2026年11月18日 · AI 早报

OpenAI 发布 GPT-6.2 Astra 与 Sol,Google 展示 Gemini 4 Hydrogen 的机器人应用。本轮另一条主线是智能体的实际效率:模型适配、推理加速、验证质量和任务编排,正在共同决定成本与可靠性。

11月17日 23:00 — 11月18日 11:00(日本时间) · 131 条 X 原帖

查看本期全部 131 条原帖 ↓

本期目录

新模型发布与产品进展

GPT-6.2 上线,视觉表现与编码评价出现分歧

OpenAI 官方宣布 GPT-6.2 Astra 和 Sol 已可用,强调效率与现实任务表现。社区首批反馈集中在视觉、三维建模与成本:Super Niu 展示 Astra 先生成肖像、再提取面部关键点辅助建模的过程;SuSu 转述的一组同提示词比较中,Astra 用时约四十分钟、花费八点二美元,Fable 5.2 接近两小时、八十一美元,但后者画面更精细。这些都是具体案例,不能直接推广成通用性能结论。

编码评价则有不同声音。Leo 转述 Artificial Analysis 的编码代理指数,称两款新模型均低于前代;他进一步猜测视觉强化影响了编码能力,但本批没有支持这一因果关系的证据。

来源 · 4 条原帖

Gemini 从对话延伸到机械臂

Google DeepMind 官方介绍,Gemini 4 Hydrogen 配合 Antigravity 新机器人插件,可以连接受支持的机械臂,在 MuJoCo 中练习,再通过自然语言执行真实任务。另一个官方演示则把 Sheets 单元格变成动画像素画。

桌面操控方面,开发者 milo 解包 Antigravity 2.33.0 后称发现独立桌面子智能体、实时画面流和用户接管机制,但相关能力默认关闭、受内部实验开关控制,不能视为已经向用户开放。

来源 · 3 条原帖

区域服务与产业消息

Michael Gately 宣布开放使用澳大利亚 AI 服务 Agentaus 1,自报模型规模二千二百亿参数、MMLU-Pro 得分百分之八十四点四,并承诺数据留在澳大利亚、不用于训练;帖子没有说明开放权重。ZeroDrift 作者则宣布接入 ChatGPT 与 Codex,主打把法规和企业政策纳入初稿生成。

产业层面,CodingNoobie 的社区新闻汇总称 Isomorphic Labs 融资五十亿美元、AMD 完成收购 World Labs。本批仅有该账号转述,缺少公司公告或具名媒体原帖,消息来源的局限需要保留。

来源 · 3 条原帖

开放模型的适配与推理加速

针对 Pi 训练,比单纯换模型更具体

Clem 发布开放权重的 Qwen3.8-27B-pi-adapted,使用强化学习与 LoRA 专门适配 Pi 编码智能体。按其公布的 pass@1,SWE-bench Pro 的 Hard 51 子集从百分之二十九点四升至六十二点七,Terminal-Bench 2.1 从三十九点五升至五十五点八,工具调用减少百分之五十五至六十二。这里体现的是特定模型与工具环境的共同适配,不能当成基础模型全面升级。

来源 · 1 条原帖

消费卡长上下文,需要把系统内存算进去

Isagii 转述的本地测试称,Qwen3.8 Flash-Next 的一千二百五十亿参数 GGUF 在单张 RTX 4090 上实现二十五万上下文、约每秒二十一输出 token。条件是使用实验版 llama.cpp,把专家层卸载到系统内存:显存占用约十八点三 GB,同时使用一百一十 GB 系统内存。这是混合卸载方案,并非整模型装进二十四 GB 显存。

来源 · 1 条原帖

投机解码开始针对真实流量优化

Tech with Mak 转述 Sciforium 的公司测试:保持 DeepSeek-V4.1-Flash 主模型权重不变,为客户流量训练草稿模型,再结合算子融合与跨卡通信优化,四张 AMD MI355X 在长上下文智能体负载下达到 P90 每秒一千一百二十九输出 token。

FHILY 另介绍 NVIDIA 面向 Qwen3.8-27B 的四十亿参数草稿模型,称其 token 接受指标比内置预测头改善百分之四十六;作者明确提醒,这不等于生成速度提高同样幅度,还要计入草稿生成、验证与显存开销。

来源 · 2 条原帖

研究开始追问:分数、判断与自我改进是否可靠

评测高分可能掩盖错误过程

Sol Messing 介绍的 NeurIPS 论文指出,提示词、评判模型和 harness 的噪声会让标注与基准低估误差。山景城小路转述的 SciResearch-Eval 则让智能体面对可能缺失物理机制的模拟器:多数模型仍给出高置信度错误答案,标称百分之九十五的区间,实际覆盖率只有百分之二十至六十二。

同一论文汇总还介绍,真实信息片段也能通过排列方式诱导多智能体接受错误叙事,多数投票或复核并未消除问题。可靠性因此不仅取决于来源真假,还取决于证据如何组织、判断如何验证。

来源 · 2 条原帖

形式化验证有效,迁移判断仍是难题

据山景城小路整理,Turing Agent 通过证明分解、并行验证与经验积累形式化研究级数学,在 Formal TCS Bench 上取得百分之五十七至六十二的忠实形式化成功率。另一项小模型研究却发现,精确验证器可以教会固定检查程序,却未必带来对未见策略的判断能力;可验证任务上的成功,需要与泛化能力分开讨论。

来源 · 1 条原帖

自动研究的进展,首先体现在持续执行

Arnaud Mercier 引述的 AutoRL 论文,把训练监测、诊断提案和执行控制分给不同智能体。在持续十四点七天的编码训练中,系统处理十一次运行故障,SWE 相关任务 pass@1 从百分之五十二点二九提高到六十五点一三;论文同时承认,当前系统更擅长修补配方,尚难提出根本方法突破。

DAIR.AI 介绍的 CompoundingKernel 则用千余智能体搜索 GPU 内核,外部档案保存正确代码及谱系,模型外检查负责判定正确性。其转述结果是在 B200 上,二百零九道公开题相对参考基线加速四点四倍,扩展规模前先提高可靠评估吞吐量,是这项实践的关键。

来源 · 2 条原帖

创意与机器人动作,都在尝试改变输出方式

Alon Albalak 的 Creative Priming 不直接训练模型模仿创意成品,而是训练引导创作过程的模型;作者报告,迁移到前沿 API 模型后,语义不同的回答数量增加四点五至十倍。Victor 介绍的机器人研究则让 Qwen-VL-2.5-3B 直接输出整数动作序列,配合预测集成与动作字符遮蔽,在其转述的无机器人预训练设置下,LIBERO 平均成功率为百分之九十四点七。

来源 · 2 条原帖

Harness 实践:少空转、留依据、查过程

DeepSeek 编排与多智能体的成本账

AYi 转述的写作实践,用 DeepSeek 做调度,通过 Agent Teams 与 ACP 串联多家命令行工具,自动交接选题、大纲、长文和配图。与此相对,Mai Yang 的实际教训是:降低定时检查频率,减少群聊,让单个 Bot 尽量独立完成一件事,公共工具直接共享。

本批两篇近乎相同的 Grok Bot 节流清单合并来看,实质方法是按事件唤醒、缓存结果、精选上下文、限制失败重试。所谓三百个 Bot 每周每个不足两美分,只有社区转述与估算,不能当成普遍成本承诺。

来源 · 3 条原帖

迁移框架时,保存修改原因

Rohan Paul 介绍的 SIM,把每次 harness 修改对应的问题、理由与测量结果写入记忆,换模型后据此保留或修订旧优化。其转述实验中,迁往 Gemini 3.8 Flash 工作区任务后,带记录方案得分八十四点七五,无记录方案为七十三点四二。Unicorn 转述的 Runta 框架比较也提醒:同一模型换框架,成本可能相差数倍,应该看每次成功的总成本,而非只删短提示词。

来源 · 2 条原帖

没有下单,不代表完成了搜索

Xudong Han 转述 Vercel 的咖啡下单评测:没有店铺满足条件时,验证器只查是否未下单,导致根本没搜索就放弃的运行也被算成成功。这个反例把验证要求说得很清楚:既要检查结果,也要检查必要行为。Viking 推荐的 Claude Code 验证文章则把端到端、性质测试和差分测试列为重点投入方向。

来源 · 2 条原帖

创作工具把分镜、编辑与验证连起来

视频工作流开始保留可修改工程

本批多帖介绍的 OpenMotion 是同一项目:根据参考视频拆镜头,用品牌素材重做,导出成片同时保留源码;MOTION.md 用来固定镜头、节奏、转场与声音规范。所谓百分之九十五还原度仍是作者宣传,已有转述者明确尚未实测。

其他项目强调制作前的可审阅步骤:Stickman Video Director 自动模式先给约二十秒有声预览;Claude Video Editing Skills 先转录并生成可调整的剪辑计划;Shlcy 介绍的动画工作台则整合一百零九种风格,但目前只支持 Claude Code,部分风格依赖 Blender。

来源 · 5 条原帖

从专业技能接口到游戏编辑器闭环

启明破晓用 Codex 配合 Blender Skills 制作“火锅星球”,体验重点是直接调用建模、材质和渲染操作。黑箱之外介绍的 Estella 则把游戏编辑器变成 MCP 服务器,开放一百六十七个接口,让智能体修改场景、运行模拟并截图检查。Josh 使用 Paper 的体验也指向同一方向:先在专用设计工具里审阅故事板、批注迭代,再推进实现。

来源 · 3 条原帖

知识库与学习:减少反复解释背景

资料先变成 Wiki,再围绕它工作

多篇 Obsidian 与第二大脑清单可以归为一条工作流:原始资料入库,智能体整理成相互关联的 Markdown 页面,维护索引与记忆,再生成日常简报。Claudian 等工具把编码智能体放进笔记库;如意介绍的 LLM Wiki 桌面项目还能聚类页面,并兼容 Obsidian。整理后的页面仍可能遗漏信息,重要结论需要回查原文。

来源 · 3 条原帖

科研与课程资源,保留动手和核验环节

AI 百球哥转述的一名食品科学研究生经验,比泛泛提示词清单更具体:用 Zotero 分批导出摘要,交给 Qwen 筛选和提取结果,再按机制组织章节,最后人工逐项核对数字与术语。学习资源中,MiniMind 覆盖小模型预训练到强化学习的完整链路;Zhaoyang 推荐的 CS329Z 则讨论智能体设计模式,以及多智能体何时有帮助、何时反而拖累。

来源 · 3 条原帖

安全与可用性:外部文本正进入执行链

Issue 与 PR 评论成为注入入口

Pritesh 转述,Depi 展示了通过 GitHub Issue 诱导 AI 标签工具执行隐藏代码、触及 Codex 的 npm 发布令牌的路径,并称 OpenAI 当天修复,目前没有已知恶意版本。Aleksandr 另自述在一个开源编码智能体服务器中,发现任意 GitHub 用户可经 PR 评论向拥有管理工具和 shell 的会话发送指令;帖子未交代修复状态。这两件事都把外部讨论文本与执行权限之间的边界推到台前。

来源 · 2 条原帖

自动攻击案例与地区限制,需要按来源理解

Real Or Rumor 对 Google 威胁报告的讨论称,已知一个工具能自行读取补丁、定位漏洞并执行攻击,其他案例仍属部分自动化;该账号也承认无法打开完整报告,部分目标信息来自 SiliconANGLE,不能据此推演成普遍现象。

服务可用性方面,AB Kuai.Dong 转述 Witten 在香港科技大学使用 Claude 时遇到地区不可用提示。这能说明一次访问受限经历,不能推断账号随后被封,也不能据其他检测网站的结果确认 Claude 的具体风控机制。

