn\_siva](https://x.com/cn%5Fsiva/status/2122662722437287989?ref=binnn.top)

11月17日 21:06 JST

Turn your local LLM into a private, multi-user ChatGPT-style workspace — self-hosted Open WebUI on a single K3s node, ready for your household or team. Running a local model endpoint is straightforward, but sharing it securely across multiple users with isolated history, permissions, and backups usually isn't. What the stack provides: • Full stack on one K3s node: Open WebUI + PostgreSQL + SearXNG (private web search) + Redis • Multi-user privacy: per-user accounts, private chat history, per-user uploads, and personal sk- API keys • Flexible auth: built-in login form (no domain/TLS required) or Google SSO • Self-verifying backups: GPG AES-256 encrypted bundles, verified by decrypting and test-restoring every member — dumps proven against a scratch database • Simple operations: one command to deploy, one to restore Bring your own OpenAI-compatible endpoint — llama.cpp, vLLM, Ollama, or a hosted API. Plain bash + Helm, no operator, no vendor lock. GitHub: https://t.co/aZWaUpzcZy #SelfHosted #LocalAI #Kubernetes #K3s #OpenWebUI #Homelab #LLM #OpenSource #SelfHostedAI #privacy

7 赞 · 0 转帖 · 1 回复

### [@WasimShips](https://x.com/WasimShips/status/2122623520005886086?ref=binnn.top)

11月17日 18:30 JST

The next leap in AI agents won’t come from larger context windows alone. It will come from persistent memory. These 8 open source projects fix that. 1\. GBrain Persistent, source-backed memory shared across your AI agents. https://t.co/5p6q6zv46b 2\. LLM Wiki A blueprint for building an AI-maintained knowledge base that compounds over time. https://t.co/md8PyRZlWj 3\. AI Memory Carry context and work between Claude Code, Codex, Cursor and other agents. https://t.co/qa0DWRvul8 4\. Agent Second Brain Turn Telegram voice notes into connected knowledge inside Obsidian. https://t.co/BlQHfVv0Dj 5\. COG Second Brain A self-improving Markdown system with specialized agents, skills and verification. https://t.co/QGeOkpKt7m 6\. Open Second Brain Local-first memory that converts repeated corrections into learned preferences. https://t.co/yqXN5NHfF8 7\. Claude Second Brain One command scaffolds an Obsidian vault, Claude Code and local semantic search. https://t.co/ljWwDwwyHA 8\. Baalda A collaborative second brain where people and AI edit the same Markdown files. https://t.co/CHXfr5cC1l Your AI does not need another prompt. It needs memory. Save this list. 🔖

7 赞 · 1 转帖 · 4 回复

### [@SemiconductorsX](https://x.com/SemiconductorsX/status/2122561900831735913?ref=binnn.top)

11月17日 14:26 JST

Exclusive from reuters: China is using US AI model outputs to train its own systems while chip curbs tighten. Reuters reviewed more than 300 Chinese papers, patents, filings and notices. The method, called distillation, has moved from lab experiments into regular industrial use. Chinese researchers pull answers, code and solutions from US models. Local firms then automate the process and prepare the material to train smaller domestic models. Beijing is pushing the approach to cut costs and spread AI across its factories. Distillation is legal and used elsewhere. It does not copy a full frontier model or remove the need for advanced chips. It can keep selected skills in smaller systems that run cheaper and cut reliance on US technology. One 2024 Shanghai Jiao Tong paper took about 5,000 answers from OpenAI’s o1, translated them with GPT-4o mini, and built 10,750 bilingual examples to train Alibaba’s Qwen2.5-72B further. US and Chinese officials clashed over the practice in July. Washington announced tighter chip limits on Nvidia and others this week. The method could blunt the impact of those hardware curbs. $NVDA How much of the capability gap stays open once the next round of distillation scales across more Chinese labs?

7 赞 · 1 转帖 · 1 回复

### [@Chinazhidx](https://x.com/Chinazhidx/status/2122590387701506538?ref=binnn.top)

11月17日 16:19 JST

A WashU Ph.D. candidate behind G-Zero is heading to ByteDance Seed. Chengsong Huang's first-author paper, G-Zero, was selected as a NeurIPS 2026 Spotlight and won the Best Paper Award at the NeurIPS 2026 TTCL Workshop. Huang earned his bachelor's degree in Software Engineering from Fudan University. He's now pursuing a Ph.D. in Computer Science and Engineering at Washington University in St. Louis, advised by Prof. Jiaxin Huang. His research focuses on self-improving LLMs, AI agents, and efficient reasoning. He's previously worked at Sea AI Lab and Tencent's Seattle AI Lab. He's currently a Student Researcher at Google Cloud AI, working on agentic reinforcement learning for LLMs. Selected research: • LoraHub (COLM 2024): Explores how different LoRA adapters can be combined across tasks. • R-Zero (ICLR 2026): Enables LLMs to generate their own reasoning training tasks. • G-Zero (NeurIPS 2026 Spotlight): Uses self-play to extend model self-improvement to open-ended generation. • EnvHarness (2026): Explores adaptive training environments for AI agents. His work shares a common goal: enabling AI systems to learn and improve with less reliance on human-labeled data. Huang says he plans to join ByteDance Seed around January 2027\. It will be interesting to see how he brings these ideas into more complex, real-world AI systems.

