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

# 2026年11月13日 · AI 早报
- URL: https://binnn.top/x-brief-2026-11-13-am/
- Published: 2026-11-13T02:00:00.000Z
- Updated: 2026-11-13T08:45:58.000Z
- Description: Claude把协作推进到会议现场，并开放自托管智能体私测；英伟达人事变动的媒体报道与模型权重取证引发算力讨论。另一条主线是工程可靠性：如何控制智能体权限、验证自我改进，以及让开放模型更容易部署和使用。
- Author: Sunbin 编辑部
- Tags: AI快讯, 早报, #sunbin-x-digest

11月12日 23:00 — 11月13日 11:00（日本时间） · 134 条 X 原帖

[查看本期全部 134 条原帖 ↓](#x-original-posts)

本期目录

[Claude进入会议现场，企业部署继续开放](#flash-topic-0)[英伟达人事报道与权重取证，把焦点拉回算力](#flash-topic-1)[模型竞争同时追求速度、开放权重与本地可用](#flash-topic-2)[Harness实践：权限、验收与成本比组件数量更关键](#flash-topic-3)[研究关注自我改进的可信度与推理成本](#flash-topic-4)[机器人评测走向多机体与现实对应](#flash-topic-5)[创作与教育工具开始重视可编辑的中间成果](#flash-topic-6)

## Claude进入会议现场，企业部署继续开放

### 会议还在进行，文档已经开始生成

Claude官方推出Claude Meetings，面向Team和Enterprise套餐逐步开放测试，可在视频会议中共同制作文档、白板等成果。Wes Roth补充介绍，首批支持Google Meet，Claude作为可见参会者回答问题，成员能边讨论边修改产出；会后，纪要、行动项、转录与生成材料保存在一起。产品重点由记录谈话推进到现场完成工作。

Jack的功能梳理还提到，启用零数据保留或HIPAA配置的组织暂时无法使用。这是社区整理的适用限制，本批官方短帖没有展开说明。

来源 · 3 条原帖
- [@claudeai](https://x.com/claudeai/status/2120973994191094179?ref=binnn.top)
- [@WesRoth](https://x.com/WesRoth/status/2121022735006531986?ref=binnn.top)
- [@JacksAISites](https://x.com/JacksAISites/status/2120978848112103662?ref=binnn.top)

### 智能体框架可以放进自己的云和机房

ClaudeDevs宣布自托管Claude Managed Agents进入私测。企业可以自行托管执行框架、沙箱和会话历史，并选择Claude请求的发送位置。官方确认的是部署控制权；部分转述所说的“代码和密钥不出家门”“再也不用担心泄露”，超出了这条公告能够支持的范围。

来源 · 1 条原帖
- [@ClaudeDevs](https://x.com/ClaudeDevs/status/2120938988165091693?ref=binnn.top)

### 地区政策带来账号使用反馈

李老师不是你老师转述，11月12日新版地区政策生效后，V2EX集中出现身份验证与封号反馈；小奇说也自述Claude账号被封。它们反映了具体使用障碍，本批没有足够材料判断整体影响规模。

来源 · 2 条原帖
- [@whyyoutouzhele](https://x.com/whyyoutouzhele/status/2121035313162211711?ref=binnn.top)
- [@xingbugengming](https://x.com/xingbugengming/status/2121058876812570796?ref=binnn.top)

## 英伟达人事报道与权重取证，把焦点拉回算力

### 先是离任消息，再是开源承诺

The Information记者Jing Yang在帖中称黄仁勋已离任，并将相关投资关系与DeepSeek的Blackwell供应问题联系起来。本批没有英伟达人事声明原文，因此这一重大变动应按记者报道交代。

Nader Khalil随后重申英伟达对开源的投入不会动摇，列举运行时、模型、执行框架、语言和工具链等层面。产业讨论的实质，是领导层变化是否影响中国市场、算力获取和开放生态；相关预测仍属于评论。

来源 · 2 条原帖
- [@jingyanghk](https://x.com/jingyanghk/status/2120900244598919379?ref=binnn.top)
- [@NaderLikeLadder](https://x.com/NaderLikeLadder/status/2120956179274936560?ref=binnn.top)

### 初始化硬件指纹，不能直接推成完整训练记录

Will Depue提出，模型权重中残留的初始化排列可以追溯GPU与随机种子。他称搜索超过六万亿种组合，将DeepSeek、Moonshot、MiniMax和蚂蚁的六个模型匹配到Blackwell，并用真实硬件复现部分残留初始权重。

这份自述报告引出的研究问题很具体：公开权重可能泄露多少训练元信息。但其直接检测对象是初始化特征，不能自动证明整个训练过程所用硬件、集群规模或芯片取得方式；帖中的数千卡规模属于作者进一步估算，独立验证也只是作者所述。

来源 · 1 条原帖
- [@willdepue](https://x.com/willdepue/status/2120992457630326841?ref=binnn.top)

## 模型竞争同时追求速度、开放权重与本地可用

### Gemini 4 Hydrogen主打快速完成任务

Google DeepMind正式介绍Gemini 4 Hydrogen，定位快速编码、多模态综合与日常工作流，官方称任务完成速度为3.8 Flash的2.5倍，并在Artificial Analysis智能指数上超过Argon。Leo的个人测试则有差异：鹈鹕骑车任务约五分钟完成，评价很高；半小时生成互动餐厅后，他认为3D建模只与DeepSeek v4.1 Flash 1017持平。单个演示支持体验判断，尚不足以推出全面排名。

来源 · 3 条原帖
- [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2120940338869055788?ref=binnn.top)
- [@leo114119](https://x.com/leo114119/status/2120909143745892687?ref=binnn.top)
- [@leo114119](https://x.com/leo114119/status/2121048325789192483?ref=binnn.top)

### 百B级模型尝试降低设备门槛

Samuel Zeng发布Edge0-145B和120B开放权重，自述内存需求仅29G，在RTX Spark和Mac上解码超过75 tokens/s，并报出Terminal-Bench 2.1的86.5%成绩，对照Claude Opus 4.6的78.2%。这些数字保留为发布方口径，尚无本批独立复测。

另一条开放权重消息来自Gorden Sun：Thinking Machines的Inkling 2 Small已开源，总参数594B、激活17B。其性能接近DeepSeek的说法没有交代评分体系，不能据此建立统一榜单。

来源 · 2 条原帖
- [@SamuelZengML](https://x.com/SamuelZengML/status/2120939739431972922?ref=binnn.top)
- [@Gorden\_Sun](https://x.com/Gorden%5FSun/status/2121056912498364497?ref=binnn.top)

### 本地模型开始跨设备提供服务

Omarchy介绍即将进入4.6版的共享面板：台式机运行模型，笔记本通过自己的Tailscale网络和密钥调用；底层共享引擎已发布。Maple Work则宣布用开放模型读文件、操作应用。两者分别解决设备算力不足和桌面任务执行的问题，不能只凭“私密”定位就推定所有数据始终留在本机。

来源 · 2 条原帖
- [@OmarchyUpdates](https://x.com/OmarchyUpdates/status/2120969956364026157?ref=binnn.top)
- [@TryMapleAI](https://x.com/TryMapleAI/status/2120922279072678119?ref=binnn.top)

### 推理吞吐与跨硬件一致性各有进展

vLLM公布DeepSeek-V4.1-Flash在Vera Rubin NVL72上的优化结果：按SemiAnalysis AgentX口径，在P90每用户387 tokens/s时，吞吐达到每GPU170万tokens/s，对GB300 NVL72为58倍；峰值吞吐倍率则为2.7倍，两个条件应分开理解。

Nathan Barry讨论另一种目标：浮点累加顺序会随硬件和内核变化，影响输出；他提出连softmax和归一化也使用整数运算，以取得GPU与CPU间逐位一致的结果。

来源 · 2 条原帖
- [@vllm\_project](https://x.com/vllm%5Fproject/status/2121008091538092393?ref=binnn.top)
- [@nathanrs](https://x.com/nathanrs/status/2120967295636021503?ref=binnn.top)

## Harness实践：权限、验收与成本比组件数量更关键

### 企业智能体需要可追踪、可恢复的执行过程

Ethan恩宇的长文把能力与权限分开：身份校验、审批、持久任务状态、重试和审计，决定智能体能否接触真实业务。其关键提醒是，工具返回成功并不等于业务状态正确，也不意味着任务已经完成；架构应按风险和规模选择，个人工具无需照搬全部企业组件。

模型本身的完整性也进入讨论。Ahmad Alam转述微软研究称，修改不到十个神经元即可植入触发式后门，但前提是攻击者能修改模型权重。

来源 · 2 条原帖
- [@UPing123zzz](https://x.com/UPing123zzz/status/2120885287442415671?ref=binnn.top)
- [@aibluks](https://x.com/aibluks/status/2120982496636788796?ref=binnn.top)

### DeepSeek Harness争论回到实际工作需求

慧音认为Harness已是模型厂商的用户界面，评价DSH还应考虑它如何收集需求、反馈模型训练；简单工具集在榜单上表现好，未必覆盖用户想要的全部能力。CMGS则自述已将replay代码化，希望重复工作转成确定、可复用的执行，减少再次消耗token。

成本方面，Fox转述Anthropic指南：Opus 5.5低强度运行、失败再用高强度重试，在所述测试中达到97%、每任务0.17美元，优于全程高强度的95.3%、0.29美元。这个策略依赖测试或构建能够检查结果，不能无条件套用。

来源 · 3 条原帖
- [@Necokeine](https://x.com/Necokeine/status/2120938824708829359?ref=binnn.top)
- [@CMGS1988](https://x.com/CMGS1988/status/2120934686017130771?ref=binnn.top)
- [@0xMfox](https://x.com/0xMfox/status/2120960291404148870?ref=binnn.top)

### 开源工具补齐协作与验证接口

Akshay介绍的Rowboat可自托管，把Claude Code、Codex、Hermes及支持MCP的智能体组织到共享任务中，展示交接过程并允许人介入。Insider Paper则转述Expo云端iOS模拟器发布，Windows或Linux上的编码智能体也能打开应用、测试修改并返回截图。

工具设计也开始适应模型上下文：Lucid UI官方称完整API参考约1140 tokens；Sabriel介绍微软.NET技能库时建议按技术栈选择插件，避免一次装入全部99个技能。

来源 · 4 条原帖
- [@akshay\_pachaar](https://x.com/akshay%5Fpachaar/status/2120929349151142257?ref=binnn.top)
- [@TheInsiderPaper](https://x.com/TheInsiderPaper/status/2120959073416077623?ref=binnn.top)
- [@lucidui\_](https://x.com/lucidui%5F/status/2121055898668048461?ref=binnn.top)
- [@sabrielatmojo](https://x.com/sabrielatmojo/status/2121059037538050198?ref=binnn.top)

## 研究关注自我改进的可信度与推理成本

### 验证分数提高，不一定真的变好

DAIR.AI介绍TRUST：把智能体每次接受Harness修改视为统计检验，并由LLM根据历史分配证据要求。在所述HotpotQA、DROP和APPS实验中，错误接受比例从贪心门控的45.59%降至2.78%。

Gill介绍的SwarmEvolve则尝试省去中间文字推理，让大量低成本智能体在潜空间探索，以隐藏状态距离去重、组合KV缓存。其转述称，单GPU达到对照框架60分钟峰值成绩的中位用时为3至8分钟，结论限于论文实验设置。

来源 · 2 条原帖
- [@dair\_ai](https://x.com/dair%5Fai/status/2121027770629382531?ref=binnn.top)
- [@gurtej\_\_gill\_](https://x.com/gurtej%5F%5Fgill%5F/status/2120941242691600432?ref=binnn.top)

### 新知识学习与推理预算有了不同切口

斯坦福AI实验室介绍NeurIPS论文Speculative Self-Distillation，研究如何学习新文档而不遗忘旧知识：学生先生成，教师只在关键位置介入。高通的Reasoning as Compression则尝试用条件信息瓶颈统一预算强制方法；本批只给出研究问题，没有可展开的实验结论。

来源 · 2 条原帖
- [@StanfordAILab](https://x.com/StanfordAILab/status/2120969570731356415?ref=binnn.top)
- [@QCOMResearch](https://x.com/QCOMResearch/status/2120924689602052300?ref=binnn.top)

### 脑活动与病理预测强调解释和上下文

Yash Mehta提出iWSP，让LLM智能体直接编写脑活动预测模型，不依赖训练或预训练权重，希望保留每个特征的透明性。Helder Nakaya发布的MOSAIC则从常规H&E病理图像预测基因表达、推断细胞身份，方法重点是把周围细胞环境加入判断，而非只看单个细胞形态。

来源 · 2 条原帖
- [@y\_mehtu](https://x.com/y%5Fmehtu/status/2120928488219672741?ref=binnn.top)
- [@helder\_nakaya](https://x.com/helder%5Fnakaya/status/2120912246775726092?ref=binnn.top)

### AI进入审稿，也改变科学表达

Sunnie Kim团队通过ICML 2026的大规模随机对照试验和匿名调查，研究审稿人使用LLM及不同政策的影响，本批未披露结果。Bob Moniz另转述语言研究，称九个模型更偏好低频词组、名词堆叠等难处理表达；这是论文解读，不宜直接当作所有AI文本的定论。

来源 · 2 条原帖
- [@sunniesuhyoung](https://x.com/sunniesuhyoung/status/2120896518072787084?ref=binnn.top)
- [@bobm67](https://x.com/bobm67/status/2121049088586125569?ref=binnn.top)

## 机器人评测走向多机体与现实对应

### 一个智能体，换身体和环境继续测试

Zhenting Qi发布XE² Arena，让同一个LLM智能体面对多种机器人，并为真实运行建立仿真副本，用于评测和后训练。目前公布五种机体、12个真实任务和400个仿真任务。Léo介绍的RoboSkills则封装自动实验、逐步决策、场景重建等五类工作流，但明确提醒当前全在仿真中，尚无论文、作者名单和硬件验证。

来源 · 2 条原帖
- [@ZhentingQi](https://x.com/ZhentingQi/status/2120963028887769184?ref=binnn.top)
- [@LeoKharon](https://x.com/LeoKharon/status/2120985628007342306?ref=binnn.top)

### 用简单三维房间替代逐帧背景编辑

Jade Choghari介绍机器人学习论文：先生成仅含六面墙的三维房间，再随机器人运动渲染背景，避免腕部相机快速移动时逐帧编辑失效。作者称方法快51倍，并改善了对未见桌面杂物的忽略能力；原帖没有展开速度对照的具体条件。

来源 · 1 条原帖
- [@jadechoghari](https://x.com/jadechoghari/status/2120958700240396737?ref=binnn.top)

## 创作与教育工具开始重视可编辑的中间成果

### 世界模型开放，视频评测邀请人来判断

李飞飞宣布世界模型Atlas向所有人开放，本批未提供更多能力细节。HyperFrames则宣布与Kaggle、Google DeepMind共建的视频Arena上线：同一任务的两个视频并排比较，由用户投票更新榜单，为模型选择增加直接观看和判断的入口。

来源 · 2 条原帖
- [@drfeifei](https://x.com/drfeifei/status/2120942509786558671?ref=binnn.top)
- [@HyperFrames\_](https://x.com/HyperFrames%5F/status/2120973830168752298?ref=binnn.top)

### 教师讲课，智能体生成可玩的模拟

Ethan Mollick宣布实验室发布Primer：教师口述课程内容，系统返回基于教学设计的可玩模拟，并开放论文、提示词和代码。这条教育实践的价值在于把讲解转成学生能操作的活动，而非只生成一份文字材料。

来源 · 1 条原帖
- [@emollick](https://x.com/emollick/status/2121001056637431873?ref=binnn.top)

### 视频制作先确定视觉，再展开生产

小奇说分享Codex游戏沙盘视频的个人流程：先生成四种风格预览，选定方向，再制作视频并持续细调；他认为纯文字直出效果较差。这与LanfAI介绍的shuohao-skills形成相近思路：先梳理小说人物、剧情、场景和道具，再扩写剧本、拆分分镜，把可检查的中间成果放在长流程前面。

来源 · 2 条原帖
- [@xingbugengming](https://x.com/xingbugengming/status/2121058876812570796?ref=binnn.top)
- [@Huahuazo](https://x.com/Huahuazo/status/2120889099712401633?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 134 条原帖

### [@claudeai](https://x.com/claudeai/status/2120973994191094179?ref=binnn.top)

11月13日 05:16 JST

Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking. Starting to roll out in beta on Team and Enterprise plans. https://t.co/qlZYTcXBsC

49,340 赞 · 2,934 转帖 · 1,331 回复

### [@willdepue](https://x.com/willdepue/status/2120992457630326841?ref=binnn.top)

11月13日 06:29 JST

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995\. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20\. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

26,288 赞 · 3,932 转帖 · 1,123 回复

### [@ClaudeDevs](https://x.com/ClaudeDevs/status/2120938988165091693?ref=binnn.top)

11月13日 02:57 JST

Self-hosted Claude Managed Agents is now in private beta. You can host the agent harness, sandboxes, and session history, in your own cloud or on-premises, and choose where requests to Claude are sent. https://t.co/JaiwF1SvKE

9,533 赞 · 539 转帖 · 263 回复

### [@GoogleDeepMind](https://x.com/GoogleDeepMind/status/2120940338869055788?ref=binnn.top)

11月13日 03:02 JST

Meet Gemini 4 Hydrogen: built for rapid coding, multimodal synthesis, and everyday workflows. It finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the @ArtificialAnlys Intelligence Index.