来源 · 2 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 131 条原帖

@OpenAI

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. https://t.co/qr2iP0GpfX https://t.co/jLCyVIws2Y

@theo

https://t.co/M034Qyf2d5

@ClementDelangue

Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: https://t.co/aH4cl2qnED Methodology: https://t.co/avSfF93K4V

@_FORAB

太惨了。美国物理学家 Edward Witten,今天来香港科技大学,分享黑洞热力学,结果期间习惯性地打开 Claude,被提示你所在的地区不可用 🤣 希望教授的账号,别后续被封了。 https://t.co/bqpv0Ma9ek

@irongiantXBT

holy f*ck this is smart a 16-year-old cs student in china worked out how to run 300 grok bots for under $0.02 per bot per week the same swarm would burn $4,500 a week without it he cut pointless status checks, stopped rereading full chat histories and killed every endless retry loop now most bots stay idle until there's real work to do here are the 20 github repos he used to bring the cost down: 1. markitdown > convert pdfs, office files and html into clean markdown before analysis so the bot reads the content, not the layout https://t.co/b2pSBDgxm4 2. llmlingua > compress long inputs before the model reads them while keeping the facts, constraints and qualifications the task depends on https://t.co/UuJrvQHhHx 3. firecrawl > collect web content as markdown or structured data so the workflow keeps only the fields it needs https://t.co/VgGWJwCpIn 4. rtk > compress common command outputs before they enter the conversation so repetitive logs don't swallow useful context https://t.co/WsEmtbRQgm 5. chonkie > split long documents into chunks so retrieval pulls the relevant passages instead of the whole file https://t.co/EoyNfi4a9c 6. flagembedding > rerank retrieved passages before generation - strong evidence stays, irrelevant matches leave the prompt https://t.co/5k3Fcp5fmn 7. mem0 > store and retrieve relevant facts across runs so each task starts with selected memory instead of the entire conversation https://t.co/cJOKzLEJTg 8. prefect > cache successful task results and reuse them when configured inputs match, skipping reruns of steps that already finished https://t.co/jvvXvMK262 9. gptcache > serve saved answers to similar requests instead of generating again - set matching and freshness rules around the task https://t.co/hTJa8c5bpB 10. graphiti > keep changing facts in a searchable graph and pull relevant context instead of replaying the full event history https://t.co/aZeprpl93O 11. portkey gateway > cache matching api responses behind one shared gateway so connected workflows reuse answers across repeated requests https://t.co/fqrRXs1nge 12. hishel > add http caching and revalidation to your fetching tools so unchanged responses stay out of repeated download-and-analysis cycles https://t.co/XmX1XNegPc 13. celery > put background jobs on a worker queue - code handles waiting and dispatch, model calls run only inside tasks that need them https://t.co/4iANsEmi0Q 14. requests-cache > store fetched responses with expiry rules so your scripts reuse valid data before asking the bot to look again https://t.co/IGoIfWpq3W 15. apscheduler > run lightweight checks on a schedule in python and batch nonurgent work before passing useful updates to the agent https://t.co/lRTTMCkR0H 16. routellm > send simpler requests to a cheaper model and keep the stronger one for work that needs it https://t.co/GINxKmUHDR 17. phoenix > trace your ai workflow, inspect repeated calls and compare changes against output quality before keeping an "optimization" https://t.co/OMOdvbfqqQ 18. pybreaker > pause calls to a service that keeps failing once it hits a set threshold, so one broken dependency can't loop forever https://t.co/XYot1jHNTp 19. semantic router > match requests to predefined workflows with local embeddings instead of paying a large model to pick every next step https://t.co/evxJcAbrMs 20. litellm > track api spending and set budgets before one runaway workflow eats the whole allowance https://t.co/rdRKZ4girP connect the tools your workflow needs through the terminal, mcp or your own code - route api calls through the gateway you picked so its controls actually apply start with the biggest leak: oversized inputs, repeated research, empty checks or failed operations that keep restarting filter before reading > reuse before generating > check for changes > stop failed loops then measure finished work against total usage, including the tools running around your bot the full course below shows how to put the workflow together ↓

@ChrisSlacker

30 个用起来像"非法"其实完全合法的 AI 网站 1. https://t.co/d9kdLEcwKl — 免费无限 AI 生图,质量直逼 Midjourney 2. https://t.co/ry9RjlTSSC — 实时 AI 生图,边画边出图 3. https://t.co/cpFAhEYCi2 — AI 无限放大图片,细节自动补全 4. https://t.co/D0DDmgOrdx — AI 一键去背景/补光/擦除 5. https://t.co/vpmn30zW44 — AI 语音克隆,5 秒模仿任何声音 6. https://t.co/XsmRvT9M85 — 输入歌词自动生成完整歌曲 7. https://t.co/vepZVwuETI — AI 视频生成先驱,Gen-3 免费试 8. https://t.co/usRX7kMMMR — 快手可灵 AI,中文视频生成最丝滑 9. https://t.co/kDpwdiCmD8 — 一张照片 + 一段语音 = 说话的数字人 10. https://t.co/zNyevNCo3f — 静态照片变说话视频 11. https://t.co/lKSguUTF71 — AI 写代码,免费额度够日常用 12. https://t.co/zJU8eIeXps — 说话就能建网站,零代码 13. https://t.co/9i2MeMTIqp — Vercel AI 前端生成器,描述即页面 14. https://t.co/POOiCJr3q9 — 浏览器里写代码 + AI 辅助 + 一键部署 15. https://t.co/K42GU6VXA8 — 粘贴文字自动生成信息图 16. https://t.co/Ce4wJX9ojL — AI 一键生成 PPT,告别 PPT 地狱 17. https://t.co/jbynStX4fK — Notion AI 写作、总结、翻译 18. https://t.co/dpQBw8vaXT — 画草图 AI 自动生成真实网页 19. https://t.co/eeQseYw1fQ — AI 搜索引擎,问任何问题秒给答案 + 来源 20. https://t.co/oDiZlpDc0G — 开发者专属 AI 搜索,代码问题一搜即解 21. https://t.co/ilU3hO6I85 — 会议实时转文字,免费每月 300 分钟 22. https://t.co/okmhOXSuLY — AI 剪辑短视频,自动找高光片段 23. https://t.co/u5xMfCKQam — 一个网站用 GPT-4o / Claude / Gemini 24. https://t.co/eTTnu21a1u — 开源 AI 模型免费试玩场 25. https://t.co/W2Z0Jo6vsE — 一键 AI 去背景,1 秒出图 26. https://t.co/DoQaUm1WGf — AI 擦除照片中任何物体 27. https://t.co/7eBGojIMsK — AI 换脸,1 分钟出片 28. https://t.co/opidD4kHVM — AI 音乐标签 + 推荐 29. https://t.co/m08eUzvrJZ — AI 写作助手,免费版够日常 30. https://t.co/Vw5PteH8Ed — Anthropic 免费 AI 助手,长文处理天花板

@MGately_au

We built Australia’s own AI model, and today it’s open to the world. Agentaus 1 (@AgentausTD): 220B parameters. 84.4% on MMLU-Pro. Ranked 43rd in the world. It runs on Australian servers. Nothing ever leaves Australia, and nothing you type is used to train AI. Delete a chat and it’s gone for good. Most of the world’s AI answers to somebody else. This one answers only to you. Free to start: https://t.co/pkgBYQlpkZ

@GoogleDeepMind

Teaching robots new tasks just got a lot faster. 🤖 With Gemini 4 Hydrogen and the new robotics plugin for Google @Antigravity, you can connect a supported arm, practice in @MuJoCo simulation, and run tasks in the real world using plain conversation.

@GoogleDeepMind

Gemini 4 Hydrogen turned Google Sheets into a pixel-art canvas. Using @Antigravity, the model shrank thousands of cells to create animated scenes. https://t.co/BHtrN5CSrt

@RahulTechX

🚨 BREAKING NEWS! The engineer who created Claude Code from scratch just posted a 28-minute video that's pure gold: how to write prompts that actually deliver ridiculously good results. I've seen $300 courses that don't even come close to what he explains in the first 10 minutes. CLAUDE.md files, memory shortcuts, parallel sessions, and prompting patterns that almost no one uses… All in one video. Completely free. No fluff. Whether you're a developer, just starting out now, or already using Claude for months: this changes the game for you starting today. Save. Watch. Share.

@techNmak

These are literally the best visual resources for learning how Transformers, LLMs, embeddings, diffusion, and inference work ↓ 1/ Transformer Explainer - Watch GPT process text through embeddings, attention, MLPs and next-token prediction. https://t.co/vkyxOqDVll 2/ Brendan Bycroft’s LLM Visualization - Explore an LLM from architecture down to tensors and operations. https://t.co/ZFGH0pRtP3 3/ 3Blue1Brown - Visual intuition for linear algebra, neural networks, backprop, attention and Transformers. https://t.co/hTqyRAvVqw 4/ The Illustrated Transformer - One of the clearest visual explanations of embeddings, Q/K/V and attention. https://t.co/tYrJCm17L6 5/ TensorFlow Playground - Watch neural networks learn as you change layers, activations, features and learning rate. https://t.co/u0KAiXWjAg 6/ Google PAIR AI Explorables - Interactive explainers on LLMs, generalization, interpretability and model behavior. https://t.co/u0KAiXWjAg 7/ Distill - Exceptional visual essays on t-SNE, feature visualization, GNNs and interpretability. https://t.co/SNqhHxuniD 8/ Abhik Sarkar’s Transformer Visualizations - RoPE, KV cache, FlashAttention, MQA, GQA and more. https://t.co/u0KAiXWjAg 9/ Visual Guide to Attention Variants - MHA, MQA, GQA and MLA visually compared. https://t.co/nbpKMa0cpT 10/ Visual Guide to Mixture of Experts - Routing, experts, sparse activation and load balancing. https://t.co/3gu7xq5e3i 11/ Modular LLM Inference Handbook - Prefill, decode, KV cache, batching, quantization and speculative decoding. https://t.co/itMIfZnPP5 12/ Apple Embedding Atlas - Explore clusters, neighborhoods and outliers in large embedding spaces. https://t.co/hxKdXSp5I2 13/ Diffusion Explainer - Follow Stable Diffusion step by step. https://t.co/YVHEG1zslC 14/ Neuronpedia - Explore features, activations, SAE latents and attribution graphs inside real models. https://t.co/L09bXWCTmA 15/ Seeing Theory - Probability, Bayes, distributions, regression and inference made interactive. https://t.co/Jk33dueK14 16/ CNN Explainer - Visualize convolutions, feature maps, activations and pooling. https://t.co/sCwv59d2S6 17/ GAN Lab - Train a GAN in your browser and watch its generated distribution evolve. https://t.co/G94RDWMNak Save this. There’s a serious AI curriculum hiding inside these links.

