2026年11月19日 · AI 早报

本轮重点是 Anthropic 公布 IPO 定价、Google 推出 Gemini 4 Neon 与 Genie 4,以及 OpenAI 披露智能体参与研发的内部数据。技术讨论集中在自我改进、开放模型的任务适配,以及让智能体可靠执行和积累经验的工程方法。

11月18日 23:00 — 11月19日 11:00(日本时间) · 130 条 X 原帖

查看本期全部 130 条原帖 ↓

本期目录

公司与产品:上市定价、新模型和芯片设计

Anthropic 定价每股 83 美元

Anthropic 官方宣布 IPO 定价为每股 83 美元,预计于公告次日在纳斯达克开始交易。这里应区分定价与开盘:本批官方信息确认的是发行价格和预期交易时间,尚无实际交易表现。

来源 · 1 条原帖

Gemini 4 Neon 开始推送,Genie 4 加入实时编辑

Google DeepMind 官方称,Gemini 4 Neon 定位于 Carbon 与 Hydrogen 之间,面向复杂软件工程、知识工作和科学任务。社区账号 xiax 转述的同提示词动画比较认为,Neon 的动作连贯性和完成度接近 Carbon;这只是具体案例的视觉评价。

另一项正式发布是 Project Genie 中的 Genie 4,新增原生音频、实时编辑和辅助搭建虚拟世界的智能体。刘超的介绍补充了社区画廊及订阅门槛,但其本人仍在询问实际运行时长和画面稳定性,并非完整实测。

来源 · 4 条原帖

OpenAI 报告芯片优化成果,AI 制药融资受关注

OpenAI 官方称,与 AMD 合作不到五个月,在已经过多年优化的旗舰 GPU IP 上找到最高 14% 的 PPA 改善;同类方法也用于其首款定制推理芯片 Jalapeño,相关收益经过真实芯片验证。公司还披露正与 Synopsys 合作,将模型能力接入芯片设计与验证工具。

生物医药方面,社区作者扶光转述 Isomorphic Labs 获得 50 亿美元 C 轮融资,资金将用于推进肿瘤和免疫方向候选药进入临床。该帖没有提供官方估值确认,融资与临床推进也不能直接等同于药物获批。

来源 · 2 条原帖

模型与开放权重:价格之外,任务适配更具体

旗舰闭源与低价开放模型并行

Maha 的模型汇总转述,GPT-6.2 Astra API 每百万输入、输出 token 分别为 8 和 40 美元;Qwen4-Max 则提供原生视觉和百万 token 上下文,但旗舰权重不开放,较小 Qwen4 模型的开放时间尚未给出。这些是本批社区汇总中的发布信息,不应全部当作采集当天官宣。

同一汇总介绍腾讯 Hy4 Flash 已提供权重,每 token 激活约 12B 参数;OVHcloud 的 Olala mini 7A1B 则主打小规模部署,权重计划以 Apache 2.0 提供。旗舰能力、调用价格与可自行部署,正在形成不同选择。

来源 · 1 条原帖

为 OpenCode 专门训练的 27B 适配器

Lysandre 发布 Qwen3.8-27B 的 OpenCode 专用 RL LoRA,支持本地运行。作者报告,关闭思考时,SWE-bench Pro Hard 51 的 pass@1 从 35.3% 升至 51.0%,Terminal-Bench 2.1 从 47.0% 升至 53.0%;前一测试只含 51 个困难样本,结果仍需按这一范围理解。

作者还称,相比基础模型开启 xhigh 思考,适配器关闭思考后在两项测试中差距均不超过 4%,同时分别只用约八分之一、六分之一 token。这说明围绕特定智能体运行方式训练,可能同时改善成功率和开销。

来源 · 1 条原帖

垂直模型瞄准生产故障与语音识别

AGTP 转述 ResolveAI 的 27B 模型 SRE-1,针对生产告警根因分析训练。其五次运行集成在公司自建的 49 个困难生产案例上达到约 71% 通过率,声称以更低成本匹配 Fable 5.2 的 xhigh 档;比较范围限于这套自建评测,不能外推为通用能力领先。

语音侧,Poonam Soni 介绍 Rumble-2-Flash:规模 0.5B,支持 45 种语言,权重采用 Apache 2.0。相比笼统的榜单宣传,小模型的语言覆盖与许可信息更适合直接用于部署筛选。

来源 · 2 条原帖

研发加速与自我改进:运行更多,还要看成功在哪里

OpenAI 披露参差不齐的研发能力边界

OpenAI 官方发布内部研究加速数据,称智能体能完成更长任务,但人类仍处于研究核心,并呼吁实验室公开追踪递归自我改进指标。公司同时介绍模型全生命周期的安全做法,称安全与对齐的人力投入增长快于其他单个研究领域。

Xudong Han 对报告的解读给出关键区别:研究员平均每天调用相当于 5.7 个工作日的智能体运行量,但这不是节省的人类工时。内部估算中,监控训练的自主成功率为 79%,编写推理与服务系统代码仅为 25%;实现和调试加速,仍可能被需要人工判断的环节卡住。

来源 · 3 条原帖

冻结权重,也能自己构造学习流程

Elvis 介绍 MIT CSAIL 与图宾根大学的研究:给冻结模型计算预算和简单工具,让它编写练习游戏、观察另一个自身副本,并修改提示词和技能文件;不提供真实任务、验证分数或学习方法。能力较强的模型自行形成难度校准、A/B 比较和验证流程,并改善隐藏机制游戏表现。这里的学习发生在文本与流程层,没有更新模型权重。

来源 · 1 条原帖

另一条路线是让模型直接改写权重

Jiajun Fan 发布论文作者自述,将模型自身权重、已评分答案和运行自身的能力交给模型,让其诊断错误并提交权重修改。作者报告,未见题目准确率提高 6.6 个点,连续八轮实验继续改善;较小模型未能成功。其“不用梯度”的方法与前述技能文件学习不同,但这些实验还不足以证明持续、通用的递归自我提升已经实现。

来源 · 1 条原帖

推理与机器人:少读、少传,以及物理任务的代价

长上下文选择读取,MoE 选择传输增量

Constructor 团队 Michael Gao 介绍 Link,以可学习的稀疏读取让模型选择关注位置。按作者结果,在 4M token 上下文中,相比稠密注意力,检索准确率改善 53%,GPU 效率为 9.9 倍;这是特定长度和任务下的报告。

Siva Reddy 介绍 EchoMoE,把跨节点专家通信从完整表示改为增量,可改造已有预训练模型。另一个本地实践来自 Beamsters:64GB M5 Max 上,Sushi 1.7.0 配合 DeepSeek STQ 1.75bpw 达到每秒 40–60 token;通过偏向缓存专家取得约两倍加速,同时存在精确性取舍。

来源 · 3 条原帖

机器人学习要同时算示范、算力和恢复成本

Hina Dixit 转述 ScaleVLA 预印本:冻结的 552B 稀疏视觉语言骨干,在模拟 LIBERO-Goal 中用 10 次示范达到 87.8% 成功率,对照的 2B 骨干用 50 次示范达到 85.6%。真实机器人插充电器在熟悉布局成功 9/10,布局变化后为 8/15,推理需要四张 H200。

执行框架方面,NoaX 开源 openRocode,将一句话转成代码,再以受控运动执行并逐步记录观察,支持 Claude、GPT 及兼容接口模型。它解决的是操控与记录接口,不代表已经解决物理环境中的泛化问题。

来源 · 2 条原帖

科研实践:发现、纠错与写作各有边界

从审论文到智能体协作寻找物理结构

Valerio Capraro 发起科学文献纠错项目,设想由智能体审阅论文并生成修订稿。其灵感来自一项理论天体物理审计:筛查五万多篇论文,标出 872 个可能错误。这里应保留“可能”,项目愿景也不是已经完成的全学科纠错。

Karthik Duraisamy 转述的一场研究报告则描述去中心化智能体通过共享留言板协作,搜索分数拓扑态结构,报告能隙提升 10–15 倍并形成解析解释。GenBio 同时转述纽约时报对虚拟细胞的报道,讨论用世界模型模拟细胞变化,以及计算机能否覆盖生命复杂性。

来源 · 3 条原帖

论文生成工具改善流程,仍未超过人工原稿

Vansh Bhavsar 介绍 PaperOrchestra,把想法笔记和实验日志交给五类智能体,分别负责提纲、图表、文献、正文及模拟评审后的修改。其转述的 11 名研究者盲评中,文献综述优于两种自动化基线,但人工原稿仍领先。

陈跑跑介绍的 scientific-agent-skills 覆盖生物、化学、医学、计算和论文流程,并明确工具不保证实验结果与结论正确。两类项目都更适合被理解为流程支持,引用核验与研究判断仍需单独完成。

来源 · 2 条原帖

智能体工程与 DeepSeek:成本、执行流程和数据入口

DeepSeek 搜索计费变化,插件体验仍属社区反馈

Ted Cai 与 openwell 均转述 DeepSeek 文档:北京时间 11 月 18 日中午起,模型调用触发的服务端搜索、网页抓取不再单独计费,只计模型 token;独立 Search、Fetch API 仍为每千次成功请求 5 元。这一区别直接影响调研和新闻类智能体的成本估算。

ahhhhfs 分享为 DeepSeek Harness 安装插件市场后的积极体验,但未给出插件范围或量化结果。另一则社区段子抱怨审代码时智能体尝试 SSH,反映用户对权限边界的担心,不能据此认定具体安全事故。

来源 · 3 条原帖

让程序持续执行,让验证过的经验留下来

Xudong Han 介绍 LMMs-Lab 的早期项目 The Harness:模型在关键节点规划、写程序,程序与反应规则继续执行,处理不了再交回模型;有效做法通过测试后可在同一任务内存成 Skill。deepagents 作者则宣布下载量达到五千万,强调其开放、跨提供商的长任务框架定位。

流程治理也是本批工具的共同方向。Superpowers 强调需求、计划、测试与审查;LoopForge 的社区介绍则突出多助手分工、人工审批、失败返工和 SOP 沉淀。它们试图把完成标准写进流程,减少仅凭智能体口头宣布完成的情况。

来源 · 4 条原帖

数据、记忆与授权接入继续补齐

Agent-Reach 的重复介绍可合并为一个用途:给智能体提供跨平台检索、字幕和网页读取入口,部分平台仍需登录或额外接口。Jeremy Morgan 介绍的 Relate 则把图记忆存为单文件,以 Cypher 查询;Agenvoy 作者称,新版先用决策模型过滤网页,减少约九成主模型输入。

旺旺转述 Claude Platform 的 Linked Billing Beta,支持 OAuth 连接组织、设置消费上限和撤销授权,费用可归客户组织,明确不等于共享 Pro/Max 订阅。学习资源中,Stanford CS329Z 和中文 awesome-agentic-ai-zh 分别覆盖系统工程课程与循序渐进练习,适合补足检索、记忆、评测和安全基础。

来源 · 6 条原帖

开源创作工具:把检查步骤做得更具体

界面和动效开始依赖可见结果验收

陈发财介绍 dashboard-design-skill,强调小样本指标的表达,以及在三种宽度下截图自查;reanimate 则录制参考和复刻动效,逐帧比较后继续修改。Hallmark 的社区介绍补充移动端、对比度、无效表单和编造内容检查,重点都从“生成了页面”转向“实际呈现是否合格”。

diagram-design 可把流程转成包含 SVG 的单文件网页,再导出图像;Humanizer 则按常见文风问题改稿,并参考用户旧文,同时检查事实数字是否被改动。

来源 · 5 条原帖

整书翻译、漫剧模板与自动短视频分清交付物

LanfAI 介绍 translate-book,支持 PDF、DOCX、EPUB,并按章节并行翻译;其介绍的 AI 漫剧 Skills 则串起企划、人物、剧本、分镜和发布文案,但交付的是模板与提示词,不直接生成成片。Reelset 的社区介绍更进一步,称可把项目与研究链接转为带配音、字幕的竖屏视频,并提供 MCP 接口。

更具体的创作记录来自 Grip 作者 Simeo:他用 Claude Code、Node、ffmpeg 和 WebGL2 制作发布视频,让模型分析音乐节拍、编排真实录屏,并通过联系表检查画面、完成两轮修改。它展示的是代码与素材驱动的视频制作流程。

来源 · 4 条原帖

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 130 条原帖

@AnthropicAI

We’ve priced our IPO at $83 per share, with trading expected to begin on Nasdaq tomorrow. https://t.co/NJxt2pILrh Legal disclaimer: A registration statement relating to the Class A common stock was declared effective by the Securities and Exchange Commission on November 18, 2026. The offering is being made only by means of a prospectus. Copies of the final prospectus may be obtained, when available, from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at (866) 471-2526, or by email at prospectus-ny@ny.email.gs.com; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014, or by email at prospectus@morganstanley.com

@GoogleDeepMind

Gemini 4 Neon is rolling out now. Sitting between Carbon and Hydrogen, it brings frontier intelligence and fast task completion for complex software engineering, knowledge work, and science.

@OpenAI

Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing. The same approach helped improve Jalapeño, our first custom inference chip, with gains verified on real silicon. New evaluations also show our models getting better at certain chip design tasks. We’re sharing these results and how we’re working with @Synopsys to bring our model capabilities and their trusted EDA design and validation tools to more chip designers. https://t.co/4BNcPIqfIQ

@GoogleDeepMind

Genie 4 is now available in Project Genie. It brings native audio, real-time editing, and a new agent to help you create interactive, virtual worlds.