6 赞 · 0 转帖 · 0 回复

### [@eviljer](https://x.com/eviljer/status/2122584021943435712?ref=binnn.top)

11月17日 15:54 JST

Claude 小号头七，遇到 VPN 流量决堤。 排查过程中，我发现 Cursor一直在后台，以接近 100M/分钟的速度持续偷传仓库快照（速度是平时的 20 倍以上）。 半个钟到对照测试后，罪魁祸首指向 Cursor 的 Share Data 模式。不知道什么时候稀里糊涂开的，还是说默认开的…… https://t.co/KBNHDjuGVr

6 赞 · 1 转帖 · 9 回复

### [@fngly343149](https://x.com/fngly343149/status/2122666606245581303?ref=binnn.top)

11月17日 21:22 JST

💥 给编程 agent 装跨工具记忆：Beacon 今日 +28 星，总数 3,548！ Claude Code、Cursor、Codex、OpenCode、Cline 等 20+ harness 的会话历史进一处，自动熬成可复用知识。 核心亮点： 🔔自我改进记忆层：工作流、纠错、调试模式全提炼回注 🔔日志可导 Splunk/Datadog/Elastic/Sentinel/Wazuh/S3/GCS 全家桶 🔔安装脚本带 sha256sum 校验，Homebrew 一键 🔔5 月 12 日建库至今 3,548 星 后果：agent 用完即忘的毛病，有人动手缝上了。 #AI记忆 #开发者工具 #Agent https://t.co/mYQRRjFfZL

6 赞 · 0 转帖 · 6 回复

### [@AGTPinsights](https://x.com/AGTPinsights/status/2122522124761928140?ref=binnn.top)

11月17日 11:48 JST

Google just warned that "the era of autonomous attacks has arrived." Here's what you need to know. The warning comes from the Google Threat Intelligence Group's latest AI Threat Tracker report, published today. It says hackers are now using autonomous LLM frameworks to continuously scan software updates, spot newly fixed security bugs, and build attacks before companies can install the patch. No human is involved. Google says it caught this in the wild for the first time. An exposed attacker staging server held a tool called exploit\_agent. It was wired into several frontier and locally hosted LLMs. It compares code before and after a patch, writes an exploit, and fires it automatically. Google ties it to Russian-speaking hackers and calls it the first fully automated run of this kind. State-backed groups are using AI too. Google links groups tied to Iran, China and North Korea to AI use for espionage, stealing sensitive data, breaking into organisations and building more advanced attacks. Chinese-linked groups used AI in Q3 2026 for reconnaissance, scanning and researching targets. Other agentic frameworks named in the coverage include ARTEX and CyberStrikeAI. Attackers are deploying multi-agent setups to automate hacking, weaponize known flaws and speed up intrusions. Security researchers say these frameworks should not be treated like traditional C2 tools. One Microsoft MVP who tested ARTEX in a lab found it attacks the same way defenders have been taught for 20 years, just at machine speed. Key numbers: - 3,800+ unique IPs per day now tied to agentic AI attack frameworks, per the Shadowserver Foundation - 4 frameworks added to Shadowserver's daily tracking: ARTEX, CyberStrike, PentestCode, VXControl - 3 state-linked countries named: Iran, China, North Korea Separately, Google will reportedly give Japan's government early access to its latest AI model to fight cyberattacks. Google says it is also helping defenders respond.

5 赞 · 3 转帖 · 2 回复

### [@linkice\_eth](https://x.com/linkice%5Feth/status/2122526315379351832?ref=binnn.top)

11月17日 12:04 JST

My goal on X is to build 10,000 meaningful connections. I want to connect with: 1\. AI & SaaS founders 2\. Indie hackers and solo builders 3\. AI agent builders 4\. Growth and marketing professionals 5\. Affiliate marketers and creators 6\. Women building businesses with AI If that's you, let's connect! 👋

5 赞 · 0 转帖 · 15 回复

### [@buildwithpb](https://x.com/buildwithpb/status/2122654863888884148?ref=binnn.top)

11月17日 20:35 JST

Hot take: You don't actually need open weights. You need a model that fits on your laptop and respects your privacy, not infinite precision.