@waveking1314

求你了,现在就把这段丢进 Claude Fable 5.2…… 它会给你搭一个 24 小时不停的交易机器人,通宵把整个市场研究一遍,帮你挑出最能打的交易策略 趁大家还没开始用,赶紧抄走: https://t.co/OXPTSjiCLK

@akshay_pachaar

Train a 64M-parameter LLM from scratch in just 2h! MiniMind covers the full pipeline in plain PyTorch, from pretraining and SFT to LoRA, DPO, RL, and distillation, on a single consumer GPU. 100% open-source. https://t.co/qQah5QPIr7

@vikingmute

Addy Osmani 的长文质量太高了,太干了,如果你打算搞什么起号,直接把他的文章翻译一下加点点评绝对哪里都成爆款。 这篇是《How the Claude Code team thinks about verification》 https://t.co/dnagpkqztP 讲的是Claude Code 团队如何通过验证闭环让 AI write 和 review 大量代码 最值得投入的三种测试: E2E,Property 测试(孤陋寡闻 我居然是第一次听所)Differential 测试 后面结合这三种测试将了 verify 的步骤,特别特别细

@AYi_AInotes

说个残酷的现实,Anthropic 掌门人达里奥最近被问到一个问题,刚高中毕业的年轻人现在最该干什么,他的回答直接把初级律师、初级金融分析师和初级顾问全判了死刑, 原话特别直白,AI 已经学会了怎么干初级律师的活、初级金融分析师的活、初级顾问的活,我们没法否认它会淘汰大量岗位,而在短短几分钟里,他只给年轻人指了一条能继续留在牌桌上的生路,学会指挥 AI, 我刷到原帖整理的那份 24 个全免费 AI 与智能体学习清单时,第一反应是这波资源确实够全、连一分钱学费都不用花,但转念一想,如果大家真像以前考研考公那样,把这 24 个链接从头刷到尾,大概率还是会沦为被 AI 淘汰的那 95%, 原因特别简单,过去学校教我们的是老老实实当一个高级执行者,花一年时间背知识点、练基本功,再去跟别人抢一个初级岗位,但在智能体已经能 24 小时批量干活的今天,跟 AI 比谁更会做基础分析、谁写初稿更快,纯属拿肉身去撞高铁,真正能留下的那 5%,从来不是把课程刷完的人,而是直接拿这些工具搭出自己 AI 员工团队、负责下指令和验收结果的包工头, 我把这 24 个零成本官方与顶级开源资源,按实战用途重新理了一遍,大家千万别贪多,缺哪块直接点进去上手跑, 第一组,先把嘴替练明白,学会用提示词精准下指令 1. Anthropic 提示词工程互动教程,边练边改最快掌握指令结构 https://t.co/dvC60UeX2z 2. Anthropic 官方系统课,从第一条提示词一路过渡到调度智能体 https://t.co/BWQlup9ZoH 3. Anthropic 实战食谱,现成的代码与业务配方拿来就能改 https://t.co/5w02XZzoma 4. https://t.co/2C2ux6EgXj 提示词工程指南,全网流传最广的提示词百科全书 https://t.co/oryBJXpKL7 5. OpenAI 官方食谱,覆盖绝大多数日常高频任务的现成代码模板 https://t.co/4wptRLfIEh 第二组,零基础补齐大模型与工程底层常识 1. 微软生成式 AI 入门课,整整 21 节课,完全不需要计算机学位 https://t.co/IWu6Y4Diub 2. Google Cloud 生成式 AI 实战笔记本,直接动手跑通云端案例 https://t.co/rNLRKAvGtd 3. Andrej Karpathy 神经网络从零到精通,想彻底搞懂底层原理看这一套就够 https://t.co/zcIpfim408 4. https://t.co/lMNHSavZln 短期课程,全是行业一线专家讲的一小时精华课 https://t.co/VS5j0VBBND 5. 优秀大模型应用合集,几百个真实可运行的开源项目直接拆着学 https://t.co/vkuDpoFU9e 第三组,直接上手搭建你的全自动 AI 员工团队 1. 微软 AI 智能体入门课,手把手教你造出能替你干初级工作的智能体 https://t.co/YRKeUbCads 2. Hugging Face 智能体全套课程,从零到一免费学习还带官方认证 https://t.co/GYcHuELvgU 3. CrewAI 多智能体框架,直接让好几个 AI 角色分工协作跑起来 https://t.co/hCiaY8og0D 4. 微软 AutoGen 框架,让多个智能体自动对话协作啃下复杂任务 https://t.co/oczlTkWfeM 第四组,Claude 全家桶官方专项实战课 1. 与 Claude 一起掌握 AI 基础 https://t.co/VXKcrZhM4G 2. 与 Claude 一起实现 AI 自动化 https://t.co/OpbbBXrcaA 3. 与 Claude 一起协作使用 AI https://t.co/yE6AhInViw 4. 使用 Claude API 进行工程开发 https://t.co/4O3kej35V2 5. Anthropic 模型上下文协议 MCP 进阶专题 https://t.co/V5MtuFhdpH 6. Anthropic 官方 AI 熟练度教学指南 https://t.co/9zCcJ0FpMd 7. 人人适用的现实世界 AI 专项课程 https://t.co/r0RaJiHwl6 8. 市场营销场景实战 Claude AI https://t.co/axnkZBJnIa 9. 零基础从头学 Claude https://t.co/4o4LD9hxBN 10. Claude Code 用智能体做现代软件工程 https://t.co/5DTm3HOvCW 说实话,这 24 个资源全部免费,一分钱不用花,但最值钱的从来不是链接本身,而是你今天就把手头最繁琐的一件重复工作丢给智能体去跑,在毕业季和职场洗牌真正砸下来之前,早一天学会当 AI 的指挥官,就少一分被算法替代的焦虑。

@opentensor

After Tencent Cloud, Bittensor subnet @chutes_ai has partnered with @alibaba_cloud. GLM-5.3 and DeepSeek V4.1 Flash are 35% off on Chutes through November 23.

@AYi_AInotes

卧槽牛逼,这绝逼是最近看到的DeepSeek Harness 的最佳实践! 有人竟然已经用开源底座把 DeepSeek、Grok、Claude、Gemini 和 Codex 编成了一支全自动写作团队了,丢进一句话,带配图的万字长文直接自动交接出炉, 用低成本高逻辑的 DeepSeek 当调度总监,配合四家各有所长的命令行工具本地打通,整套内容生产被彻底重构成一条自动化流水线, 彻底实现一个人干出一家深度内容工作室的产能,秘诀就在于彻底砍掉跨窗口复制粘贴的脏活, 通过 Agent Teams 插件和 ACP 协议把五大模型在本地串成自动流转的任务看板,从扒热点、搭大纲、写长文到代码画图全自动跑完。

@swapnakpanda

Stanford's TEN FREE Courses on AI & ML: ❯ CS336 - LLM from Scratch ❯ CS221 - Artificial Intelligence ❯ CS229 - Machine Learning ❯ CS230 - Deep Learning ❯ CS234 - Reinforcement Learning ❯ CS236 - Deep Generative Models All links inside:

@frankyecom

Easy 3 adjustments for all Creative Strategists. >Opus 5.5 on Medium not High. (Better than Fable 5.2) >Take initial ad scripts from Claude and run them through Gpt/Grok again asking for improvements in realistic UGC vocabulary (This is called LLM stacking. I don't see enough people talk about this, but it has made some of my top performing scripts) >Omni 1.1 instead of 1.2 for yaps and run 15 seconds of script instead of 10 for 100% natural UGC pace. I also use Grok for all Omni prompting, it's by far the best. These are all micro adjustments you find along the way through trial and error, if any of you guys have any other gems please drop em below. Happy scaling.

@xDinoDeer

抛开立场问题,我恰恰认为 Dario 很有种,他的意识形态是一贯的,没有摇摆不定的,他的产品策略也是极其坚定聚焦,不做大众市场,不做所谓的多模态,坚定认为 coding 是通过 AGI 的坦途;这意味着他的认知一定是高于做大模型的同行的,在这点上,梁文锋也是后知后觉,最终也认可这一点,全力追赶 coding 的智能。 他的 blog 非常值得一看:https://t.co/V7QumH7JsI

@wquguru

https://t.co/jAhKeHLVod

@cnyzgkc

分享 8 个 做游戏开发 必须知道的 Skill,先收藏再装~ 从生成素材、写玩法、做界面、调手感,再到多人联机,链路基本齐了: 1. higgsfield-websites:生成游戏素材,顺便把浏览器游戏搭起来、部署出去。原来的 higgsfield-game-generation 已经并进这个 Skill 了。 https://t.co/ZNnYKRN9vy 2. game-engine:游戏循环、物理碰撞、2D / 3D 渲染,底层先立住。 https://t.co/0vJum09QSM 3. game-developer:Unity / Unreal 的玩法开发和性能优化,偏引擎侧实作。 https://t.co/a0c826llB1 4. game-ui-design:菜单、血条、背包这些界面,别再是能玩但不能看。 https://t.co/OCzVeBk9xL 5. game-design-theory:玩法、难度、奖励怎么设计,专门治「能跑,但不好玩」。 https://t.co/r1GF00ymMl 6. game-feel:打击反馈、屏幕震动、音效,手感是另一回事。 https://t.co/OSTY27yhAu 7. multiplayer-game(历史版):玩家匹配、房间、实时状态同步。当前主分支已经没有这个目录了,要用的话看这笔历史提交,依赖得自己核对一遍。 https://t.co/EFegPYpEv5 8. develop-web-game(OpenAI 历史版):让 Codex 边写边测,看截图、查报错、修 Bug。今年 4 月 23 日已从 OpenAI curated 目录移除,同样是历史提交。 https://t.co/t3pbfv0eVd 如果只是想先做网页小游戏,也不用一口气全上,先用 game-engine + game-feel + develop-web-game 这一组就够起步。

@AlbalakAlon

Do you want your AI model to be more creative? How would you train it for creativity? What if I told you that directly training a model to be creative can actually prevent the model from reaching its full creative potential? Think about this: using past examples of creative success as training examples makes similar outputs more likely in the future. But to be creative, reproducing past examples is exactly the opposite of what you want! This thought experiment led us to what we call the Creative Paradox: "emulating examples of past creative success (even if they were your own successes) can actually prevent you from reaching your full creative potential." In "Creative Priming: Decoupling Process from Product to Escape the Creative Paradox", we take a different approach: train a creative priming model to guide the creative process, rather than directly generate responses. We trained a priming model to make Qwen3-8B more creative, but the principles it learns transfer to frontier API models! With creative priming, API models generate 4.5-10x more semantically distinct responses, with 2.5-4.6x more diversity! A massive thank you to the whole open-endedness team at @LilaSciences, to everyone who contributed to the paper (Sebastian Gabriel, @Evo_Daveo, @joelbot3000, and @kenneth0stanley ), and to LILA for giving us the space to explore big questions like this (and more!)

@s4yonnara

THIS IS F*CKING WILD... you're wasting hours running your social media while a Grok Bot army could do the work of an entire content team on autopilot. One Bot finds viral trends. Another studies competitors. One writes in your voice. Another creates visuals. The rest handle publishing, engagement and analytics. 15 GitHub projects to build your own AI social media agency: GROK BOT TEMPLATES & INTEGRATIONS 01 / Awesome Grok Bot Templates → Ready-made Bot profiles for content, marketing, research, sales and more. https://t.co/gWQJ3QAbM4 02 / Awesome Grokbot → Copy-paste Grok Bot profiles, including Social Queue, Viral Tweet Scout and Competitor Watch. https://t.co/XCnfzaYVSM 03 / So-me Studio Grok Plugin → Connect Grok Bot to a social media studio for drafts, media uploads, scheduling and publishing. https://t.co/1YxwAwmyr7 04 / Genviral MCP → Let Grok Bot create, schedule, publish and analyze content across social platforms. https://t.co/JBjZPg63Xi 05 / Socialync MCP → Connect Grok Bot to cross-platform publishing, scheduling and performance analytics. https://t.co/ujexnoChjq AI CONTENT CREATION & STRATEGY 06 / Social Media Skills → AI skills for content strategy, writing, hooks and performance analysis. https://t.co/Ptcd63eZSY 07 / 106 Social Media Skills → A massive toolkit for research, competitor analysis, viral content, video, design and growth. https://t.co/OktpCPgeOm 08 / Social Media Agent Skills → Turn one idea into platform-native content for X, LinkedIn, Instagram, TikTok and more. https://t.co/HOhDddiAHv 09 / Awesome Social Media Skills → A curated collection of 189 agent skills and 27 MCP tools for social media workflows. https://t.co/t0l4MH8c6r 10 / Social Media Skills by SocialRobot → Skills for content calendars, publishing and social media management across 9 networks. https://t.co/UzJZZ5DV8o PUBLISHING & AUTOMATION 11 / PostQueen → AI-powered scheduler with MCP support for publishing across 30+ social networks. https://t.co/q39GykTNxT 12 / Postiz Agents → Connect AI agents to social media scheduling and publishing workflows. https://t.co/q2GZ2K3pJP 13 / TryPost → Open-source social media scheduler built specifically for AI agents, with MCP and REST API. https://t.co/qY4r3kGh72 14 / Postsider → Publish and schedule content across 30+ platforms through an API, SDK and MCP server. https://t.co/hZKZIp8B8P 15 / DemandBird → AI-native content pipeline for creation, repurposing, scheduling and analytics. https://t.co/ezoQwX8oUo THE 6-BOT SOCIAL MEDIA TEAM: Scout → Finds viral trends and content ideas. Strategist → Studies competitors and plans campaigns. Writer → Creates posts in your personal voice. Designer → Produces images, graphics and videos. Publisher → Schedules and distributes content. Analyst → Tracks performance and improves the next batch. One Chief of Staff. Five specialist Bots. Your entire content operation working together. You bring the ideas. They do the heavy lifting. Bookmark this. Your future social media team is already on GitHub. Follow for more Grok Bot setups, AI agents and open-source tools.