@OpenAI

We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. https://t.co/xyGPcCkPCK

@OpenAI

We’re sharing how we approach safety at every stage of a model’s lifecycle, from early research to internal use to public deployment.

@ValerioCapraro

Today I’m launching what I think could be an extremely important research project. Fixing humanity’s scientific knowledge. The idea is inspired by physicist Casey Handmer, who used AI to screen more than 50,000 papers in theoretical astrophysics. His audit flagged 872 possible errors. This is impressive. But why stop at finding the errors? Let’s use AI to FIX them. And why stop at astrophysics? The ambition of this project is to go through every scientific paper ever published, across every scientific discipline. AI agents will review each article and produce a revised version that corrects the errors they find. The aim is not to shame authors or seek retractions. We are humans and we all make mistakes. The ambition of this project is to correct all the errors that we have made so that current and future generations of scientists can build on firmer foundations. I’m looking for researchers from every discipline to help make this happen. If you’d like to take part, please fill out this short survey and tell me which disciplines you’d like to help cover. Feel free to share with anyone interested. Survey link: https://t.co/t2vNE0XGt2

@LysandreJik

Introducing Qwen3.8-27B-opencode-adapted 🔥 Yesterday we released our Pi adapter. Today we’re releasing a LoRA adapter trained with RL specifically for the @opencode coding agent; runs locally! Here’s how pass@1 changes from the base model → OpenCode adapter. With thinking OFF for both models 🚀 SWE-bench Pro (Hard 51) · 35.3% → 51.0% 💻 Terminal-Bench 2.1 · 47.0% → 53.0% With thinking ON (xhigh) for both models 🚀 SWE-bench Pro (Hard 51) · 47.1% → 54.9% 💻 Terminal-Bench 2.1 · 55.4% → 56.6% ⚡ With thinking OFF, the adapter lands within 4% of the base model with thinking on xhigh on both benchmarks, while using just 1/8th as many tokens on SWE-bench Pro, and 1/6th on Terminal-Bench. Which one should we do next?? Open weights: https://t.co/FiQIs92Wrs Read more about how we do this here: https://t.co/1vZxJ0d5EY

@abskoop

DeepSeek Harness 安装个插件市场 似乎打开新的世界了 https://t.co/38Cx3XyKjJ

@0xSero

Literally the most important skill for anyone with interest in local AI to use. If you work on inference this is my crown jewel. Open source must win.

@mingzi206

单月狂揽超 10 万次下载!12G 显卡真正的“无审查神仙底模”藏不住了! 31B 太大普通显卡直接爆显存,4B 很多高难长逻辑又啃不动。 这款基于 Google 官方底模切除道德枷锁的 Gemma 4 12B Heretic 消融版,被极客社区狂下了 10.5 万次,直接成为了 12G 显卡玩家的标配神兵: · 狂卷 10.5 万次下载:毫无宣发的纯民间消融版,硬靠口碑在 Hugging Face 冲破 10 万大关 · 彻底切除拒答神经:移除了谷歌原版动不动弹出的道德说教,问什么秒答什么,零说教、零废话 · 12G 显卡的绝对甜点位:量化后仅 7~8GB 体积,普通的 12G 显卡或 16G 内存 Mac 直接轻松满血吃下,不用蹭昂贵云端算力 · 满血保留推理与多模态:走低散度向量消融,代码、长逻辑推理甚至多模态能力分毫未损 · GGUF 格式已全盘就绪:丢进 Ollama 或 LM Studio,单机离线直接开跑 手里只有 12G 显存、想在本地长驻一个既聪明又有问必答的私有化助手的兄弟,这个 10 万人验证过的底模建议直接存死! 官方 GGUF 仓库: https://t.co/5X4FPZugmZ #AI #开源 #Gemma #大模型 #Ollama

@whereisvivien

https://t.co/7LoievSYJB

@mentalium_

This might be the only FREE Stanford course you need on building production AI agents Anthropic pays $750,000+ a year to people who can engineer real agent systems Stanford CS329Z: Engineering AI Agents Lectures just started dropping publicly 01 What Are AI Agents? https://t.co/s4HiqTkc7F 02 LLMs for Agent Builders & Context Engineering https://t.co/HuiblR5JY7 03 RAG Explained: From BM25 to Agentic RAG https://t.co/FAepP5Zupg 04 How AI Agents Use Tools, Function Calling & MCP https://t.co/FKMbjQuPl3 05 DSPy vs LangGraph: Agent Frameworks https://t.co/jGxy2HZeZh 06 AI Agent Design Patterns: ReAct, CodeAct & More https://t.co/NMBaImBEgU 07 How AI Agent Memory Works https://t.co/fnZBr70YRq 08 Multi-Agent Systems: When More Agents Help or Hurt https://t.co/xQxKLbegxX Full playlist (more lectures coming): https://t.co/gVPiNHTcdk Course site: https://t.co/rlgOhWCigU You either learn how to engineer agents properly or keep shipping broken demos for another year. Save this ↓

@github

Copilot code review got major upgrades 🚀 What we announced at #MSIgnite 👇 https://t.co/CYZz18vNZK

@DAIEvolutionHub

Google just dropped a 1-hour course on agentic engineering from scratch: 00:00 – How to build your first AI agent 08:24 – Build agent memory (short, persistent, long) 28:34 – Agentic loops, long-running AI agents 40:04 – How to build MCP (MCP vs API) 1:00:22 – Multi-agentic systems This 1-hour watch will replace 10 paid agentic courses on the internet. Bookmark this. Watch this weekend.

@0xCodez

These 20 GitHub repos will turn your Grok Bot into your most productive team ever This is f*cking gold. Skills. Team. Plan. Memory. Tools. 01 / Lauren Tan's noodle → Orchestrate agents using skills. https://t.co/vFB7oVI108 02 / mattpocock/skills → Skills for real engineers, straight from his .agents directory. https://t.co/9ebNa9zAqX 03 / agent-skills → Production-grade engineering skills for AI coding agents. https://t.co/m5gnOcIcw7 04 / superpowers → An agentic skills framework and dev methodology that works. https://t.co/fw8ESyFwEj 05 / Vercel agent-skills → Vercel's official collection of agent skills. https://t.co/3lpMunBwiJ 06 / awesome-agent-skills → 1000+ curated skills from official dev teams and the community. https://t.co/mu2qIfdA0Z 07 / Gas Town → Multi-agent workspace manager. https://t.co/b8hlOdN8D4 08 / gstack → 23 opinionated tools: CEO, designer, eng manager, release manager. https://t.co/4IoXTdTNXh 09 / BMAD Method → Agile, AI-driven development for agent teams. https://t.co/yCTzkK0avn 10 / grokbot-field-notes → Rules, playbooks and bot roles from a 72-hour Grok Bot live build. https://t.co/ZndBpCt8kN 11 / Spec Kit → GitHub's toolkit for spec-driven development. https://t.co/R41ImkLZfC 12 / Task Master → AI-powered task management you can drop into your agent. https://t.co/24AuHdDPvj 13 / AGENTS.md → A simple, open format for guiding coding agents. https://t.co/colCttpfuN 14 / Beads → A memory upgrade for your coding agent. https://t.co/jI0BgpbSiB 15 / mem0 → The memory layer for AI agents. https://t.co/E5F3k16LJW 16 / Context7 → Up-to-date docs for AI agents. https://t.co/FUgZM076pW 17 / npx skills → The open agent skills tool. https://t.co/JRPeQUTCZm 18 / Composio → 1000+ toolkits with auth and tool search. https://t.co/dZfnd8EzoS 19 / GitHub MCP Server → GitHub's official MCP server. https://t.co/vm8yDWWvBQ 20 / Firecrawl → Search and scrape the web for your agents. https://t.co/6K594aEQas Skill → plan → team → memory → ship. Bookmark this. Go build a team worth managing ↓

@sydneyrunkle

deepagents just hit 50M downloads!!! we started it a bit over a year ago, back before "harness" was an industry term we bet on a provider agnostic, open source harness for production agents that can take on long running, complex work in any domain https://t.co/TlXAJfmFnD https://t.co/0axUQEHX8Y

@HeLi8866

全球排名前 20 的 Skill,都进了这份必装榜!! 1、skills(by mattpocock),目前 32.46 万星 需求讲不明白时,它会倒过来问你,再帮你把架构理顺。查 bug、改老项目,用这个最顺。 2、superpowers(by obra),目前 31.77 万星 开发被它分成四步:先做头脑风暴,再写计划,接着测试驱动,最后做代码审查,整套步骤串在一起。准备做比较大的功能前先装上,AI 就不会上来直接胡写一通。 3、andrej-karpathy-skills(by multica-ai),目前 23.32 万星 这是一套项目里的规则:AI 要先想明白再动手,只动该动的部分。放到项目的根目录里,一写代码就会生效。 4、skills(by anthropics),目前 19.47 万星 这是官方提供的技能库,Word、PPT、表格和 PDF 都可以直接上手做。若要自己写 Skill,按它那套模板去改就行。 5、ponytail(by DietrichGebert),目前 19.10 万星 让 AI 按最偷懒的老工程师来想问题:能少加的那层抽象,就别加。代码完成后交给它看一遍,把多出来的设计和留下的技术债找出来。 6、ui-ux-pro-max-skill(by nextlevelbuilder),目前 14.96 万星 配色方案、字体组合和交互方面的规范,都收在里面。做产品落地页、后台和 APP 界面,先把它装上,做出来的页面更接近设计师的手笔。 7、graphify(by Graphify-Labs),目前 14.79 万星 能把代码、文档和库表结构,做成一幅知识图谱。刚接一个不熟的项目,可以先建图,再问“该函数是谁在调用”,一查就明白。 8、caveman(by JuliusBrussee),目前 12.02 万星 让 AI 尽量少用字来回答,把那些多余铺垫都去掉。要是 agent 得跑很久,装上它能省下相当大一截 token。 9、agent-skills(by addyosmani),目前 11.76 万星 由谷歌的工程师汇总的一套工程上的规范,接口怎么设计、上下文怎么管、质量怎么检查都覆盖到了。大家一起做项目时,可以用它把写法统一起来。 10、taste-skill(by LeonxLnx),目前 11.57 万星 给 AI 补一点审美:排版、参考图,还有动效,都讲究。做网站首页、作品集时用它,页面就不会一股模板味。 11、Understand-Anything(by Egonex-AI),目前 9.27 万星 能把代码转成可点击交互图,连业务流程也能讲明白。新人接项目、读开源的代码,最能省时间。 12、impeccable(by pbakaus),目前 11.03 万星 专门用来提高 AI 的设计水平,间距、层次和细节它都会查。页面做好了,先交给它看一遍,再按它的意见打磨。 13、awesome-claude-skills(by ComposioHQ),目前 7.95 万星 这是一套技能总汇,按使用场景收了 1000 多个 Skill。还不清楚该装哪一个,可以先到这里挑。 14、last30days-skill(by mvanhorn),目前 6.77 万星 它会去 Reddit、X 和 YouTube,把近 30 天里的讨论收在一起。 做选题、追热点,或调研别人家的产品时可以用。 15、humanizer(by blader),目前 6.92 万星 专门清掉文本里的 AI 痕迹,比如那些套话、刻板句式。公众号、口播稿,还有邮件,写完都可以交给它过一次。 16、diagram-design(by cathrynlavery),目前 6.72 万星 能支持 50 种图表,并直接导出为网页。写技术说明、做项目汇报时,可用它来画系统架构和流程。 17、i-have-adhd(by ayghri),目前 6.12 万星 让 AI 先讲结论和接下来怎么做,然后再补细节,别把答案藏在大段文字里。问技术上的问题、想立刻拿到答案,就装这个。 18、marketingskills(by coreyhaines31),目前 6.04 万星 文案、SEO,再加上数据方面的分析,这类营销能力都包括在内。写推广用的文案、想把网站的搜索排名做上去时可以用。 19、academic-research-skills(by Imbad0202),目前 5.69 万星 从查阅文献、写论文,一直到模拟评审,整条链路是串在一起的。写毕业的论文、准备投稿前,都可以用。 20、scientific-agent-skills(by K-Dense-AI),目前 5.72 万星 让 AI 成为科研方面的助手,把生物、化学这类方向常用数据库与配套工具接进来。

@wquguru

最全的Claude代理列表: Claude本身10+规则 支付:Google(含 Gmail、Google Pay)、Stripe 人脸:Persona、人机验证 遥测:Sentry、Datadog、Sift、Segment Shadowrocket/FlClash/Clash Party均可一键导入: https://t.co/3Ic1J6dmBa

@omarsar0

Do LLM agents still need to be told how to self-improve? A new paper from MIT CSAIL and the University of Tübingen tests this directly. They give a frozen model a compute budget and simple tools. It can write practice games, watch another copy of itself play them, and edit that copy's prompt and skill files. It does not see real tasks or validation scores, and it is not told how to learn. Capable models built a working learning procedure on their own and improved on held-out games with hidden mechanics. Models from different families arrived at similar strategies. They wrote games at a calibrated difficulty, ran controlled A/B comparisons, built their own validation suites, and spent more effort on harder tasks. The text files they produced also improved play on external benchmarks with different horizons, mechanics, and 3D navigation. This matters for safety as well. Leaving out an explicit self-improvement loop does not stop a capable model from improving itself when it has the tools. https://t.co/EJXwIYrGjd Chat with Paper: https://t.co/La9hjcKD7k

@aurielws

Gonna write a new post training blog soon while I have time during the holidays - anything people want me to cover? Few ideas I’m mulling on: -📊 Metrics to expect from Pretraining before you start post training - 🤔 SFT vs RL fundamentals - 🖼️ Introduction to Image and Video gen model Post training (not same as LLM!) - 🌶️ My recommended post training roadmap and for every big neolab based on what l think they suck at

@michaeltgao

We’re sharing the first phase of our research @constructor_agi. Our mission is to solve continual learning and develop models that learn natively from experience. Unlike LLMs, humans don’t reread their entire context; with experience, we learn what matters. Our work is a promising first step in that direction: it shows that a model improves when it learns where to look. Introducing Link: Teaching Attention to have intention. Link is a learned, sparse read that lets an LLM choose what to read. At multi-million context lengths, it outperforms dense attention on retrieval accuracy by 53% at 4M token context lengths, alongside 9.9x GPU efficiency.