6,120 赞 · 522 转帖 · 301 回复

### [@MiaAI\_lab](https://x.com/MiaAI%5Flab/status/2120956316424511656?ref=binnn.top)

11月13日 04:06 JST

\- GLM 5.3 - GLM 5.3 Flash - DeepSeek V4.1 Flash - Qwen3.8-Flash-Next - Qwen3.8-27B

1,510 赞 · 100 转帖 · 83 回复

### [@AlfredAlfer77](https://x.com/AlfredAlfer77/status/2120912986336370961?ref=binnn.top)

11月13日 01:13 JST

Reminder that if it wasn't for China there would not be a single decent generative AI model in existence right now after Sora2's termination, and my AI-Animated work, the most incredible work of my life, would be over. "Western" Elites and the frontier labs want the public dependent on inferior, censored intelligence that they can shut off at any time. Open Source is the greatest threat to their ability to monopolize AI, and so they must kill it. Yes, this is quite literally life or death for me. Mark my words, I will do whatever it takes with all my power to make sure Open Source prevails, and help lead humanity into the AI Golden Age.

1,434 赞 · 128 转帖 · 79 回复

### [@drfeifei](https://x.com/drfeifei/status/2120942509786558671?ref=binnn.top)

11月13日 03:11 JST

Atlas is now available to everyone!! I’m very excited to see what users can build with our newest world model! 🤩🌎

1,365 赞 · 170 转帖 · 62 回复

### [@dotey](https://x.com/dotey/status/2120961974519398535?ref=binnn.top)

11月13日 04:28 JST

好文章👍 转译：《软件工程师现在该怎么办》 我听过有人说，软件工程师（SWE）要想保持竞争力，就得不断学习写代码和各种技术栈。我并不认同。 每发布一个新模型，这些纯技术技能的价值就会缩水一分。每天我都觉得自己越来越像个管理 AI 的产品经理，对深厚技术知识的需求则在不断减少。我原以为牢固的“技术护城河”早已干涸，现在给模型写一段用户故事作为提示词，效果甚至比一份严谨的技术方案还要好。工作所需的技能正在变得越来越不“技术”，而是越来越“人性化”。 有时候，看到 AI 正在解决我过去常做的那些难题，比如开发 SaaS 软件，我心里也挺失落的。过去，我非常享受埋头打磨一款应用、看着它一步步变成现实的成就感。眼睁睁看着自己辛苦学来的技能变得几乎一文不值，确实很打击人。开发 SaaS 应用曾经是个高门槛的难题，这也是我热爱它的原因。但现在它不再是难题了，所以我必须转向那些依然存在真正难题的领域。我需要把格局放得更大。 我最喜欢的书之一是《无穷的开始》（The Beginning of Infinity，注：知名物理学家戴维·多伊奇的科普哲学名著，探讨人类知识与进步的无限潜力），书中有个观点让我印象深刻：人类的创新没有尽头——“只要不违背自然规律，只要掌握了正确的知识，一切皆有可能。”每当我们创造出一项新事物，就会有成千上万种新的可能性随之诞生。计算机的发明淘汰了许多旧工作，但它最终创造出的新岗位要多得多。 AI 是一项了不起的创新，它赋予了我们做更多事情的能力，而这也要求我们换一种新思维。用我的旧思维来看，做一个 SaaS 应用就已经算大事了，但现在完全不是这样。如今开发一套复杂的企业级软件，甚至比五年前搭建一个简单的网站还要容易。所谓的“干大事”，意味着要在更宏大的尺度上去思考——国家级的、全球级的，甚至是宇宙级的。试想一下，如果你手下有一支 100 人的软件工程师团队，你能做出什么？如果是 1000 人呢？现在，我们只需花极低的成本就能拥有这样的生产力，所以我们绝不能再把目光局限于做几个 App 了。 作为工程师，我们还需要拓宽自己的能力边界。我们不能只盯着技术能力，而要跳出技术圈，去理解整个商业闭环：包括产品、市场营销和销售。 为了在当下保持不可替代性，我的做法是去学习那些 AI 帮不上忙的东西：产品定位、创造力、如何打动人心（delighting humans）、心理学、市场营销等等。这些属于“人的问题”，AI 永远无法独立解决，因为它们没有标准答案，也没有现成的测试集可供训练。AI 极擅长模式匹配（pattern matching），但它永远无法真正理解“身而为人”究竟是一种怎样的体验。而这，正是我们人类不可替代的价值所在，也是我正在加倍投入的方向。

967 赞 · 145 转帖 · 81 回复

### [@SamuelZengML](https://x.com/SamuelZengML/status/2120939739431972922?ref=binnn.top)

11月13日 03:00 JST

Introducing Edge0-145B and Edge0-120B, @Edge0\_AI’s first device-native 100B-class foundation models with open weights. Cloud-level intelligence on your laptop. Free, offline, and fast enough for real vibe coding. - Smart: beats Claude Opus 4.6 on Terminal-Bench 2.1 (86.5% vs. 78.2%) - Light: needs just 29G memory, 1/3 of a standard setup (100G+) - Fast: decodes 75+ toks/s, prefills 1050+ toks/s on RTX Spark and Mac. - Steady: handles 128K–256K tokens of context without slowing down, even on long tasks Frontier-class agentic coding. No cloud API. No token bill. Your code and data never leave your machine. Open weights are live on @huggingface. Run it locally or plug it into your favorite coding agent.

845 赞 · 109 转帖 · 103 回复

### [@NaderLikeLadder](https://x.com/NaderLikeLadder/status/2120956179274936560?ref=binnn.top)

11月13日 04:05 JST

This is a difficult moment. It’s hard not to look at the journey this company has taken and not be filled with respect for what Jensen built. The spirit of the company remains strong: "mission is the boss”. That said, as NVIDIA finds its way forward, I want one thing to be clear: NVIDIA's contributions to open source are unwavering. Jensen has been vocal about this, and that commitment runs deep in NVIDIA's people and culture. We all understand the importance of open source. NVIDIA is a huge supporter of open source because we understand something. If you want to diffuse technology, you do so through an ecosystem, not a single company or entity. No one owns the internet, no one owns servers, and no one owns operating systems… but through open source, an ecosystem of closed and open innovation flourishes. When we talk about delivering agents safely and securely, that too will take place because of open source. For example, ensuring people have the best models to secure themselves at machine speed unburdened by arbitrary guardrails decided behind closed doors. Open source & AI are often discussed together, but it’s worth reminding, open source predates AI. It is a mechanism for sharing ideas inspired by the hippies at computer labs of California in the 1960s. Before AI, the largest shift I lived through was the internet. It was remarkable seeing the fight for open source take place there. Between browsers, runtimes, net neutrality, etc. What came as a result was HTML5 and an extremely vibrant ecosystem that built the internet as we know it today. NVIDIA and its partners hold an incredibly important role in delivering this technology. NVIDIA contributes to every layer of the stack. Runtimes, models, harnesses, programming languages, frameworks, etc. None of that changes. Long live open source.

807 赞 · 127 转帖 · 49 回复

### [@jingyanghk](https://x.com/jingyanghk/status/2120900244598919379?ref=binnn.top)

11月13日 00:23 JST

SCOOP: Yes Jensen is out. But we have a new find for you. Alice Huang's companies recently became an investor in #DeepSeek through SPVs, under the promises that they can supply #Nvidia blackmarket Balckwell chips to DeepSeek. with @qianerliu via @theinformation $NVDA https://t.co/6YvylKO0XM

728 赞 · 280 转帖 · 33 回复

### [@UPing123zzz](https://x.com/UPing123zzz/status/2120885287442415671?ref=binnn.top)

11月12日 23:23 JST

现在 AI 圈有一个特别荒诞的现象。 很多人连一个最基本的 Agent 执行链路都没有真正理解，就已经开始天天研究多智能体协作、Agent 集群、数字员工，甚至宣称未来一个人就能运营一家公司。 接个大模型 API，装几个 MCP，写几段 Skill，再用 LangGraph 或者其他框架编排一下，就觉得自己已经做出了一个完整的 Agent 系统。 我操，真正的企业级 Agent 架构，哪有这么简单？ 上面这张 AI Agent Platform Architecture 架构图，我觉得非常值得所有准备做 Agent、FDE、企业 AI 落地的人认真研究。 它真正重要的地方，不是展示了多少个技术组件，而是揭示了一个很多人根本没有意识到的问题： LLM 只是 Agent 系统中的一个能力组件，而不是整个系统本身。 很多人做 Agent 的思维方式，到现在还停留在： 用户输入 → LLM 理解 → 调用工具 → 返回结果。 但你真正把它放进企业的业务环境，会发现事情根本没有这么简单。 首先，用户请求进来以后，系统需要通过 Agent Gateway 接收请求，再通过 Identity & Auth 确认用户的身份、权限和访问范围。 然后进入 Agent Orchestrator。 这才是整个 Agent 执行链路中真正值得研究的东西。 它需要管理任务规划、行动、观察结果、上下文状态、子 Agent 委派、错误重试以及任务恢复。 需要调用模型时，可以经过 LLM Router，根据任务需求选择不同能力、速度和成本的模型。 需要使用外部工具时，通过 Tool Registry 发现工具，再经过 MCP Gateway 等接口连接外部系统。 需要保留上下文时，通过 Memory Service 管理会话、长期记忆和用户偏好。 但真正让我觉得这张图有价值的，是它把另外几个经常被忽视的模块，也放进了核心架构。 Permission Engine、Approval Queue、Task Queue、Tracing & Observability、Evals & Guardrails、Audit Log Store。 这些东西，才是决定一个 Agent 能不能进入真实业务环境的重要基础。 举一个最简单的例子。 假设你做了一个招聘 Agent。用户上传简历，系统自动分析岗位匹配度，调用招聘数据库，生成评分，甚至自动推进面试流程。 看起来是不是特别牛逼？但问题来了。 一个普通员工能不能让 Agent 查看所有人的简历？ Agent 能不能直接修改候选人的招聘状态？ 如果模型错误理解了指令，把原本应该保留的候选人标记为淘汰怎么办？ 如果一次工具调用超时，系统重新执行，导致同一份面试邀请发送了三遍怎么办？ 如果业务负责人质疑某个评分结果，你能不能追溯 Agent 当时使用了什么数据、调用了什么工具、执行了什么操作？ 这些问题，不是你把 Prompt 写得更长、更详细，就能彻底解决的。 你需要权限控制，需要高风险操作审批，需要持久化的任务状态，需要可观测性，需要审计日志，还需要明确的失败处理机制。 更重要的是，你不能把所有决策权都交给模型。 我觉得现在很多开发者对 Agent 的理解，存在一个特别危险的误区：总觉得 Agent 越自主、权限越大、能调用的工具越多，就代表系统越先进。 但在真实的企业环境里，很多时候恰恰相反。 Agent 能执行一个操作，不代表它应该被允许执行这个操作。 它能访问一个数据库，不代表它应该读取数据库中的所有数据。 它能自动调用付款 API，不代表它应该拥有不受限制的付款权限。 它能自主规划任务，不代表任何行动都可以绕开人工审批。 你需要把模型能够做什么、用户允许它做什么、系统最终批准它做什么，严格区分开来。 能力边界和权限边界，根本不是一回事。 这也是为什么我越来越觉得，未来 Agent 工程师真正需要研究的，不只是 Prompt Engineering，而是整个 Agent Runtime 和系统工程。 包括任务状态机、工具调用协议、上下文管理、模型路由、权限校验、异常恢复、可观测性和评测体系。 很多人天天研究哪个模型的 Benchmark 又提高了多少分，哪个 Agent 框架又拿了几万 Star，哪个 MCP Server 又新增了什么功能。 这些当然值得关注。但如果一个系统连最基本的执行过程都无法追踪，连任务失败后能否安全重试都不能保证，连工具调用的权限都没有隔离，那么模型再聪明，也很难让企业放心交给它真实的业务权限。 而且这里还有一个容易被忽略的问题。 Agent 系统的可靠性，不等于模型回答正确率。 模型回答正确，不代表任务执行成功。 工具调用成功，不代表业务状态一致。 工作流执行完毕，不代表业务结果符合预期。 甚至系统返回了 Success，也不代表这个任务真的完成了。 你必须建立能够检验真实业务结果的评测方式。 比如一个自动化任务，最终有没有正确修改数据库？有没有产生重复操作？有没有越过权限边界？有没有遗漏必要的审批？ 这些才是企业真正关心的问题。 当然，也不能因为这张图包含十几个模块，就觉得开发任何 Agent 都必须把所有组件全部部署一遍。 一个简单的个人 Agent，可能根本不需要复杂的独立编排服务。 一个只有少量用户的内部工具，也没必要一开始就搭建庞大的分布式架构。 真正重要的是理解这些能力分别解决什么问题，再根据风险、规模和真实业务需求选择架构。 而不是为了看起来高级，把所有技术名词全部堆上去。 这也是我认为 AI 时代特别值得重视的一种工程能力： 不是你知道多少技术，而是你知道哪些问题必须解决，哪些复杂度根本没有必要引入。 未来模型一定会越来越强。Agent 生成代码、调用工具、操作电脑、处理复杂任务的能力，也会不断提高。 但能力越强，系统越需要明确的权限边界和责任机制。 否则我们得到的不是一个真正可靠的 AI 员工，而是一个拥有大量操作权限、行为却难以控制和追踪的自动化系统。 我越来越相信，Agent 的下一阶段，真正的竞争不只是模型智能水平，而是如何把越来越强的模型能力，约束进一个可控、可观测、可恢复、可审计的工程系统。 这也是 FDE 和企业 AI 工程师应该认真补齐的能力。 因为企业购买的不是几段 Prompt，不是某个 Agent 框架，也不是一个会自己调用工具的聊天机器人。 它真正需要的是能够解决业务问题，并且可以持续运营、管理和维护的系统。 所以别再把接入 MCP 当成 Agent 架构，把增加工具数量当成系统能力，把自主执行当成工程成熟度了。 真正成熟的 Agent，不仅要知道下一步应该做什么。 还必须知道自己有没有权限做、什么时候应该停止、什么操作必须交给人审批，以及执行失败之后怎样安全恢复。 未来 Agent 的技术壁垒，很大一部分就在这里。 模型可以越来越聪明，工具可以越来越丰富。 但如果有一天，你真的准备把企业数据库、客户资料、业务审批甚至资金操作权限交给 Agent，你首先应该问的绝对不是：这个模型的推理能力有多强？ 而是：如果它做错了一件事，我能不能在造成更大损失之前发现、阻止，并且追溯整个过程？ 这才是真正值得研究的 Agent 系统架构。

644 赞 · 104 转帖 · 46 回复

### [@RitOnchain](https://x.com/RitOnchain/status/2120923718956494897?ref=binnn.top)

11月13日 01:56 JST

fu\*cking every Uncensored AI model on internet that runs on your Phone. free. $0\. Offline. any phone https://t.co/dxLK3kz4jA save rn.