@BardhylBytyqi

If you design with AI, save this 1. https://t.co/BVLdRK7UqB — best vibecoding tool 2. https://t.co/18k94vwBgP — top notch landings 3. https://t.co/qfXlaziR51 — free templates 4. https://t.co/SiyE5sMFEh — 1 prompt 1 website 5. https://t.co/GwoiHkfiYN — world best icons 6. https://t.co/bfeoJy1e99 — fonts behind brands 7. https://t.co/3J5MJsvWDD — design community 8. https://t.co/VXNLXhmsoD — award websites 9. https://t.co/7raDCl7bq1 — mobile UI animations 10. https://t.co/Zyngx75BhQ — footer inspiration 11. https://t.co/xTjAWnt4X4 — next level design 12. https://t.co/Q3Fm8mtHNx — backgrounds 13. https://t.co/dUWlOXqWPa — big brands posts 14. https://t.co/FcRtygMWKH — animated components 15. https://t.co/HnfTgTWTw8 — figma designs 16. https://t.co/O4YDfuQjwz — real app screens 17. https://t.co/ZvLLkeIuwU — premium logos 18. https://t.co/pOtJrp5Hiw — paper anim bookmark this

@IamKhanPhD

Claude is a game changer for academic research. These 13 prompts help you review literature, uncover debates, identify research gaps, and synthesize evidence from uploaded papers 👇👇 Professors don’t want you to know. (Save this for your next research project) https://t.co/p1cNCYgcf6

@SolomonMg

The evals are not ok New NeurIPS paper: Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking. Evals drive deployment, safety standards, and research. Yet we ignore noise from prompt, judge, agent harness etc, & so underestimate error in benchmarking & annotation, & it gets worse with more data. I'll get to what this means for CIs & how to fix, but first let's talk about how this makes benchmark hacking easier 🧵

@stephenhaney

Paper with Linear agents is really promising to me You could have a canvas ready to review before you pick up a ticket, or create visual audits of work that has been done by agents. You could keep the canvas in sync with code changes. What else? It feels like new ground

@Zhaoyang___

The Stanford NLP @stanfordnlp Group's CS329Z course is absolutely brilliant! Michael @michaelryan207 has delivered two of the best lectures I have ever watched: AI Agent Design Patterns (Lecture 6) and Multi-Agent Systems: When More Agents Help or Hurt (Lecture 8). Michael's teaching is truly exceptional! He explains these two important topics with remarkable clarity, insight, and logical structure, offering a thoughtful and systematic introduction to the core concepts and design patterns, their strengths and limitations, as well as the challenges involved. If you are currently working on AI agent design or research, these are two lectures well worth your time! https://t.co/U2MxJSBEkk https://t.co/XGKyYl9uCz

@EngMoElgaraihy

🚨 مشروع GitHub ممكن يغيّر طريقة وصول AI Agents إلى الإنترنت بالكامل! 👀 اسم المشروع Agent Reach، وفكرته بسيطة لكنها قوية: بدل ما يكون الـ AI Agent محدودًا بأدوات البحث التقليدية، تمنحه القدرة على الوصول إلى منصات ومصادر محتوى متعددة وقراءة بياناتها والبحث فيها. 🔥 ماذا يستطيع أن يفعل؟ → 🐦 البحث وقراءة المنشورات على X. → 🔴 البحث في Reddit وقراءة النقاشات. → 🎥 مشاهدة محتوى YouTube واستخراج الترجمات. → 💻 البحث في GitHub وقراءة المشاريع والـ Issues. → 📺 الوصول إلى Bilibili. → 📕 التعامل مع منصات مثل Xiaohongshu. → 🌐 قراءة صفحات الويب وتحويلها إلى محتوى منظم يفهمه الـ AI. والأجمل أنه يعمل مع أدوات مثل Claude Code وCodex وOpenClaw وغيرها. بدل ما تضيع وقتك في تثبيت أدوات منفصلة وضبط كل منصة بنفسك، Agent Reach يتولى التثبيت، وفحص القنوات، واختيار طرق الوصول المناسبة. المشروع مفتوح المصدر ومجاني، لكن بعض المنصات تحتاج تسجيل دخول أو إعدادات إضافية.

@NFT_Chen

🚨突发!智谱 GLM-5.5 疑似灰测! 有开发者在 GitHub 项目里蹲到了疑似 GLM-5.5 灰测的关键迹象: 🔹一个要求 Agent 提交 PR 时必须标注自己模型和 harness 的 GitHub 项目中,Agent 自动写上了 “Authored by a coordinator subagent (GLM-5.5 session, 2026-11-17)”,项目本身和提示词从未提及过 5.5。 🔹对应 PR 文件内容与 attribution 高度一致,且已合并(PR 地址:https://t.co/MfhmVGpAub)。 目前公开可用仍是 GLM-5.3 系列,这次出现在真实 Agent 工作流里,极可能是智谱内部/灰度放量的信号,继续蹲后续 PR 的 LLM attribution 吧! #智谱 #GLM #GLM55 #ZCode #大模型灰测 #AIAgent #唐杰

@cwolferesearch

While writing my blog on score centering, I found many resources on RL stability, training-inference mismatch (TIM), and importance sampling. Here are some favorites: Training-inference mismatch (TIM) and RL stability. - Isolating numerical TIM and its impact on RL stability: https://t.co/ab7VT713pv - Numerical mismatch versus policy staleness: https://t.co/3Ki2yiilHK - When Speed Kills Stability (blog): https://t.co/QFg17v25Ze - Training-inference router mismatch in MoEs: https://t.co/dXyx33npim - Non-determinism in LLM inference: https://t.co/oUQJCh48lL Score centering and off-policy drift. - Score centering paper: https://t.co/BR1vLD5e0y - Notes on off-policy drift: https://t.co/RXr083Hodh - Nathan's score centering writeup (RLHF book): https://t.co/iT9QqFwu7M - StableREINFORCE docs (Tinker): https://t.co/Jk7NlR7g13 - My blog: https://t.co/9xYItEfjRo Importance sampling and related techniques. - IcePop (sec. 2.3.2): https://t.co/NZBY16xcf4 - TIS blog: https://t.co/eIJGBitRrg - Calibrated importance sampling (CIS) and truncation bias: https://t.co/WW26Hghgqy - CISPO (sec. 3.1): https://t.co/mZj9UQsrnC - Token-, cumulative-, and sequence-level importance sampling: https://t.co/TGYjkPlPSZ - Trust region masking: https://t.co/738YvEKFkE - TOPR: https://t.co/A33lc8y3R7 - GSPO: https://t.co/C5lIi7zepv - DPPO: https://t.co/aAAbgFSrSr

@kumesh_a

Today we’re launching ZeroDrift for @ChatGPT and Codex. @zerodrift_ai knows your regulations and company policies, so your teams can move faster with your rules built in from the first draft. Watch the film below. Book a demo: https://t.co/kzblHsBaVI https://t.co/gXYZN0uvay

@benroy

NEW post riffing on creativity and what makes human thinking special compared to machine intelligence https://t.co/lcJPLkAgoD

@NFT_Chen

😱天呐!一枚小小的LoRA适配器,把这个Qwen 3.8 27B在Pi Agent上彻底激活了!最难工程任务从29%飙到63%! 直观战绩 🔹最难的软件工程任务(SWE-Bench Pro Hard):成功率从29.4%直接干到62.7% 🔹终端任务(Terminal-Bench 2.1):从39.5%提升到55.8% 🔹平均少调用**55-62%**的工具,跑得更快、更省、更稳 🔹真实例子:原模型跑到100步还失败,适配器只用8次工具调用就搞定 5分钟搞定: 1️⃣去Hugging Face下载适配器:harness-hub/Qwen3.8-27B-pi-adapted 2️⃣用vLLM加载模型和适配器 3️⃣在Pi里把模型指向本地服务 4️⃣直接开始用:pi --provider local --model qwen3.8-pi 5️⃣也支持llama.cpp、https://t.co/pAtwoaw4Nl,或者直接把适配器合并进模型 开放权重已上线,完整训练方法周五公开! 别再硬扛通用模型了,现在就换上这个,体验直接起飞! 模型地址:https://t.co/FQTcR2ok4J
 #Qwen38 #PiAgent #开源AI #本地Agent #SWEBench #阿里巴巴 #刘大一恒

@eth_apple

CT is dead 🥲 Launched a dedicated AI educational content account this week. - Hit 1M views in just one day. - Language: Korean - Content: LLM applications (ChatGPT, Claude, Gemini), AI agent development, prompt engineering, AI workflow automation, and practical AI tutorials. https://t.co/O5XCDDixx9

@dair_ai

Banger paper from Amazon on self-improving agents. (bookmark it) CompoundingKernel writes GPU kernels from scratch with over a thousand agents: - 235 problems, every one with a verified correct kernel - First place on the public L1, L2 and Quantization boards of SOL-ExecBench on NVIDIA B200 - 4.40x faster than the reference baseline on the 209 public problems The agents start from each problem's PyTorch definition, with no reference kernels and no official scores. An archive outside the model keeps every correct kernel and its lineage. An LLM controller reads the archive, picks a parent, and gives each child agent one direction. Checks outside the model decide correctness. The controller beat a UCB-style tree search on 35 of 38 problems at equal wall time. It chose parents about as well, but made 2.4x as many correct attempts. One restart fix raised the fleet's evaluation rate 3.6x. If you scale an agent search, measure how many reliable evaluations per hour it runs before adding agents. https://t.co/CgU1PWlYlJ Chat with Paper: https://t.co/boLrcIzIha