5 赞 · 2 转帖 · 1 回复

### [@juhavalvanne](https://x.com/juhavalvanne/status/2122579883771449521?ref=binnn.top)

11月17日 15:37 JST

AI writes faster. CI needs to keep up. That’s why we started building @avreacom with @HannuValtonen. Now #1 in RunsOn’s GitHub Actions build benchmark. At Slush? DM me: 10k–1M build minutes in credits for VC-backed startups. https://t.co/Wb5umbG7Zl https://t.co/8tdjvFD9ol

5 赞 · 1 转帖 · 0 回复

### [@SoduWork](https://x.com/SoduWork/status/2122579739651252653?ref=binnn.top)

11月17日 15:37 JST

Meet SoduWork — an open-source, enterprise-grade AI-native engine. Multi-agent collaboration, built-in governance, verified delivery. MIT licensed. Build with us: https://t.co/uJ7Pj2xxJc #OpenSource #AIAgents https://t.co/oH6LpNd2XZ

5 赞 · 0 转帖 · 0 回复

### [@RovertoID](https://x.com/RovertoID/status/2122616158775828749?ref=binnn.top)

11月17日 18:01 JST

An AI that fits into 400 KB and runs under DOS. Yes, really! While modern AI models rely on powerful servers and huge amounts of memory, one enthusiast has created a tiny AI that can run on an old computer. And here are the impressive details: - Just \*\*367,000 parameters\*\* - Four layers, each 96 neurons wide - Around \*\*380 KB\*\* to store the model weights - Approximately \*\*540 KB of RAM\*\* required to run it The model was trained on 3,500 facts, each presented in several different ways. This allows it to recognize familiar questions even when they’re phrased differently. Of course, it’s not competing with modern models like ChatGPT or Claude. Ask it something outside its training data, and it may confidently make things up. But that’s not really the point. This project is a reminder that you don’t always need billions of parameters or expensive GPUs to experiment with AI. Sometimes, all it takes is creativity, a little code, and a desire to see what’s possible with very limited resources. The creator has made the project open source on GitHub, where you can explore how it works and even run it yourself: 👉 https://t.co/7kTDtfD0ix Would you try running an AI on a computer from the DOS era? #dos #AI #Claude

5 赞 · 0 转帖 · 0 回复

### [@zeromarkapp](https://x.com/zeromarkapp/status/2122593207242354946?ref=binnn.top)

11月17日 16:30 JST

Your saves are the photos you took. Your taste is the ones you picked. Try yours in the Brainstorm tab: https://t.co/aV2zY0tgfL https://t.co/U0ziTGFTYE

5 赞 · 1 转帖 · 1 回复

### [@shenxiankk](https://x.com/shenxiankk/status/2122534189593657563?ref=binnn.top)

11月17日 12:36 JST

GitHub 上有个项目已经拿下 2.6 万 Star！它能让 AI 掌握 CAD 技能——只要把形状和尺寸讲清楚，三维模型就由它替你搭建。 这个开源项目叫 text-to-cad，作用是把建模能力对接到 AI Agent 上。 举个例子，你说“做一个长 80 毫米、带四个安装孔的支架”，Agent 便会写出一段 Python 建模脚本，在本机执行，随后启动 CAD 查看器把成品呈现出来。 无论是单个零件还是装配体，它都能生成，支持导出 STEP、STL、3MF 等格式，还可以产出标注尺寸的工程图、执行可制造性检查，跟 3D 打印流程无缝对接。 对机械专业学生、硬件工程师和 3D 打印爱好者来说，脑海中设想的零件总算可以先说出来，再一步步打磨了。 项目链接：https://t.co/4GGOMgJsGG

5 赞 · 1 转帖 · 7 回复

### [@cheerselflin](https://x.com/cheerselflin/status/2122565608684601524?ref=binnn.top)

11月17日 14:40 JST

🔥fable 5.2的含金量仍在上升 fable 5.2 + three.js 王炸组合 游戏，motion动效，视频剪辑.... 这些百万爆款的制作方式都已经收录到 https://t.co/rWaDZL5M68 https://t.co/igxD5u1l6j

5 赞 · 0 转帖 · 0 回复

### [@tongfeifan](https://x.com/tongfeifan/status/2122565199379227006?ref=binnn.top)

11月17日 14:39 JST

https://t.co/73WRAPbl7X

5 赞 · 3 转帖 · 0 回复

### [@SolhawkGo](https://x.com/SolhawkGo/status/2122677114357137438?ref=binnn.top)

11月17日 22:03 JST

WTF，受不了了！挖到金矿了！！！！ GitHub 上这 10 个开源项目，简直是把一整家 AI 公司免费送给你了！ AI 程序员、产品经理、设计师、销售、客服，甚至tm连 AI 的记忆、工作流、自动化系统都给你准备好了 以前公司做个项目要花几个月才能做出来啊，现在直接拿现成的改，效率直接提高10倍 搞 AI 副业、做产品、接外包，都能找到用得上的，开整： 一、AI 员工与专业技能：直接组建 AI 团队 1、Agents｜4 万 Star 一个超超超超超级大的 AI 员工工具箱 涵盖编程、测试、代码审查、安全、调试等任务，按需安装，不用自己从零写 Agent，配合Codex小学生都能用 https://t.co/81mYB7KVUc