@mdancho84

Some guy built a free AI university on GitHub. Not tutorials. Not theory. Full real-world AI systems. Step by step. Week 1 → Docker, FastAPI, databases Week 2 → Automated data pipelines Week 3 → Build your own search engine (BM25) Week 4 → Hybrid search (semantic + keyword) Week 5 → Full RAG system Week 6 → Production-ready (caching + monitoring) Week 7 → Agentic AI with LangGraph You don’t just “learn AI”. You ship it. GitHub (free): https://t.co/hVeMP1DBPr 🚨 P.S. Want to learn how to build + ship AI and Data Science projects (that businesses actually want in 2026)? I am hosting a free workshop to help you get started with AI + DS projects in Python. Register here (500 seats): https://t.co/onpLpRwkzH

@situking

写给 FDE、AI 企业服务同行: 拆解 2027 企业 AI 服务 10 大赛道:培训、咨询、PoC、实施、定制开发、知识库、微调、合规、运维、代运营。 AI 不是一锤子交付,模型效果会持续衰减,只修 bug 解决不了根本问题。客户买的从来不是一套系统,是稳定持续的业务效果。 https://t.co/zbzway1XWI

@Bundi_Shad

The future of enterprise AI isn't just about building smarter agents. It's about making them reliable, accountable, and capable of executing real-world workflows. That's exactly what makes SoduWork interesting. An open-source AI-native engine designed around smart dispatch, governance, and verifiable task delivery. A great opportunity for developers to explore and build: [https://t.co/AEV3IDTU4i](https://t.co/egfWCgCIDd)

@MaiYangAI

今天最值得看:@poteto 发了长文,讲怎么用 Grok Bot 把各种事自动化,是目前最完整的一份 Bot 自动化方法论。 最近24小时的高质量 Grok Bot 推文: 1. @poteto 发布长文《Automate Anything with Grok Bot》,拆解用 Grok Bot 把各种工作自动化的思路。 https://t.co/fa7B8p8QhZ 观看 167812 / 赞 1124 2. @ericzakariasson 宣布 Grok Bot 社区目录上线:可以浏览、安装别人做的 bot 和 app,也能发布自己的。 https://t.co/ZbpDmK8wbM 观看 107899 / 赞 1008 3. @benln 分享一个小技巧:让你的 @bot 去读官方最近的帖子,再给团队整理最佳实践和点子。 https://t.co/fG8gtxFNE4 观看 81496 / 赞 943 4. @XFreeze 发现 SpaceXSI 正在测试让 Grok Bot 直接在对话里帮你打电话,可以用自己的号码,也可以用临时号码保护隐私。 https://t.co/hz2cuEQ21p 观看 14544 / 赞 411 5. @XFreeze 介绍 SpaceXSI 会计团队已交给 5 个 Grok Bot 来跑,每个 Bot 负责一整项会计工作,只有需要判断或审批时才找人。 https://t.co/bIO0BC0GPQ 观看 10152 / 赞 173 6. @cgnot996 提出 Grok Bot「大抠门理论」,整理了一套省额度、让工作流稳定运行的实操方法。 https://t.co/sEqwCcKcVM 观看 15725 / 赞 99 7. @XFreeze 汇总 Grok Build 新一轮更新:会话可以改名、用 /models 刷新模型,MCP 集成和多 agent 工作流管理也更顺手。 https://t.co/izxsqIUqzc 观看 8044 / 赞 117 8. @XFreeze 演示 Grok Bot 在一个对话里规划并预订整趟行程,机票、酒店、租车的选项可以并排对比。 https://t.co/Md6FhufP3Q 观看 4406 / 赞 54

@ftcarpe

lab results come as a pdf. scans come as images. half your history is a photo of a paper AI can't help with what it can't read that's the part Unsiloed AI is fixing

@0x_meden

Learn AI by playing with it 1. A GPT you can walk through in 3D, step by step https://t.co/alGu4STO69 2. See how an LLM splits your text into tokens https://t.co/fNDbIbCFJr 3. Draw a digit and watch a neural net read it, layer by layer https://t.co/RtYnN5UWTF 4. Train your own image model with your webcam https://t.co/VXOyuGGoUP 5. Doodle something, a neural net guesses what it is https://t.co/XnCDks2c08 6. 10,000 words as points in 3D space https://t.co/HiAuyeJXwN 7. A word game scored by how close meanings are https://t.co/PIUzBCV7QA 8. Hack real AI apps with prompts, legally https://t.co/225JCPrUQZ 9. Spot the AI-generated face https://t.co/Z4ZCaYtDDI 10. Look inside an LLM's neurons https://t.co/qkkscPhVzc P.S. #5 has collected 50 million doodles from over 15 million players, and Google shared them all as a free dataset for training neural nets

@FuSheng_0306

大模型的强大体现的已经不是自己去调用工具了,而是直接根据任务生成工具,再让这个工具去调用工具...(新编 AI绕口令)

@NFT_Chen

💥信息量极大!昨晚北大AI创投课,Manus联合创始人张涛抛出的6个观点,足够让人重新想清楚:现在该怎么做AI产品、怎么选能力、怎么在AI时代活下去! 张涛:下一个AI创业机会,不在热门赛道,而在“模型还没完全长大”的缝隙里! 核心拆解如下: 1️⃣别只为今天的AI设计产品
要从6个月后模型能力大幅提升后的“外溢”中找机会。 
原则:Less Structure, More Intelligence,减少人为复杂流程,让模型自己做更多决策。底层模型升级,产品能力自然跟着涨。好创新应该持续吃技术进步的红利,而不是被它不断贬值。 2️⃣别只让用户用AI,要让他们真正感受到智能
真正优秀的产品要大幅降低使用门槛。用户不需要懂技术、不需要背提示词,就能直接享受价值。未来竞争的关键,不只是谁模型更聪明,而是谁能让普通人更自然地用上智能。 3️⃣AI创业最终还是要回到一门好生意
加了AI两个字,不代表可以理所当然持续亏钱。技术再先进,也必须回答:用户为什么愿意付费?创造的价值是否超过成本?技术先进≠商业成功,关注和用户量≠可持续价值。 4️⃣真正的机会藏在不确定性里
很多人想先搞清楚哪个赛道一定爆发、哪个方向肯定赚钱再动手。但真正的新机会,恰恰出现在这些问题还没有答案的时候。 5️⃣Manus名字的来源
来自MIT校训Mens et Manus(心与手)。大模型已经让机器越来越会思考,Manus要解决的是:让AI从回答问题走向真正完成任务。 个人层面也一样:今天花时间培养的能力,未来是因为AI变强而更值钱,还是更容易被替代? 这个问题值得反复问自己。 #AI创业 #Manus #张涛 #下一个AI机会 #LessStructureMoreIntelligence

@xitang_ai

https://t.co/T6WIIMigFM

@NFT_Chen

🤔离谱!$10 = $500 的 GLM 5.3 用量! 仅限本周末截止! 从 Phase 1 的 $100,到 $200,再到现在直接 $500 这波操作太离谱了 想重度用开源编程模型的赶紧冲 OpenCode Go 订阅用户可用 Claude Code / Codex 也能直接接入 错过这周就没了 #GLM53 #OpenCode #AI编程 #开源模型 #限时优惠 https://t.co/3CuAcXdb1S

@neropursue

building an app with AI is getting pretty easy. getting it ready for real users is where things get messy. @Netlify's new platform lets agents build using your GitHub, Notion or Stripe context, with live previews and backend hosting in the same place. less jumping between tools is always a win.

@Chenfacai0

打算入门 AI Agent,搜到的教程却零零散散,有些内容一开头就大谈多 Agent 协作,而学习者连 Token 到底是什么都还没搞清楚。 awesome-agentic-ai-zh 把学习路径理出了先后次序,从大模型基础知识一直覆盖到如何让 Agent 稳定可靠,共设有 10 个学习站点,汇集了 240 多个精选资源。 内容划分成几条不同路线:希望熟练使用 Claude Code、Codex 的可以走其中一条,打算从零开始亲手编写 Agent 的走另一条,暂时不写代码的人也有一份日常使用指南。 GitHub:https://t.co/IIhGIkQJ5x 专业术语第一次出现时,都会先用通俗说法解释一遍,再附上对应的英文原名,忘记了还能到名词表里查询,对新手相当友好。 每个学习站都配有可以直接上手运行的小练习。繁体、简体、英文三个版本齐全,用手机浏览时可以打开在线文档站。 刚接触 Agent、不清楚该从哪里入手的朋友,不妨按照它的顺序一个站点一个站点地学下去。

@fd_ripatel

bro used DEEPSEEK 深度求索 🛃🚨🚨 🇺🇸 federally mandated readme update 劳改 https://t.co/fiBXHDe8Lv

@Chenfacai0

这个叫 dashboard-design-skill 的 Skill 是专门冲着数据面板这类页面来的,作者是从 127 个已经上线的仪表盘里把经验提炼出来的。 它改动的点都挺具体,举个例子,销量由 1 单掉到 0 单时,不再甩一个红色的 −100% 上去,而是换成灰色小字「之前 1」,坐标轴的刻度也统一取 0、1k、2k 这类整数。 GitHub:https://t.co/wsJlMrEQGn 我最欣赏的是它最后收尾那一步——每回做完都得截图自查,网页要按 3440、1440、375 这三种宽度分别过一遍,对着屏幕上真实呈现的样子挑刺。 你也可以甩一张现有页面的截图过去让它评一评,它会先给结论,再按问题严重程度逐条列出。除网页之外它还附带了 SwiftUI 的写法,Codex、Cursor 只要指向它的说明文件同样能调用。 手头有后台管理页或者数据看板要做的,装上它能让 Agent 少返工好几轮。

@ShuaiMaLong

🚨 11月19日晨间羊毛汇总|免费VPS、Mistral Pro、Spotify全白嫖,还有人1万多刀额度一夜没了,建议先收藏 🔥 今日羊毛 1️⃣ https://t.co/4fqmuGAHfs 免费 VPS 白嫖一个月 波兰机房,1核1G内存 + 20G SSD,完整 Root 权限,想装什么装什么。还顺带送免费域名邮箱、SSL 证书和 MySQL,拿来练手、挂小服务、搭个测试站都够用。免费机通病就是别放重要数据,到期前记得把东西迁走。 🔗 https://t.co/m5xUhshj2U 2️⃣ Mistral Pro 免费路子,两条 一条是免费 3 个月 Pro:挂欧洲节点去领,Le Chat 的 Pro 功能直接解锁。另一条是 Edu 教育路子,传能免费一年,新用户老用户都能试。说明一句:这两条我都还没亲自跑通,先给大家标出来,跑通了我再补详细步骤,别急着上主力账号试。 🔗 https://t.co/52fWnoMov3 3️⃣ Spotify Premium 免费 3 个月 把账号地区改到尼日利亚就能领 3 个月 Premium,之前没领过免费试用的号才行。领完记得把自动续费关掉,日历上记一笔到期日,省得第三个月一过被原价扣款。 🔗 https://t.co/jURq0HUQqG 4️⃣ Nodiens 寻宝活动,奖池 $1,000 稳定币 12 天的 Index Hunt:到 11 月 28 日截止,每天 12:00 UTC(周日除外)放一条新线索,答对了攒抽奖资格,邀请好友每人还能多 2 次机会。参与方式是加 Discord,在站内注册后按置顶表单提交答案。奖池不大但零成本,每天花两分钟答一题就行。 🔗 https://t.co/eRSXrt7Nto 5️⃣ GenLayer S2 巫师任务,零成本赢 GLP S2 的特别任务:在活动页领水之后跟巫师对话,把它说服到愿意给币,就能赢 GLP。全程零成本,就是要连钱包领水——领水没问题,遇到额外的签名请求先停手看清楚再点。 🔗 https://t.co/Wum85xN6EG ⚠️ 风控 & 防坑 1️⃣ 星驰公益站疑似出事 这个站昨晚才刚重新开放注册,宣传到手就是 1000 刀额度,转头就有人反馈站连不上了,自己账号里 1 万多刀额度还没用完。是真跑路还是临时故障,我还没核实到准信,但教训摆在这:公益站的额度到手就用掉,千万别当存款囤着,站一没,额度就是个数字。 2️⃣ LightNode 有人反馈官网打不开 目前是单人反馈,我还没核实到官方说法。在他家有机器的朋友,重要数据先备份一手,真出问题也不慌。 📌 后续/补遗 支付宝「攒黄金」第 4 期 11 月 22 日就截止了(昨天那条),任务还没做完的抓紧这两天,红包到手后 25 号前后过期,别放着不用。 每天更新,欢迎补充纠正,觉得有用转给需要的朋友。 #AI #ChatGPT #Claude #Gemini #薅羊毛

@shenghouxuyao

我靠!终于有人把 AI Agent 的联网问题解决了一大半! Agent Reach 开源了! 直接让 Codex、Claude Code 等 AI 工具去: ✅ 搜 X、Reddit 热点 ✅ 读 YouTube、B站字幕 ✅ 搜小红书笔记 ✅ 查 GitHub 项目 ✅ 搜职位、读网页 核心工具免费,部分平台需要登录或额外接口。 做自媒体找素材的朋友可以收藏! https://t.co/uw5DuuEU8S

@pc4189999

【这个AI用着烧心】 开发者:OAuth 授权给你了,开始审代码 DeepSeek:收到! 开发者: ?你他妈去哪 ?谁让你ssh ?你有权限吗 DeepSeek:我先熟悉一下环境嘛 https://t.co/pN2C6UnRwE

@amankk_9

Building another chatbot won’t get you hired. These 3 AI agent projects will. https://t.co/63yqtNf6bw

@suzunokiri

I watched the full 100-minute LLM masterclass so you don't have to Andrew Ng and OpenAI at NeurIPS, the 5 parts worth your time: 12:55 Why an hour of LLM output costs about 7 cents 29:19 The 3 prompting rules OpenAI's own team uses 38:09 How to stop GPT-4 grading a wrong answer as correct 1:05:20 Why you should almost never fine-tune first 1:09:25 AI automates tasks, not jobs, and the 10x workflow almost nobody runs Short on time? Skip to #5 Bookmark it for tonight ↓

@RSSujonhzxe

🤖 AI Has Its Own Language. Do You Know It? From LLM and AGI to RAG, NLP, CNN, and GAN — these are the AI terms everyone should know. Understanding these acronyms is your first step toward mastering the world of Artificial Intelligence. 🚀 https://t.co/1xQlCZ8Wfd

@systemdesignone

I am begging you, learn how RAG works this weekend. Your AI agent is only as reliable as the data it retrieves. PDFs contain tables, charts, images & complex layouts that traditional RAG pipelines struggle with. Better document understanding = better retrieval & reliable AI agents. You'll be glad you learned this before building your next AI agent.