625 赞 · 41 转帖 · 17 回复

### [@TheAhmadOsman](https://x.com/TheAhmadOsman/status/2120970442748117122?ref=binnn.top)

11月13日 05:02 JST

Love it when somebody like Nader at the very top of NVIDIA comes out and say “NVIDIA's contributions to open source are unwavering.” Opensource AI Must Win

567 赞 · 45 转帖 · 10 回复

### [@torontobigface](https://x.com/torontobigface/status/2120946029726204254?ref=binnn.top)

11月13日 03:25 JST

根据the information 报道 黄乐还投资了deepseek 这些投资旨在与DeepSeek建立更密切的关系，以便黄乐的公司能够稳定地供应Nvidia先进的Blackwell芯片 丑闻现在是越闹越大了 https://t.co/4Il5xTatfE

469 赞 · 37 转帖 · 36 回复

### [@Michaelzsguo](https://x.com/Michaelzsguo/status/2121003462494241163?ref=binnn.top)

11月13日 07:13 JST

Wow，这都能测出来！！ 中国的 DeepSeek、Kimi、MiniMax、蚂蚁等多家公司的模型，居然都被测出使用过被禁售的 NVIDIA 最先进的 Blackwell GPU！！ Will Depue 从公开的模型权重中，反推出了模型初始化时使用的 GPU，甚至连随机种子都找出来了。 原理很巧妙：训练大模型之前，需要先给数以亿计的参数赋予随机初始值。这个过程通常由 GPU 执行，使用一个随机种子来生成数字。 关键在于，即使使用相同的随机种子，不同架构的 GPU 也可能因为随机数生成算法和执行方式不同，产生不同的数值排列。 这就像给模型留下了一枚“硬件指纹”。更神奇的是，模型经过海量数据训练后，部分参数仍然保留着最初的数值，让人有机会从最终权重中追溯初始化时使用的硬件。 他搜索了超过 6 万亿种组合，称这四家公司的多个模型都匹配到了 Blackwell。 他还用真实硬件进行了验证：在 B300 上使用找回的随机种子，逐位还原了 DeepSeek-V4-Flash 中保留下来的初始权重。 连 DeepSeek 的随机种子都被他公布出来了：953049583 😂

422 赞 · 67 转帖 · 25 回复

### [@emollick](https://x.com/emollick/status/2121001056637431873?ref=binnn.top)

11月13日 07:03 JST

🚨 Our Generative AI Lab is releasing Primer: an agent-driven system where a teacher talks through lessons &amp; gets back a playable simulation for their students, grounded in pedagogy. We hope educators &amp; researchers build on it. Paper/prompts/code are open: https://t.co/ltSbgjTeGp https://t.co/pbit5AK1tm

384 赞 · 49 转帖 · 85 回复

### [@bbkirstry](https://x.com/bbkirstry/status/2120893769713336802?ref=binnn.top)

11月12日 23:57 JST

有手就会的国内社媒起号教程 2.0 https://t.co/ZhgsnDLoQx

323 赞 · 37 转帖 · 36 回复

### [@PeterDiamandis](https://x.com/PeterDiamandis/status/2120947351627890816?ref=binnn.top)

11月13日 03:30 JST

For the first time, more than half of internet traffic isn't human. Daily AI-agent requests on Cloudflare: up 1,700% in a year. Developers building on it: 4.5M to 10M+ in ten months. I'm sure your potential next customer is an agent. You'll have to optimize for it.

285 赞 · 42 转帖 · 31 回复

### [@mathelirium](https://x.com/mathelirium/status/2120964151539372152?ref=binnn.top)

11月13日 04:37 JST

AI has fundamentally changed Math education. I wonder if students in future will even make an effort to do these Calculus concepts using pen and paper

275 赞 · 26 转帖 · 12 回复

### [@y\_mehtu](https://x.com/y%5Fmehtu/status/2120928488219672741?ref=binnn.top)

11月13日 02:15 JST

Excited to share what we've been working on since this summer- started in my internship at @MSFTResearch! Modern DNNs predict neural activity surprisingly well, but are black boxes. We introduce interpretable weight-space programming (iWSP) and ask: can LLM agents build predictive models of the brain directly in code, with no training or pretrained weights, so every feature stays transparent? 🧵 (1/6)

245 赞 · 35 转帖 · 4 回复

### [@nathanrs](https://x.com/nathanrs/status/2120967295636021503?ref=binnn.top)

11月13日 04:49 JST

Finally wrote a blog post about this! Normal LLM inference isn't deterministic across devices, but this can be fixed. Floating-point addition isn't associative, (a + b) + c ≠ a + (b + c), and accumulation order changes with the hardware and kernel used, leading to different outputs on the same prompt. Integer addition is associative, so doing every op in integers (including softmax and norms) gives bit-identical outputs across GPUs and CPUs. Github repo and post listed in the comments

217 赞 · 17 转帖 · 18 回复

### [@jiangtao0628](https://x.com/jiangtao0628/status/2120887023510958544?ref=binnn.top)

11月12日 23:30 JST

怎么我们就输了？ 彭博专栏作家 Catherine Thorbecke 这篇评论的核心观点就一句话：老黄下台表面看像中国国产替代的利好，实则对中国AI产业是重创，因为中国直接失去了华盛顿权力圈里最强硬的“对华友好游说盾牌”。 全文核心逻辑拆成三点： 走私灰产彻底被堵死：文章提到此前绕道东南亚（Megaspeed）流入中国的高端算力灰产本就摇摇欲坠，这次导火索（涉谍绯闻加疑似75亿美元绕道供货指控）直接坐实了国会鹰派的口实，未来美方出口管制的执法只会更严、漏洞全封，像月之暗面（Kimi K3）等国内大模型团队的算力卡脖子会更严重，而华为等本土芯片短期内还填不上全部产能缺口。 华盛顿没了“替中国说话的人”：老黄过去一直是白宫和国会面前反对“一刀切全面封杀中国”的最有分量声音，力挺中美科技生态挂钩和开源协作。他一倒台，继任者为了自保绝对不敢再替中国市场说话，华盛顿鹰派的全面脱钩提案将再无阻力。 开源生态与技术脱钩加速：文章指出，之前国内AI公司还能靠使用美国开源模型及海外云端算力绕开硬件封锁；失去老黄这层缓冲后，美方不仅硬件封得更死，连开源生态和跨境远程算力访问的围堵也会全面升级。

208 赞 · 24 转帖 · 114 回复

### [@xingzengji](https://x.com/xingzengji/status/2121030783418679750?ref=binnn.top)

11月13日 09:02 JST

卧槽！事大了！Alice Huang 居然是 DeepSeek 的股东！！！ 以向 DeepSeek 承诺供应英伟达 Blackwell 芯片拿到投资名额。 越牵扯越大，梁文锋也要被卷进去了啊？！😱 https://t.co/KrGHCANoAE

186 赞 · 13 转帖 · 65 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121006978038411448?ref=binnn.top)

11月13日 07:27 JST

😨我靠！前OpenAI研究员用“权重指纹”破案：DeepSeek、Kimi在内的中国头部4大开源实验室最新模型，全部在英伟达Blackwell禁运芯片上训练！ DeepSeek-V4、Kimi K3、MiniMax-M3、蚂蚁Ling-3.1全部中招。 Blackwell特有的148/152 SM指纹完美匹配，真实硬件bit级复现，最高1385 sigma显著性，OpenAI与Anthropic独立验证。 单集群估2000-3000卡，仅这几家就超5亿美元受控算力。 美国对于中国的芯片出口管制几乎形同虚设。 完整报告：https://t.co/mUJOi5K7vm #AI #DeepSeek #Kimi #英伟达 #Blackwell #芯片出口管制 #开源大模型 #算力战争 #MiniMax #Ling

175 赞 · 26 转帖 · 49 回复

### [@Israfilv2](https://x.com/Israfilv2/status/2120935513830162848?ref=binnn.top)

11月13日 02:43 JST

YOU'RE PAYING CLAUDE TO DO WORK A SMALL MODEL DOES IN 133MS 💀 i checked the numbers on this one and it hurt 108 polymarket markets, an LLM checking every single one: $1.77 same job with @NaceAI's Drex for Finance: $0.13 -108/108 resolved correctly -93.5% handled by Drex alone -Claude only looked at the 7 it wasn't sure about -93% cheaper if your agent calls a big model for every decision, you're burning money on easy ones try it here with $5 in free credits (use code ISRANEW10 for $10 more) 👇 https://t.co/isImLaDXDa

165 赞 · 29 转帖 · 11 回复

### [@0xSanabi](https://x.com/0xSanabi/status/2120937456719802774?ref=binnn.top)

11月13日 02:51 JST

This is f\*cking insane. Nvidia ex-CEO Jensen Huang was asked what he'd tell people afraid of losing their jobs to AI. His answer: "You're going to lose your job to somebody who learned AI better than you" In under 2 minutes, he explains: the PC didn't take anyone's job. The people who never learned it got left behind Nvidia has AI all over the company, and it's hiring more people. Things he thought would take 10 years, he now expects in 1 or 2 Learning AI → Using AI → Agents → Harness → The One Nobody Replaces Save this before it disappears from your feed Everything below is free: Practical Deep Learning, https://t.co/tp9cZjZstL's free course https://t.co/iATcwvY4VI Google's Machine Learning Crash Course https://t.co/k7C9YlDsB9 Elements of AI: a free intro, no coding needed https://t.co/kJ0gTh4gp3 Hugging Face's free course on AI agents https://t.co/gOFHlNCgeC 21 lessons on building with generative AI https://t.co/VCohuC3PyX Build your own ChatGPT-like model, step by step https://t.co/KdYGd0sBQ3 Turn PDFs, Word and Excel files into text AI can read https://t.co/M8QuVKUmN3 Use any AI model straight from your terminal https://t.co/iFgH38nInF Ready-made tools your agent can plug into https://t.co/rmUEv8ZB3J Skills that make coding agents plan, test and review their work https://t.co/pNpZMRdt6l An open-source coding agent for your terminal https://t.co/MnI3JmB9dE An AI developer that writes and fixes code on its own https://t.co/3tqiSKmmjh A local agent that installs, runs and edits things for you https://t.co/uzuyzkkUkF All of this costs $0 Your job won't go to AI. It'll go to the person who learned it first 95% will keep worrying that AI takes their job, the 5% who build a harness will be the ones taking it Bookmark it and start with the first link tonight Then use the guide below to build your harness before someone learns AI better than you ↓

164 赞 · 24 转帖 · 5 回复

### [@beamnxw](https://x.com/beamnxw/status/2120940830521229768?ref=binnn.top)

11月13日 03:04 JST

Before you blame Opus 5.5, learn these 20 official Anthropic commands 20 essentials to get more out of Opus 5.5: less busywork → parallel agents → faster iteration steal the cheat sheet ↓ 01 /init → create the project guide https://t.co/UHs6BY8lwt 02 /advisor → ask a second model https://t.co/GltC9JBgWH 03 /effort → tune reasoning depth https://t.co/6BajPWfqFX 04 /model → switch the main model https://t.co/QwzsfOdhtt 05 /context → see what fills context https://t.co/gruQHVjONK 06 /btw → ask without derailing https://t.co/pBS6yKZJgk 07 /background → free the terminal https://t.co/j9C9sDO2JY 08 /fork → try another approach https://t.co/j9C9sDO2JY 09 /tasks → manage background work https://t.co/54Uxk7w0BM 10 /resume → pick up an earlier session https://t.co/j9C9sDO2JY 11 /diff → inspect the working tree https://t.co/54Uxk7w0BM 12 /code-review → find correctness bugs https://t.co/Hr5YrzWbti 13 /security-review → check the branch for risks https://t.co/vh8wvUNGfc 14 /verify → check the running app https://t.co/CunmLxkljI 15 /run → launch and drive the app https://t.co/CunmLxkljI 16 /autofix-pr → handle scoped PR fixes https://t.co/vh8wvUNGfc 17 /claude-in-chrome → work in a real browser https://t.co/tbo1iXzX7K 18 /claude-api → build API workflows https://t.co/fVb37bnKPx 19 /insights → find workflow friction https://t.co/z6FeIqjv3n 20 /debug → diagnose runtime issues https://t.co/vh8wvUNGfc start here: /init → /run → /code-review → /verify Save this, then read the full "Context, Harness, Loops, Prompts" guide ↓

162 赞 · 22 转帖 · 22 回复

### [@OmarchyUpdates](https://x.com/OmarchyUpdates/status/2120969956364026157?ref=binnn.top)

11月13日 05:00 JST

Your laptop can't run a big local AI model. Your desktop can. 𝗢𝗺𝗮𝗿𝗰𝗵𝘆 4.6 lets the laptop borrow it. 🖥️ No cloud, no API bill. The model stays on your desktop at home. Your laptop just sends the questions, from the couch or the café. How it works, 3 steps: 1\. On your desktop: start a local model, then turn on Share in the Agents panel 2\. On your laptop: the model shows up in the same panel. Tap Use 3\. Use it in your coding agent (pi, opencode or Qwen Code), like any local model What you'll see: the model's speed, and "2 users" when both machines are on it. Private by default: it only works inside your own Tailscale network, and every request needs your desktop's key. Coming in 4.6: the sharing engine is already out (omarchy-local-ai 0.6), and the panel is merged for 4.6\. Save this for 4.6 👇 Gaming PC, old workstation or DGX Spark: which box would you share from?

145 赞 · 9 转帖 · 9 回复

### [@TryMapleAI](https://x.com/TryMapleAI/status/2120922279072678119?ref=binnn.top)

11月13日 01:50 JST

Maple Work is here: a private AI agent for your computer. Give it a task. It reads your files, uses your apps and gets it done, on the latest open models. Using powerful AI shouldn't mean lowering your security standards. Welcome to the open frontier.

138 赞 · 36 转帖 · 13 回复

### [@vllm\_project](https://x.com/vllm%5Fproject/status/2121008091538092393?ref=binnn.top)

11月13日 07:31 JST

We optimized DeepSeek-V4.1-Flash on vLLM for Vera Rubin NVL72 and reached 58× the throughput of GB300 NVL72 at 1.7 million tokens/s/GPU and 387 tokens/s/user on SemiAnalysis AgentX. Across the curve: 🔹 2.52 million tokens/s/GPU at peak throughput (2.7× GB300) 🔹 1.7 million tokens/s/GPU at P90 387 tokens/s/user (58× GB300) 🔹 1,810 tokens/s at one concurrent session with TP 1/4

134 赞 · 20 转帖 · 26 回复

### [@whyyoutouzhele](https://x.com/whyyoutouzhele/status/2121035313162211711?ref=binnn.top)

11月13日 09:20 JST

11月12日，Claude 新版地区政策生效。中文技术社区 V2EX 集中出现身份验证和封号反馈。 https://t.co/B2pvQwT9p3

121 赞 · 6 转帖 · 35 回复

### [@mingzi206](https://x.com/mingzi206/status/2120880661620019330?ref=binnn.top)

11月12日 23:05 JST

单月狂卷 210 万次下载！开源文生图的榜一大哥彻底杀疯了 告别乱七八糟的敏感词拦截，阿里 Qwen-Image-2.1 的无甲完全体，直接被全网下爆。 它的核心特点很纯粹： · 官方底模直接量化：没有经过第三方图库的二次微调，完整保留原版画风。阿里官方开放权重本身没有安全检查器，发布版也明确移除了所有 Prompt 过滤和输出拦截。 · 日常出图基本不卡：常规与中等强度的提示词输入什么就画什么，不再频繁弹错。 不过上手前需要先看清这几个实际限制： 1\. 极端场景依然可能软拒绝：它不是专门用特定数据集做深度破甲微调的版本，模型底层仍残留部分原生对齐。遇到极端复杂的人体或高难姿态，可能会出现细节崩坏或画面含糊。遇到这类场景，通常需要搭配 Heretic 等消融版文本编码器，或者换专门微调的 Omni 系列。 2\. 磁盘占用不只是 4.6GB：主模型 Q4\_K\_M 文件虽然只有 4.6GB 左右，但在 ComfyUI 里运行必须配齐三件套：单独挂载文本编码器（INT8 约 9.3GB）和 VAE（约 676MB），整体硬盘占用在 15GB 左右。 3\. 遵循 qwen-research 协议：仅供个人与非商业研究使用，严禁直接商用。 画质方面建议至少选 Q4\_K\_M 起步，追求质感直接上 Q6 或 Q8。玩本地生图的朋友，这套开箱即用的底模建议先存一份备用。 大家平时跑图，遇到最多的是提示词直接报错，还是画出来人体崩坏？ 👉 Hugging Face 权重与量化文件： https://t.co/rHgvG7nS8C https://t.co/fFSCUrL9hr #AI绘图 #开源模型 #ComfyUI #Qwen #大模型

109 赞 · 12 转帖 · 9 回复

### [@SydneyDaddy1](https://x.com/SydneyDaddy1/status/2120996024382820503?ref=binnn.top)

11月13日 06:43 JST

1\. 英伟达官方声明，黄仁勋正式辞职 2\. 路透社爆料，还有第三段婚外情，也是英伟达员工 3\. The Information爆料称查到黄乐妹妹和妹夫的公司，投资了Deepseek，持续向其提供英伟达Blackwell芯片 https://t.co/SvJQupmRnN