@Zen_with_AI

Jane Street公开说-- AI时代。面试有变化,但是一点点。 算法还要考,系统设计加码。 以下是全文: 我们的软件工程面试正在改变(一点点) 过去几年,大语言模型改变了软件的构建方式,在 Jane Street 也不例外。如今我们大多数工程师每天都在用它们,不仅写代码,还读代码、调试代码、做代码评审,以及做很多与代码无关的事。 我们最近在一期播客里聊过 AI 在开发中的用法,也在一篇关于形式化方法的文章里谈到 agent 式编程如何重塑我们的工作方式。 但我们的编程面试一直没变。今天我们问软件工程师的问题,和多年来问的还是同一种风格:和一两位面试官坐下来,一起编程解一道题,边做边聊。 这种方式一直很管用,核心目标也没变:我们想看你怎么解决问题、怎么组织工作、怎么沟通,以及怎么在合作中追求优秀的软件开发。但工作的形态已经变了,我们觉得面试也该跟着变。既然生成代码比以往任何时候都更便宜,工作的其他部分就显得更重要了:对“该做什么”的品味、选择能经得起时间的抽象、确保做出来的东西确实正确,等等。 所以今年夏天,我们开始在校园招聘的 onsite 里试行几种新的面试形式。大体上,它们写代码的部分更少,更偏高层面的东西:软件设计、验证代码的质量、结构和意图,以及用 LLM 生成代码。大多数轮次仍将是传统编程面试,不借助 LLM 进行。 我们知道这听起来很含糊,但这是故意的:这些形式还在演变,我们不想今天详细描述一个六个月后可能就变了样子的东西。 有几点值得了解: 不是每次 onsite 都有这些新轮次,来了之后我们会提前告诉你会遇到什么。 其余流程不变。如果你读过我们以前关于面试的文章,那些内容依然准确。 我们觉得 LLM 在日常工作中非常有价值,但也明白不是每个人都有丰富的 LLM 使用经验。你不需要有一套很成体系的 agent 工作方法也能在我们的流程里表现好。归根结底,我们要找的是对构建软件有出色直觉的人,而评估这一点我们觉得不需要 LLM。 不同的人把 LLM 融入工作流的方式和程度各不相同,我们不要求你有特定的配置,也不要求你事事都用它们。我们在意的是你对自己用的工具想得清楚:它们擅长什么、在哪里不足、什么时候该换别的工具。 和一段时间以来一样,更有经验的候选人的面试会更多样,涵盖从产出软件、设计系统到讨论过往工作等多种形式。 和往常一样,我们会继续迭代,有更多可说的会再写。 如果你有兴趣加入我们,欢迎在官网申请! https://t.co/UNaj0YiE1A

@plusxbt

HOLY F*CK, a full token diet for your AI agent in 15 repos 15 repos that make your agent read less, say less and pay less, 880K+ stars combined, every link checked Count. Squeeze. Skip. Shorten. Cache. Local. 01 tiktoken → count tokens before you send https://t.co/8IbZomRjBk 02 code2prompt → codebase into one prompt, tokens counted https://t.co/9PVsCJteIQ 03 LLMLingua → compresses prompts up to 20x https://t.co/ZzVHVbwRqK 04 TOON → JSON for prompts, fewer tokens https://t.co/545mG7c7FQ 05 MarkItDown → files and Office docs to Markdown https://t.co/T25RmvKfWr 06 Jina Reader → any URL to LLM-friendly text https://t.co/SKlFDGQYOx 07 Repomix → a whole repo in one file https://t.co/w3jvCV4Ji6 08 Serena → reads symbols, not whole files https://t.co/SNlFPycYAE 09 Claude Context → code search, ~40% fewer tokens https://t.co/ClIrWLdIo1 10 caveman → talks like a caveman, cuts 65% of tokens https://t.co/KA79jQm5XR 11 RTK → 60-90% fewer tokens on dev commands https://t.co/pfPTL8GPjJ 12 GPTCache → a repeat question hits the cache https://t.co/60u1fcbPyM 13 Semantic Router → decides by meaning, no LLM call https://t.co/qcf1Moy6Ne 14 Ollama → open models on your own machine https://t.co/dGwh1R8w47 15 llama.cpp → local LLM inference in C/C++ https://t.co/ttsxJScrPW Count → squeeze → read less → say less → cache → run local Bookmark this. How to run an AI employee for almost nothing, in the article below ↓

@leo114119

6.2 Astra 和 Sol 在 Artificial Analysis 的编码代理指数上都比其前代有所退步,落后于所有最近的 Claude 模型 我的天啊,6.2 的 Astra 和 SOL 这是通过后训练强化了前端和 3D 建模,导致编码能力下降了? https://t.co/H4GGv0O8BA

@eng_khairallah1

F*ck do it ASAP... this prompt builds your first team of AI agents with Claude Fable 5.2 in under 5 minutes how: Claude Fable 5.2 + 5 agents read your job → split it into lanes on your "yes" → run it as a team while YOU do nothing paste it now, before everyone starts using it, here's guide: step 0 → what they are: AI agents with 1 job each (Scout, Builder, Critic, Shipper, Scribe) and Fable as the boss that checks their work. They work on YOUR job: no hiring, no meetings, nobody quits, nobody asks for a raise step 1 → open Claude Fable 5.2 in Claude Code or the API and paste the prompt (no code? paste it in Claude chat) step 2 → it reads your job and shows 1 line: the roster + what each agent owns, reply "yes" or change any role step 3 → Scout digs, Builder makes it, Critic tears it apart, Shipper verifies it, and anything weak goes back. Nothing ships without your OK step 4 → go to sleep, wake up to 1 brief from Scribe: who did what, what passed, what's left the result: your own AI team with no salaries that works while you sleep and hands the hard calls back to you Send it now. Then read below how to give your team its first real job:

@lez99999

THIS IS AN ACTUAL F*CKING GOLD Stop opening Claude Code like it’s a blank chat every time The file that actually matters isn’t your code — it’s CLAUDE.md. And Claude can write it for you in one short interview. It captures who you are, what you’re building, and how your folder is structured. That’s the difference between starting from zero and having a teammate who already knows the context. How to set it up: 1. Open Claude Desktop → Code tab 2. Create an Obsidian vault with 3 folders:
→ raw (your sources)
→ wiki (Claude’s notes)
→ output (finished work) 3. Point Claude at the folder and say: “Interview me one question at a time” 4. Answer like you’re onboarding a new person. It writes CLAUDE.md (keeps it under ~200 lines) 5. Drop a file into raw and say “ingest this” Want to skip the manual setup? Two ready options: → second-brain-starter comes with a BOOTSTRAP.md that interviews you on first launch, fills your memory files, then deletes itself https://t.co/RyfFhJUz7m → obsidian-skills by kepano (Obsidian CEO) — 50k+ stars. Claude writes proper wikilinks and notes from day one https://t.co/prqN5qTePa Next time you open the folder, Claude already knows the context. No more re-explaining yourself. Bookmark it and open the first link now.

@WolfTrainer_101

近期Cloudflare追踪 观测了50 多个 影响Cloudflare客户的攻击智能体,并列出了渗透率与利用能力最强的前五名,其中包括Artex、LuaN1aoAgent、Cairn,都是我们比赛中的优秀项目,国货之光。 https://t.co/6BZJMBm8Vn

@noisyb0y1

f*ck yes this works, Fable 5.2 built me a roller coaster right in my living room and saved me $5,000 you can build anything and sell it for $1,000, without spending a cent or writing a line of code all of it comes from 10 free GitHub repos paired with Claude Code 01 Open Brush → draw with your hands in the air, in 3D, at room scale. > https://t.co/JBoJKFeFOZ 02 MCP for Unity → build a VR game in words, Claude creates the scenes and writes the code. > https://t.co/LQS69bazgU 03 Blender MCP → type "make a cozy room" and it builds it in Blender. > https://t.co/mYGfVnTeK6 04 SAM 3D Objects → one photo, and any object in it becomes a 3D model. > https://t.co/OCjQtVKs4I 05 Hunyuan3D-2 → type "red windmill" and get a textured model in seconds. > https://t.co/4HlmAd12Mh 06 MR Utility Kit → the headset finds your tables and walls so you can put objects on them. > https://t.co/4NANMJrOSy 07 Interaction SDK Samples → grab, pinch and press virtual things with bare hands, no controllers. > https://t.co/jdzz5SfxOM 08 The World Beyond → your walls dissolve and another world opens behind them. > https://t.co/Zzuv29kF3q 09 Passthrough Camera API → give the AI the headset's eyes so it sees the world around you. > https://t.co/JmkIT8bZ35 10 React XR → the easiest way to vibe code your own VR game with Claude. > https://t.co/Y4FoFULYva ten free links that can close you $7,000 a month save it, your first client is already looking for someone who can do this ↓

@MelihaSerching

This is f*cking insane. Google put out a free 8-minute lesson on building your first AI agent It's the clearest explanation of agents and loops out there Prompt → Tools → Memory → Loop → Your First Agent People pay $500 for courses that teach less than this 8 minutes. No excuse not to watch it Bookmark and watch it today Then read the step-by-step loops guide: Harness, Loops, Graphs ↓

@NFT_Chen

💥冰之圣殿同一提示词对比:Fable 5.2 vs GPT-6.2 Astra Claude Fable 5.2 🔸冰层纹理真实,光影与极光融合自然 🔸水面与鲸出水更逼真,船细节清晰 🔸整体电影级质感领先 🔸成本高、速度慢($81 / 近2小时) GPT-6.2 Astra 🔹成本约1/10,速度快近3倍($8.2 / 40分钟) 🔹冰拱结构清晰,船造型尚可 🔹水面,尤其鲸周围有明显瑕疵,质感偏人工 🔹精致度稍逊,但差距因人而异 点评:复杂视觉Fable仍更优;Astra性价比极高,适合快速迭代。编码选Astra,纯视觉选Fable。你更看重哪边? #AI #Claude #GPT #模型对比 #Dario #SamAltman https://t.co/FwudO6wlcq https://t.co/JAj8Ez5XeN

@Joyingeth

THIS F*CKING INSANE Two prompts built six agents from one empty folder. I ran build from this one line: mkdir gta6-machine && claude 1. 5 directories dropped, one CLAUDE.md capped at 200 lines, 4 rules that make every agent cite the vault first 2. 6 subagent files written. Scout, librarian, writer, editor, seo, publisher. One job each 3. Nothing publishes unless a source file confirms it. Rumors stay tagged 15 minutes. 6 agents. 2 prompts. Save the build log

@wwardenn

GITHUB JUST RELEASED A FREE COURSE ON BUILDING AI AGENTS @github has published Copilot App for Beginners, an open-source course that teaches developers how to build and manage AI-agent workflows using GitHub Copilot App. The repository includes 8 chapters with practical exercises covering everything from parallel agent execution to custom automation workflows. WHAT YOU'LL LEARN: Running multiple AI agents in parallel. Isolating tasks using Git Worktrees. Working with Plan and Autopilot modes. Creating custom Skills. Connecting external tools through MCP servers. The interesting part is that this goes beyond using AI for isolated coding tasks. These workflows make it possible to delegate different tasks to specialized agents, run them independently, and coordinate their output. The same approach has applications outside software development. Marketing teams can use agent-based workflows for market research, competitor analysis, data processing, content preparation, and repetitive operational tasks. Instead of asking AI to complete every task in one conversation, you can build systems where multiple agents handle different parts of the process. The real shift is moving from using AI as an assistant to building workflows around autonomous agents.

@0xForce_

THESE 10 GITHUB REPOS TURN YOUR AI AGENT INTO A FULL TEAM Instead of one tool doing one job, these repos let your agent handle 3D, video, web research, and office work at the same time. Here’s how the main ones work: → DreamLoop: agent imagines the final scene, builds it in Blender, then a critic agent checks and loops until it matches → Trellis.2 (Microsoft): drop in one image, get back a detailed 3D mesh with realistic materials → HyperFrames: agent writes HTML and turns it into a full high-quality video so your promo is a file, not a week of editing → BrowserUse: agent uses the web for you — researches, fills forms, and clicks through dashboards → Google Workspace CLI: puts the agent inside Gmail, Drive, Calendar, Sheets, and Docs One agent. Multiple specialists. No extra hires. Bookmark this.