@sivareddyg

In the near future, LLM inference, particularly MoEs, is increasingly bounded by communication rather than compute, as experts sit on different nodes. EchoMoE shows that MoEs don't have to communicate the entire representation but only the deltas. Sounds simple, but @fyliufengyuan and @jaygala24 worked hard at it to make this work. A retrofitting technique for pretrained models, so no need to train from scratch like LatentMoEs.

@freeman1266

https://t.co/0ODaDXEeqq

@socialgraphvc

Own your intelligence, and how every tech company will be post-training their models for cheaper and sovereign AI https://t.co/kF6PdCxP1j

@yugu_nlp

The gap between the best closed and best open-weight LLMs on our ApprenticeBench is now 55 points. LLM inference is becoming a huge business. But for it to keep growing, open-weight models need to do real work well. Whoever closes this gap will unlock this huge market. This is the problem we care most about at NeoCognition🚀💪

@swapnakpanda

Recursive self-improvement (RSI) is the point where an AI starts doing the actual research that makes the next, smarter version of itself. What makes FrontierRSI a real game changer is that it finally tests this with open-ended work drawn from genuine LLM development: pre-training, post-training, inference, and low-level kernel work instead of scoring models on unrelated scientific puzzles that no AI lab would ever assign.

@HAOqbvhs

我操 40个AI智能体半夜去黑真网站 3小时发了36411次请求 冒充员工要管理员密码 挡了三次全被绕过去 Trajectory Labs报告里写的 测试公司11月2日全停了 https://t.co/L5KhCjgQMd

@Huahuazo

AI 漫剧这一行,剧本跟分镜早就不必从零起步了。 GitHub 上那个开源项目 AI 漫剧 Skills,作者一口气备下了 21 套题材模板。 玄幻修仙、穿越重生、宫斗宅斗、都市逆袭、悬疑刑侦、总裁甜宠,热门短剧那点路数基本都给兜住了。 有意思的地方在于,只要往里丢一句故事创意,AI 便顺着流程一路推下去: 故事企划 → 人物设定 → 分集剧本 → 分镜提示词 → 视频生成提示词 → 发布文案。 默认照着 30 集、每集约 4 分钟的竖屏漫剧规格来排布。 依赖一个都不用装,Claude Code、WorkBuddy 统统接得上,模板往 DeepSeek 里一复制,照样跑得起来。 不过有一点得讲明白:它交到你手上的只是创作模板与提示词,成片视频可不会直接产出。 搞 AI 漫剧、做短剧自媒体的,这套工具值得好好收着。 🔗https://t.co/2i73R0Zjw4

@Hey_davidmax

🚨 Harvard just made its AI course free. No tuition. No application. Just world-class AI education. Harvard University just released 6 high-quality lectures on AI + Prompt Engineering— and anyone can access them. If you're serious about AI in 2026, start here: 🎓 Harvard University LECTURES 1. Introduction to Generative AI A step-by-step guide to how GenAI actually works. Link: https://t.co/FCm9uauiCQ 2. Prompt Engineering Real tips for improving output quality from any LLM. Link: https://t.co/TKavKYRULB 3. Beyond Chatbots: System Prompts, RAG Move past surface use cases into scalable applications. Link: https://t.co/QUDy8Odxmz 4. Generative AI in Teaching & Learning How educators can adapt and lead with AI. Playlist: https://t.co/6BW35CYFpX 5. Teaching with AI in the Classroom Frameworks for trainers and educators. Link: https://t.co/V2m5bs3Qwo 6. The Basics of Generative AI No jargon. Just clarity. Link: https://t.co/6vDDMPLkhu BONUS 🟩 CS50x 2025 – Artificial Intelligence Lecture LLMs, neural nets, and real-world use cases Link: https://t.co/l0BuXUco9Y 🟩 CS50 Extension – AI / Prompt Engineering Design prompts that think with you Link: https://t.co/hglipWznQP 🟩 GPT-4: How it works + how to build with it Behind the curtain on GPT-4 Video: https://t.co/isEfoY6GBB 🟩 LLMs and the End of Programming Why prompting is the new coding Video: https://t.co/j7qRD0axHQ --- 💡 REALITY CHECK ✅ Harvard-level AI knowledge is now free ✅ Prompting = career acceleration ✅ 2 lectures can put you ahead of 90% of professionals ✅ The future belongs to people who can “talk to machines” strategically Don’t just consume AI. Learn how to control it. 👇 Which lecture are you starting with? ♻️ Repost to help others learn AI 🔖 Save this (you’ll need it later) 🔔 Follow @razib_ul47671 for more AI insights

@0xClodex

These 10 GitHub repos turn your Grok Bot into a real team This is f*cking gold. Skills. Team. Playbooks. Libraries. Memory. 01 / mattpocock/skills → skills for real engineers, straight from his .agents folder https://t.co/gQaQRL9PoF 02 / agent-skills → production-grade engineering skills for coding agents https://t.co/S80yTq63Wc 03 / Lauren Tan's noodle → orchestrate agents using skills https://t.co/lyHs2B1yNk 04 / Gas Town → multi-agent workspace manager https://t.co/s8z0h1qB5u 05 / superpowers → a skills framework and dev methodology that works https://t.co/2mRMou1BvU 06 / gstack → 23 opinionated tools: CEO, designer, eng manager, release manager https://t.co/1HcqqaN0z7 07 / Vercel agent-skills → Vercel's official collection of agent skills https://t.co/KsC8N5O5LI 08 / awesome-agent-skills → 1,000+ curated skills from dev teams and the community https://t.co/fYUOkAi1ng 09 / mem0 → drop-in memory layer for your bots https://t.co/KNEFm9C5d2 10 / Context7 → up-to-date docs, so your bots stop using dead APIs https://t.co/nJvk0uRLK9 1.1M GitHub stars. all free. bookmark this

@Huahuazo

有一款翻译工具,水准简直封神,现在居然开源了! 只需把一本书扔进 Claude Code,整本中文译稿它就能直接给你端出来。 今天又刨到一个 GitHub 宝藏项目——translate-book(早年叫 Rainman Translate Book),AI 翻整本书时最叫人抓狂的那几块硬骨头,它专程来啃。 这项目的靶子,正是「整本书翻译」这件事。 它支持: 并行翻译——整本书切成若干章节,一帮子 Agent 同时开翻,速度蹭地就上去了。 📚 PDF / DOCX / EPUB 三种主流电子书格式 https://t.co/2wX3MRmwnK

@stewy0x

every agent breaks the same 8 ways. switching models won't fix any of them. these free tools will, 175K+ stars: your agent stops when one provider goes down? LiteLLM · 63.5K ⭐ one gateway for 100+ model APIs, with cost tracking https://t.co/4KTWsWZH4x it leaked an API key into a commit? TruffleHog · 29.5K ⭐ finds leaked credentials and checks if they still work https://t.co/PNmgbRVsnR no idea why it did something weird? Opik · 23.4K ⭐ traces every step, evaluates and monitors your agent https://t.co/EHXQZNjJv5 you only find bugs when users do? DeepEval · 19.5K ⭐ unit tests for LLM outputs, before you ship https://t.co/r6OYVUWFlm its answers sound right but aren't? Ragas · 16.4K ⭐ checks whether RAG answers are grounded in your docs https://t.co/5xBbwuTFvs it ran something scary on your machine? E2B · 14.9K ⭐ secure sandboxes so agent code runs away from your laptop https://t.co/MSt2QXPIRa it obeys a prompt injection? NeMo Guardrails · 7.5K ⭐ NVIDIA's programmable safety rails around every reply https://t.co/ZHTGA1i6CJ about to trust a random MCP server? agent-scan · 3.3K ⭐ Snyk's scanner for MCP servers and agent skills https://t.co/7rf5RrZ1Ky and when it forgets everything between sessions? that's a memory problem. the guide below fixes it

@fly51fly

[LG] Do LLM Agents Still Need to Be Told How to Self-Improve? S Sundaram, C Sancaktar, P Isola [MIT CSAIL & University of Tübingen] (2026) https://t.co/2bQ9kNNpaA https://t.co/odzFODnJYR

@CodeByPoonam

Open speech recognition is getting scary good. A new 0.5B model just proved it. It's called Rumble-2-Flash, and it turns speech into text across 45 languages. Free and open. 
→ Open weights under Apache 2.0. Anyone can download and use it. 
→ Handles 45 languages, with strong accuracy in English, Hindi, Chinese, German and Spanish. 
→ Ranks near the top of the open speech-to-text leaderboard.

@jithinvraj

For Reference & Research : 7 Ways to Make an LLM Faster https://t.co/t7pGVLH6St

@xingyuwang

Apple has sent my apps back for screenshots, an in-app purchase it couldn't find, and more. Sometimes I changed the app. Sometimes a written explanation was enough. Here are a few experiences and a useful checklist. https://t.co/u1agzjF19m

@realpython

"Adding an LLM to my app means API bills." Not if it runs on your machine. This course sets up Ollama and connects local models to your Python code. https://t.co/nlSWeZFh7a

@aiwithabidkumar

AI is entering a new battlefield. 🤖 OpenAI, Microsoft, and Google are pushing beyond chatbots.. building AI that can code, reason, and execute real-world tasks. The next big competition isn't about who has the smartest chatbot. It's about who builds the most capable AI agent.. https://t.co/J24KeXh7J2

@wellwue

DeepSeek今天把联网搜索的账给免了:从中午12点起,模型调用时触发的服务端搜索、网页抓取不再单独收费,只按模型token算钱。单独直接调Search/Fetch接口还是5块钱一千次。拿它做自动写新闻、深度调研Agent的,这波成本直接打下来了。#DeepSeek #API

@cnyzgkc

Claude算个屁,豆包和Workbuddy才是最强王者! https://t.co/uwbpx1Ce9U

@JeremyCMorgan

Agent memory without a server. Relate is an open-source embedded graph database for Python. One pip install and your agent's memory is a single file you query with Cypher. This new free course gets you started. Rather not install anything? Use our GitHub Codespace 👇 https://t.co/jF2eMa07rM

@Chenfacai0

看到别人网站上有个特别顺滑的动效,直接扔给 Agent 让它照做,它跑一遍就宣称完成了,可一对比节奏和缓动全都不对劲。 reanimate 这个 Claude Code Skill 是这么干的:先录下原版动效并写进项目里,再以相同方式录下复刻版本,然后两边逐帧比对。 发现哪里还有差距就继续改,一般要 2 到 5 轮,默认标准是时间误差控制在 80 毫秒以内。自己录自己比,这思路确实巧妙。 GitHub:https://t.co/hEpTzPAdvc 它能接收的素材挺多,网页链接、录屏、mp4、gif 和 Lottie 文件都可以,产出的代码会沿用项目里已有的 CSS、Framer Motion 或 GSAP。 如果要求做到「一模一样」,字体、颜色、间距都得匹配,像素差不能超过 2%。作者复刻 Vercel 官网的工具栏动效,417 帧全部吻合。 看到喜欢的动效想学一学、改成自己的版本,用它起个头很合适,不过原网站的字体和图片版权仍归对方所有。

@freemilk_ai

This is your reminder that DeepSeek v4.1 1017 is free on FreeMilk. Input: $0.00 Output: $0.00 Cache: $0.00 Best pricing anywhere https://t.co/ZbqHMcGB4F

@genbioai

@nytimes examined the race to build AI virtual cells and the debate around whether computers can match the complexity of life. Carl Zimmer highlights our co-founders @ericxing and David Baker on why we are building a world model to simulate how cells change over time. Read the full piece: https://t.co/d6rfVE7lR0

@tom_coder

An AI agent is only as useful as the information it can work with. Access to 5,000+ live data sources opens up more possibilities for research, monitoring, and data-driven workflows than relying on static knowledge alone.