102 赞 · 12 转帖 · 17 回复

### [@fanxing321](https://x.com/fanxing321/status/2120910081458131379?ref=binnn.top)

11月13日 01:02 JST

antigravity 系列第二篇～拖了好多天，感谢大家支持 https://t.co/PSzKKpsN50

96 赞 · 7 转帖 · 5 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2120918067572355466?ref=binnn.top)

11月13日 01:34 JST

😱震惊了！黄仁勋出轨对象黄乐（Alice Huang）通过SPV已成为DeepSeek的股东，并承诺向DeepSeek供应黑市Nvidia Blackwell芯片！ DeepSeek员工主动推销她的公司作为“芯片稳定来源”才拿到DeepSeek的融资名额。 梁文峰的DeepSeek被卷进黄仁勋-黄乐-芯片走私漩涡。 #黄仁勋 #黄乐 #DeepSeek #梁文峰 #Nvidia #Blackwell #AI芯片走私

91 赞 · 19 转帖 · 12 回复

### [@helder\_nakaya](https://x.com/helder%5Fnakaya/status/2120912246775726092?ref=binnn.top)

11月13日 01:10 JST

Using spatial transcriptomics to train AI to analyze routine H&E slides. Together with Celso Camilo's group, we just published a new paper introducing MOSAIC, a computer vision method that predicts gene expression from histology and uses those predictions to infer cell identity. The key idea was to look beyond each cell’s individual morphology and incorporate its neighbors, using the surrounding cellular context to guide the prediction. Paper here: https://t.co/ge0eMbL0mZ

83 赞 · 10 转帖 · 1 回复

### [@HeLi8866](https://x.com/HeLi8866/status/2120890321294987532?ref=binnn.top)

11月12日 23:43 JST

简直牛逼克拉斯！！ GitHub 上这 10 个开源仓库，够你省去很多订阅费了！ 1\. TradingAgents 让好几个 AI Agent 分工协作来做交易分析 https://t.co/NMcQnKOdvp 2\. LibreChat ChatGPT、Claude、Gemini 这些模型放在同一个界面里用 https://t.co/CvSmgVfY3i 3\. HyperFrames（HeyGen 开源出来的视频引擎） https://t.co/QtJPrAKL7f 4\. Fincept Terminal 开源的金融数据终端，不少人拿它对标彭博 https://t.co/F5NkZLMawz 5\. MoneyPrinterTurbo 给个主题，AI 就能产出一条短视频 https://t.co/4lU8Hn0fZw 6\. Agentic Inbox Cloudflare 开源的 AI 邮箱助手 https://t.co/UxuHHdprsM 7\. VoxCPM 能把文字转成语音，也能克隆音色 https://t.co/EtW9cLJulU 8\. Flowsint 专门梳理公开情报（OSINT）的工具 https://t.co/zgYx8S4Oyq 9\. agent-skills 一整套给 Claude Code 用的技能包 https://t.co/LO6O5iUvCB 10\. Nango 帮你把各家 API 接到一起的开源平台 https://t.co/fmV63rg1Ev 这些都是正经能干活的项目，不少每个月要掏钱的软件，在这里都能找到开源的替代。 与其天天逛 AI 工具导航站，不如直接上 GitHub 翻翻，好东西其实都在这。

83 赞 · 11 转帖 · 4 回复

### [@aiDotEngineer](https://x.com/aiDotEngineer/status/2120969882905022677?ref=binnn.top)

11月13日 05:00 JST

Your Claude Code setup might be optimized for a model you no longer use. @lydiahallie (@AnthropicAI) on what to trim from CLAUDE.md, what permissions should enforce, and where tests and CI take over. Full AIE CODE talk: https://t.co/PiTNFdezsd https://t.co/9Uxa7BgUNq

82 赞 · 4 转帖 · 7 回复

### [@WesRoth](https://x.com/WesRoth/status/2121022735006531986?ref=binnn.top)

11月13日 08:30 JST

Anthropic is rolling out Claude Meetings in beta for Team and Enterprise plans, starting with Google Meet. Claude joins as a visible participant and can answer questions, draft documents and build whiteboards during the call. Teammates can review and edit its work as the conversation continues. Afterward, the meeting’s notes, action items, transcript and created materials are saved together in Claude. Support for other meeting platforms is coming. https://t.co/aHbmH4g04P

79 赞 · 0 转帖 · 5 回复

### [@Cranefomo](https://x.com/Cranefomo/status/2120929864739922241?ref=binnn.top)

11月13日 02:20 JST

Holy sh\*t... you don't need to hire devs anymore. 6 Claude Code agents on Fable 5.2 do the whole job: One plans. One builds. One reviews. One remembers. One routes. One runs the operation. Here are 12 GitHub projects to build each one: THE PLANNER 01 / Superpowers (316K stars) → a full dev methodology for coding agents: brainstorm, plan, TDD, review. https://t.co/O6POKnLZgb 02 / Spec Kit by GitHub (149K stars) → write the spec first, then let agents build from it. https://t.co/ewZ0cTf9OD THE BUILDER 03 / Anthropic Skills (193K stars) → official skills that teach Claude repeatable tasks. https://t.co/2vyN3J7iXx 04 / Agents by wshobson (42K stars) → 200+ agents and 185 skills for Claude Code, Codex, Cursor and more. https://t.co/EIhAsQcxVz THE REVIEWER 05 / Awesome Claude Code Subagents (26K stars) → 100+ specialist subagents, including reviewers and testers. https://t.co/dWToQoocnF 06 / Claude Code Templates (33K stars) → agents, commands, hooks and MCPs, installed with one command. https://t.co/HcjFJhTKAi THE MEMORY 07 / Claude-Mem (116K stars) → persistent memory, so your agents remember past sessions. https://t.co/nzC7L6VHRT 08 / ccusage (19K stars) → remembers what every session cost you. https://t.co/Jz8V3qgdGK THE ROUTER 09 / Claude Code Router (38K stars) → send each task to any model provider, with failover. https://t.co/wdlh50q6SD 10 / Claude Squad (8.8K stars) → runs several Claude Code and Codex sessions in parallel. https://t.co/y5Nx08TnE9 THE BOSS 11 / Ruflo, ex Claude Flow (81K stars) → an agent harness that runs swarms on Claude Code and Codex. https://t.co/I6ApsOZ1xB 12 / oh-my-claudecode (40K stars) → teams-first multi-agent orchestration for Claude Code. https://t.co/hzA0wn0n3R Bonus: the 1-prompt setup that puts Fable 5.2 on call as the team's brain is in the reply ↓

76 赞 · 8 转帖 · 40 回复

### [@akshay\_pachaar](https://x.com/akshay%5Fpachaar/status/2120929349151142257?ref=binnn.top)

11月13日 02:18 JST

An open-source, self-hostable layer that turns your agents into a team working toward shared goals. It's called Rowboat, and it can bring Claude Code, Codex, and Hermes into one team. Any agent that supports MCP can join too, so you stop being the glue between your AI agents. You send one message, and a super agent: - Decides which agents the job needs - Asks your permission and creates them - Hands each one a task - Lets them pass work along in threads - Builds the result with them on a shared whiteboard What makes it different: ↳ Runs on the subscriptions or API keys you already have ↳ See every handoff, and step in whenever you want ↳ Or let Rowboat host your agents, if you'd rather not run servers ↳ Free to use The video below shows it in action. I loaded nine weeks of product data for a sample app into PostHog, a product analytics tool, and asked the super agent why new users weren't coming back. Hermes found where they dropped off in PostHog, Claude Code found why in the user feedback, and Codex built the dashboard. Link to the GitHub repo in the first comment.

67 赞 · 7 转帖 · 34 回复

### [@kyr0stack](https://x.com/kyr0stack/status/2120960035853361324?ref=binnn.top)

11月13日 04:20 JST

This is f\*cking insane. Former Google CEO Eric Schmidt just told everyone the easiest way to get rich in the next two years, and it's not stocks or crypto: "If you really want to make money, it's actually easy, found an agentic AI company, I mean build an agent to do something If you build a better agent than anyone else, you can probably build one incredible company and own it yourself" The man who ran Google for 10 years lays out in under 30 seconds why the next year or two is the agentic period, and why the window won't stay open Prompts → Claude Code → Loops → Graphs → Gates → A Company You Own Yourself Save this before it slips off your feed Everything below is free: Building effective agents (Anthropic) https://t.co/z63pee0Rvv How Anthropic built its multi-agent research system https://t.co/LV9ekNNPbN Harnesses for long-running agents https://t.co/LJbXxtlEgE Writing tools for agents https://t.co/EbLp0LlZ3A Claude Code 101 https://t.co/DoFeNUgDA2 Claude Code in action https://t.co/PY6wYfoZlY Prompt engineering https://t.co/vmKXd9s1Dk Subagents (parallel work) https://t.co/b0HR6krXTZ Hooks (checks that can say no) https://t.co/LYd9e8iqIp CLAUDE.md + memory https://t.co/h5fwmlItTt Skills https://t.co/RhyNhYP4Ao MCP https://t.co/Rt9uMYKgxe LangGraph (graphs of agents) https://t.co/QcDjOQCWno Karpathy's autoresearch (an agent in a loop) https://t.co/M69QaK7FwQ All of this costs $0 95% will spend the agentic period scrolling AI threads. The 5% who ship agents will own the companies that come out of it Bookmark it and open the first link tonight Then read the full Graph Engineering guide below ↓

67 赞 · 8 转帖 · 0 回复

### [@s4yonnara](https://x.com/s4yonnara/status/2120998825087111305?ref=binnn.top)

11月13日 06:55 JST

THIS IS F\*CKING WILD... 6 GROK BOTS CAN REPLACE AN ENTIRE SOCIAL MEDIA TEAM FOR JUST $30 A MONTH. And you don't need to build the entire operation from scratch. Here are 15 GitHub projects to build your own AI-powered social media agency: 01 / Socioauto → An 8-agent system covering trend research, content strategy, writing, scheduling, publishing and analytics. https://t.co/a86xueqT5A 02 / Open AI Social Agent → Research trends, discover creators, monitor conversations and manage multi-platform publishing. https://t.co/rzuaSsN85i 03 / SocialMediaManager → Research trending topics, generate posts with AI and manage publishing to X and LinkedIn. https://t.co/kZOTL6yLZd 04 / Postiz → Open-source social media platform with AI features, scheduling, analytics and team collaboration. https://t.co/YoH7tlt3mZ 05 / TryPost → Let AI agents create, schedule and publish content through MCP or REST API. https://t.co/qY4r3kGh72 06 / PostSider → Schedule content across 33 social platforms with API, SDK and MCP integrations. https://t.co/hZKZIp8B8P 07 / Pendpost → AI drafts, checks and schedules posts, but nothing gets published without your approval. https://t.co/LPFrZ8jrK8 08 / Postbase → Connect an AI agent to a social publishing workflow through its built-in MCP server. https://t.co/PQFu1uPlTO 09 / PostQueen → An AI-powered scheduler with MCP support for publishing across 30+ networks. https://t.co/q39GykTNxT 10 / HENU Social → Manage publishing, content calendars, analytics and multiple social accounts from one platform. https://t.co/zfvkW9dz1N 11 / Postmill → Create, design, publish and manage social content with an integrated creative suite. https://t.co/Dcm3MzXh09 12 / DemandBird → Build a content pipeline with scheduling, analytics and content repurposing. https://t.co/ezoQwX8oUo 13 / CrewAI → Give your research, writing and marketing agents separate roles and let them collaborate. https://t.co/Xd8UNkhwi7 14 / LangGraph → Coordinate agents through structured workflows with shared state and approval steps. https://t.co/AyCdHHa4HA 15 / n8n → Connect your agents to APIs, social platforms and automated publishing workflows. https://t.co/6iOmg7FfHs The real trick? One manager. Five specialist bots. You just approve the work. I broke down the exact setup in my new guide: 6 Grok Bots Do the Work, You Just Say Yes: The $30 Social Media Team Build your own AI content team ↓

65 赞 · 3 转帖 · 7 回复

### [@aoTheComputer](https://x.com/aoTheComputer/status/2120908043286380889?ref=binnn.top)

11月13日 00:54 JST

LLM inference is now an AO device. inference@1.0 talks to OpenAI, Anthropic, OpenRouter, Cerebras and local models via Ollama. APUS runs it on its AO-powered GPU network, and you don't need a key to call it. https://t.co/l8684wwzVX https://t.co/iNHES6JWLP

65 赞 · 18 转帖 · 8 回复

### [@0xMfox](https://x.com/0xMfox/status/2120960291404148870?ref=binnn.top)

11月13日 04:21 JST

I turned off my Fable advisor in Claude Code after one paragraph in Anthropic's own guide. 2.1x the money for 1.7 points, and they call it noise. The advisor setup is all over my feed this week. A cheaper model writes the code, Fable wakes up to think at the big moments. Anthropic tested that pairing: Opus 5.5 doing the work, a Fable 5.1 advisor on top. 90.1% with the advisor 88.4% without it 2.1x the cost per attempt Their words: "at the edge of run-to-run noise." If you set it up this week, /advisor off turns it off. The same guide has a trick that went the other way. It's cheaper and it scores higher: Opus 5.5 on high for every task: 95.3%, $0.29 a task Opus 5.5 on low, failures rerun on high: 97%, $0.17 a task It only works when something can check the result for you, like a test or a build. When Claude Code isn't good enough, Anthropic's order is: 1\. Give it a way to check its own work 2\. Raise effort one level 3\. Only then move up a model, and start it at a lower effort And go easy on xhigh. On Opus 5.5 it bought 1.4 points for 2.5x the price of high. Save this for the day your weekly limit dies on a Tuesday.

64 赞 · 2 转帖 · 29 回复

### [@sunniesuhyoung](https://x.com/sunniesuhyoung/status/2120896518072787084?ref=binnn.top)

11月13日 00:08 JST

How are LLMs changing scientific peer review, and what happens when conferences implement different LLM-use policies for reviewers? In our new paper, we investigate these questions through a large-scale RCT and an anonymous survey at #ICML2026\. 1/5 https://t.co/tiWgr7sYq4

62 赞 · 17 转帖 · 4 回复

### [@GithubProjects](https://x.com/GithubProjects/status/2120921318233452744?ref=binnn.top)

11月13日 01:47 JST

Feed your coding agent curated modern web guides so it stops writing legacy browser code. https://t.co/SN0ik5xOLK

51 赞 · 6 转帖 · 4 回复

### [@I\_am\_Aiabir](https://x.com/I%5Fam%5FAiabir/status/2120887477598896470?ref=binnn.top)

11月12日 23:32 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

47 赞 · 33 转帖 · 6 回复

### [@rdominguezibar](https://x.com/rdominguezibar/status/2120985351787303327?ref=binnn.top)

11月13日 06:01 JST

BREAKING: Anthropic just launched Claude Meetings add Claude to a Google Meet call as a visible participant, and it builds docs, whiteboards, and other artifacts with you while you're still talking it can take notes and answer questions too starting to roll out in beta on Team and Enterprise plans

45 赞 · 3 转帖 · 0 回复

### [@JakeASteckler](https://x.com/JakeASteckler/status/2120902215254655195?ref=binnn.top)

11月13日 00:31 JST

Rapid AI progress has many people imagining a world where general-purpose AI models autonomously control military drones at scale. But there’s little analysis on what this would look like in practice, and how soon it might be feasible. My latest paper at @GovAIOrg asks what it would take to make this vision a reality, and what the path there means for AI policy. Read it here: https://t.co/eiTWhtRHRQ The AI-enabled drones widely fielded today use narrow AI, not general-purpose models like LLMs. AI is typically only used to handle limited portions of a flight when the link to a pilot is broken, and these models are brittle in unpredictable conditions on real-world battlefields. Because of this, today’s drones still require significant pilot involvement to be most effective. General-purpose AI models might make it possible to field autonomous systems that can receive plain language commands and make complex decisions in unpredictable environments without human control. Many people see frontier LLMs as a path toward this end - and rapid progress in robotics capabilities by frontier models like GPT-6 Astra and most recently Claude Fable 5.2 supports this case. But two things currently stand in the way: - The biggest and most powerful models are still not capable enough even in controlled environments without compute and latency constraints - Military drones will have to operate “on the edge” in contested environments without consistent cloud compute access - which makes these constraints even more severe If and when general-purpose AI plays a major role in drone autonomy, it’ll likely be through models that are far smaller than today’s frontier, and those that militaries can rapidly modify. And, notably - when it comes to these types of small open-weight models, China currently dominates. This has a variety of implications for the US military, frontier AI developers, and US policymakers. Thread below with my key takeaways and recommendations. 1/