@Chenfacai0

既想不露脸又打算做科普类内容,火柴人动画确实是条可行路径,只不过完成一集要经历分镜设计、配音录制、字幕制作等环节,往往得耗费好几天工夫。 Stickman Video Director 是一款专门制作火柴人视频的 Skill,只要提供给它一个话题、一篇文章或者一份脚本,它便能输出带有配音、字幕与音效的完整成片。 使用方式分为两种:导演模式会产出分镜方案以及每个镜头对应的视频提示词,供你拿到视频生成工具中产出素材并自行剪辑;自动模式则直接交付整条视频。 GitHub 地址:https://t.co/wfOLYc6vD7 自动模式更适用于时长超过 10 分钟的长视频,在 Codex 或 Claude Code 环境中运行,它会先给出约 20 秒的有声预览,待你确认之后才制作全片。 该工具内置 10 种画风,其中包括经典黑白、复古电视、洞穴壁画等,横屏、竖屏与方形三种画幅均可输出。 作者建议优先使用 Codex 来运行,其本人提供的示例均按此方式完成。如果你运营的是知识类频道,不妨挑一个小话题先试做一条。

@rohanpaul_ai

New Google and Penn State paper finds that when one AI rewrites the code that gives another AI its instructions and tools, keeping notes on why each change was made helps the gains carry over to a new model. LLMs can now rewrite an agent's harness, the code that runs prompts and tools. After a model or benchmark switch, it can score below the original, and its code hides why edits were made. Self-Improvement Memory, or SIM, logs each edit with the problem it targeted, its reasoning and its measured result. The optimizer reads this log on the new setup to decide which old edits to keep or revise. When a Claude Opus 5.5 harness moved to Gemini 3.8 Flash workspace tasks, SIM lifted it from 68.41 to 84.75, versus 73.42 without the log. Across 10 repeated coding runs, SIM finished ahead 9 times, by 2.36 points on average. Store why each harness edit was made and how it scored, and carry that log to new models.

@NFT_Chen

🔥IPO前夜,OpenAI选择突袭阻击Dario! GPT-6.2 Astra一夜杀回AA Intelligence Index 第2名(62分),直接反超Gemini 4 Carbon(60),逼近Claude Fable 5.2(65)! 🔹对比Fable 5.2:只差3分,但成本优势明显,Pareto前沿几乎被OpenAI包圆 
🔹对比Gemini 4 Carbon:直接反超,重夺第二,科学+3D+机器人多项领先 Dario的IPO估值要重新算了……这波突袭太狠! #GPT62 #OpenAI #Anthropic #ClaudeFable52 #Gemini4Carbon #AIRanking #SamAltman #Dario

@Onefly

中转下游对接平台已上线 最新模型 gpt-6.2-sol gpt-6.2-astra 对接分组 0.07x Codex 0.07 pro500原生不降智:https://t.co/lpHPIiJCXc 卡网兑换码:https://t.co/f1OLICYJDI 点击链接加入群聊【中转下游联盟】:https://t.co/urMu8jsxYR https://t.co/xGGepBCLiJ

@RealMarvelX

Obsidian’s official agent skills, Karpathy’s LLM Wiki pattern, and a self-hosted AI second brain are all free on GitHub. yet people still start their second brain with an empty folder. here are 7 repos with 175K+ combined GitHub stars to help you build an AI-powered knowledge system. 01. Obsidian Skills : 52.7K ⭐ official skills from the Obsidian team. teach AI agents to work with Markdown, Bases, JSON Canvas, and the Obsidian CLI. https://t.co/NnHE8Ne7q6 02. LLM Wiki : Karpathy’s knowledge-base pattern Andrej Karpathy’s idea for building a persistent wiki that AI can maintain. turn raw documents into connected Markdown knowledge instead of repeatedly asking AI to rediscover information. https://t.co/bYBIi676D5 03. Khoj : 38.1K ⭐ a self-hostable AI second brain that answers questions from your documents and the web, with custom agents, automations, and deep research. https://t.co/ygOOsXdmkk 04. Claudian : 16K ⭐ an Obsidian plugin that brings Claude Code, Codex, and other coding agents directly into your vault. your vault becomes their working directory. https://t.co/d3D94N60yp 05. Obsidian Copilot : 8K ⭐ run agents like Claude Code, Codex, and OpenCode inside Obsidian. let them inspect notes, create files, and handle multi-step knowledge work. https://t.co/TBxiRQfin9 06. Kepano’s Obsidian Vault : 6K ⭐ the Obsidian CEO’s personal vault template. copy its folder structure, templates, and organization instead of building everything from scratch. https://t.co/utx0BWaKgq 07. Basic Memory : 4K ⭐ a local-first Markdown memory system that lets AI read, write, and search your knowledge across conversations. works with Claude Code and other MCP-compatible clients. https://t.co/oXJy8bExaK HOW TO BUILD YOUR SECOND BRAIN you don’t need all 7 tools running at once. start with this workflow: step 1: download Kepano’s vault template and open it in Obsidian. step 2: install Obsidian Skills so your AI agent understands Obsidian’s Markdown, Bases, Canvas, and CLI. step 3: install Claude Code and open your vault as its working directory. step 4: give Claude Karpathy’s LLM Wiki gist and ask it to adapt the knowledge-base pattern to your vault. step 5: use Basic Memory if you want persistent Markdown memory accessible across AI conversations. step 6: add Claudian or Obsidian Copilot to bring your coding agents directly into your vault. THE MEMORY WORKFLOW → raw: collect articles, documents, links, research, and ideas. → process: let your AI agent summarize and extract useful information. → connect: create related notes, backlinks, and knowledge pages. → index: maintain a clear map of everything inside your vault. → update: feed new information into the existing knowledge base instead of starting over. the goal is simple: stop making AI rediscover the same information every time you start a conversation. build a knowledge base that gets more useful as you use it. bookmark this and build your second brain tonight. 🔖

@dr_cintas

I made this week’s AI roundup as two completely different decks. 
Both came from one prompt with the new Agent in Gamma 5 👇 • Claude Fable 5.2: faster and cheaper • Meta’s Muse Ember at $0.10 per million tokens • Mira Murati’s Inkling 2 Small, with open weights promised • Tencent’s Hy4 • Pruna’s 4MP images in 2.6 seconds Which one would you send, left or right?

@hiwise_aigc

卧槽,同样是让 AI 操作 Blender,Computer Use 和 Skills 完全是两种体验! 我给 Codex安装后让做了个《火锅星球》!整个过程由 Codex 配合 Blender Skills 完成。 它把 Blender 的建模、材质、灯光、动画、渲染等专业操作封装成 Skills,Codex 可以直接调用,不用再靠鼠标一步步点。 开源地址: https://t.co/Q6782QlqTB 以前是让 AI 像人一样操作软件,现在是直接给 AI 装上专业技能。 这才是 Agent 工具该有的进化方向吧!

@_akhaliq

Accelerate Bench Evaluating Agents for Application-Level Program Optimization paper: https://t.co/0qABu4HB7H https://t.co/LRT935ly7Q

@danikkk_wqs

It's f*cking insane: a plugin with 15.8K stars puts Claude Code inside Obsidian, and most people still copy-paste their files into ChatGPT one by one the loop: → you drop anything into one inbox folder → Claude files it, links it to what you already know, flags what's outdated → every morning it writes you a brief from your own vault the 1 line version for beginners, run it on a schedule: claude -p "read my inbox folder, file everything, and write today's brief in today.md" --permission-mode acceptEdits 8 repos that build the full system: Claudian (15.8K stars) Claude Code inside Obsidian as a sidebar, no terminal needed https://t.co/HPxYXUss0j claude-obsidian (3.3K stars) drop in a source, Claude files it and fixes dead links https://t.co/WGEn1s4qKs llm-wiki-agent (2.2K stars) ingest, ask, clean up: the whole auto-wiki loop as a Claude Code skill https://t.co/vx3fxx9vy3 obsidian-mind (1.6K stars) hooks load your goals, projects and open tasks every time Claude starts https://t.co/zvv6KKMIsu claudesidian (2.6K stars) ready commands like /daily-review and /research-assistant https://t.co/8ZUYrlwrfi qmd by Shopify's CEO (31K stars) local search over your whole vault, plugs straight into Claude https://t.co/uMBHGeD24H Basic Memory (2.3K stars) Claude saves what it learns, and it shows up in your Obsidian graph https://t.co/bLo3bDvxka Agent Client (2.5K stars) Claude Code, Codex or Gemini in a side panel that sees what you have open https://t.co/c53O7SSUdo you stop organizing. you just open the brief save it so you don't lose it

@Oluwaphilemon1

Your current GPU might not be powerful enough to run every open-weight AI model today. That doesn’t mean you shouldn’t preserve the models you want to experiment with tomorrow. Four models I’d consider adding to a personal AI archive are: Qwen3.8-27B Qwen3.8 Flash-Next GLM-5.3-Flash DeepSeek V4.1 Flash 1017 Each offers a different opportunity for future local inference experiments, coding workflows and model comparisons. Why download them now? Because the local AI ecosystem changes quickly. Model repositories get updated, older checkpoints become harder to find, download links change and community-maintained quantisations sometimes disappear. A model that’s easy to download today might require considerably more effort to locate a year from now. Preserving the weights gives you the option to revisit these models whenever your hardware or inference software improves. Four models worth preserving 1. Qwen3.8-27B A dense 27B model worth keeping for general-purpose inference, reasoning and coding experiments. It also provides a useful reference point when comparing dense architectures against larger Mixture-of-Experts models. You can use it to investigate quantisation quality, inference speed, context handling and different local deployment configurations. 2. Qwen3.8 Flash-Next This model is particularly interesting for exploring Mixture-of-Experts architectures. MoE models use different expert networks for different tokens, activating only a portion of their parameters during each computation. This creates an interesting balance between total model size, active parameters and inference efficiency. It’s worth preserving for future experiments with speculative decoding, memory management and different inference engines. 3. GLM-5.3-Flash Another model worth keeping for coding and agentic workflow experiments. Local inference improvements could make it interesting to compare how different architectures handle tool use, multi-step reasoning and software development tasks. 4. DeepSeek V4.1 Flash 1017 This is another checkpoint worth adding to the archive for future comparisons. As inference engines, quantisation techniques and hardware continue to improve, preserving models from different releases gives you more opportunities to evaluate their capabilities under new conditions. Don’t overlook quantised and abliterated versions I’d also consider preserving selected community-made variants. Different quantisation formats can significantly reduce checkpoint sizes, making models easier to store and potentially more practical to run on consumer hardware. For example, you might keep Q4, Q5, Q6 and Q8 versions where available. I’d also preserve abliterated variants separately from the original checkpoints. Abliteration modifies a model’s refusal behaviour, but it doesn’t automatically improve its reasoning or coding capabilities. These versions should be evaluated independently because the modification can affect other behaviours too. Keep the original weights whenever possible so you always have a reference for comparison. Why download models you cannot run yet? Your current hardware shouldn’t determine every model you preserve. Perhaps you currently have a single GPU with limited VRAM. In the future, you might upgrade to a more powerful GPU, build a multi-GPU system or gain access to more system RAM. Inference software is improving as well. Techniques such as: GGUF and other quantisation formats. CPU and system RAM offloading. Speculative decoding. Multi-GPU inference. KV-cache compression. Improved memory management. These can change which models are practical to run. A model that currently exceeds your hardware budget might become usable after improvements to the inference runtime or a change in your hardware configuration. Having the weights already downloaded means you can start experimenting without having to locate the original files again.