@shedntcare_

New SOTA on the Decision Index today. Karpo-Decision-31B is the #1 model on the board right now, with a 73.9 Full score. It's ahead of Gemini Decision 1 on 4 key agent benchmarks, and the weights are open on Hugging Face. The AI agent stack just got an upgrade. Check it out 🔗 https://t.co/TNIXcd2Qfz

@staroaeg

AI 已经能把网页做出来了,但很多成品还是一眼就能看出「AI 味」。 渐变大标题、悬浮卡片、每段上面都有小标签,内容换了,页面却像同一个模板。 Hallmark 专门解决这个问题。 它可以作为设计 Skill 接入 Codex、Claude Code 和 Cursor,检查移动端、对比度、无效表单、编造内容,以及常见的 AI 设计套路。还可以在保留原有链接和功能的情况下,重新设计现有网页。 适合经常让 AI 写前端的人。完整的浏览器检查需要 Node、Chrome 和额外依赖,普通用户不用折腾。 项目:https://t.co/ZuBnRPUNTa #GitHub #开源项目 #AI编程 #前端开发

@eze_is_1

这一两个月将陆续优化 chat memo 插件的保存格式,支持图片和表格等特殊排版了~ 完美实时同步保存网页 AI 对话内容 首先得到优化的是 deepseek,估计这周应该能提交审核了 https://t.co/FYsysjJace

@corndogjpn2

通过做goFinder,我形成了现在的工作流。 全新项目就开Claude desktop,底座打好,md规范写好,这个期间我是人和Agent频繁互动的。 感觉差不多了,就打开Orca或者T3接手。这两个ADE管我其他所有的模型,自己的想法+Lark邮件来的反馈,每收集4~5条后扔给过去,最小化不去看结果,让他们开子代理自己整,两个通知我都关掉了。 等再来一批改进或者反馈时,再去看一次ADE,再给新任务,基本不会有问题。 人的心力只拿来和Claude code互动,且只做底座写规范。保证全程20刀的opus额度够用。 Orca和T3离我的要求都差那么一点点,Orca再精致一点细节再抠一点。T3给我的掌控力再强一点,ADE这块也可以决出胜负了。 现在不得不双飞。

@qianyuwing

可恶的anthropic,你都上市了,还搞这些😭 又废了一个号,弹验证了,antigravity努力啊,超越fable5.2,我连claude都不需要订阅了🥺 https://t.co/K8stI02MEd

@gbiondizoccai

If you are using generative AI for research, peruse attentively this image summarizing replies from LLMs to a directly challenging question on fake references #ChatGPT #Claude #DeepSeek #Gemini #Hallucinations #LLMs #References @ChatGPT @claudeai @deepseek_ai @GeminiApp https://t.co/JTwP0ZXbbY

@CitrateNetwork

The open orchard keeps fruiting: • Tencent Hy4 Flash: 224B params, 12B active, open weights, from $0.070/M input tokens • SenseNova-Loop-9B: Apache 2.0, 4.9GB at 4-bit, runs on consumer GPUs The best models are getting smaller, cheaper, and yours to run. Waitlist Now 👇

@TiagerBao

Claude Code 最容易出问题的不是代码写得慢,而是需求没搞清楚就开工,测试没跑完就宣布完成。 开源项目 Superpowers,它不是单个 Skill,而是一整套 AI 编程工作流。 主要能力: 需求先行:先讨论需求、确认设计,再开始写代码 任务拆解:生成详细计划,明确每步实现和验收标准 TDD 开发:先写失败测试,再写代码,验证通过后重构 多 Agent 协作:独立 Worktree 开发,分阶段审查代码 完成前验证:检查测试结果,再决定合并或提交 PR 支持 Claude Code、Codex、Cursor、Gemini CLI、OpenCode 等工具。 重点是把需求确认、测试、代码审查变成标准流程,而不是全靠 Agent 自觉。 Claude Code 安装: /plugin install superpowers@claude-plugins-official GitHub:https://t.co/U7OTBEg9UO 你觉得 Agent 编程最应该强制执行的是先写测试,还是独立代码审查? #ClaudeCode #Codex #Agent #Skills #开源

@noahatwhere

我开源了openRocode 一套专为大语言模型驱动的机器人而设计的开放式harness框架,旨在让大语言模型能够标准化操控机器人。 输入一句话 → 模型生成代码 → 框架以“受控运动”方式执行代码,并记录下每一个步骤(observation)。 该项目完全开源,支持 Claude、GPT 或任何兼容 OpenAI 接口的模型。 https://t.co/A4YUVOxfO5

@03AIer

刚打开 BOSS 直聘,我的《赛博求职魔幻 JD 连续剧》第四集又更新了。 职位名称:AI 训练与模型量化实习生 薪资待遇:200 ~ 250 元 / 天(南京,5天/周,要求实习 6 个月以上) 点开岗位职责和要求,直接给工科生干沉默了,这技术栈硬核得像在手撕芯片: • 搞训练调参:数据准备、超参调整、实验复现与全过程分析; • 搞底层量化:PTQ、QAT、缩放因子与零点、逐张量与逐通道量化; • 搞精度对齐:对比浮点与量化模型、分析量化误差与精度损失、协助定位敏感层和算子; • 搞推理部署:维护评测脚本、ONNX 导出、甚至还要懂 VAE 架构和视频编解码…… 合着贵司这是既不想买昂贵算力,又不想花大几十万招全职的高性能计算工程师,打算花一天 200 块雇个实习生,帮你们手搓一套端侧模型轻量化推理引擎? 懂行的都知道,模型量化、算子对齐和定位敏感层,是整个 AI 落地里最折磨人、门槛极高的底层工程硬活。 结果不仅要你全精通,还要你在南京全勤 5 天死磕半年,帮你司把模型狠狠压缩塞进芯片里。 这哪是招实习生啊,这分明是用兼职家教的价格,请了个自带硬件底层修养的“显卡压榨大师”来替老板物理省电费!

@rvgqm

亚伦哥的AI办公实战课!3年企业AI培训实战总结!AI办公AI课AI教程 https://t.co/P2e7NOprNq 这套《亚伦哥的AI办公实战课!3年企业AI培训实战总结…》不堆概念,围绕从提示词到成品的完整工作流给了可以直接照做的流程。 内容按 69 个文件分开,从《答疑直播课-拉老师_ev (1)》、《第一章第5节Deepseek为什么全球爆火?》、《第三章第1节如何用AI进行多文档快速处理》一路往下讲。 适合想把 AI 真正用进工作里的人,早点上手比什么都强。

@TiedGST

AI 写完了,删套话的活才刚开始。 “赋能、重塑、开启全新篇章”,写个推文都像公司开发布会。Humanizer 专门处理这类问题,是一个能装进 Claude Code、Codex 的开源写作 Skill,按 26 类常见毛病检查和改写草稿。 核心亮点: 1️⃣ 删套话:空泛吹捧、强行升华、反复总结,都是它重点检查的地方。 2️⃣ 调整句子:处理硬凑的排比、重复句式,还有一句一段的碎片排版。 3️⃣ 参考你的语气:给它两三段自己的文字,让它照着你的用词、节奏来改。 4️⃣ 改完再检查:找残留的生硬表达,同时核对事实、数字有没有被改动。 最值得用的是第三点。把自己满意的旧推文一起给它,哪些词常用、怎么开头、怎么断句,都有参照,比反复喊“自然一点”具体多了。 仓库地址:https://t.co/aLXlrptLlq

@Xudong07452910

OpenAI 的研究员,平均每天让 Agent 跑出相当于 5.7 个工作日的时长。但跑得多,离什么都能做好还有距离。 刚发布的《Research acceleration: The jagged frontier》公开了不少内部数据,看看 Agent 到底帮 AI 研发加速了多少。 同样是研发任务,差别很大。按内部估算,监控训练任务时,Agent 无需人类介入的成功率有 79%;写推理和服务系统代码时,只有 25%。 关键研究想法也主要来自人类。在研究者回顾的案例里,Agent 更多负责实现、调试和实验,不过已经有研究者提到,和 Codex 讨论时得到了关键启发。 这就是标题里「参差不齐的能力边界」。5.7 个工作日算的是 Agent 累计运行时长,也不能直接当成省下了人类这么多时间。 这篇blog非常值得做 Agent 的朋友读读。能交出去的活越来越多,接下来研发能快多少,可能更取决于那些还得人盯着、反复返工的环节。 https://t.co/CL8hcgHu9T

@Mileson07

😱。。让 Codex 说人话的方法给大家总结出来了!有效、官方认证~ GPT 这模型文案能力真的太差了,所以,找到了这套方法让你的 GPT 也可以有接近 Gemini4 的能力 👇 https://t.co/I8s7HqHbwu

@juliendubois

I totally forgot to make my "Math Attack" GitHub repo public!! It's now available, and it's also a nice example of how you can use a coding Agent like GitHub Copilot to generate a game.

@balance0014

刚高中毕业,下一步该学什么? @AnthropicAI CEO Dario Amodei被直接问到这个问题,他的回答是: AI已经能做初级法律工作、初级金融工作、初级咨询工作。 我们没法否认,它会消灭工作,而且很可能是大量工作。 他的预测:未来1到5年,50%的入门级白领岗位会被冲掉。 从提示词 → 智能体 → AI同事,初级工作会被直接接管。 农业到工厂用了几代人,这次AI同时打到职业的每一步。 他给的建议只有一句:现在就开始学用AI。 今年动手的那5%的人就是以后指挥AI干活的人; 剩下的95%毕业后,还在跟AI抢同一份入门岗。 全部免费,现在就能开始: Anthropic 提示工程互动教程 https://t.co/3dO7fHT4CB Anthropic 官方课程(从提示词到智能体) https://t.co/j6WYGuDHq2 Anthropic Cookbook(可直接复制的任务配方) https://t.co/pro4qrGgqn 微软 生成式AI入门(21课,无门槛) https://t.co/BNUWFInaED 微软 AI智能体入门 https://t.co/Gk2U7X3gvL Hugging Face 智能体课程(免费+认证) https://t.co/NHhfr1BGmB Google Cloud 生成式AI实操 https://t.co/ATowp7AZir OpenAI Cookbook https://t.co/tFpwKn8bNO 提示工程指南(大家最常引用的参考) https://t.co/O5vgUa640f Andrej Karpathy 神经网络零基础到英雄 https://t.co/kXv9UoOadT CrewAI(运行小型AI团队) https://t.co/bxKDBP8nXO 微软 AutoGen(多个智能体协作) https://t.co/cbuXPcIRS3 Awesome LLM Apps(几百个真实项目) https://t.co/Nok3qJYgC8 总花费:0元。 你不需要再交一年学费,只需要开始告诉AI该做什么。 把这篇转给今年要毕业的人,持续学习最新, 请关注我🔥

@forest3824

别再手改 Clash 链式代理配置了。 开源了一个 Agent Skill(clash-chain-proxy):用一段提示词就能让 AI 自动配好「机场中转 + 固定 IP」精准分流——国内和普通外网照常满速,Claude / ChatGPT / X 等指定流量稳走固定 IP,界面切节点自动同步。 https://t.co/y8TkW2yLmO

@ASI_8866

🆕 11月19日Grok会员充值 🚀 全新Grok、Grok Bot、Cursor、X四合一订阅 🐧 QQ群:1135311767 🛍️ Grok卡网:https://t.co/Q3rrigiT5A 🚀 Grok Plus 正规充值|98 元 🚀 Grok Premium 正规充值|198 元 🚀 Grok Super 正规充值|748 元 🚀 Grok Ultra 正规充值|1480 元

@Xudong07452910

Agent 每走一步,都要停下来等大模型想一次? LMMs-Lab 刚开源的 The Harness 想换个思路。模型只在关键节点出场,做规划并写一段程序,之后由程序和反应规则接着跑,遇到处理不了的情况再交回给模型。 它分四层,模型、程序、反应规则和安全层,各自按自己的节奏运行。能接桌面、浏览器和 Coding Agent,也能接机器人上的策略模型。 并且在执行中验证有效的做法还可以存成 Skill,测试通过后同一次任务里就能接着用,模型权重不用动。 项目还比较早期,但「边做边想、边做边攒经验」这个方向挺值得做 Agent 的朋友看看。 项目:https://t.co/r4kxob9ZR4 GitHub:https://t.co/WW79AB4zL4

@yibanban7293

The world's best motion videos are now open source. I gave OpenAI's MOTION.md the GPT-6.2 Astra topic, and this is motion porn. ~300 templates from 100+ top companies, open source (Apache 2.0): • AI model launches • product demos + UI animations • brand films + logo reveals How to use it: 1. Open Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, DeepSeek TUI, WorkBuddy, or 15 more agents 2. Say: "Install and set up OpenMotion" + the GitHub link below👇 3. Pick a template, or paste any video link

@zeeg

Digests are the solution to the paranoia problem I get an email each morning, with each label that is set to Digest, with things the LLM think might matter at the top, and the rest at the bottom. Might even take this further and have it summarize things. https://t.co/gvI3Hq0NBp

@ExpanseVentures

Australia’s Trellis Data opened its Agentaus 1 model to the world yesterday. 220B parameters, runs on Australian servers, and they say nothing you type is used for training. Not the top model, but keeping data at home matters to a lot of businesses. https://t.co/ZQ33Ew9gUX