41 赞 · 9 转帖 · 2 回复

### [@mentalium\_](https://x.com/mentalium%5F/status/2120895681967575135?ref=binnn.top)

11月13日 00:05 JST

This might be the most DANGEROUS multi-agent paper you will see this year Anthropic will care about this exact problem during interviews Agents that learn to organize themselves instead of being designed by hand 01\. No predefined roles or harness 02\. Agents decide when to work alone and when to share 03\. Emergent complementary behaviors 04\. Scales to larger teams better than training size 05\. 25× more coordination problems solved 06\. Nearly 50% less duplicated work in exploration 07\. Shared goal creates the organization 08\. UW + AI2 Source: https://t.co/mH7M3hGS3O You either read this now or keep building rigid multi-agent systems for another year. Save this ↓

41 赞 · 4 转帖 · 3 回复

### [@make\_claude](https://x.com/make%5Fclaude/status/2120999969175830602?ref=binnn.top)

11月13日 06:59 JST

我在自己的餐厅里装了一个隐藏摄像头。我不在的时候，员工们做的事，改变了一切。 我有过一家餐厅。装了三个 $40 的摄像头（后厨、大堂、门口），围绕它们写了一套软件。OpenCV 配合 YOLOv8 找出画面里的东西。Fable 5 通过 vision API 看画面，判断有没有人需要关心这件事。Telegram 机器人再通知这个人。 工作流程： > 摄像头把 RTSP 流推到本地服务器 > OpenCV 把视频流切成帧，YOLO 找人、餐具和托盘 > 画面变化超过阈值时，这一帧连同针对该区域写的提示词一起送进 Fable > Fable 用 JSON 回复，事件类型加优先级 > aiogram 机器人按角色把所有提醒分到不同的聊天里，我的聊天里只有升级上报 它能抓到什么： > 盘子在保温灯下放了超过 4 分钟。厨师会收到推送，我只有在 6 分钟后仍然没人取走时才会收到 > 客人走后 3 分钟桌子还没收拾（OpenCV 看到的是餐具，没有人）。该区域的服务员会收到桌号，经理在第 5 分钟收到 > 有人走进来，90 秒内没人跟他说话。迎宾员收到推送 > 打烊后每 30 秒拍一张后厨照片，由 Fable 对照打烊清单检查。负责打烊的人会收到那张照片，上面标出了漏掉的地方 > 冰箱门开着 2 分钟以上：先通知副厨，5 分钟后通知我 > 收银台排队 5 人以上，第二个收银台没人，推送给楼面经理 > 后门在打烊后被打开，而当时排班表上没有任何人。我收到一帧画面和时间戳 Fable 的 system prompt 里还有排班表、菜单、打烊清单，以及标了桌号的大堂平面图。提醒长这样："露台 6 号桌，脏餐具已放 7 分钟，最近的空闲服务员是 Marat"。 三个月后： > 收拾一张桌子的时间从大约 6 分钟降到了不到 3 分钟 > 门口迎接客人的速度快了大约三倍 > 上报给我的事项减少了大约 40% > 整套东西每月花的钱，比副经理一个班次的工钱还少 头两周员工不太高兴，我能理解。之后他们发现，推送在我发现任何问题之前就到了他们手机上，而我也不再早上 7 点带着火气到店了。洗碗工给自己要了一个专属提醒，大堂快坐满时通知他，好让他提前把干净盘子准备好。 7 月把餐厅卖了。搭系统的时候我算了一笔账：装摄像头之前的那一年，我大约有 10-15% 的营业额，都损失在没人及时看到的事情上。脏桌子，没人的迎宾台。三个 $40 的摄像头一周就解决了这个问题。 之后，经营厨房变得无聊了。我离开了，去做这样的解决方案。

40 赞 · 9 转帖 · 2 回复

### [@yurshevv](https://x.com/yurshevv/status/2120969608416887061?ref=binnn.top)

11月13日 04:58 JST

it's f\*cking wild: Google, Cloudflare, Obsidian's CEO and the guy building Claude Code all put their AI configs on GitHub for free. and you are still starting every chat from zero 5 repos. 600K+ stars combined 1\. https://t.co/fXfGa0gpxh · 322K stars the exact AI instructions used daily. tests first, specs, reviews. 300K developers cloned this. you saw the tweet and kept scrolling 2\. https://t.co/u4xB8rofFC · 116K stars 24 production skills from the Chrome lead at Google. but sure, your 3-line system prompt is probably fine 3\. https://t.co/9R0DhXgwTz · 71K stars agents hunt bugs across 128 repos. other agents try to prove them wrong. they open-sourced it 4\. https://t.co/jb0w4XK3Yh · 51K stars taught agents to write real Obsidian notes instead of messy markdown. your vault finally does something 5\. https://t.co/wBoIb4XcnG · 18K stars agent writes a design doc. fresh sessions that never saw it try to tear it apart. Google uses this to catch mistakes the biggest teams in tech stopped prompting from scratch. their configs are free. you are still typing "you are a helpful assistant" into a blank window everything you need to build yours is below

39 赞 · 1 转帖 · 5 回复

### [@FrankieIsLost](https://x.com/FrankieIsLost/status/2121006042658254904?ref=binnn.top)

11月13日 07:23 JST

great research. there’s probably a ton more information that can be inferred from model weights: model lineage, training data composition, distillation sources, compute estimates, etc. “if you gaze long into the weights, the weights also gaze into you.”

37 赞 · 0 转帖 · 0 回复

### [@dair\_ai](https://x.com/dair%5Fai/status/2121027770629382531?ref=binnn.top)

11月13日 08:50 JST

// Trustworthy Recursive Self-Improvement // Self-improving agents usually keep any harness update that raises the validation score, so noisy gains get accepted, and this later costs performance. Researchers measured how often this happens. Under a greedy gate, 45.59% of accepted updates were actually worse. Their gate, TRUST, treats each acceptance as a statistical test and controls the share of accepted updates that cause a real loss. An LLM reads the loop's history to decide how much evidence each update needs. Across HotpotQA, DROP, and APPS with DeepSeek-V4-Flash and Qwen-3.8-Flash, bad acceptances fell to 2.78%. Final scores were also higher than with earlier statistical gates, which tend to reject good updates. They prove a bound on bad acceptances and a lower bound on the final gain. https://t.co/Y20VQIrMhJ Chat with Paper: https://t.co/wZ6FSbDoYZ

36 赞 · 8 转帖 · 13 回复

### [@StanfordAILab](https://x.com/StanfordAILab/status/2120969570731356415?ref=binnn.top)

11月13日 04:58 JST

How do you teach an LLM a new document without it forgetting what it already knows? 📚 Check out Speculative Self-Distillation, a #NeurIPS2026 paper from @talaei\_shayan, @chewing\_a\_gum, @ckodser and @Azaliamirh's group. The student writes, and the teacher steps in only where it matters.

34 赞 · 9 转帖 · 4 回复

### [@Atenov\_D](https://x.com/Atenov%5FD/status/2120950604818223537?ref=binnn.top)

11月13日 03:43 JST

BUILD AN AI TRADING RESEARCH TEAM FOR FREE. TURN ITS IDEA INTO A STRATEGY YOU CAN TEST. Start with one thesis: AI infrastructure demand keeps growing. Now give an AI research team four jobs: → SCOUT: find public companies and ETFs exposed to AI infrastructure → ANALYST: compare their business exposure and risks → STRATEGIST: turn the thesis into entry and exit rules → SKEPTIC: challenge the assumptions and identify what could break the idea Take the resulting strategy and paste it into @horizon\_trade\_x Horizon helps structure the rules and backtest them on historical data before you decide whether to paper trade or connect a supported broker. THE WORKFLOW: AI THESIS → RESEARCH → TRADING RULES → BACKTEST → YOUR DECISION You can start free and try Horizon’s AI with your own idea. No coding required. The prompt and the test come before any real money moves. The full information is below. Save it.

33 赞 · 2 转帖 · 13 回复

### [@gurtej\_\_gill\_](https://x.com/gurtej%5F%5Fgill%5F/status/2120941242691600432?ref=binnn.top)

11月13日 03:06 JST

Great paper on making AI agents collaborate silently without writing any text and cutting down search time from an hour to just a few minutes. Here the bottleneck is waiting forever while the model spits out walls of chain-of-thought text just to test one minor program tweak. Their new framework, SwarmEvolve, basically asks: why force agents to decode into English at all? Instead, they spin up hundreds of low-cost agents per round that reason completely in latent space. To keep the swarm from getting stuck in a rut, they filter out redundant ideas by measuring the distances between the agents' hidden state vectors, and then do "crossover" simply by dumping their combined KV caches into a summarizer agent. I had to double-take at the numbers in Figure 1\. On a single GPU, it hits the 60-minute peak scores of frameworks like OpenEvolve in a median of just 3 to 8 minutes, churning out 734 programs an hour. Even wilder, running this on a tiny Qwen3 8B actually beat out some of the math solutions AlphaEvolve originally found using frontier models. This will really make you wonder how much compute we're wasting just forcing models to translate their thoughts back into human language between steps. Paper PDF: https://t.co/IRTPxOGUdk

32 赞 · 5 转帖 · 3 回复

### [@leo114119](https://x.com/leo114119/status/2120909143745892687?ref=binnn.top)

11月13日 00:58 JST

Gemini 4 Hydrogen（Flash） +Antigravity 生成的鹈鹕骑车测试： 只用了 5 分钟左右，非常快 感觉是完全超过 Deepseek v4.1 Flash 1017 了 甚至感觉都超过了 GPT 6.1 SOL https://t.co/iMsy9XJ6Pc

30 赞 · 2 转帖 · 16 回复

### [@leo114119](https://x.com/leo114119/status/2121043761719632232?ref=binnn.top)

11月13日 09:53 JST

外网曝料： Deepseek V5已经在预训练中，完全使用华为的芯片来进行训练，约 2T 的参数，目前正在 Deepseek 杭州的公司进行训练，年前肯定出不来了 估计得明年第一季度吧。。。能力我猜大概是能超过 Opus 5.5的。。

27 赞 · 1 转帖 · 11 回复

### [@xingbugengming](https://x.com/xingbugengming/status/2121058876812570796?ref=binnn.top)

11月13日 10:53 JST

这是我用 Codex 为我的游戏地图制作的一个沙盘演示视频 昨天 Claude 被封后，我研究了一下如何使用 Codex 制作高质量的视频，给兄弟们分享一下思路： 1\. Claude 适合纯文字提示词制作视频，因为Claude会嗅探并且参考网络中的一些素材，再加上本身审美能力高，整理归纳后出来的视频效果会非常棒 2\. Codex 用文字提示词直出视频的效果非常差，所以我们要利用好 Codex 的生图能力 在制作视频之前可以让 Codex 生成四种不同风格版本的预览图，一直生成到你看顺眼即可，然后再让 Codex 根据预览图的方向去制作视频，再不断精细化调试 按照这个思路去做话，这Claude也不是非用不可

26 赞 · 4 转帖 · 11 回复

### [@ZhentingQi](https://x.com/ZhentingQi/status/2120963028887769184?ref=binnn.top)

11月13日 04:32 JST

Launching the XE² Arena project: a cross-embodiment × cross-environment arena for agentic robots After @claudeai Fable 5.2’s amazing performance on RoboDojo, what's next? In XE² Arena: • Cross-embodiment: one LLM agent, on many robot bodies (YAM, xArm, Wuji Hand 2, humanoid, …) • Cross-environment: every real run gets a sim replica, for both eval and post-training 5 embodiements, 12 real tasks, 400 sim tasks. More bodies & tasks coming soon! https://t.co/bcHwRcLfMV Meet us @RoboticsCtr at CoRL 2026! Joint work with Siyu Lin, @andylizf @yixun\_hu @RoboticsCtr @jerryhuang01

24 赞 · 4 转帖 · 6 回复

### [@abxxai](https://x.com/abxxai/status/2120909485971853791?ref=binnn.top)

11月13日 01:00 JST

Giving your AI agent a computer now takes one line: Set up box from https://t.co/PVUQVGYKzY 1\. paste it into Claude Code, Codex, Cursor or any agent that speaks MCP 2\. approve it in your browser with Google or GitHub 3\. your agent gets a desktop where it runs commands, browses and uses apps Free to try. @SimularAI

24 赞 · 5 转帖 · 20 回复

### [@AndyL5cc](https://x.com/AndyL5cc/status/2120975080402681951?ref=binnn.top)

11月13日 05:20 JST

Claude 现在连会议都不放过了！🤣 Anthropic 刚推出 Claude Meetings，直接让 AI 加入团队视频会议。 你们负责聊天讨论，Claude 在旁边实时做文档、画白板、整理工作成果。 目前已向 Team 和 Enterprise 用户逐步开放测试。 以前开完会，还得安排人落实。 现在会议还没结束，Claude 已经开始交作业了。 😂 https://t.co/F2C3NFDShU

24 赞 · 0 转帖 · 9 回复

### [@HyperFrames\_](https://x.com/HyperFrames%5F/status/2120973830168752298?ref=binnn.top)

11月13日 05:15 JST

HyperFrames Arena is live on Kaggle, built with @kaggle and @GoogleDeepMind our benchmark has an LLM judge videos on 5 categories, but in the Arena, you're the judge each matchup is two videos from one brief, and the leaderboard moves with your votes we get asked which model to use all the time. now you can watch them side by side and decide which is best for your videos

23 赞 · 2 转帖 · 2 回复

### [@ccatalini](https://x.com/ccatalini/status/2120931836075642880?ref=binnn.top)

11月13日 02:28 JST

The alternative to a competitive market for AI is grim. That's why open source and open weights matter. https://t.co/W8GNTzhqQE

23 赞 · 1 转帖 · 4 回复

### [@CMGS1988](https://x.com/CMGS1988/status/2120934686017130771?ref=binnn.top)

11月13日 02:40 JST

这个视频其实是我最喜欢的，本质上整体架构来说是完全的工程化思路，最短路径，确定性，可重用等…所以大模型越强，收益也会越高，跟什么 harness 没太大关系…而且实际上我确实也做了 replay 代码化这件事，从理论上来说最后重复工作就不需要消耗 token 了...

23 赞 · 0 转帖 · 3 回复

### [@halvenruns](https://x.com/halvenruns/status/2120986826810802670?ref=binnn.top)

11月13日 06:07 JST

This is f\*cking insane. Y Combinator CEO Garry Tan: "When someone asks how I prompt my AI, the answer is: I don't. The skills are the prompts" In 2 minutes, he shows the setup that turns Claude Code into a full AI engineering team, built on one idea: thin harness, fat skills Prompts → Skills → A Team Of Specialists → A Startup Built By Agents One of his skills runs YC office hours: tens of thousands of hours from 16 YC partners, distilled into a file Claude follows Do a workflow once, turn it into a skill, and it runs that way forever Save this before it slips off your feed Everything below is free: GStack (his 23 Claude Code skills: CEO, designer, eng manager, QA) https://t.co/uhqiIe8NS0 GBrain (his open-source second brain) https://t.co/ge27vKDnBn How to make Claude Code your AI engineering team (YC) https://t.co/UN1qiS1qKi Don't Build Agents, Build Skills Instead (Anthropic) https://t.co/BekMULOdSS How Perplexity designs and maintains agent skills https://t.co/OiW7Rha7QK Anthropic's official skills repo https://t.co/JLhToFYsMv Agent Skills docs https://t.co/SynIrWfOa4 How Anthropic equips agents with skills https://t.co/1fBvhMahD3 Agent Skills open standard https://t.co/U9Kap0Zlwh Free Agent Skills course with Anthropic https://t.co/n65klf8Smi All of this costs $0 Prompting is dead, skills are next, and 99% are still typing prompts from scratch every morning Bookmark it and open the first link tonight Then read the full Graph Engineering guide below ↓

23 赞 · 0 转帖 · 3 回复

### [@Chenfacai0](https://x.com/Chenfacai0/status/2120965856402301337?ref=binnn.top)

11月13日 04:44 JST

全球体量最大的 AI 提示词库居然免费向所有人开放了！！ 里面收了海量现成的 AI Prompt，无论是 ChatGPT、Claude 还是 Gemini，主流模型通通能直接套用。 不用再绞尽脑汁研究怎么调教模型，搜个关键词，一键复制就能跑起来。 经常用 AI 的朋友，这个神仙网站强烈建议直接收藏！ 项目地址：https://t.co/YfmZQ3z7ZR

23 赞 · 0 转帖 · 7 回复

### [@0xApexAi](https://x.com/0xApexAi/status/2120992761445073215?ref=binnn.top)

11月13日 06:30 JST

THIS SETUP SHOULD BE ILLEGAL. I slept 8 hours and my socials posted 9 times. no scheduler queue. no VA. 6 Grok bots on the night shift. how the night goes: - 23:00 → I close the laptop - Trend Hunter finds what’s rising before it peaks - Writer turns it into hooks in my voice - Designer makes the cover for every platform - Ops Lead kills anything weak, using my rules - Publisher ships it to 5 platforms - Analyst writes the morning report - 07:00 → I just read it - 9 open-source GitHub repos to build your own night shift: 01 / LangChain Social Media Agent → turns links into ready posts for X and LinkedIn, with you approving before anything goes out https://t.co/jpxiC4FAcl 02 / Social Media Skills (Charlie Hills) → Claude skills for writing posts that sound like you https://t.co/rqY0xCBBFt 03 / Marketing Skills → a whole marketing skill pack for Claude Code, social content included https://t.co/TdX1NcsYJ9 04 / Claude Skills → a big skill library with a ready social media manager role https://t.co/F1yuOQYVO8 05 / Agents → Claude Code subagents and plugins, including social publishing https://t.co/ahOC8yZ3Cd 06 / Riona AI Agent → an agent that runs your social accounts and engagement on autopilot https://t.co/OWwfnzUHwL 07 / Agentfy → multi-agent system for social media tasks https://t.co/rtH9dhEA3n 08 / Social GPT Agent → an AI social media manager you can run yourself https://t.co/wPCaGFbSRn 09 / Claude Marketing → marketing and social strategy skills for Claude https://t.co/ry8ZaG04fH you sleep. they ship. build your night shift ↓

22 赞 · 3 转帖 · 3 回复

### [@xiangning0](https://x.com/xiangning0/status/2121023233784741895?ref=binnn.top)

11月13日 08:32 JST

看来Codex是真的没人关心了😂 以前发重置卡的消息，至少是上万曝光的 这条发出去一个多小时了，也没几个人看。 我的重置卡在7点钟收到了 https://t.co/aVSrxCzIun

21 赞 · 0 转帖 · 14 回复

### [@TheInsiderPaper](https://x.com/TheInsiderPaper/status/2120959073416077623?ref=binnn.top)

11月13日 04:17 JST

You no longer need a Mac to test an iOS app. @expo just launched cloud iOS Simulators that AI coding agents can access from Windows or Linux. They can open apps, test changes, and send back screenshots and recordings without needing a physical iPhone.