@yyyole

这对不懂专业名词的人也太友好了吧!! 有人把视觉动效,做成了一套可视化词典! 240 种专业视觉效果,全部带动态演示和提示词,一目了然,想要哪个直接copy就行! 项目叫 Motion Prompt,GitHub 已经突破 2k Stars! 以前用 AI 做视频、设计网页,经常遇到一个特别蛋疼的问题: 脑子里明明有画面,却不知道该怎么描述。 比如镜头怎么推进、文字怎么丝滑入场、网页滚动时怎么产生高级的交互动效…… 最后只能跟 AI 说:高级一点、丝滑一点、有电影感一点。 现在好了!! 作者直接把 240 种效果做成了可以预览的动态案例库,覆盖 10 大类别: 镜头运动、转场、MG 动画、UI 组件、网页动效、摄影光影、艺术风格、PPT 图表等等。 最爽的是,它还做成了Skill!! 一句话直接安装到 Claude Code 或 Codex: npx skills add showheyohtaki/motion-prompt 做网页,或者用AI做视觉时,看中哪个效果,点击 Pick,再点 Send to AI,就能把选中的效果交给 Agent。 AI 会读取对应的效果规范和参考代码,再根据项目实现。 项目地址放下面了,有兴趣可以让Agent看一看!

@0xGenAi

HOLY F*CK... ONE GROK BOT TEAM CAN RESEARCH, BACKTEST, PAPER-TRADE AND REPORT ON YOUR ENTIRE PORTFOLIO. No analysts. No terminal subscription. No staring at charts all day. Just 6 AI agents, each with one job on the desk. Here are 12 GitHub projects to build your own AI trading desk: THE AGENT TEAMS 01 / 𝗧𝗿𝗮𝗱𝗶𝗻𝗴𝗔𝗴𝗲𝗻𝘁𝘀 → Analysts, a research manager, a trader and a portfolio manager work every ticker together. https://t.co/8SLv1SvumA 02 / 𝗔𝗜 𝗛𝗲𝗱𝗴𝗲 𝗙𝘂𝗻𝗱 → A fund of investor agents with paper trading and backtests. Nothing trades before an approval step. https://t.co/lrpGwJTLT7 03 / 𝗙𝗶𝗻𝗚𝗣𝗧 → Open-source financial language models, with the trained models on Hugging Face. https://t.co/C1sxQJ8Ete THE DATA FEED 04 / 𝗢𝗽𝗲𝗻𝗕𝗕 → An open data platform for analysts, quants and AI agents. https://t.co/McWFapPrHs 05 / 𝘆𝗳𝗶𝗻𝗮𝗻𝗰𝗲 → Pulls market data from Yahoo Finance. https://t.co/ZlPulCld8e THE QUANT LAB 06 / 𝗤𝗹𝗶𝗯 → Microsoft's AI quant research platform, from first idea to production. https://t.co/Hi5RxArHdX 07 / 𝘃𝗲𝗰𝘁𝗼𝗿𝗯𝘁 → Tests thousands of trading ideas in the time others test one. https://t.co/d1tQtQQMZN 08 / 𝗟𝗲𝗮𝗻 → QuantConnect's algorithmic trading engine, in Python and C#. https://t.co/JHgTaJdJZV RISK AND REPORTS 09 / 𝗥𝗶𝘀𝗸𝗳𝗼𝗹𝗶𝗼-𝗟𝗶𝗯 → Portfolio optimization: how much goes into each position. https://t.co/dMTHF6PxUk 10 / 𝗤𝘂𝗮𝗻𝘁𝗦𝘁𝗮𝘁𝘀 → Portfolio analytics and full reports for quants. https://t.co/AbOe22zQAa PAPER TRADES FIRST 11 / 𝗔𝗹𝗽𝗮𝗰𝗮-𝗽𝘆 → The official Python SDK for Alpaca's trading API, paper accounts included. https://t.co/G9XwWRTQm6 12 / 𝗡𝗮𝘂𝘁𝗶𝗹𝘂𝘀 𝗧𝗿𝗮𝗱𝗲𝗿 → A production-grade trading engine written in Rust. https://t.co/4o6pVm5DC2 These 12 projects have over 438,000 stars on GitHub. THE SETUP: Researcher reads filings, earnings calls and the news. Quant turns every idea into a backtest. Risk manager sizes each position and sets the stops. Trader places paper trades first. Analyst writes the Friday report. Manager brings every real trade to you for a yes. One person. Six bots. Nothing trades with real money until you say so. Paste the 12 links into one Grok Bot chat and type: "Hire 6 bots for my trading desk, give each one the repos it needs, paper trades only" Save this, then read the full guide to setting up a team of Grok Bot agents from 0 in the article below.

@EianLu

我平时用 Claude 做视频,没想到图文也能这么玩啊! 一个下午,Fable 5.2 + Postiz 搞定一整周的 AI UGC 图文。 完整流程和每一条提示词都在文章里,严肃学习🫡 https://t.co/HOTbT9Wnhs

@siddhantgarg33

𝗔𝗜 𝗔𝗚𝗘𝗡𝗧 𝗔𝗥𝗖𝗛𝗜𝗧𝗘𝗖𝗧𝗨𝗥𝗘: 𝟵 𝗖𝗢𝗥𝗘 𝗖𝗢𝗠𝗣𝗢𝗡𝗘𝗡𝗧𝗦 An AI agent is much more than an LLM. To build agents that can reason, remember, act and improve, you need several components working together. 01 — 𝗟𝗟𝗠 | 𝗧𝗛𝗘 𝗕𝗥𝗔𝗜𝗡 Understands instructions, reasons about problems and generates responses. 02 — 𝗠𝗘𝗠𝗢𝗥𝗬 | 𝗥𝗘𝗧𝗔𝗜𝗡𝗦 𝗖𝗢𝗡𝗧𝗘𝗫𝗧 Stores conversation history, user preferences and past experiences so the agent can make better decisions. 03 — 𝗥𝗔𝗚 | 𝗞𝗡𝗢𝗪𝗟𝗘𝗗𝗚𝗘 𝗔𝗖𝗖𝗘𝗦𝗦 Retrieves relevant information from external knowledge sources using embeddings and vector databases. 04 — 𝗧𝗢𝗢𝗟𝗦 | 𝗔𝗖𝗧𝗜𝗢𝗡 𝗘𝗫𝗘𝗖𝗨𝗧𝗢𝗥 Lets agents interact with the real world through search, APIs, databases, calculators, email and other services. 05 — 𝗣𝗟𝗔𝗡𝗡𝗜𝗡𝗚 | 𝗗𝗘𝗖𝗜𝗗𝗘 𝗦𝗧𝗘𝗣𝗦 Breaks complex goals into smaller actions, executes them and observes the results. 06 — 𝗥𝗘𝗙𝗟𝗘𝗖𝗧𝗜𝗢𝗡 | 𝗦𝗘𝗟𝗙-𝗜𝗠𝗣𝗥𝗢𝗩𝗘 Evaluates results, identifies mistakes and retries or improves the approach when needed. 07 — 𝗠𝗨𝗟𝗧𝗜-𝗔𝗚𝗘𝗡𝗧 | 𝗖𝗢𝗟𝗟𝗔𝗕𝗢𝗥𝗔𝗧𝗜𝗢𝗡 Multiple specialized agents can work together, such as research, coding and QA agents. 08 — 𝗠𝗢𝗡𝗜𝗧𝗢𝗥𝗜𝗡𝗚 + 𝗚𝗨𝗔𝗥𝗗𝗥𝗔𝗜𝗟𝗦 Adds input validation, output filtering, safety checks, rate limits, logging and auditing. 09 — 𝗛𝗨𝗠𝗔𝗡-𝗜𝗡-𝗧𝗛𝗘-𝗟𝗢𝗢𝗣 Humans review, approve, edit or reject important decisions before the agent continues. 𝗧𝗛𝗘 𝗕𝗔𝗦𝗜𝗖 𝗔𝗚𝗘𝗡𝗧 𝗟𝗢𝗢𝗣: Goal → Plan → Use Tools → Observe → Evaluate → Improve → Repeat 𝗧𝗛𝗘 𝗕𝗜𝗚 𝗜𝗗𝗘𝗔: A powerful AI agent isn't just a smart model. It's a system that combines: 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 + 𝗠𝗲𝗺𝗼𝗿𝘆 + 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 + 𝗧𝗼𝗼𝗹𝘀 + 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴 + 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸 + 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 Save this cheat sheet if you're learning AI agents or building agentic workflows. Repost it for someone learning AI engineering. ❤️ Like 🔁 Retweet 🔖 Bookmark

@bkdgiffug

Rust 重造的浏览器你见过吗?GitHub 上有人搞了个 Moli,Star 数已经冲到 1.67 万,专供 AI Agent 使用。 过去让 Claude Code、Codex 碰网页,总得先拉个 Chrome 起来,吃内存不说,还慢得让人着急。 Moli 的打法相当狠:网页不用看,那就干脆不渲染! 网页抓取、JavaScript 执行、自动化操作它都干得了,Playwright、CDP、WebDriver 也一并兼容。 按项目自己的测试,Moli 的内存占用比 Chrome 低出一截,AI 爬虫、Agent 自动化、批量网页处理这些活儿都挺对口。 往后给 AI 挑浏览器,Chrome 不再是唯一答案了。 GitHub: https://t.co/NIpYii0OLU

@AYi_AInotes

Anthropic 首席执行官 Dario Amodei 最近在回应刚毕业的高中生该如何规划未来时,给出了一个毫不留情的冷酷判断:人工智能已经能够胜任入门级的法律、金融与咨询业务,并且将在短期内直接消灭一大批初级岗位。 这不是他第一次拉响警报,他此前就公开预警过,未来一到五年内,整整一半的入门级白领工作都可能遭到彻底颠覆。 大家默认以为历次技术革命就像当年人类从农场走进工厂、再从工厂跨进互联网一样,每次转变总会留出几十年甚至几代人的缓冲期让普通人慢慢适应。 但这一次的技术切割不仅刺进了脑力劳动的深水区,而且渗透速度快到根本不给传统教育留出调整学制的窗口期。 他给出的破局建议听起来平平无奇,只有学会用工具四个字,但绝大多数人根本没意识到,会用这两个字背后其实横着一道从聊天到指挥官的六级天堑: 第一层是日常聊天,第二层是琢磨提示词,第三层是调用现成智能体,第四层是搭建专属的工作流底座,第五层是让系统自动化循环运转,最高层则是人彻底退居幕后做全局指挥。 市面上超过九成的人都死死卡在前两层,每天依然停留在把对话框当成高级搜索引擎的阶段,真正拉开个人产能百倍差距的分水岭,在于有没有跨过第四级的 Harness 门槛,也就是亲手给模型配齐固定工具、长效记忆与标准流程,让它按你的业务规矩反复干活,而不是每次都从零开始重新聊。 留个清醒,搭建自动化体系虽然能成倍放大单兵作战效率,但在金融、法律等容错率极低的严肃业务里,系统依然无法替人类承担最终的合规与道德责任,越是自动化循环运转,人类在验收节点上的判断力就越不能缺席。 未来职场最残酷的淘汰从来不是不会打字,而是当别人已经坐拥一整支随时待命的自动化执行团队时,你还在对着对话框一字一句地重新解释背景。 https://t.co/8dAyR0ztFO

@huang18825

解包了下antigravity 2.33.0的更新内容发现了几块官方更新日志没提到的部分: 1.原生桌面 GUI 操控 Agent(Computer Use / CUA): 包含独立桌面操控子智能体、桌面悬浮 HUD 与遮罩、最多 12 路(默认 6 路)WebRTC 实时桌面画面流、应用级授权锁以及用户随时接管/交还控制权.但是该功能目前默认关闭,受内部实验开关控制。 2. MCP Apps 交互式微应用体系(SEP-1865)正式落地: 官方日志仅在 Bug 修复里顺带提了一句 "custom file viewers from plugins and MCP apps",完全没有宣传整套新增的 MCP Apps 机制: 后端新增 McpAppTool、进度流 reportProgress 与交互式反问 elicit,并随客户端下发了完整开发文档技能 mcp-app-development。 第三方 MCP Server 现在可通过 ui:// 资源返回单文件 HTML 应用,在独立沙箱 iframe 中渲染为对话内联卡片、右侧辅助面板 Tab(aux_pane)或左侧导航栏全屏应用(global),甚至可以声明仅允许前端 UI 调用、对模型隐藏的专属工具。 3.新增 update_steps 实时步骤清单工具与「运行时监督器(Runtime Supervisor)」