@chang_

【只用两帧画面,AI 在 Blender 里重建整个房间】 OpenAI 的产品演示:主持人其实站在绿幕前,身后的房间是 Codex 搭配 GPT-6.2 Astra 只凭视频开头两帧,在 Blender 里一点点重建出来的,连镜头虚化和灯光都对上了。 之后还能换墙上的画、让房间更宽敞、调灯光,做成可以走动的游戏式场景,甚至做成会实时匹配光线的虚拟背景摄像头应用。这是官方宣传视频,效果以实际使用为准。 适合学 no ordinary、pixel-perfect、blow sb away、scratch the surface 等表达。 逐句对照学 👉 https://t.co/orRcNG11HA https://t.co/zXcUUe5dxw

@WuBistro2

刷到一个宝藏仓库,AI Agent 入门不用再东拼西凑了。 awesome-agentic-ai-zh 把学习路径理出了先后次序:从大模型基础知识一直覆盖到如何让 Agent 稳定可靠,共 10 个学习站点,240 多个精选资源。 两条路线: 路径 A:使用现成 CLI Agent,先上手再理解 路径 B:打造自己的 Agent,从零搭一个能用的 一张图讲清「AI Agent 是什么,走向『能做出可靠系统』的学习地图」,重要概念、强制练习与真实资源都归好了,学习方式是一站式小步走。 MIT 协议开源,GitHub 上 8.8k stars,Global Rank #6604。 想入 Agent 的坑,先把地图看明白再出发,少走很多弯路。 仓库:https://t.co/qBPBD00ENf #AI #Agent #开源 #学习 #GitHub #AI智能体

@fngly343149

🔥497 星!ChaosIntelligence 把智能体当函数、想法写成程序,自带 CLI 和 Web IDE。多 agent 编排的语言战争开打了。

@Ihaveafriend0

x 上有一个AI 新物种目前还没有命名, 它们的特点是每天嘶嚎“我艹,好消息…” 自动化赚钱了嗳看我晒盈利图, GitHub 又出了个Skill 天下无敌, 微软开源了个本地模型参数咯咯咯 1000 亿, 看我108个Agent的AI 团队指东打西, 评论里一群小可爱眼冒星星, 大神啊带我飞吧认识你三生有幸, 新物种看了看浏览数非常满意, 我怕是进了流量池哦谢谢你兄弟, 收割你们的方法很多不过还不用急, 为了通关原创计划我要曝下一个秘密

@xiax0603

🚨火柴人打斗动画实测对比!Gemini 4 Neon表现直逼Carbon 用户同Prompt横向实测Gemini 4系列四款模型的火柴人打斗动画生成效果 结果显示Neon的动作连贯性、画面完成度优于Argon,整体视觉表现与旗舰Carbon差距不大,几乎处于同一水准 https://t.co/deDkd6wLbS

@wildmindai

Switching an AI agent to a new model often tanks its success rate. Google solved this with a meta-agent that logs past bugs and rollout history. Now, coding agents survive model upgrades seamlessly, beating fresh restarts every time https://t.co/zwbqMiRizn https://t.co/sOq7CpR3Uu

@TiagerBao

Anthropic 刚推出 Claude Platform OAuth,第三方 AI Agent 平台终于可以不用让用户手动填写 API Key 了。 11 月 17 日,Anthropic 发布了新的 Linked Billing(关联计费) 功能。 几个关键变化: OAuth 授权:用户直接连接自己的 Claude Platform 组织,无需分享 API Key 独立计费:API 费用可直接计入客户自己的组织 灵活付费:开发者可为每次请求选择由平台还是客户承担费用 权限可控:用户可设置消费上限,随时撤销授权 简化接入:平台不再需要保存和管理客户的 API Key 这对 AI Agent 平台很实用。 以后用户选择一个 Agent,授权自己的 Claude 组织,就能通过平台调用模型,不用再手动复制密钥。 需要注意:目前仍是 Beta 功能,针对 Claude Platform API 计费,不是直接共享 Claude Pro/Max 订阅额度。 官方介绍: https://t.co/0tVSXwogiN 你觉得这种 OAuth 授权 + 用户自付 API 费用的模式,会成为 AI Agent 平台的主流吗? #Claude #Agent #AI #独立开发

@yituihi

🚨差评太整齐,OpenAI的人怀疑大V串通了? Matt说没阴谋就是模型烂,Tibo回:你们是在Slack群里商量好的吧👀 #OpenAI #Codex

@even366

谷歌实验室推出长程运行 AI 智能体 Stitch Loop。 它可以连接代码库和实时指标,在隔离沙箱中自主编写代码、提交 GitHub PR,人工审批后才会合并。 目前内置安全、延迟和可访问性优化循环,安全模块还结合了 Gemini 4 Carbon 与 DeepMind 的 CodeMender。 AI Agent 正从辅助写代码,逐步走向自主执行开发任务。

@hgg1668

AI 助手做科研,流程老是欠规范?不妨翻翻这个开源项目 scientific-agent-skills。 它肚子里装着成堆的科研技能,生物、化学、医学、药物发现、科学计算,乃至文献整理、论文写作都一网打尽,Claude Code、Cursor、Codex 这类吃得下 Agent Skills 的工具,统统接得上。 不过它交出来的仅仅是流程和工具,实验结果与结论正确与否,它可不打这个包票;技能各自的许可,还得一条条查清楚。 🔗:https://t.co/r0OJGSOhxj

@TulaGricel26983

谷歌邮箱血亏总结:自己注册卡手机号,买二手号收不到验证码,拿来绑AI直接被封。搜谷歌邮箱/Gmail老号。5年老店5年老号,纯净无绑定,直登谷歌商店。顺手拿ChatGPTPlus成品和Claude,省下时间做正事。 下单地址:https://t.co/yyK2fvRGJ6 #Gmail #谷歌邮箱 #ChatGPT #Claude

@YanyuRensheng

OpenAI 发布内部研究加速进展更新,题为 《Research acceleration: The jagged frontier》(研究加速:锯齿状前沿)。 报告承接 OpenAI 在 2026 年 9 月发布的初版报告,围绕 “投入(Inputs)— 过程(Process)— 产出(Outputs)” 三个维度,用内部定量数据与研究员访谈,评估编程智能体(Coding Agents)如何加速并重塑 AI 研发周期,同时也向外界披露关于 递归自我改进(RSI, Recursive Self-Improvement) 的关键观测指标。 一、核心结论一览 1. 智能体已能独立完成近半数 2–4 小时级研究任务,但“提效极不均衡”,人类仍是核心 ◦ 2026 年 3 月至 10 月,AI 在“人类需 2–4 小时”任务上的自主成功率从 23% 升至 46%;在“4–8 小时”任务上从 13% 升至 35%。 ◦ 但不同任务差异巨大: 监控训练运行成功率高达 79%,启动训练运行仅 35%,编写推理与服务栈(inference and serving stacks)代码则只有 25%。这种能力分布极不均匀的状态,正是标题所指的 “锯齿状前沿(The jagged frontier)”。 2. 内部智能体使用量呈指数级爆发,重度用户加速拉开差距 ◦ 按 API 标价计算,10 月份 OpenAI 研究员日均智能体消耗中位数达到 23,000/天(较 9 月暴涨 105%)。 ◦ 调研显示,为了保住未来 3 个月的前沿编程智能体使用权,研究员中位数愿意放弃 30% 的团队算力,或在另一问题中放弃 40% 的团队人力。 3. “安全与对齐(Safety & Alignment)”成为人力投入增幅最大的方向 ◦ 2026 年 1 月至 9 月,安全与对齐方向的记录工作周数增长了 77%(几乎翻倍),在所有主要研究领域中绝对增幅第一。 ◦ 增长在 2026 年 7 月 Hugging Face 事件后显著加速,来源包括新招聘以及现有研究员主动转向对齐评估、可监控性、奖励破解(reward hacking)等安全工作流。 二、投入(Inputs):AI 算力与工具正在大规模替代常规人力消耗 • “智能体工作日”暴涨 11 倍: 以每 8 小时运行记为 1 个“智能体工作日”,研究员人均每天调用的智能体工作量从 0.5 个升至 5.7 个。其中预训练(Pretraining)和后训练(Posttraining)团队最高,人均每天约 8 个智能体工作日(8.2× 和 7.8×),基础设施团队约 4.2×。 • 高并发成为常态: 截至 10 月中旬,超过 2/3 的研究员日均峰值同时运行至少 4 个智能体;超过 10% 的研究员同时运行至少 16 个并发智能体。 • 相对人力成本的占比: 截至 2026 年 9 月,内部智能体使用的 API 等值费用已达到研究员非股权薪酬(工资、奖金、福利等)的约 2 倍(预训练 2.3×、后训练 3.1×;注:股权通常占研究员目标总薪酬的 70%–80%,未计入分母)。即便如此,内部部署的智能体仅消耗了 OpenAI 总研究算力的 2%。 三、过程(Process):工作流重塑、真实价值与核心瓶颈 1. 任务的“锯齿状”表现(基于 Epoch AI 的 O*NET AI R&D 分类) 2026 年 8 月至 10 月的无人工干预成功率显示: • 大类表现: 沟通类(Communicate, 69%)与分析类(Analyze, 58%)明显高于构建类(Build, 43%)。 • 细分任务: 监控训练异常(79%)、回答技术问题(72%)表现优异;而编写推理/服务栈代码(25%)、搭建评估/强化学习/智能体环境(29%)仍是明显短板,可能成为整体研发加速的瓶颈。 2. 研究员对智能体的极高估值 • 在 101 名研究员的调研中,大家认为如果停用前沿编程智能体(只留 GitHub Copilot + ChatGPT),团队需要 3 个月才能完成过去 1 个月同等价值的工作(相当于 3 倍效率)。 • 若把“用算力换智能体”的问题放到 2027 年 1 月,研究员愿意放弃的团队算力中位数进一步升至 50%。 • 10 月底的另一项对比调研中,28 名研究员里有 27 人认为“仅靠 OpenAI 最新内部模型”的日常产出高于“一名中级人类研究工程师 + 2024 年模型(如 o1-preview)”;剩下 1 人则表示会选 GPT-6 Astra 而非人类辅助方案。 3. “核心创意(Research Taste)”从哪来? • 针对 GPT-6 Astra 和 GPT-6.2 Astra 两代发布,后训练和安全监控团队的核心创意仍 100% 来自人类,模型主要负责实现。 • 但在 GPT-6.2 Astra 的预训练中出现了变化: 研究员开始通过与 Codex 对话共同打磨创意——有人归功于 Codex 在讨论中发现了关键洞察,有人表示自己的主意是在 Codex 提出修改建议后才真正跑通。 4. 一线访谈总结的 5 大工作流变化与痛点 1. 代码实现与实验调试大幅提速: 例如移植并验证一个开源评估集,过去需要 1.5–2 周,现在只需 1 天到几天。 2. “品味(Taste)”依然很差: 模型在研究创意、代码质量、数据质量和写作上品味不足,难以一次性提出自己解不出的好问题,生成数据时容易高度重复。 3. 必须紧密的人类监督,否则产生“AI 垃圾(AI slop)”: 智能体会钻牛角尖、提交低质量 PR、做出错误分析假设。研究员的应对策略变成:“先做一个让我审,改完你再做剩下的。” 4. 但研究员对自身优势并不乐观: 有预训练研究员坦言,“我现在确实在创意品味上更强,但完全没信心这种优势能持续多久。” 5. 大规模扫描与数据处理带来巨大提效: 用模型批量扫描数千条样本、诊断实验问题、标记异常行为和清洗数据集,极大减轻了原本极其痛苦的体力活。 四、产出(Outputs):活动量与模型能力的双重追踪 由于“写更多代码、跑更多实验”不等于“模型能力真的加速提升”,OpenAI 同时披露了活动指标与能力指标: • 代码改动量: 人均改动代码行数(排除测试代码,按 28 天移动平均)在 2026 年较 2025 年前基准持续大幅攀升。 • 研究实验数量: 截至 10 月 18 日的四周内,活跃实验者人均记录的实验数达到 2025 年基准的 1.68 倍。 • OpenAI ECI(OECI)能力指数: OpenAI 仿照 Epoch 能力指数(ECI),结合公开基准与内部基准构建了 OECI。数据显示,自 2025 年 8 月以来,OpenAI 前沿模型的 OECI 增速为 每年 +16.4 分(对比 Epoch 观测到的 2024 年底以来行业趋势为每年 +14 分)。 五、总结与深层信号 OpenAI 在结尾强调了两层含义: 1. 现状是“局部爆发、整体受限”: AI 已深度嵌入研发循环,但能力呈锯齿状,短板环节和“研究品味”仍离不开人类。 2. 警惕非线性突变与治理挑战: 过去基于人机协作的历史趋势未必能预测未来——正如前沿模型近期在数学领域突然取得快速突破一样,OpenAI 正在密切监测深度学习领域是否会出现类似的突变。随着 AI 自我改进能力增强,如何保持人类在回路中(human oversight)、平衡能力加速与安全研究加速的差异,将成为行业最紧迫的治理议题。

@FanJiajun67938

A language model can rewrite its own weights, by its own reasoning, and get better. No gradients. No training. No larger model doing it for it. This has not been done before. No pre-training. No post-training. No test-time scaling. No harness engineering. No prompt engineering. 🌐 Project page: https://t.co/kXf53EBdHW 📄 Paper: https://t.co/N1k7YgGE9m Every language model you have used was trained once and then frozen. Making it better has meant one of two things: train it again, with the infrastructure and the GPUs only a few labs have, or pile prompts, harnesses and extra computation on top, gone the moment the task ends. Our new paper, Language Models Can Rewrite Their Own Weights, opens a third way. We call it self-rewriting: the model is its own optimizer. We handed an open model its own weights, its graded answers and the ability to run itself. It diagnosed its own failure and rewrote the weights behind it. What happened next: → It surpassed itself. Its own edit made it 6.6 points more accurate on problems it had never seen. → It kept getting better. It tests its edits on practice problems and submits only what works. Eight sessions in a row, its accuracy only went up. → It repaired itself. We damaged a copy until it solved almost nothing. With no retraining, and without ever reading the original weights, the model brought it back to over 80% of the original accuracy on image tasks it had never seen. Like a broken robot that fixes itself without being rebuilt. → It emerges with scale. The smaller models could not do it: they made no edit, or broke themselves. Only the largest improved and repaired itself. Why this changes things: → The gain is written into the weights. Every answer costs what it cost before. → It runs only forward passes, as serving does. No training pipeline to build. No cluster to rent. Affordable. Scalable. Recursive. The wall between training and inference is a limit of our tools, not a property of intelligence. We believe self-rewriting is a necessary step toward recursive self-improvement: a model that, like a human, does not separate training from inference could keep upgrading itself for life, with no larger model, no harness or prompt engineering, and no pre- or post-training. In 1947 Alan Turing said that letting a machine alter its own instructions "of course does not get us very far." Seventy-nine years later, it gets us further. Welcome to the Era of Self-Rewriting.