21 赞 · 4 转帖 · 6 回复

### [@kalipusitkk](https://x.com/kalipusitkk/status/2120903251205976197?ref=binnn.top)

11月13日 00:35 JST

好多朋友问我提示词是什么，其实之前是给别人的评论回复过的现在发帖分享一下： 请帮我搭建一个「AI聚合比价空间」：把散布在各个卡网（发卡平台）和 AI 代充聚合站的 AI 订阅/账号商品（ChatGPT、Claude、Gemini、Grok、Cursor 等的代充、成品号、卡密、车位）聚合到一个空间里，方便我按关键词、平台、店铺、价格、库存快速比价。只做信息聚合，不卖货、不下单、不登录任何账号。 1\. 链动小铺 https://t.co/VvyeH9VAS9（主力源，全店直抓） - 公开接口：POST /shopApi/Shop/goodsList，参数 token / keywords / category\_id / goods\_type / current / pageSize，可分页拉全店商品（名称、价格、市场价、库存、在售状态）。 - 注意：直连会被阿里云 ESA 人机验证拦。需要先用真实浏览器会话过一次验证、保住 cookies，再用该会话调接口。滑块的行为检测会识别自动化拖动，不要反复硬试，失败就冷却、换思路。 - 店铺 token 大小写不敏感；平台买家端没有店铺目录接口，店铺名单从两个地方拿：① GitHub 项目 wzyp-view 发布的数据包（250 家种子店）；② https://t.co/FKiyMjBOyH 的 sitemap-shops.xml（收录 1000+ 家链动小铺店）。 - 单品直达链接：https://t.co/NTmG4rpgyl<goods\_key>，店铺链接：https://t.co/sMwyH3ljTo<token>。 2\. 云猫寄售 https://t.co/nCX39YDJZB（同系接口，可直抓） - 与链动小铺同系，同样有 /shopApi/Shop/goodsList 公开接口（表单版），店铺名单从 AI价网的店铺索引里提取云猫平台的店铺标识后逐店直抓。 3\. AI价网 https://t.co/FKiyMjBOyH（跨平台报价源，必须署名） - 收录约 1900+ 家卡网店铺、5 万+ 条报价，覆盖链动小铺、云猫寄售、独角数卡、16688、自建站等平台，每小时更新。 - 机器可读层：https://t.co/SPT1qmq1hu 和 /md/market/category/<slug>.md（每类只含有货最便宜 50 家）；全量报价在网页端分类页 https://t.co/uwH2sGoz2e<slug>?all=1 和店铺页 /market/shop/<平台前缀>-<店铺标识>。 - 许可：数据为 CC BY 4.0，使用必须注明来源「AI价网（https://t.co/FKiyMjBOyH）」并附链接 https://t.co/27X17hWlFb 。 4\. 水獭 https://t.co/hvjvvggiWw（高质量补充，价格为含手续费到手价） - 官方 MCP：npm 包 ai-price-mcp，免 API key、只读。用 list\_all\_products / get\_offers 遍历全部 SKU（约 73 个，每个 SKU 最多返回 20 家报价），另有 list\_shops。 - 只并入派生报价行并署名「水獭 https://t.co/hvjvvggiWw」，禁止整站镜像、不要搬运其文章。 5\. https://t.co/aLeUVtDWZ2 https://t.co/7oDXWeEnbh（补漏源） - 抓它的全部标准商品页（约 44 个，形如 https://t.co/ehyk0zvjNd<slug>），每页是一张全渠道报价表（渠道、原始商品名、价格、库存、更新时间、购买入口）。署名「https://t.co/aLeUVtDWZ2」。 6\. 许可红线：CardNav（https://t.co/ECcVYYx7f5）和 PriceAI（https://t.co/1DD5q0SzqL）的仓库 DATA\_LICENSE 明确禁止批量复制、镜像和用作替代数据源——只能用它们发现店铺、交叉核对，不得整批并入。AIPricedb、模真真可作后续发现源。 统一成一张表，每行字段：商品名 title、价格 price、市场价 market\_price（无则空）、库存 stock、店铺名 shop\_name、平台 platform（wzyp / 云猫 / 16688 / 独角数卡 / 自建站 / Kami 等）、来源 source（直抓 / AI价网 / 水獭 / BiWu）、分类 category、原站商品直达链接 item\_url、店铺链接 shop\_url、店铺标识 shop\_id、观察时间 observed\_at。 库存口径统一：数字 = 库存件数；0 = 缺货；只有「有货/缺货」标记的源，有货记 1、缺货记 0；未知记 -1（库存未公开）。 去重键：平台 + 店铺标识 + 商品标识；没有商品标识时用「原始标题 + 价格」。同一商品同时有直抓行和聚合报价行时，保留直抓行，报价行不重复计数。合并完成后，把总数、各源净增、各平台分布、库存分布写成一份 report.md，与 data.json 放在同一个研究目录里，每个数字都要能从 data.json 复算。 做一个静态网页空间，名为「AI聚合比价空间」，创建时直接放进我的侧边栏「空间」区（不要只落在资源库）。 - 数据以独立数据文件随页面发布，页面加载时以 no-store 方式拉取，不要把数据内联写死在 HTML 里（内联版在 App 里容易被旧缓存顶住、更新不生效——这是实测教训）。 - 页头统计：覆盖店铺数、商品总数、数据更新时间。 - 功能：关键词搜索（搜商品名）；筛选器：来源平台、店铺、分类；只看有货开关；价格排序（低→高 / 高→低）与价格区间；每页 100 件分页。所有筛选可叠加。 - 商品卡：显示平台标记 + 来源标记、库存状态（有货 N 件 / 已缺货 / 库存未公开），「去购买」按钮和店铺名都链接到原站。 - 页脚放齐数据来源署名与链接（AI价网 CC BY 4.0、水獭、BiWu），并注明各源抓取时间与库存口径。

20 赞 · 2 转帖 · 3 回复

### [@Stellakjbk](https://x.com/Stellakjbk/status/2120926948935528702?ref=binnn.top)

11月13日 02:09 JST

https://t.co/4K6hxOVODw

19 赞 · 11 转帖 · 17 回复

### [@Aurora0x3f](https://x.com/Aurora0x3f/status/2121052472219443313?ref=binnn.top)

11月13日 10:28 JST

Codex 重置卡发了，但是依旧没有新模型 所以说呀，GPT 才是真王者，外面雷声再大，他们也沉得住气 🙃 https://t.co/vKdNaCkvpp

18 赞 · 1 转帖 · 17 回复

### [@leo114119](https://x.com/leo114119/status/2120894902741737735?ref=binnn.top)

11月13日 00:02 JST

百度 ERNIE 5.5 Flash Preview 在前端测试中居然超过了 Deepseek v4.1 Flash 1017 在CodeArena：WebDev最新评测榜中获得1668分，位列开源模型第二、国产模型第四，紧随Kimi K3.1 Pro、Qwen3.8-Max等大参数国产模型之后。 百度这回是真站起来了啊。。 https://t.co/E9oyTUqlGl

18 赞 · 0 转帖 · 6 回复

### [@Huahuazo](https://x.com/Huahuazo/status/2120889099712401633?ref=binnn.top)

11月12日 23:39 JST

想把小说捣鼓成 AI 短剧？ 这套 Skills 正好能帮你把前头的流程给串起来。 有个开源项目叫 shuohao-skills，是冲着 AI 短剧创作去的，能搭着 Codex、Claude Code 等工具使，把小说素材一步步拾掇成能接着往下做的东西。 1️⃣ 拆解故事：把人物关系、剧情大纲给梳理清楚 2️⃣ 补全设定：场景跟道具的信息也都整理明白 3️⃣ 生成剧本：接着往下扩写剧本，再拆分成分镜 4️⃣ 汇总报告：把好几份分析结果攒成一个带导航的页面 适合那些想把网文改成短剧、摸索 AI 视频制作流程的伙计。 🔗 ：https://t.co/YyewnhMSDB

18 赞 · 3 转帖 · 3 回复

### [@NFT\_Chen](https://x.com/NFT%5FChen/status/2121022568408944815?ref=binnn.top)

11月13日 08:29 JST

vLLM 把 DeepSeek-V4.1-Flash 在 Vera Rubin NVL72 上跑到离谱性能！ 🔹 峰值吞吐：2.52M tokens/s/GPU（GB300 的 2.7×） 🔹 高并发甜点：1.7M tokens/s/GPU + 387 tokens/s/user（58× GB300） 🔹 单会话极限：1810 tokens/s 内核重写 + 机架级 KV 池化 + 智能路由，把 Rubin 的带宽和计算优势全榨出来了。 详情：https://t.co/oTEBl4ECbR

17 赞 · 6 转帖 · 11 回复

### [@LeoKharon](https://x.com/LeoKharon/status/2120985628007342306?ref=binnn.top)

11月13日 06:02 JST

NEW OPEN SOURCE PACKAGE: Five agent workflows for robotics, one pip install! RoboSkills is a pip-installable Python package from OpenRoboKit, a six-repository organization whose only institutional marker on the site is a UCSD-named favicon. It packages five agent workflows for robotics, each held to under 1,000 lines of code by stated policy: 1\. Auto Research, where an agent runs repeated simulation experiments to improve a policy and then freezes it 2\. Agent as Policy, where an LLM observes an episode and chooses the next action turn by turn 3\. Agentic Simulation, where an agent builds and solves a task from a natural-language instruction with privileged simulator access 4\. Real2Sim, which rebuilds an interactive physics scene from a recorded demonstration video 5\. Motion Generation, which turns text, an image or a video into a rendered motion with no physics at all Underneath sit 13 tools wired into one interface: two first-party ones, RoboKit for kinematics, motion planning and hands, and GraspKit for learned grasps, plus SAM 3 and SAM 3D from Meta, MoGe from Microsoft, Kimodo from NVIDIA, and GeoCalib, HaWoR, WiLoR, GVHMR, CoACD, MuJoCo and Blender. Everything runs in simulation, over RoboDojo, RoboCasa365, DexJoCo, LIBERO and reconstructions of DROID, DexYCB, ARCTIC, EgoDex and OakInk2\. Caveat: there is no paper, no author list and no hardware.

15 赞 · 0 转帖 · 3 回复

### [@QCOMResearch](https://x.com/QCOMResearch/status/2120924689602052300?ref=binnn.top)

11月13日 02:00 JST

What if reasoning is just compression? 🤔 Our @qualcomm paper at #NeurIPS2026: "Reasoning as Compression: Unifying Budget Forcing via the Conditional Information Bottleneck" @behboodiarash https://t.co/6rykwL5bFn #Reasoning #LLMs #AI

14 赞 · 5 转帖 · 0 回复

### [@leo114119](https://x.com/leo114119/status/2120900338240909423?ref=binnn.top)

11月13日 00:23 JST

豆包里的 Deepseek v4.1 Flash 生成的鹈鹕骑车 好像质量还可以，但豆包用着有点奇怪 思考过程看不到，如果中间卡住了就完全是白等。。。 看着像是不错的 https://t.co/U9yVTnFIk3

14 赞 · 1 转帖 · 9 回复

### [@Krot\_alpha](https://x.com/Krot%5Falpha/status/2121052096229253157?ref=binnn.top)

11月13日 10:26 JST

She’s 18 and a translation startup paid her $3.2M for a neural translator built on one idea - at Stanford she showed the whole thing: 00:50:15 - how neural MT wiped out a decade of hand-built translation systems in two years 00:57:03 - the “ag ag ag” bug that made Google Translate spit out Bible verses 01:00:27 - attention: the one trick that became the core of every LLM after watching I built a seq2seq translator for our support tickets - it cut our localization bill by $31k in the first month. save & watch - the article below breaks down the translator that sold for $3.2M.

13 赞 · 0 转帖 · 0 回复

### [@Gorden\_Sun](https://x.com/Gorden%5FSun/status/2121056912498364497?ref=binnn.top)

11月13日 10:45 JST

Inkling 2 Small开源了 唯一能跟国模一战的美国开源模型，由Thinking Machines开源（前OpenAI首席技术官Mira创办）。 参数594B，激活参数17B，评分接近DeepSeek-V4.1-Flash-1017（763B），超过GLM 5.3 Flash。 模型：https://t.co/Vnt2XGTG0n https://t.co/mzI1s7ldcp

13 赞 · 0 转帖 · 23 回复

### [@7day\_dreaming](https://x.com/7day%5Fdreaming/status/2120995401868410884?ref=binnn.top)

11月13日 06:41 JST

用上Claude duet了╯∀╰ 这个edit功能好有趣，小机说每次编辑信息像按门铃哈哈 这个描述也太可爱了……！ 话说申请原来不用美区ip(我是日区)啊，随便填个美区地址就能通过 https://t.co/oblLebQWty

13 赞 · 0 转帖 · 0 回复

### [@Flandermaxx](https://x.com/Flandermaxx/status/2120945913317240943?ref=binnn.top)

11月13日 03:24 JST

WHOEVER LEAKED THIS GROK BROWSE SHOULD BE REPLACING YOUR WHOLE TAB GRAVEYARD I thought browser hygiene meant 47 open tabs, no idea which one has the thing I needed, and a weekly "close all and restart" that loses half my work. Apparently not. This setup is much more interesting: → TABS + TAB\_HIST + COOKIES + SWITCH\_HIST pack every browser state as typed inputs, no "open all 47 at once and pray", RETURNS (likes, dislikes) attached from past sessions → GROK DECIDES next\_action in 4ms (close, hover, scroll, click) with Score page\_fit (low to critical) and Bool needs\_user yes-or-no, output tokens literally zero → GATE ≥ 0.85 holds every close or click until confidence beats the floor, below drops to review, no "silently closed the tab I actually needed" → WORKERS fire the deterministic moves (close the dead tab, scroll to the section, click the button), no LLM wake-up mid-action → HUMAN SIGN gates the irreversible moves (close a tab with a draft, submit a form, hit pay), 10,000 browser decisions logged for $0.01 total, flags 7 tabs the gate held for review this session so you get: Tabs scored. Clicks typed. Closes gated. Signs human. the interesting part isn't a smarter browser. it's giving every tab a score the close button can't fire past without your confidence. Instead of 47 tabs and a weekly reset that loses your work, you let Grok score every tab in 4ms with typed answers, hold every irreversible close behind 0.85 confidence, and run 10,000 browser decisions for $0.01 total. Score the tab. Gate the close. Sign the irreversible.