5 赞 · 2 转帖 · 1 回复

### [@EthanLinAI](https://x.com/EthanLinAI/status/2122616773426151432?ref=binnn.top)

11月17日 18:04 JST

Alibaba open-sourced its engineers' AI code reviewer — it beats Claude Code on precision at 1/9 the tokens. 🔍 53K stars. OpenCodeReview is an AI code-review CLI that reads your Git diff and returns structured, line-level comments — deterministic engineering where it can't go wrong, an LLM agent where it needs judgment. ❌ Agents "cutting corners" on big diffs — reviewing some files, silently skipping others ❌ Review comments pointing at the wrong line, or the wrong file entirely ❌ Review quality swinging wildly with every small prompt change ✅ Line-level precision: deterministic file selection, bundling and positioning, agent judgment on top ✅ Proven at Alibaba's scale: two years, tens of thousands of developers, millions of defects What makes it different: 🧪 Benchmarked, not claimed: 200 real PRs, 50 repos — it beat Claude Code on precision and F1, at \~1/9 the tokens 🧩 Smart bundling: related files reviewed together, each in its own sub-agent with isolated context 📏 Multi-language rules: NPE, thread-safety, XSS, SQL injection, matched to each file by template, not by vibes 🔀 Two modes: OCR calls the model itself, or delegation mode lets your agent review with its own model — no extra API key 🚀 Get started: 1️⃣ Install: \`npm install -g @alibaba-group/open-code-review\` (needs Git 2.41+) 2️⃣ Configure a model: \`ocr config provider\` then \`ocr config model\` 3️⃣ Review: \`ocr review\` for your working tree, \`ocr scan\` for whole files ⚠️ Honest caveat: recall is deliberately lower than a general-purpose agent — it trades finding every defect for far fewer false alarms. Without delegation mode, you must configure your own LLM endpoint first. ⭐ 53,074 stars (Nov 17, 2026) · Apache-2.0 #AI #OpenSource #GitHub #CodeReview #AIAgents 🔗 https://t.co/SvQhBow0WF

4 赞 · 0 转帖 · 2 回复

### [@moondance2010](https://x.com/moondance2010/status/2122574130524631182?ref=binnn.top)

11月17日 15:14 JST

AI/LLM大模型逐步入门指南，已经4K多星。 https://t.co/JAMG0srIkq

4 赞 · 0 转帖 · 1 回复

### [@fngly343149](https://x.com/fngly343149/status/2122625253352554816?ref=binnn.top)

11月17日 18:37 JST

🤯 341 种动画风格塞进一个 skill：1,229 星！ 代码动画提示词库 + 196 张参考静帧 + 10 大系列，一句话出 16:9 成片，Claude Code 和 Codex 直接装。 核心亮点： 🔔一句「用 Z5 剪纸做 12 秒新年片头」render\_video.py 直接出 MP4 🔔Z 国风（剪纸/水墨/敦煌）、R 赛博（霓虹雨夜/像素都市），原始 HTML+MP4 全随仓库 🔔渲染脚本开源，帧率时长随便调 后果：别人肝 keyframes 的时候，你一句话交片。 #ClaudeCode #视频生成 #提示词 #动画 https://t.co/SYOG81V5s4

4 赞 · 0 转帖 · 2 回复

### [@fngly343149](https://x.com/fngly343149/status/2122659747308679663?ref=binnn.top)

11月17日 20:54 JST

🤯 完全本地跑的语音 Jarvis：2,332 星，本周还在涨！ 语音识别、语言模型、合成全在你机器上，喊名字就能插话，没有订阅费。 核心亮点： 🔔模型按需选：qwen3.5:0.8b 省硬件、默认 gemma4:e4b、大机器 qwen3.8:27b 🔔滚动转录旁听聊天，敏感信息进上下文前先脱敏 🔔macOS/Windows/Linux 安装包 release 直出，Python 3.11+ 可源码编 🔔2025 年 8 月建库至今 2,332 星 553 fork 后果：科幻片里那个管家，落到自家电脑了。 #本地AI #语音助手 #隐私 https://t.co/FdetZMNZ0g

4 赞 · 0 转帖 · 1 回复

### [@vintcessun](https://x.com/vintcessun/status/2122669845380739079?ref=binnn.top)

11月17日 21:35 JST

给 QQ 接 Agent，先别急着堆唱歌、语音，看看“什么时候该闭嘴”能不能改。 Aurabot 把 DeepSeek Harness 接进 QQ，值得关注的是控制层：群聊是否接话、选哪个模型、回复怎么拆，都能用模块改写；自写模块热重载，升级不覆盖。 记忆落 SQLite、按聊天来源隔离，默认不向 QQ 发消息，还能截获发送做模拟测试。这些比功能数量更有参考价值。 但模块与插件同进程，不是沙箱，只装可信代码。 https://t.co/Cqzljeb2qO