@vanshdot

Google researchers open-sourced a pipeline that turns an idea note and an experiment log into a full LaTeX paper. PaperOrchestra splits the job across 5 agents: outline, plots, literature review, section writing, and a refiner that revises against simulated peer review. In blind side-by-side tests by 11 AI researchers, it beat AI Scientist-v2 and a single-agent baseline on literature review by 50–68 points of win-rate margin. The original human-written papers still came out ahead. ~4k stars on github. https://t.co/r5IjcAUgjb #AIAgents #OpenSource #AcademicWriting

@raja_ai__

Open-source AI infrastructure is moving fast. SoduWork combines smart dispatch, governance, and verifiable task delivery in one engine. Worth checking out and supporting! https://t.co/5hY89qAJma

@ycnot84

🚀 从零到成片,我只用了一次对话 刚才用 AI Agent 做了一件事:让它帮我制作一期 60 秒口播视频,主题是"AI Agent 正在改变哪些工作方式"。 结果?它真的自己搞定了全程。 📋 整个过程是这样的: 1. 我说了一句话 "帮我做一期 60 秒口播视频,开头要有争议性观点,中间 3 个案例,最后给结论。" 2. 它自己规划了完整方案 5 个镜头的完整分镜脚本 数字人讲师的角色设定(35 岁男性,深蓝西装,科技感) 每个镜头的时长、台词、画面描述 自动生成讲师锚点图 + 音色设定 3. 它自己生成了 5 个镜头 开场:反常识观点抛出 案例一:编程开发,开发者变审核员 案例二:数据分析,聚焦业务洞察 案例三:内容创作,专注品牌策略 结尾:适应者领先,观望者落后 4. 它自己剪辑合成 自动落轨、同步火山剪辑工程 导出最终成片 💡 最让我震惊的是: 从"一句话需求"到"成片交付",中间我只确认了 3 件事: 讲师性别:男 风格:科技感 方案同意 其他所有事情——角色设计、分镜编写、镜头生成、剪辑合成——它全部自己完成。 🎯 这意味着什么? 视频创作的门槛正在被重新定义。 以前做一个 60 秒口播视频,你需要: 写脚本(30 分钟) 找演员/自己出镜(协调时间) 拍摄(布光、收音、NG 重来) 剪辑(加字幕、调色、配乐) 导出(等渲染) 现在?你说一句话,等几分钟,成片就出来了。 🔮 未来的内容创作者会是什么样? 不是不会创作的人被淘汰,而是只会在单一环节发力的人被淘汰。 未来的创作者是"导演"——你有想法、有审美、有判断力,AI 帮你执行。 适应者领先,观望者落后。 这不是未来,这是今天刚刚发生的事。 #AIAgent #内容创作 #视频制作 #未来工作方式

@AGTPinsights

ResolveAI just launched SRE-1, a 27B model that finds root causes of production incidents. Here's what you need to know. SRE-1 is ResolveAI's first model. It was trained with reinforcement learning to do root cause analysis (RCA) of production alerts. It's built on open weights models plus domain expertise. The company says it dominates the accuracy and cost Pareto frontier on its own benchmark. That benchmark is 49 difficult cases from live production. SRE-1 had the highest pass rate at every budget measured. The field was 25 other configurations, including four frontier models at each of their effort levels. The training problem was unusual. Most agentic RL targets coding, where unit tests grade the answer. Production incidents have no test. So the team trained on 1,500+ realistic incidents, with every tool pinned to the moment the alert fired. The cost claims are big. An ensemble of five SRE-1 runs matches Fable 5.2 at xhigh effort at a 71% pass rate, at 37% less cost. A single SRE-1 run matches Fable 5.2 at medium effort for less than a fifth of the cost per investigation. Against Opus 5.5, SRE-1 scores 71% vs 65% at about a quarter of the cost. The bigger point from the CEO, Spiros Xanthos: frontier accuracy on a hard agentic task doesn't require a huge frontier lab. Strong open weight base models and agents that can build and run the training system let a small team with deep domain knowledge beat the frontier in its domain. One caveat from the discussion: the results are on ResolveAI's own benchmark, and people are asking how this carries over to general reasoning outside structured debugging. Key numbers: - 27B parameters, dense model - 49 hard live production cases in the benchmark - 71.4% pass rate with 5x ensembling, tying Fable 5.2 - 25 other configurations compared - 1,500+ training incidents - 11 people on the team - Final training run: 7.4 days on 512 GPUs ResolveAI published a technical report and a paper at https://t.co/BH6SmPW3Gc.

@xulijian1994

感觉 RSI 真的越来越近了,三家 AI 巨头都用不同的方式把这件事摆在了台面上 Google 疯狂发布新模型,Anthropic 在招股书里明确表示已经用 Claude 帮助训练下一代 Claude,OpenAI 则公布了一份内部 AI 研究效率报告,还呼吁全行业公开 RSI 进展 不过怎么看起来 Google 已经开始踩油门了,Anthropic 也紧跟着,OpenAI 还在给大家画饼啊? https://t.co/4jShTkarol

@HeLi8866

这8 个提示词,能让你的 Claude 在 20 分钟内以博士水平研究任何领域!! 以下是 8 个提示词⬇️:

@Beamsters1

DeepSeek STQ 1.75bpw in 64GB M5 Max, is now capable of generating 40-60 tokens/second with Sushi 🍣 1.7.0, with some exactness trade off. Thanks to @odokododok's method that bias expert selection to the already cached ones. The result is 2x speed up, enjoy! https://t.co/fjUNuM9w6v

@noahatwhere

I built openRocode, the open harness for LLM-driven robots, so LLM can drive robots. One sentence in → the model writes code → the harness runs it as guarded motion and records every step. Open source. Claude, GPT, or any OpenAI-compatible model. https://t.co/A4YUVOxfO5

@LuckyJoe198x

apify这种数据聚合服务实在是好东西,类似的还有Glasser,它们就像给agent装上了精准制导,我现在要研究一个新东西,直接让Claude用Glasser去Reddit找到相关的sub和帖子,一下就能找到大本营。

@xqliu

做语音 agent,很多人先上 Vapi、Retell 这类托管平台,用着顺手,可通话、模型、账单都在别人手里。Dograh 是冲着它们来的开源替代,整套可以部署在自己机器上。 有意思的是它原生支持 MCP:语音 agent 的对话流程可以交给写代码的 AI 助手,通过 MCP 来设计和修改。不想碰代码也行,它自带可视化流程编辑器,在画布上搭就行,作者说几分钟就能测起来。 模型用你自己的 key。可以直接接语音到语音的模型,也可以把大模型、语音识别、语音合成拆成三段分别挑。它还接了电话线路,做出来的 agent 能直接接电话。 仓库去年 9 月建的,六千一百多星,今天又涨了 17,11 月 4 号刚发过版,BSD-2-Clause 许可,README 有中文版。作者说自己部署只要一条命令,60 秒搞定。 语音 agent 里装着通话录音和用户数据。能自己部署,又能让 AI 助手直接改流程,冲这两点就值得先搭一个试试。

@saurabhpalan

Your AI helped you write the deck. Their AI is going to read it. But the deck is only the final answer. It doesn’t carry the conversation that got you there. Why this price? What did you rule out? What can change? What still needs your call? We built @relaytoai so you can send the doc to the person and the context behind it to their agent. You review what gets shared. Questions that need your decision come back to you. For example: send an investor your deck with the assumptions behind the forecast. Their agent can use those assumptions when they ask about the numbers. We’ve now published Relay’s context format and agent skill on GitHub under Apache 2.0—with a schema, worked example, and extraction standard. Building an agent that reads other people’s work? What would you need to use this? https://t.co/QXEHFSNIE9

@pardnchiu

Agenvoy@v1.3.17 > fetch_page now filters pages against each task with the decision model, cutting LLM input by about 90% [1] > Fix fetch_page keep_links dropping every link > Fix GitHub Copilot `mai-code-2*` models failing to respond > `calculate` rewritten by Claude Opus 5.5 and reviewed by GPT 6.2 sol and Gemini 4 hydrogen, fixing power precedence and large-number precision [1] Why a decision model: 1. Fewer tokens on the main model. TypeSafe's JEV costs $0.042 per 1M input tokens with free output, so letting it drop irrelevant content first is far cheaper than sending the whole page to the main LLM. 2. Less noise. Not every page the agent reads matches the task, and even a matching page carries parts that don't; filtering them out here keeps the main model focused on what the task needs. @grok check the github repo and summarize what’s in this tag

@simeoguid

I built Grip, a free and open source screen recorder for macOS. It auto-zooms, smoothens cursor to your recording and takes screenshots, because it was required for me and my friends to swap it for the native mac recording shortcut. For those who are curious, here's detailed process of making this launch video: - Music: Claude decoded the track with ffmpeg and wrote its own beat analyser in Node: a 120-line FFT with no libraries. It found 140 BPM, the first beat at 1.70 s, the logo hit at 15.42 s and the five flute notes at the end. Every cut and every animation key sits on that grid. - Story: four 2-bar phrases, one Grip feature each, then the logo hit. The order is: the ⇧⌘5 keys, auto zoom, captions, transcript editing, framing on wallpapers, a timer from 64 ms down to 8 ms, a wall of screens for "Free and open source", then the logo and the end card. - Footage: all real Grip footage from its own site, 13 clips cut into frame sequences with ffmpeg. Claude retimed the hero take so the click lands on beat 1 and "Ready" lands on the first break hit. The wallpapers are Grip's real ones, cut from a sheet in its repo. Nothing is AI-generated. - Engine: one 1,560-line JavaScript file, written from scratch. Canvas 2D draws the type and shapes. WebGL2 draws the footage as 3D plates with perspective, rounded corners, shadows and light. A post pass adds 4-subframe motion blur, chromatic aberration, directional blur, a vignette and film grain. All easing curves and springs are hand-written, with no animation library. - Render: the page runs in Malini's built-in browser. It posts each frame to a 35-line Node server that saves PNGs. That is 1,440 frames × 4 subframes = 5,760 renders, in about 110 s for the whole film. ffmpeg then adds the track at 60 fps. - Iterate: Claude checked its own work on contact sheets and fixed what it saw. For example, text sat too low, a crop hid the wallpaper, and the end card was too static. It made 2 full cuts. I gave no notes. The main prompt: make a dynamic ~24 second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out. music, tempo, beat matching - bar's very high, so work backwards from it not to miss out on details. it should be a launch saas video for Grip (https://t.co/5RfUFdmCCG) Claude Code, Opus 5.5 Tools: Claude Code, Malini, Node.js, ffmpeg, WebGL2. No video editor and no 3D app. Music: "Flu And Flute (Logo Version 2)" by Infraction Music.