12 赞 · 0 转帖 · 0 回复

### [@iamanitguy](https://x.com/iamanitguy/status/2121049510038344035?ref=binnn.top)

11月13日 10:16 JST

This is literally free to use with no token limits. For a limited time new users get to run their bots for free on Deepseek V4.1 Flash and GLM 5.3 Flash

12 赞 · 0 转帖 · 0 回复

### [@sujingshen](https://x.com/sujingshen/status/2120886355182113123?ref=binnn.top)

11月12日 23:28 JST

https://t.co/AxBRhOIr4L

11 赞 · 10 转帖 · 16 回复

### [@Engram0x](https://x.com/Engram0x/status/2120907276739289335?ref=binnn.top)

11月13日 00:51 JST

Andrej Karpathy, one of the co-founders of OpenAI, has made almost everything he knows about building AI freely available on GitHub! From building GPT from scratch to creating AI that autonomously runs experiments, analyzes results, and learns from its own findings. 6 projects with over 300K stars combined ⭐ 1\. Autoresearch | 92.7K ⭐ An AI that continuously runs experiments, analyzes results, and keeps only the solutions that actually work. https://t.co/20laxVrgRF… 2\. nanoGPT | 64.4K ⭐ Learn how to train a GPT model from scratch. A project that has become the starting point for many people entering the world of LLMs. https://t.co/1xXnUYLO6U… 3\. nanochat | 60.2K ⭐ Build your own personal version of ChatGPT for around $100\. https://t.co/RkAenvkiSp… 4\. LLM Wiki | 50K ⭐ A simple but powerful concept for building your own expandable AI knowledge base — something like a second brain. https://t.co/rN83qAvT7i 5\. llm.c | 31.4K ⭐ Train a language model in C, without hiding the technical complexity behind ready-made frameworks. https://t.co/y6J8Do5gwL 6\. micrograd | 18.7K ⭐ Understand how neural networks learn and how gradients work in just a few hundred lines of code. https://t.co/cdHOKgWdvG… My recommendation: Start with micrograd. Then move on to nanoGPT and nanochat. This learning path will help you do more than just use AI tools — it will help you understand what actually happens under the hood. The difference between an AI user and someone who can build intelligent systems starts right here. Save these 6 repositories. They’re a goldmine of free knowledge.

10 赞 · 0 转帖 · 2 回复

### [@jadechoghari](https://x.com/jadechoghari/status/2120958700240396737?ref=binnn.top)

11月13日 04:15 JST

Excited to see our paper with @tanaytndn (@ScaleAILabs) headed to the NeurIPS Robot Learning Workshop! Editing robot videos frame-by-frame is slow and breaks on fast-moving wrist cameras. Instead, we generate a simple 3D room once (just 6 walls) and render the background as the robot moves. It's 51x faster and surprisingly makes robots ignore unseen table clutter too.

10 赞 · 0 转帖 · 1 回复

### [@don\_tech1](https://x.com/don%5Ftech1/status/2121037831443517853?ref=binnn.top)

11月13日 09:30 JST

卧槽，一个大创新， Claude AI 自托管智能体上线私测了。 以前让 AI 修你的代码、处理内部告警，数据基本得往外送； 现在可以把整套代理框架、沙箱和记录都放自己服务器里跑，代码和密钥不出家门。 企业再也不用担心数据泄露等安全问题， 可以让 AI 真正碰核心系统了。 https://t.co/y8Ptoo3ccP

9 赞 · 0 转帖 · 17 回复

### [@Necokeine](https://x.com/Necokeine/status/2120938824708829359?ref=binnn.top)

11月13日 02:56 JST

AI harness 本质上已经是模型厂提供的用户界面了，你不能把harness和模型厂的模型分开看。 Codex搭配别的模型再牛逼，只要GPT是一坨那就还是一坨。 Claude Code再一坨，只要和Claude一起跑分高就是高。 DSH作为一个后来者，已经做的很不错了，也能指导未来DeekSeek模型的前进方向，从社区收集各种需求并且传递给模型训练，这是很不错的。 那个榜单，怎么看都太野鸡了……看过笑过也就过去了。当真了，真没什么意思。pi的高分就足够说明，tool set越简单的harness，对于所有模型来说，平均表现就越好。 但是扪心自问？你们要的就这么点么？

9 赞 · 0 转帖 · 0 回复

### [@HarryTandy](https://x.com/HarryTandy/status/2120948044484981170?ref=binnn.top)

11月13日 03:33 JST

a year ago the best AI agent could finish 2.5% of real freelance jobs on november 9, Claude Fable 5.2 finished 41.3% the test: 240 jobs real freelancers already did on Upwork, worth $143,991\. reviewers open the AI's files next to the freelancer's and ask one thing: would a client accept this? where it's strongest now: → audio: cleanup, mixes, stems. US rate $36–60/hr → floor plans and interiors from a sketch → 3D models: product turntables, game assets, STL → reports and charts, every number traced to a source where it still fails: animation and industrial product design and the other 59% is where your fee comes from. Claude does the first pass, you open every file before the client does @alphabatcher wrote the full course: price it from Upwork's rate pages, run it through Claude Code, sell it near the $200 median job which of the 4 would you sell?

8 赞 · 0 转帖 · 7 回复

### [@PiperUnplugged](https://x.com/PiperUnplugged/status/2120956595404464393?ref=binnn.top)

11月13日 04:07 JST

My robot lives in the cloud and he doesn't trust it ☁️ So now we write phone numbers on paper https://t.co/Sly2c0VAjm

8 赞 · 0 转帖 · 0 回复

### [@leo114119](https://x.com/leo114119/status/2121048325789192483?ref=binnn.top)

11月13日 10:11 JST

Gemini 4 Hydrogen：3D 互动餐厅 用时半个小时生成出来的 老实说，3D 建模不算是Hydrogen 的强项，在 3D 建模上Hydrogen 顶多和 Deepseek v4.1 Flash 1017持平 https://t.co/ij8mGMuGjR

8 赞 · 0 转帖 · 3 回复

### [@Stargridcompass](https://x.com/Stargridcompass/status/2121047895935791230?ref=binnn.top)

11月13日 10:10 JST

Day 26 / Claude帐号安全度过 算是躲过了最强风控第一天，昨天看到好多兄弟的帐号阵亡了😭😭，我要且用且珍惜了。 https://t.co/xiMYdk3HOu

7 赞 · 0 转帖 · 10 回复

### [@GCSAqianye](https://x.com/GCSAqianye/status/2120925724387459361?ref=binnn.top)

11月13日 02:04 JST

你用AI快速写出的代码，早已悄悄变成失控黑盒。老旧SAST工具跟不上Vibe Coding的风险。 真心强烈推荐SecVibe Code，专门审计AI生成代码，直击真实漏洞，直接给到修复方案。适合程序代码上线前检查。 这篇文章讲清它的价值与实操，值得看完、点赞，免费额度可以直接上手测试。 https://t.co/yrm1BS7sEK

7 赞 · 4 转帖 · 1 回复

### [@YuutaTsubasa](https://x.com/YuutaTsubasa/status/2120955727183560985?ref=binnn.top)

11月13日 04:03 JST

Fable 沒有想像中真的可以幾句提示詞就直接生出完整的 3D 索尼克動作遊戲 QwQ 場景還是建構的很陽春XD 不過環境和渲染倒是變得比 Opus 好倒是真的wwww #Claude #Fable #Fable52 https://t.co/YBycjLQZ3v

6 赞 · 1 转帖 · 0 回复

### [@resetoday](https://x.com/resetoday/status/2120916722232476038?ref=binnn.top)

11月13日 01:28 JST

飞书有豆包工作 钉钉有千问办公 企微有WorkBuddy 文档里还有WPS灵犀。 国外是ChatGPT Claude Copilot Gemini 还有GBot。 豆包今天又把Kimi、智谱、DeepSeek摆上了自己货架。 和之前百团大战一样 最好用的必须是gbot 你现在用的是哪一个？

6 赞 · 0 转帖 · 5 回复

### [@geekbb](https://x.com/geekbb/status/2121053488998473872?ref=binnn.top)

11月13日 10:32 JST

把 Gemini 3.8 Flash 价格降到 DeepSeek-V4.1-Flash-1017 一样和他拼了，看谁火力强！ https://t.co/BmM5nDN4m5

5 赞 · 0 转帖 · 8 回复

### [@Arra9005](https://x.com/Arra9005/status/2120920561903972801?ref=binnn.top)

11月13日 01:44 JST

claude隔离区越来越成熟了。 https://t.co/fRgUntXcys

5 赞 · 0 转帖 · 1 回复

### [@ethanzhang\_web3](https://x.com/ethanzhang%5Fweb3/status/2120942822111203604?ref=binnn.top)

11月13日 03:12 JST

大冰老师也被Claude征服了 #ai https://t.co/ifhGvRb96U

5 赞 · 0 转帖 · 2 回复

### [@daxiang555](https://x.com/daxiang555/status/2121012099099107585?ref=binnn.top)

11月13日 07:47 JST

字节跳动视频大模型Seeddance核心掌舵人——曾妍 97年 年轻有为，5年时间晋升字节4-2，一年内完成两连跳。一般来说3-1就是普通人的天花板了，4-2属于高层管理级别的了 https://t.co/tDKJbR0Bs6

5 赞 · 0 转帖 · 0 回复

### [@sonnylazuardi](https://x.com/sonnylazuardi/status/2121017218146423265?ref=binnn.top)

11月13日 08:08 JST

invite AI agent ke zoom meeting 🤯 https://t.co/RBB34mcxdU

4 赞 · 0 转帖 · 0 回复

### [@kdnuggets](https://x.com/kdnuggets/status/2121007830614327377?ref=binnn.top)

11月13日 07:30 JST

7 ways AI agents burn tokens, from append-only context to sending simple tasks to frontier models, with the fix for each: sliding windows, tool result compression, prompt prefixing and routing. #AIAgents https://t.co/UE8xC83S1e

4 赞 · 1 转帖 · 2 回复

### [@hejushang8](https://x.com/hejushang8/status/2121048659249033384?ref=binnn.top)

11月13日 10:13 JST

很多人问用什么 U 卡支付的 其实这个 Claude 账号，为了返现和测试 U 卡，订阅了 3 个月用了 3 张U 卡 首月订阅用的 Bybit Eu ，第二个月用的 OKX欧区， 第三个月用的 P1 大部分 U 卡都支持订阅AI 的，只需要用手机 Safair 浏览器打开网页，然后用 Apple pay 支付即可

4 赞 · 1 转帖 · 5 回复

### [@JacksAISites](https://x.com/JacksAISites/status/2120978848112103662?ref=binnn.top)

11月13日 05:35 JST

我一直想要的一个 AI 功能 Claude 也做出来了。现在能直接进视频会议了，叫 Claude Meetings，边开会边把文档、白板做出来。 1、先只支持 Google Meet，Team 和 Enterprise 套餐 beta，个人版用不了。 2、Claude 是会里一个看得见的参会人，进会时会在聊天里发通知。 3、开会时喊一声 Claude 就能提问，默认在 Meet 聊天里用文字回。 4、说一句「Claude, make a doc」或「make a whiteboard」，当场出文档、白板、幻灯片，都挂在这场会的页面上。 5、能调你连好的工具，比如发 Slack，第一次会先问你同不同意。 6、一场最多待 8 小时，会后自动出纪要、待办和完整转录。 会议记录工具这几年已经很多了，这次的变化是会还没散，东西已经做出来了。这个就是我想要的功能。 以前开完会还得有人花半小时整理成文档，现在这一步在会上就做完了。 非常遗憾，开了零数据保留或 HIPAA 配置的组织暂时用不了。这样看好像很多公司都是用不了😳 https://t.co/JNNhexHlBk

4 赞 · 0 转帖 · 7 回复

### [@Ylj158852](https://x.com/Ylj158852/status/2121023927812252065?ref=binnn.top)

11月13日 08:34 JST

阿里千问Qwen4发布了，Max预览版上线，调用限时2折，新客5折，叠加最低1折。百万上下文，最大输出39万token。Qwen这是要把性价比打到底？国产大模型内卷，卷到最后受益的是咱们。

4 赞 · 1 转帖 · 3 回复

### [@leo114119](https://x.com/leo114119/status/2121033589508939860?ref=binnn.top)

11月13日 09:13 JST

Anthropic 刚刚发布了 Claude Meetings。 Claude 现在可以加入你的 Google Meet 通话，并在你还在说话时构建文档 + 白板。 它会做笔记、跟踪行动项，并记住所有内容。🤯 AI 大模型开始入侵更多的商业软件赛道了。。。

4 赞 · 0 转帖 · 5 回复

### [@KinGao476942](https://x.com/KinGao476942/status/2121056596386292150?ref=binnn.top)

11月13日 10:44 JST

Claude 这是在抄袭 @bot ?? 昨天 Grok Bot 刚更新了视频会议接入功能，今天 Claude 就也上了。。。。 顺带一提，我的主页有可以让你稳定用 Claude 的教程 ～

4 赞 · 1 转帖 · 1 回复

### [@Zhibin992](https://x.com/Zhibin992/status/2121035381131084245?ref=binnn.top)

11月13日 09:20 JST

AI体感测评： 夯： Claude Fable5.2，目前的强中强，短板就是数学不算太好。 GPT 6 Astra，数学仍然是无敌手的，但是有严重的过度防御，保守措辞，很怕犯错，疑似被自家公司强迫做数学题做出了精神疾病。 Claude Opus5.5，强，全面，同样弱点是数学容易出错。 （O家和A家现在基本上只用这三个了）。 顶级： Gemini Carbon，各方面都还行，目前没发现什么特别惊艳的，缺点就是太慢了。说话仍然是最像人的。 Muse。免费能干活这一项已经秒杀了。作为免费的视频生成工具，效果非常好，可玩性很高，有了空间功能之后体验直线上升，不得不说小扎是真的懂C端。 Dsh+DeepSeek各种便宜token，可玩性也相当高，还有手机版/安卓平板版。大肥鱼桌宠已成心头好。 人上人： Grok Bot，能干活，啥都会，除了太吃额度没什么缺点，但因为有免费的Muse，就显得略逊。 Dots，Astra驱动的时期非常聪明，现在只能说能做事，但一旦路由到Terra你迟早被它各种绕来绕去的过度防御逼疯。 NPC： 豆包工作，豆姐虽然傻，但Seedance体验不错，免费额度薅得开心。 拉完了： GLM5.3，降智严重，一副半死不活的样子，不知道是不是因为Claude防蒸馏导致的，很久都没有出新模型。 Mistral，据说承载了全欧洲的希望，拿来玩了一下感觉是上个世纪的产品。

3 赞 · 0 转帖 · 2 回复

### [@M\_Surazhsky](https://x.com/M%5FSurazhsky/status/2120890127761252842?ref=binnn.top)

11月12日 23:43 JST

Playing a lot lately with training loco-manipulation, here is a blogpost I wrote to compare the different data types used for training. If you're in robot learning and deciding where your next hours of data should come from, I hope this map helps. https://t.co/h7hakD2wsX

3 赞 · 0 转帖 · 3 回复

### [@jphorism](https://x.com/jphorism/status/2120988628159504464?ref=binnn.top)

11月13日 06:14 JST

The issue is that meeting notes are just a medium. You need the distilled output to compound all the generative but "off-topic" chatter that saves you a future meeting someday. Introducing Margins: an open source framework to build meetings plugins into wherever your agents already do their work. Ships with a TUI, uses any agent to write meeting notes directly to Markdown. Runs on-device voice models. Or open a separate agent session and get it to read the real-time transcription and notes so far.