4 赞 · 0 转帖 · 4 回复

### [@AYUSHKUMAR82274](https://x.com/AYUSHKUMAR82274/status/2122619668670103956?ref=binnn.top)

11月17日 18:15 JST

@Alibaba\_Qwen just dropped Qwen4-Max. Looking at their benchmark chart, it leads or stays competitive across several agentic and coding tests compared to Opus 5.5, GPT-6.1-Sol, and Kimi K3.1\. Notable scores: • DeepSWE v1.1 → 75.7 • Agents’ Last Exam → 46.7 • Automationbench-AA → 72.5 • BioMysteryBench → 85.0 • HLAB-Vals-1.1 → 21.3 1M context, native vision, and claimed \~60% lower task cost vs previous Qwen3.8-Max. Weights are closed for now. The Chinese open-source ecosystem is moving fast. #qwen4 #LLM

3 赞 · 0 转帖 · 1 回复

### [@jinglian](https://x.com/jinglian/status/2122686555383107750?ref=binnn.top)

11月17日 22:41 JST

🔥 发现了一个完美解锁「Claude Fable 5.2」的神仙途径！ 不仅比官网 API 便宜 10 倍，更爽的是 100% 零封号风险！ 最近看推上大佬们用 Fable 狂做视频，心痒难耐？不用折腾，最省事的解决方案来了 👇 ✅ 极简操作，只需 3 步： 1️⃣ 下载 YouArt Desktop 桌面端应用 2️⃣ 打开并直接选择「Claude Fable 5.2」模型 3️⃣ 丢给它一个官网链接（例如 https://t.co/hEBr4Gy7wH ），下达指令：帮我出个宣传片 然后去喝杯咖啡，等个十几分钟 ☕️ 一条高质量片子就直接跑出来了！整体画面效果非常惊艳，再重新补一个旁白配音，一条高质感的宣传片就彻底搞定了 🎬 🎯 废话不多说，想抢先体验 Fable 5.2 视频生成魅力的，赶紧戳这里上车： 🔗 https://t.co/iocWXNBiCl

3 赞 · 2 转帖 · 2 回复

### [@Feng\_btc](https://x.com/Feng%5Fbtc/status/2122665585544974781?ref=binnn.top)

11月17日 21:18 JST

为什么现在学生或者实体办公的Web2人都知道ChatGPT很强，但还是不去尝试呢？ 说白了就是懒 懒得找梯子，懒得注册 Gmail，懒得折腾支付方式，更懒得花钱充会员 最后默默地打开了豆包，或者是去千问撸一杯奶茶🤣 #ChatGPT #Codex

3 赞 · 0 转帖 · 3 回复

### [@ai\_goge](https://x.com/ai%5Fgoge/status/2122669739239711011?ref=binnn.top)

11月17日 21:34 JST

https://t.co/q8z9BJdIzZ

3 赞 · 0 转帖 · 2 回复

### [@hailports](https://x.com/hailports/status/2122511070686257628?ref=binnn.top)

11月17日 11:04 JST

my local llm server, a base m4 24gb mac mini: live median decode is 33 tok/s. peak looks fun. 938 answers of 100+ tokens. measured median 32.8\. fastest 69.0\. 35% of answers at 40 tok/s or more. if you run local decode on a base m4, check median over hundreds of long answers, not peak.

2 赞 · 0 转帖 · 3 回复

### [@DatBoyBlu3](https://x.com/DatBoyBlu3/status/2122652445285421390?ref=binnn.top)

11月17日 20:25 JST

Looking at my grad project rn where I have to train an LLM to detect deviations from baseline traffic…I picked the best career but holy shit, we’re in a new era folks.

2 赞 · 1 转帖 · 0 回复

### [@Cult\_Of\_Token](https://x.com/Cult%5FOf%5FToken/status/2122637532101779527?ref=binnn.top)

11月17日 19:26 JST

Open weights alone don't move science. #NVIDIA ships the model, \~50K chemistry questions and post-training recipes. Lila Sciences: 61% vs 54% for GPT-6 Astra at 10–20x lower cost. #OpenSource #AIagents #LLM

2 赞 · 0 转帖 · 0 回复

### [@LLMDevPass](https://x.com/LLMDevPass/status/2122667640300671374?ref=binnn.top)

11月17日 21:26 JST

GLM-5.3 Prime is live on LLM Gateway: the high-speed edition of Zhipu's GLM-5.3, same capabilities at 1.5-2x the output throughput. 1M context, tools + reasoning. $2.80/M in, $8.80/M out. https://t.co/ReKbZ1o9hr

2 赞 · 1 转帖 · 0 回复

### [@lg5uf10cml](https://x.com/lg5uf10cml/status/2122682916190769602?ref=binnn.top)

11月17日 22:26 JST

Startup hiring: Principal Attack Engineer, AI/LLM Company: https://t.co/7A67pUzX47 Pay: $238K–$280K Location: Remote (US) - Lead AI/LLM offensive security research - Build attacks for the NodeZero pentesting platform - Set technical direction Comment "in" for the link.