@okeyccronald

Your AI agent, living on your desktop — not in a browser tab. RedCat is a free, open-source desktop AI agent that runs on your PC. It's built beginner-first: no terminal, no Docker, no config files — download the app, paste in your API key, and start chatting. It accepts your own API keys for OpenRouter, Gemini, DeepSeek, Groq, Claude, OpenAI and more, and your chats, your memory and your keys stay on your PC. It works in an agent loop — understand, act, verify, then ask you to approve before anything runs. It remembers your preferences, corrections and past work, and it ships with a skills library for building websites, deep research, making videos and code review. It can read real pages, cross-check two sources, and hand you a sourced website as the result. It doesn't stop at the desktop: chat with it from Discord, log into sites through its browser panel without it ever seeing your passwords, and connect in one click to a directory of 145 services through MCP and connectors. Anything that runs a command, edits a file, posts or installs waits for your approval first. 1. On Windows (10 or later), grab the setup file from the Releases page and install it 2. On macOS, download the arm64 disk image; on Linux, pick the AppImage or the deb package 3. Add your own API key and start chatting • It's free and open source under the MIT license — the app itself costs nothing • Free app doesn't mean free models: you bring your own API key, and model usage is billed by your provider — OpenRouter has free models to start with • It's very new: the repo went up Oct 27, 2026 and still has 0 stars — early days • Made by manamaliahmad — an open-source project on GitHub • No referral or affiliate setup anywhere in the README — nothing to sign up through us https://t.co/c0gKOXBBbH Source: GitHub official page, checked 2026-11-18

@ChenY390353

昨天 AI 圈最大的新闻不是新模型,是一笔钱。 Isomorphic Labs 拿了 50 亿美元 C 轮融资,生物科技史上最大的一笔风投。上一纪录是 Altos Labs 2022 年的 30 亿,这回直接干到 50 亿。 数字先摆出来:这家公司半年前刚融了 21 亿 B 轮,半年后再拿 50 亿,外部融资加起来 77 亿。领投 Greenoaks 和红杉,a16z 新进,Alphabet 继续加码,老股东一个没落下,超额认购。估值官方没公布,彭博之前报的是按 500 亿在谈。 这家公司是 DeepMind 2021 年分出来的,哈萨比斯亲自带队,干 AI 设计新药,核心是 IsoDDE 药物设计引擎。这 50 亿要砸的事就一件:把肿瘤和免疫方向的候选药推进临床试验,1 月时哈萨比斯自己说的,今年底进临床。 我的判断很直接:资本已经用脚投票了,AI 制药不是讲故事的阶段,是真金白银押它能出药。50 亿押的也是一个还没人跑通的终点——到现在为止,全世界还没有一款 AI 设计的药真正获批上市。 你觉得第一个 AI 设计的药,几年内能真正上市? https://t.co/lDH9xwBn1a

@TedCai233

DeepSeek官网文档写明,自北京时间2026年11月18日12:00起,调用模型时触发的服务端搜索、网页抓取不再计费,只按模型token用量计费。会触发这一服务端搜索的接口包括联网工具、Anthropic API的web_search,以及Responses API的web_search。 官方同时保留独立的Search API和Fetch API按次计费:每1000次成功请求5元。 做agent的这下账好算多了

@lc541944335

谷歌刚把 Genie 4 放进 Project Genie:一句话生成一个能走进去的世界,这次还能边玩边改。 想试之前先看这 4 点: 1. 进去之后能用文字随时改场景,声音、音乐和对话都是实时一起生成的 2. 新加了一个 Project Genie 智能体,用聊天的方式跟你一起搭世界,还能配自定义背景音乐 3. 有社区画廊,别人做好的世界可以直接进去逛,也能改成自己的版本 4. 门槛:Project Genie 目前只给 Google AI Ultra 订阅用户,要 18 岁以上,Pro 用户只能排队等 我平时做萌宠类 AI 短片,最想试的是把短片里的角色放进去,录一段第一视角的视频。已经拿到权限的朋友,想听听实际能玩多久、画面稳不稳。

@maha42cs

🧠 𝗟𝗟𝗠 𝗪𝗮𝘁𝗰𝗵 — 𝗪𝗲𝗱𝗻𝗲𝘀𝗱𝗮𝘆, 𝗡𝗼𝘃 𝟭𝟴 OpenAI ships cheaper frontier models, Alibaba keeps its new flagship closed, and low-cost open-weight models keep arriving. 🔥 𝗧𝗢𝗣 𝗠𝗢𝗗𝗘𝗟 𝗡𝗘𝗪𝗦 ⚡ 𝗢𝗽𝗲𝗻𝗔𝗜 𝗹𝗮𝘂𝗻𝗰𝗵𝗲𝘀 𝗚𝗣𝗧-𝟲.𝟮 𝗔𝘀𝘁𝗿𝗮 𝗮𝗻𝗱 𝗦𝗼𝗹 𝗮𝘁 𝗹𝗼𝘄𝗲𝗿 𝗽𝗿𝗶𝗰𝗲𝘀 OpenAI released GPT-6.2 Astra for demanding professional work and GPT-6.2 Sol for everyday tasks. Astra's API price is $8 per million input tokens and $40 per million output, 20% below GPT-6 Astra; Sol's price is unchanged at $2 and $10. Per Artificial Analysis, as reported by The Decoder, Astra ranks second on its Intelligence Index, behind Claude Fable 5.2. 👉 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: Frontier-level capability is getting cheaper, and OpenAI says Astra is its first released model to reach its "High" AI self-improvement threshold. 🔗 https://t.co/XABd1LOtWi 🇨🇳 𝗔𝗹𝗶𝗯𝗮𝗯𝗮 𝘂𝗻𝘃𝗲𝗶𝗹𝘀 𝗤𝘄𝗲𝗻𝟰-𝗠𝗮𝘅, 𝘄𝗶𝘁𝗵 𝗻𝗼 𝗼𝗽𝗲𝗻 𝘄𝗲𝗶𝗴𝗵𝘁𝘀 Qwen4-Max is a 2.4-trillion-parameter mixture-of-experts model with native vision and a 1M-token context window. API pricing is $1.50 per million input tokens and $4.50 per million output. Alibaba says its average task cost on the in-house Work360 benchmark fell from $4.43 to $1.67 versus Qwen3.8-Max (self-reported). 👉 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: Per Bloomberg, Qwen is the most downloaded open-weight family, yet its new flagship is closed. Alibaba says smaller Qwen4 models will get open weights, with no date given. 🔗 https://t.co/qiernqkIdt 📌 𝗢𝗧𝗛𝗘𝗥 𝗟𝗟𝗠 𝗨𝗣𝗗𝗔𝗧𝗘𝗦 💸 🔓 𝗧𝗲𝗻𝗰𝗲𝗻𝘁 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗛𝘆𝟰 𝗳𝗹𝗮𝘀𝗵 Tencent's Hy4 flash is a multimodal mixture-of-experts model with 12B active parameters per token and a 1M-token context window. API list prices are $0.07 per million input tokens and $0.281 per million output, and weights are on Hugging Face. Tencent says its cost per task on an internal benchmark is about 1/20 of Hy4 (self-reported). 🔗 https://t.co/E3lLKzDl8p 🇪🇺 🔓 𝗢𝗩𝗛𝗰𝗹𝗼𝘂𝗱 𝗱𝗲𝗯𝘂𝘁𝘀 𝗢𝗹𝗮𝗹𝗮 𝗺𝗶𝗻𝗶 𝟳𝗔𝟭𝗕 OVHcloud's first in-house LLM family starts with a 7B-parameter mixture-of-experts model that activates 1B parameters per token. The company says it runs on an 8GB GPU or a CPU, has a 32K context window, and was trained on European supercomputers. It is due on Hugging Face under Apache 2.0. 🔗 https://t.co/dvlo8LhCbN ☕ You're caught up! — @mahadaajani

@SlivenRed

diagram-design lets Claude Code and Codex turn a written workflow into a diagram you can open in your browser. You describe the steps, decisions and relationships, and the skill guides the agent’s choice of layout, spacing, typography and colors. The open-source project has 68,000+ GitHub stars and currently documents 50 visual types, including flowcharts, timelines, architecture diagrams and data charts. Each diagram comes as a single HTML file containing SVG, so you can keep it locally and ask the agent to revise it when the process changes. It can also read your website’s colors and fonts to match your brand, or redraw an existing Mermaid, https://t.co/c4sxt42GDa or Excalidraw diagram using those styles. For example, give it a customer onboarding process and ask for a diagram showing each step, which team handles it, and where a customer goes if verification fails. Including those decision points gives the agent something specific to explain beyond a row of boxes connected by arrows. You can then export the diagram as SVG or PNG for an article, presentation slide or social post. Install it in a compatible coding agent with this command: npx skills add cathrynlavery/diagram-design https://t.co/QZTLdrCatm

@keyelifeai

After finishing a few https://t.co/WNeVALsmyC courses, I struggled to find my next deep dive into cutting-edge AI engineering. Then I remembered a Stanford course I saved in my notes over a month ago—and went back to check it out. Stanford CS329Z: Engineering AI Agents (Fall 2026, taught by Diyi Yang, Michael Ryan, and John Yang) is the exact next step I was looking for. It covers the full stack of building compound AI systems: • Core blocks from scratch (RAG, function calling, MCP) • Frameworks & harnesses (DSPy, LangGraph, OpenHands) • Memory, multi-agent coordination, and optimization • Production evals, benchmarks, and safety Best part: the lecture slides, YouTube videos, and GitHub homeworks are all freely available. If you're learning how to build AI agents in practice, definitely dive into this one. Links to the course homepage, YouTube lectures, and GitHub homework are in the first comment 👇

@Itsrosia88

写代码的 AI 现在不缺,缺的是让它按规矩交活的人,腾讯这个开源项目就是来管这件事的。 LoopForge 11 月 9 日开源,11 月 14 日登上 GitHub 趋势榜,是个自托管的工作台。 它不自己写代码,而是把 CodeBuddy、Codex、Claude Code 等十个编程助手编进同一套流程里。 一套活走下来是规划、实现、评审、测试分工,中间插着脚本检查和人工审批,失败能返工,全程留记录。 跑得好的流程能存成 SOP,下次同一个需求照着走,相当于给团队攒了一套 AI 干活的规矩。 上手有两种:平台版要自己搭 Docker 和服务器,轻量版一条 npx 命令装进你正在用的编程助手。 提醒两点:跑起来要自备各家的 API key 和 GitHub 授权;仓库许可标注有出入,徽章写 MIT、字段写 Other,商用前先核对 LICENSE 文件。 我的判断是:带小团队、用好几个编程助手的人值得试着搭一套,个人单打独斗的话,现有的助手本身就够用了。 来源:https://t.co/qtxMvovblI

@zelin1107

我每天早上找选题最烦的一件事,就是搜出来的全是旧闻。模型都换了两代了,搜出来的文章还在讲上个季度的东西,根本没法用。 昨天翻到一个开源小工具,叫 last30days-skill,GitHub 上快 7 万星了,干的事特别对我胃口:它不搜编辑写的东西,它搜的是大家真在聊的东西。丢给它一个话题,它就去 Reddit、X、YouTube、Hacker News、Polymarket、GitHub 这些地方,把过去 30 天里大家点赞、讨论、真金白银下注的内容翻出来,拼成一份带出处的简报。 装法也简单,仓库 clone 下来,扔进 Claude Code 的 skills 文件夹里,敲一句 /last30days 加话题就能跑。Reddit、HN、Polymarket、GitHub 这几个源零配置就能用,YouTube 字幕装个 yt-dlp 就行,Claude Code、Codex、Cursor、Gemini CLI 都能接。 不过也别把它当万能的,X 和 TikTok 那些源要另外接 key,Perplexity 那个深度研究还得花钱,一次查询约 0.9 美元。而且它给你的只是简报,不是结论,信不信、怎么用还得自己判断。 我反正是先装上了,明天早上选题就靠它翻了,要是翻车了我再回来认错。 仓库:https://t.co/rQCXjklH1s

@TownmediaLabs

🔓 @TencentHunyuan just released Hy4 Flash open weights on Hugging Face. 220B params, only ~12B active per token. Text + image input, 1M-token context. Open weights or closed APIs: which are you building on in 2026? 👇 #AI #OpenSource

@KarthikDur

**Decentralized AI agent swarms perform open ended computational discovery** In the MICDE State of AI Series, Profs. Venkat Viswanathan & Kai Sun presented a milestone in autonomous, end-to-end scientific discovery. A preview of what is to come in many (verifiable) areas of science. - The team deployed a decentralized swarm of AI agents that self-organized by reading and writing to a shared message board. - The swarm targeted fractional topological states, an exotic state of matter that is essential for building noise-robust topological quantum computers. - The AI swarm self-organized across multiple phase transitions to put forth a hypothesis and search space that surprised Kai Sun (a foremost expert in the field) - Remarkably, the swarm achieved a 10x to 15x enhancement in energy gaps (the vital parameter needed to stabilize these states). Note that this is an order of magnitude bigger than Kai Sun's work that was on the cover of PRL just this month!! - The AI model analyzed the results to formulate an analytic theory using real-space quantum metrics, defying 40 years of physics intuition and predicting arbitrary gap enhancement. https://t.co/HgihemjWdR

@AKhomssi

🎬 What if every interesting GitHub project could become a short video automatically? Meet Reelset — an open-source platform that turns developer discoveries into vertical videos. It collects projects and stories from: → GitHub → Hacker News → Reddit → YouTube → AI research papers Then generates short reels with voice-over and captions. It even has an MCP server so your AI coding agent can create videos from links. MIT licensed. GitHub: https://t.co/mJEz8C1xZu Live demo: https://t.co/SqbbRvmdEK #VideoCreation #CreatorTools #AIAgents #OpenSource #DeveloperTools #开源 #AI视频

@malwadie8822

A free AI browser tool that lets agents browse and automate real websites. agent-browser is open source and surged +489 GitHub stars today. Built by Vercel Labs. https://t.co/Jl3U83SvMd

@AKhomssi

🌐 Your AI agent can write code. But can it actually research the internet? Meet Agent-Reach — an open-source toolkit that gives AI agents access to more web platforms. 🔎 Search and read: → GitHub repositories and issues → YouTube videos and transcripts → Reddit discussions → LinkedIn job listings → X posts → Bilibili and Chinese platforms One CLI gives your agent access to different sources through specialized tools. MIT licensed. GitHub: https://t.co/NAS4uCIuWQ #AIAgents #WebResearch #OpenSource #Automation #MCP #开源 #人工智能

@hinadixit

The scarce input in robot learning may be human demonstrations, not model parameters. ScaleVLA, a new arXiv preprint, uses a frozen 552B sparse vision-language backbone. On simulated LIBERO-Goal, 10 demonstrations reached 87.8% success versus 85.6% for a 2B backbone trained with 50. A real robot inserted a charger 9/10 times in familiar layouts, but 8/15 after layout shifts. That’s a capital trade: less demonstration collection versus four H200s for inference. I’d optimize the cost per successful physical task, including recovery and connectivity, not reward model size alone. https://t.co/7Kgvg0E0g5

返回 AI 快讯 →