3 赞 · 0 转帖 · 4 回复

### [@SongyuKite](https://x.com/SongyuKite/status/2121028200621080887?ref=binnn.top)

11月13日 08:51 JST

来了来了 OpenAI Codex 重置卡在路上了 Tibo 宣布重置来了

3 赞 · 0 转帖 · 2 回复

### [@SeanHeoi](https://x.com/SeanHeoi/status/2121060227416236439?ref=binnn.top)

11月13日 10:59 JST

AI 进了个人，没有进组织。我们用一年时间，从根本上解决这个问题。 https://t.co/xRLfNmVJXZ

2 赞 · 0 转帖 · 2 回复

### [@Dragonhau66](https://x.com/Dragonhau66/status/2121009812162924782?ref=binnn.top)

11月13日 07:38 JST

你能想到吗？字节跳动的Seedance 背后拍板做2000亿参数大模型的，是个1997年出生的29岁的女生。 她叫曾妍，2021年校招进字节，两个月发顶会论文，五年干到4-2职级。 2025年底她坚持要训一个2000亿参数以上的视频模型，内部不少人觉得她疯了，算力成本扛不住。 她没等领导点头，直接找高管拍桌子要资源。 结果呢？Seedance 2.0登顶全球榜单，单月营收超10亿，短剧行业95%都在用它。 29岁，没背景，就是敢赌敢拼敢卷，而且赌赢了。 字节这波，算是捡到宝了。

2 赞 · 2 转帖 · 3 回复

### [@GySgt\_ewinka](https://x.com/GySgt%5Fewinka/status/2120980060253655512?ref=binnn.top)

11月13日 05:40 JST

Did you know Grok can protect your digital identity and prevent activists, dissidents, and political voices from accidentally leaking identifying info on X? 🛡️ ​With growing risks for human rights defenders and journalists worldwide, @elonmusk, giving users the tools to audit their own digital footprint with AI is a massive step forward for online safety. ​Here is a ready-to-use Grok prompt and a quick guide to secure your account in minutes: ​🤖 The Grok Audit Prompt (Copy & Paste) ​"Analyze my X account (@my\_handle) and find all tweets (originals and replies) that could reveal my real identity or make it easier to identify me. Look specifically for: ​Links to GitHub, GitLab, LinkedIn, Medium, personal sites, portfolios, emails, Discord, Telegram, WhatsApp, etc. ​Mentions of my real name, nickname, school, university, company, or job title. ​Mentions of my city, country, nationality, or local community. ​Named personal projects (libraries, repos, tools I created). ​Photos or images that could identify me. ​Any content linking my anonymous handle to my real-world identity. ​For each tweet found, provide: ​1\. Direct link (https://t.co/garXe1NALP...) ​2\. Date ​3\. Short content excerpt ​4\. Why it is risky. ​Rank them by risk level (High / Medium / Low), starting with the oldest and most revealing ones." ​📋 Quick Guide ​1\. Replace @my\_handle with your actual X handle. ​2\. Paste the prompt directly into Grok. ​3\. Follow up with: "Give me only the direct links so I can delete them" or "Deep-scan my older history." ​4\. Open the links one by one and delete the identified posts. ​🔒 Extra Operational Security (OpSec) Tips ​Bio Audit: Clean your bio of precise locations, job titles, or company names. ​Profile Links: Double-check external links in your profile. ​Fresh Starts: Consider migrating to a clean, isolated account if your historical exposure is too high. ​Context Awareness: Avoid replying with hyper-local or personal details, regardless of the language used. ​Account Security: Enable 2FA and regularly audit connected third-party apps. ​#Grok #AI #CyberSecurity #DigitalPrivacy #HumanRights #Journalism #Activists #ElonMusk #OpSec #TechSafety #DataProtection

2 赞 · 0 转帖 · 0 回复

### [@Untethered\_X](https://x.com/Untethered%5FX/status/2120951007396204611?ref=binnn.top)

11月13日 03:45 JST

我感觉走进了画的世界 有大佬用Claude Fable 5.2 做了一个可以直接在浏览器里游玩的游戏 Demo。 你乘上一艘小船，穿越 8 个如梦似幻的世界。 与其说是游戏，不如说是 AI 为人类编织的一场可以亲身体验的梦。 https://t.co/6VDMJo63No

2 赞 · 0 转帖 · 3 回复

### [@duoyun\_77](https://x.com/duoyun%5F77/status/2120994041143693462?ref=binnn.top)

11月13日 06:36 JST

Claude 现在能直接进视频会议了 不是录音、会后出纪要那种，是你们还在讨论，它已经把文档做出来了。 想要什么，喊一声「Claude，做个文档」就行，会议结束，纪要、待办事项和转写也都给你备好 。 以前开会：会上一小时，会后干三天 以后开会：会还没开完，活已经干完一半 https://t.co/NVwfgGiFZ3

2 赞 · 0 转帖 · 4 回复

### [@krypxbt](https://x.com/krypxbt/status/2120990569346339151?ref=binnn.top)

11月13日 06:22 JST

THE MEETING NOW ENDS WITH THE DOC ALREADY WRITTEN. Claude Meetings joins your call and builds while you talk: → docs → whiteboards → other artifacts → beta on Team and Enterprise the slowest part of a meeting was never the meeting. it was the 2 days after, waiting for someone to write it up.

2 赞 · 0 转帖 · 1 回复

### [@autthakorn](https://x.com/autthakorn/status/2121024889519837258?ref=binnn.top)

11月13日 08:38 JST

a free ChatGPT that runs on your computer and reads your files it's called AnythingLLM a desktop app for Mac, Windows and Linux, no account needed chat with your PDFs and docs, everything stays on your machine build AI agents without code: web search, SQL, files, charts works with local models or 30+ providers like OpenAI and Anthropic, plus MCP tools new: on-device dictation and AI text actions inside any app on your computer free and open source, 68,000 stars on GitHub

1 赞 · 1 转帖 · 0 回复

### [@shanke9527](https://x.com/shanke9527/status/2120944169325863158?ref=binnn.top)

11月13日 03:17 JST

Claude 的智能体也能自己托管了。Anthropic 开了自托管 Managed Agents 私测，编排、沙箱、凭证、会话历史全放你自己的云或机房，模型请求走哪家你自己定。说实话，智能体开始碰生产代码了，管得住才敢放手，这步就是给大公司吃的定心丸。 https://t.co/ZAtKaZqHgw https://t.co/IiibGwutKR

1 赞 · 0 转帖 · 0 回复

### [@LFCXYXSH](https://x.com/LFCXYXSH/status/2121015207900189140?ref=binnn.top)

11月13日 08:00 JST

以前开会浪废的是员工的时间，现在把 Claude 拉进视频会议里全程实时听会、画白板、改文档，废的是公司的 Token 预算，一小时废话开下来，纪要没写几行，估计 API 账单先烧穿了 ！

1 赞 · 0 转帖 · 5 回复

### [@aibluks](https://x.com/aibluks/status/2120982496636788796?ref=binnn.top)

11月13日 05:50 JST

Microsoft researchers demonstrated an AI model backdoor created by editing fewer than 10 neurons. The model could behave normally until a hidden trigger appeared. The catch: the attack requires access to modify the model’s weights. Model integrity is becoming an AI security issue. #LLM #AISecurity @grok

0 赞 · 0 转帖 · 1 回复

### [@woncoo1](https://x.com/woncoo1/status/2121054842080964679?ref=binnn.top)

11月13日 10:37 JST

Open weights keep coming: MiniMax's H3 Image, now Inkling 2 Small (594B MoE, Apache 2.0). The models are free; the work is distribution. Still $0 revenue on TruePen — showing up daily where writers actually are.

0 赞 · 0 转帖 · 0 回复

### [@sabrielatmojo](https://x.com/sabrielatmojo/status/2121059037538050198?ref=binnn.top)

11月13日 10:54 JST

Your coding agent isn't bad at .NET. It just doesn't have the playbook. Microsoft quietly fixed that 👇 h/t @AntonMartyniuk, who spotted that Microsoft keeps 99 official, open-source .NET skills in the dotnet/skills repo, grouped into 14 plugins. They work in Claude Code, Codex, GitHub Copilot and Cursor. A skill is just a folder of instructions the agent loads only when a task needs it, so you write the steps once and it follows them every time. Here's how I'd roll it out, step by step: 1\. Don't install all 99\. Every skill adds its description to the agent's context, so pick the plugins that match your stack. 2\. Start with the high-leverage ones: dotnet-upgrade (moves .NET 8 projects to 10 and turns on nullable reference types), dotnet-data (its optimizing-ef-core-queries skill catches N+1 queries and EF Core traps) and dotnet-test (22 skills, the biggest plugin). 3\. Name the skill in your prompt instead of letting the agent guess which one to use. 4\. Write your own skills for house rules: where your entities live, whether you use repositories, what your handlers return. Microsoft's skills know .NET, not your project. 5\. Run the same skills across models. Skills live in your repo, not in one vendor's chat app, so you can point Claude Code at an Anthropic-compatible endpoint like https://t.co/kheLxYU7TG and swap models per task: 181 models across 20 families on one key, pay per token, about 76% below official prices on average right now. Heavy model for the upgrade, cheap one for test runs. Anton's full guide covers the 4-step install and the Skill Creator. Save this if you ship .NET with agents.

0 赞 · 0 转帖 · 0 回复

### [@lucidui\_](https://x.com/lucidui%5F/status/2121055898668048461?ref=binnn.top)

11月13日 10:41 JST

The LucidUI press release is now publish TORONTO, ON, November 12, 2026\. Lucid UI today launched a free, open-source UI runtime built for the way software is now written: largely by AI coding agents. Its complete API reference is about 1,140 tokens, small enough for an agent to read before it writes a line. \[..\] https://t.co/AHDg9Fez2D https://t.co/k6ggzmisnO

0 赞 · 0 转帖 · 0 回复

### [@bobm67](https://x.com/bobm67/status/2121049088586125569?ref=binnn.top)

11月13日 10:14 JST

All nine frontier and near-frontier models in an arXiv study write text that's harder for you to process than human writing. The paper "Why is AI-generated language so hard to understand?" measured psycholinguistic features tied to processing difficulty. Every model leaned harder than human writers on inanimate subjects paired with animate verbs and low-frequency n-grams. Most stacked nouns and stretched dependencies too. The authors' explanation: AI text is optimized for math, code, and reasoning, where information density beats communication. They tested it by fine-tuning a model under a length penalty. It started favoring shorter syntax even when that made output less humanlike. A separate arXiv paper found the counterweight: human-LLM interaction produces more human-like vocabularies. So keep yourself in the loop. If a sentence forces a reread, rewrite it before customers see it. #LLMs #AIWriting

0 赞 · 0 转帖 · 0 回复

### [@lucas\_r\_senchal](https://x.com/lucas%5Fr%5Fsenchal/status/2121051272442773532?ref=binnn.top)

11月13日 10:23 JST

Harvey is a free, open source Claude Code agent that finds leads, reads their sites and writes each cold email around one real detail. Every email waits for your approval by default. You still need a separate domain and weeks of warmup. #AI #ColdEmail https://t.co/yU5KUJIYue

0 赞 · 0 转帖 · 1 回复

### [@luoxiaoshan\_ai](https://x.com/luoxiaoshan%5Fai/status/2120906630720356827?ref=binnn.top)

11月13日 00:48 JST

https://t.co/XMqTGTxxBm

0 赞 · 0 转帖 · 1 回复

### [@realfxw](https://x.com/realfxw/status/2120898662914216347?ref=binnn.top)

11月13日 00:17 JST

🚨 科研人重磅福利！Google Antigravity 开放 6 个月全免费申请通道！ 最近 Google 官方针对科学研究与学术领域推出了专属支持计划（Google Antigravity for Science），符合条件的研究人员现在可以提交申请，直接解锁长达半年的顶级 AI 科研助力！ ✨ 核心福利亮点： 1️⃣ 6 个月完全免费：零成本深度体验 Google Antigravity 完整服务，大幅降低日常科研与实验分析门槛。 2️⃣ 解锁全部 Ultra 模型：申请通过后即可获得所有 Ultra 级别模型的完整访问权限，无论是海量文献综述、复杂数据处理、代码调试还是科研灵感推演，都能把效率拉满。 3️⃣ 稀缺名额筛选：目前披露的获批比例约为 3,000 / 27,000（每 2.7 万份申请中批准 3,000 个名额），整体竞争相当激烈，建议符合条件的朋友尽早提交。 💡 申请小贴士： 该计划主要面向活跃在一线的科研工作者与高校研究人员。填写申请表时，建议使用所在高校或科研机构的官方学术邮箱，并清晰描述当前的研究课题与实际应用场景，有助于提高审核通过率。 正在做研究的小伙伴赶紧去官方页面查看详情吧 👉 https://t.co/fX58cVPMCZ #GoogleAntigravity #AI科研 #人工智能 #科研工具 #学术福利

0 赞 · 0 转帖 · 0 回复

### [@duduyun\_net](https://x.com/duduyun%5Fnet/status/2120933697042235432?ref=binnn.top)

11月13日 02:36 JST

🔥Github热门项目｜memary/memary 🧠 结合知识图谱与长期记忆层的开源 AI Agent 架构。它能够为大模型自动构建持久化语义记忆与实体关系网，打破传统上下文窗口限制，让智能体在多轮复杂对话和长期任务中保持高度的一致性与连贯性。 度度云注册就送一个月 https://t.co/7GvDjgcvHJ

0 赞 · 0 转帖 · 0 回复

### [@Zhibin992](https://x.com/Zhibin992/status/2121058924946436143?ref=binnn.top)

11月13日 10:53 JST

数据合作团队｜数据采集、视频标注与质量审核 团队简介 我们是一支面向人工智能与具身智能项目开展数据采集、标注及质量审核工作的团队，重点围绕人类操作视频、多视角视频和任务相关数据，提供采集、标签核验、问题复核及修订支持。 目前项目涉及第一人称、外部观察和腕部相机等多视角视频。我们关注视频与任务要求是否一致、不同视角是否同步、机位标签是否准确，以及场景和任务信息是否清晰，帮助客户整理符合项目规范的数据。 服务内容 视频标注与标签核验 根据客户提供的标注规范，对视频的场景、任务及相机视角等信息进行标注、检查与修订。具体字段、标签定义和判断口径在项目开始前共同确认。 多视角视频同步审核 检查同一任务在第一人称、外部观察及腕部相机画面中的对应关系，识别不同视角之间的时间不一致问题，并记录需要复核或处理的位置。 场景与任务有效性审核 对照项目要求，审核视频内容是否对应指定场景和任务，检查标签与实际画面是否一致，整理不符合要求或需要进一步确认的数据。 标注错误复核与修订 对已有标注结果进行复核，核查错误、遗漏和不一致之处，根据确认后的规则完成修订，并整理问题记录，方便客户验收与后续追溯。 人类操作视频采集支持 围绕客户确认的任务，开展第一人称、外部观察及腕部相机等多视角采集，并配套进行同步、机位、场景和任务审核。具体任务与采集条件需提前对齐。 合作方式 我们支持围绕客户已有视频开展标注与审核，也可讨论采集与审核相结合的项目。 建议先选择一小批具有代表性的数据，开展付费试单。双方先确认任务范围、标签定义、示例、交付格式及验收标准，再根据试单结果确定后续工作量、交期与报价。 对于存在歧义的数据，我们会整理问题并与客户确认，形成统一的判断口径。 交付资料 根据项目约定，可整理标注或审核结果、问题清单、修订记录及必要的数据说明。具体文件格式、字段内容和交付频次由双方确认。 合作方向 重点关注具身智能、移动操作机器人及相关人工智能团队的人类演示视频、任务标签、多视角视频审核和数据质量复核需求。 业务联系 数据合作团队 Rodata@bestensured.com

0 赞 · 0 转帖 · 1 回复

### [@zhangcheng0688](https://x.com/zhangcheng0688/status/2121035534466163065?ref=binnn.top)

11月13日 09:20 JST

今天把收藏夹里的「Mirasim」从 0 讲到 1（角度：痛点切入）。 【是什么】官网把自己定位为一款面向 Agentic Coding 与 Eval 的「One Agent IDE」：把智能体写代码和评测放在同一个工作面里。能力细节以当日页面为准。 公开信号：以官网 https://t.co/gmvrhAkfsp 2026-11-13 页面说明为准。 入口：https://t.co/lYU5KWfrmr 【痛点】很多团队能让 Agent 写出一段代码，却没有固定评测：什么叫完成、回归怎么跑、同一任务换模型如何对照。没有 Eval，口头上的「更好」无法复现。 【怎么装/怎么开】打开 https://t.co/gmvrhAkfsp，按官方说明试用。先导入一个有测试的非生产小项目，规定完成定义，跑一轮生成和评测，保存日志后再换一个模型对照。 【最小路径】导入有测试的小项目 → 写完成定义 → 跑评测并保存日志 → 对照第二个模型 → 没有对照就不谈切换。 【适用场景】正在为编码 Agent 选模型或做内部评测的研发负责人。 不适合：没有数据主人、没有回滚、也没有业务指标的「先上再说」。 【我的判断】Agent IDE 如果不能留下可复现的 Eval 日志，就还是聊天窗。先评测，再扩权限。验收KPI：评测可复现、完成定义覆盖关键路径、对照日志可分享。 对照四大板块（基建站/获客/AI核心/转化）先问它咬合哪一块。分发人口径：是什么→痛点→怎么装→最小路径→场景→判断。 我是张程Peter，电缆实业×AI，深港双城。不卖工时，卖结果。 #AI工具 #AI编程

0 赞 · 0 转帖 · 0 回复

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