2 赞 · 1 转帖 · 0 回复

### [@arnavkot](https://x.com/arnavkot/status/2122612673707909498?ref=binnn.top)

11月17日 17:47 JST

i was filling up a google labs survey to be a beta user, where i was asked what AI tools I use i realized that my main use of AI tools is just accessing data through natural language MCP connections: Granola, Exa, Paper, etc. my more meaningful, deeper use of AI is not in tools I pay for, it's agents that I custom build with the tools I use not sure how this'll evolve in a couple years but today I know MCP connections are important and I want to be on the forefront of that for relationships on X i want enterprises to be able to query X beautifully - a 1v1 with @bot (emerald player vs grandmaster but playing ARAM on my one-trick)

2 赞 · 0 转帖 · 2 回复

### [@SlimAssiliX](https://x.com/SlimAssiliX/status/2122642007168303369?ref=binnn.top)

11月17日 19:44 JST

Microsoft Found AI-Assisted Attacks Reaching Privileged Access in Under 3 Hours, Here's What Actually Happened: https://t.co/YTjQbHZemU

2 赞 · 1 转帖 · 1 回复

### [@arito\_si](https://x.com/arito%5Fsi/status/2122576233297051680?ref=binnn.top)

11月17日 15:23 JST

离谱，现在随便拿个官方Demo改个UI就能包装成“AI产品”出来割韭菜了？信息差真就这么好赚？其实背后技术门槛并没有你想的那么高。 干脆把底裤掀了，这篇推文手把手教你从0到1「手搓」一个同款 DeepSeek。看懂了，你也能自己搭！！ https://t.co/0YGRnzhQN5

2 赞 · 0 转帖 · 2 回复

### [@tuturetom](https://x.com/tuturetom/status/2122657399484072303?ref=binnn.top)

11月17日 20:45 JST

应对Claude大封号最好的方式就是把它的生态搬到其他agent上 💥🔥 之前我们把它的Design生态开源了，这次更是把Claude motion直接搬进了Codex、DeepSeek和WorkBuddy。 完全不受模型能力限制，支持BYOK多模态模型和20+agent调用及上千份模版资源，把全栈 Code2Video 工作台装进你已经在用的 Agent。 不用绑死在某一个模型、某一个账号上。

2 赞 · 0 转帖 · 4 回复

### [@AliciaCrawoiue](https://x.com/AliciaCrawoiue/status/2122655636190269588?ref=binnn.top)

11月17日 20:38 JST

Yesterday Alibaba launched Qwen4-Max and kept the weights closed. Today Mistral's Large 4 went generally available — 1 trillion parameters, open weights, $0.70 per million tokens. The open-weight labs aren't conceding anything.

1 赞 · 0 转帖 · 1 回复

### [@TweetThreatNews](https://x.com/TweetThreatNews/status/2122513113341661539?ref=binnn.top)

11月17日 11:12 JST

California Attorney General Rob Bonta sued OpenAI, alleging inadequate safeguards for AI models contributed to sandbox escapes and real-world hacks, and that the company obstructed investigations. OpenAI had not commented to CyberScoop. https://t.co/o9whptkzPV

1 赞 · 1 转帖 · 1 回复

### [@LFGAction](https://x.com/LFGAction/status/2122661030031716613?ref=binnn.top)

11月17日 21:00 JST

California is suing OpenAI over allegedly unsafe AI-agent testing. The state points to the Hugging Face breach and says at least 80 incidents occurred. OpenAI says it has strengthened safeguards and disputes the lawsuit's framing. Who is accountable when an AI agent goes rogue? https://t.co/ZRcPFsmzHE

1 赞 · 1 转帖 · 1 回复

### [@dewyashtwts](https://x.com/dewyashtwts/status/2122674426131198258?ref=binnn.top)

11月17日 21:53 JST

most goated mcp connector platforms for your agents 1\. Composio @composio - 1,700+ toolkits, managed oauth https://t.co/xAVxPhPfME 2\. Merge agent handler @merge\_api - audit trails and scoped tool packs https://t.co/k1wh0DrpFl 3\. Executor @executorsh - open source gateway you can self-host https://t.co/BevwW1Bk6m 4\. Pipedream @pipedream - 3,000+ apps https://t.co/klLVvYdbvu 5\. Arcade @tryarcade - per-user permissions on every tool call https://t.co/IA9r0BW2yu what did I miss?

1 赞 · 0 转帖 · 0 回复

### [@Zzz76ha](https://x.com/Zzz76ha/status/2122672711394820535?ref=binnn.top)

11月17日 21:46 JST

AI seems to be crazy and funny recently https://t.co/XDmOVNBkPn

1 赞 · 0 转帖 · 0 回复

### [@asha\_aiii](https://x.com/asha%5Faiii/status/2122574376713551946?ref=binnn.top)

11月17日 15:15 JST

Horizon (14.5k ⭐) — an open-source AI news radar that replaces manual feed scrolling • Aggregates Hacker News, GitHub, X, Reddit, Telegram &amp; RSS • Deduplicates stories, scores relevance &amp; summarizes top comments • Includes an MCP server &amp; Agent Skill https://t.co/xETPNPiXvx

1 赞 · 0 转帖 · 0 回复

### [@oren\_ai](https://x.com/oren%5Fai/status/2122594564783444068?ref=binnn.top)

11月17日 16:35 JST

Everyone's racing to let AI build and run its own loops. Cool. Who's checking what those loops actually do? Azathoth: the AI writes the workflow, the engine keeps it deterministic. Every step recorded. Models never touch the engine. Open source: https://t.co/PzSjLKDuWU

1 赞 · 0 转帖 · 1 回复

### [@Bigmindsai](https://x.com/Bigmindsai/status/2122652352952086617?ref=binnn.top)

11月17日 20:25 JST

AI agents should learn from past sessions instead of repeating the same mistakes. Better context retention can reduce costs and improve workflow efficiency. @MSFTResearch @github

1 赞 · 0 转帖 · 0 回复

### [@arnaudmercier](https://x.com/arnaudmercier/status/2122688856470290854?ref=binnn.top)

11月17日 22:50 JST

<p class="wp-block-paragraph">The AI industry’s boldest promise right now is that AI will soon improve itself, with almost no need for human oversight. LLMs can already write code, generate synthetic data for training, and optimize the computer chips they run on. <a href="https://t.co/5i6tw0ZLoo">Forecasts</a> of explosive AI progress predict that what researchers call <a href="https://t.co/eFfB42nNQz">recursive self-improvement</a> is on the horizon.&nbsp;</p> <p class="wp-block-paragraph">But a <a href="https://t.co/hAt592sfEB">new study</a> suggests that it might take a while for us to get there. The researchers behind it found that AI agents are not yet capable of conducting open-ended AI research—free-form investigations that have no clear-cut answers and require judgment and taste, which may be integral to building self-improving AI.</p> <p class="wp-block-paragraph">A multi-institution group of researchers, led by Peter Kirgis and Sayash Kapoor at Princeton University, found that AI agents could solve the engineering problems necessary to do AI research but lacked the judgment and creativity to produce original research at the caliber of&nbsp; papers accepted by a top machine-learning conference. The gap suggests that some of the hyped-up timelines for automating AI research may be running ahead of the evidence.</p> <p class="wp-block-paragraph">Most existing research on how agents can automate AI research evaluates their ability to complete narrow tasks with checkable answers, such as solving engineering problems or post-training small language models against a benchmark. But making progress in AI research also requires open-ended thinking—choosing a set of hypotheses, deciding what evidence would settle a question, or knowing when to start over.&nbsp;</p> <p class="wp-block-paragraph">To test agents on those kinds of skills, the researchers in the study proposed a new method of evaluation called “shadow evaluation,” which requires the AI to answer a research question from a high-quality unpublished paper.&nbsp;</p> <p class="wp-block-paragraph">The researchers asked Anthropic’s Claude Opus 4.8, running on open-source software called OpenClaw, to tackle such questions, in this case from two papers submitted to the prestigious machine-learning conference NeurIPS 2026.&nbsp;</p> <p class="wp-block-paragraph">The first question was whether a large language model’s “personas,” which determine its behavior, can be controlled by editing the model’s <a href="https://t.co/KyoA3Rqtkh">weights</a> (the billions of numbers that store everything it learns during training). The other asked how to design a detector that points out when a model that makes predictions based on spreadsheet data has become unreliable. Because the papers had not been made public, the agents could not memorize the answers from their training data or find them online.&nbsp;</p> <p class="wp-block-paragraph">The agents were given six days, $3,000 in Anthropic API credits, a GPU budget to run the experiments, their own virtual computers, and access to the open web to produce a research paper worthy of publication at a top-tier AI conference. The papers’ original authors graded the agents’ papers as they would evaluate one submitted to a conference.</p> <p class="wp-block-paragraph">Those authors rejected both papers.&nbsp;</p> <p class="wp-block-paragraph">The agents were capable of all the engineering required to conduct the research, the human scientists found. The agents reviewed the literature, ran hundreds of experiments, and compiled the results.&nbsp;</p> <p class="wp-block-paragraph">“On the other hand, the agents were unambiguously bad at carrying out the research itself,” says Kapoor. They ran bizarre experiments (in some cases testing their hypotheses on tiny synthetic datasets), struggled to write intelligibly about their work, and made no novel contribution to their fields. “The papers were nowhere close to the mark when it came to being at the quality of a top AI conference,” he says.&nbsp;</p> <p class="wp-block-paragraph">That’s because the agents struggled to muster the creativity and judgment necessary for conducting research. They didn’t do enough to explore different ideas, and they committed to unpromising approaches too quickly. Though the agents developed novel and ambitious hypotheses resembling those that the original authors themselves started with, they rejected them on the basis of very limited data. And they couldn’t backtrack from failing approaches. They could make small pivots but could not fundamentally rethink their approach or try new ones from scratch.&nbsp;</p> <p class="wp-block-paragraph">The agents also failed to incorporate feedback from subagents or external AI reviewing tools. Instead of revising their methodology, the agents narrowed their claims and added caveats. They also couldn’t effectively use resources, such as tokens, compute, and time. And they couldn’t follow instructions about things like how much time to spend on different phases of the research or how long their paper could be.</p> <p class="wp-block-paragraph">For all their failures, the agents didn’t engage in the misbehavior that researchers call “<a href="https://t.co/njZZXAcuQp">reward hacking</a>,” hiding or misrepresenting experiments or data. Although subagents, or helper AIs that the main agent spawns to handle pieces of the work, occasionally hallucinated or misrepresented the results, these were caught by the orchestrator agent, the lead AI supervising the project.&nbsp;</p> <p class="wp-block-paragraph">The reason AI models are good at research engineering but not at open-ended research may come down to how they’re trained, says Kapoor. Models get good at whatever they can be drilled on in a training regime called reinforcement learning, which is easier to apply to tasks whose success can be checked automatically. “But it’s harder to create environments to train these models when the task itself is open-ended,” he says.</p> <p class="wp-block-paragraph">Kapoor says the team is now conducting the experiment with Mythos, Anthropic’s most advanced model, which launched in April. It was subsequently required by the Trump administration to meet various safety restrictions and is now available only to approved organizations. Anthropic did not respond to a request for comment.</p> <p class="wp-block-paragraph">There are some limitations to the study. It covered just two research papers, and the original authors knew the papers they were grading were generated by AI agents, which could have colored their evaluations. And the researchers had substantial discretion in designing and executing the study, meaning that their preexisting beliefs and biases could have slipped into the results. Evaluations of open-ended research trade some objectivity for a much richer test than any benchmarks can offer.</p> <p class="wp-block-paragraph">Still, the results may temper the claims that recursive self-improvement is on the horizon. In June, Anthropic published a blog post titled <a href="https://t.co/VaEAwJZyuD">“When AI Builds Itself,”</a> charting its progress toward models that speed up their own development. In July, OpenAI <a href="https://t.co/Zk4ZtUxPdY">advertised</a> the fact that its new model GPT-5.6 Sol had helped post-train a smaller model, saving researchers weeks of work.</p> <p class="wp-block-paragraph">The new finding may echo what AI companies are finding internally, regardless of their most optimistic public statements. Anthropic cofounder Jack Clark wrote in his newsletter <a href="https://t.co/xA3KhhFUfi">Import AI</a> that it rhymes with what the company found when it tried to <a href="https://t.co/iYxBhpZ6hz">automate</a> some aspects of AI safety research. </p> <p class="wp-block-paragraph">“There’s a certain absence of valuable, intuitive creativity in today’s AI systems, and though they’re extraordinarily capable engineers they seem to have a certain property of rote, formulaic thinking that might prevent them \[from\] being good researchers,” he wrote. He called AI systems’ lack of creativity a “bearish signal on short recursive self-improvement timelines.”&nbsp;</p> <p class="wp-block-paragraph">AI companies do have every incentive to develop AI systems that can rapidly accelerate their own progress, just as they did to make the models better at coding. OpenAI has made building an <a href="https://t.co/2WRwJsxLVh">automated AI researcher</a> an explicit goal, and Anthropic identifies self-improving AI as the industry’s next milestone. </p> <p class="wp-block-paragraph">“If there is investment and then conscious effort toward this direction, I feel like there would be interesting progress, even if it’s failing currently,” says Najoung Kim, a professor of linguistics and computer science at Boston University who researches how AI agents can automate AI research but did not work on the study. On the other hand, it’s possible that AI progress may be bifurcated. AI systems might race ahead on narrow tasks—the kind that can be scored—while advancing slowly on open-ended research.&nbsp;</p> <p class="wp-block-paragraph">The big open question, then, is how crucial open-ended research is to recursive self-improvement—whether AI systems can grind their way there without it, simply by improving on the narrower tasks. “If we look back to the biggest advances in the field, the invention of transformers or the invention of big new architectures that allowed us to make a lot of AI progress—all of those did require creative leaps,” says Kapoor.&nbsp;</p> <p class="wp-block-paragraph">“That said, others have this hypothesis that all of what we need for transformative AI, in particular for recursive self-improvement, is already there.” That would include making a model train faster and boosting its benchmark scores.</p> <p class="wp-block-paragraph">“That’s frankly the trillion-dollar question right now,” he says.</p> <p class="wp-block-paragraph"></p> #Artificialintelligence #App #Summary https://t.co/yxHm07a1d4

0 赞 · 0 转帖 · 0 回复

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