@choopyplug1

Fable 5.2 now runs a whole design studio on 12 free tools design agencies are f*cked sketches. UI. icons. fonts. AI images. live edits 01 Excalidraw → sketch the idea like on a whiteboard before a single line of code https://t.co/ydFrFOBV5X 02 Penpot → open-source design tool: screens, prototypes, code handoff https://t.co/AAtMtpqoMZ 03 Graphite → vector and raster editing in one app, right in the browser https://t.co/k59QMKrOWR 04 Inkscape → logos and vector illustration, no license https://t.co/RPuySy49Nd 05 Tailwind CSS → one styling language for the whole product https://t.co/LSld2TigLS 06 ComfyUI → your own image AI pipeline from nodes, on your own GPU https://t.co/ngLZJ9kJpN 07 shadcn/ui → real buttons, forms and dialogs you copy in and own https://t.co/QYBZ81Owbv 08 Lucide → 1,000+ icons in one consistent style, goodbye emoji icons https://t.co/UnSX81yFBv 09 Fontsource → open fonts as packages, so it stops defaulting to the same one https://t.co/3X2thqkSvt 10 Leonardo → Adobe's open tool for palettes that pass contrast checks https://t.co/cOAA7S0xoS 11 Onlook → edit the live site like a design file, the changes land in your code https://t.co/7pDlYzCfQ5 12 Satori → turns HTML and CSS into images: covers, social cards, previews https://t.co/wHXffvXWHx everything Figma and Canva charge you for. now $0 plot twist: nobody opened a design app to make this video. Fable 5.2 coded every frame save this and ship something that doesn't look AI-made ↓

@isagi1000x

THE VRAM BARRIER IS STARTING TO LOOK VERY DIFFERENT. A 125B-parameter AI model just ran on a single 24GB RTX 4090, with a massive 250,000-token context window. And the performance is seriously impressive: - 20.97 tokens/sec decode speed - 364.29 tokens/sec prompt processing - Just 18.3GB VRAM used - 110GB system RAM - No MTP, no DFlash, no KV-cache quantization. We're getting closer to running datacenter-scale models on consumer hardware. Here's how the benchmarks break down. 4 TESTS. ONE RTX 4090. 1. HYBRID OFFLOADING — 80K CONTEXT 40 expert layers offloaded to the GPU. → 383.85 t/s prefill → 22.52 t/s decode → 23.85GB VRAM → 97GB RAM 2. FULL CPU MoE OFFLOADING — 80K CONTEXT All 512 expert layers moved to system RAM while attention remained on the GPU. → 355.72 t/s prefill → 20.84 t/s decode → 11.66GB VRAM → 110GB RAM 3. THE 180K CONTEXT TEST → 357.75 t/s prefill → 20.98 t/s decode → 15.6GB VRAM → 110GB RAM 4. THE 250K CONTEXT BREAKTHROUGH → 364.29 t/s prefill → 20.97 t/s decode → 18.3GB VRAM → 110GB RAM And there's still approximately 5.7GB of VRAM headroom. WHAT MAKES THIS POSSIBLE? Two important optimizations: 1. Larger batch sizes Using "-b 4096 -ub 4096" pushed prompt processing from roughly 150 t/s to over 364 t/s. 2. CPU MoE offloading The "-cmoe" flag moves expert layers into system RAM, freeing up GPU memory with only a modest impact on decode speed. The model itself is a 111.4GB Qwen 3.8 Flash-Next GGUF, distributed across four shards. It requires an experimental llama.cpp branch (PR #27742) to run. THE BIGGER PICTURE A single RTX 4090 paired with around 110GB of system RAM can now serve a 125B hybrid model with a 250K context window. That's a serious development for anyone building local AI infrastructure. Dense 27B models still have their advantages on single-GPU setups, but this shows just how far hybrid MoE inference can go. The gap between consumer hardware and datacenter-scale AI is getting smaller. Hugging Face model files and complete performance telemetry graphs are in the replies. One question: GLM 5.3 Flash vs Qwen 3.8 Flash Next which open-weight model takes the crown this week? 👇

@Oluwaphilemon1

NVIDIA has released a new speculative decoding draft model on Hugging Face designed to accelerate Qwen3.8-27B. The interesting part is its size. The draft model uses just 4 billion parameters to help the much larger 27B model generate tokens more efficiently. According to the reported results, it achieves a 46% improvement in token acceptance compared with Qwen3.8-27B’s built-in speculative decoding head. To understand why this matters, we need to look at how speculative decoding works. The idea behind speculative decoding Normally, an LLM generates text one token at a time. Each new token requires another generation step from the main model. Speculative decoding introduces a smaller, faster draft model to speed up that process. Instead of waiting for the main model to generate every token individually, the draft model proposes several tokens ahead. The larger model then checks those proposals in parallel and accepts the ones that match its own predictions. If several proposed tokens are accepted, the main model can produce more output with fewer sequential generation steps. The larger model remains responsible for the final output, so speculative decoding is designed to preserve the target model’s generation results while reducing the time required to produce them. Why NVIDIA’s 4B draft model is interesting Qwen3.8-27B already includes a built-in speculative decoding head. NVIDIA’s release introduces a separate 4B draft model as an alternative. The reported 46% increase in token acceptance suggests that the external draft model produces proposals that align more closely with the target model’s predictions. Higher acceptance generally means fewer rejected draft tokens and less wasted verification work. However, acceptance rate is not the same as a 46% increase in tokens per second. Actual speed improvements depend on several factors: How quickly the draft model generates proposals. How many proposed tokens the 27B model accepts. The cost of verifying those proposals. Available GPU memory and bandwidth. The inference engine and speculative decoding configuration. A draft model can achieve a higher acceptance rate but still deliver limited speed gains if its own generation overhead is too high. What needs to be tested next The most useful comparison would put NVIDIA’s 4B draft model against Qwen3.8-27B’s built-in head under identical inference conditions. Important measurements include: Actual tokens per second. Average accepted tokens per verification step. Draft model latency. GPU memory consumption. Performance across short and long prompts. Results on coding, reasoning and general text generation. This is particularly relevant for people running Qwen3.8-27B locally, where memory bandwidth and inference overhead can have a major impact on generation speed. NVIDIA’s release offers another approach to speculative decoding for Qwen3.8-27B. The 46% acceptance improvement is promising, but the real test is how much faster the complete generation pipeline becomes in practical workloads.

@aehyok

🚀 OpenMotion 正式开源! 它的还原度超过 95%,甚至不用 Fable 5.2 也能复刻爆款视频! 全网首个全栈的 Claude Motion & 剪映平替 逆向了最先进的设计、最好看的模板 14 天写了 47 万行代码。 ✅ 100+ 顶尖公司的产品宣传片 ✅ 300+ 动效版 DESIGN.md ✅ 152+ 设计系统 ✅ 支持 Codex / Claude Code / DSH 等 20+ AI Agent 🔥 我研究了一下这个项目,顺便分享一下👇 1、它是什么 丢一条参考视频或者社媒链接进去,它用你自己的素材、Logo、文案把片子复刻出来。 视频是用 HTML、CSS、JS 写出来的。Agent 负责构思、生成素材、排镜头,你在 Studio 里改时长、字幕、节奏,每一版都有历史记录。 导出的成片旁边会留着完整的源码工程,随时能接着改。协议是 Apache 2.0,现在 10.4k Star。 2、MOTION.md 这是他们主推的东西,可以理解成动效版的 DESIGN.md。 镜头顺序、文字节奏、转场、声音,都写成 Agent 读得懂的规范,再配一份 Brand Kit(Logo、配色、字体、语气、产品信息)。 一个品牌可以连着出一个系列:换产品、换文案、换素材,动效语言保持一致。 线程里还有人把星球大战、黑客帝国、F1、漫威的片头写成 MOTION.md 拿去混剪。 3、什么 Agent 都能接 作者的说法是,动效的经验已经收进 Motion harness 里了,所以不用非得上 Opus 5.5 或 Fable 5.2。 README 里列了 22 个:Claude Code、Codex、Cursor、Gemini CLI、WorkBuddy、Kimi CLI、Qwen Code、Qoder、Pi、Kiro、OpenCode、Hermes、Devin、Trae 等等。 接入方式有 Skills、MCP、插件三种。 可以在对话里提需求,也可以直接进编辑器改。 4、素材从哪来 Agent 和 API Key 都用你自己的,本地流程免费,不按渲染次数收钱。 内置 7 个模型: · 生图:Codex Image、Nano Banana 2、FLUX Schnell · 生视频:Seedance 2.0、Wan 2.2 等等 · 配乐:Suno v4 · 配音:Kokoro 底层用的是 GSAP、Remotion、FFmpeg、Three.js、Blender、Manim、Lottie 这些。 最后附上开源仓库地址:https://t.co/KI5ndnVrbN https://t.co/yOh8uKFNcJ

@CY160909

有人把 18 位创作者的动画 Skills 整合到了一个开源项目里 以前分散在不同仓库里的动画Skill,现在可以在一个本地工作台里选风格、看预览 它收录了 109 种视频风格,选好风格,输入想做的故事,Claude Code会先写一份分镜脚本,每个场景演几秒、画面出现什么、字幕和旁白写什么,都列得清清楚楚。哪里不满意可以直接改表格,再选音乐、加可选的本地配音,最后让对应的动画 Skill 渲染并合成 MP4 这个项目把原本需要挨个配置的制作流程集中起来,对想自己控制镜头和画风的人很有用 目前只支持Claude Code,电脑也要安装所需的依赖,部分风格还要用Blender GitHub:https://t.co/0KUadik5h0

@RealMarvelX

Anthropic just released a 5-part architecture that shows how to build AI agents that handle inbox events, scheduled tasks, persistent memory, approvals, and follow-ups. and the best part? the architecture is simpler than you think. an always-on agent doesn’t run 24/7. it sleeps, wakes up when something happens, does its job, saves its progress, and goes back to sleep. Anthropic breaks the architecture into 5 building blocks: 1. event inbox every trigger goes into a durable queue. emails, messages, webhooks, and other events land here. the agent picks them up, processes them, and can retry if something fails. 2. timers this lets the agent handle scheduled tasks. think daily reports, reminders, follow-ups, and recurring checks. when the timer fires, it creates an event that wakes the agent. 3. memory the agent needs to know what happened before. store its instructions, user preferences, previous messages, and task progress. load the essentials when it wakes up, then save new information before sleeping. 4. permissions don’t give an agent unlimited access. keep credentials outside the agent, control which tools it can use, and require human approval for sensitive actions. 5. outbox this is how the agent communicates what it has done. it can send a result, report that nothing needs attention, or request approval before continuing. the complete loop looks like this: event arrives → agent wakes up → loads memory → performs the task → saves progress → sends an update → goes back to sleep. want to build something similar? check out Alook, an open-source project for connecting people and coding agents through persistent rooms and inboxes. it supports Claude Code, Codex, Cursor, OpenCode, and Pi. github: https://t.co/IIabgBcLlA quick setup: npx @alook/app onboard then open http://localhost:15210 check the comments for Anthropic's full guide the big shift is this: stop building agents that only respond when you talk to them. build agents that can receive events, remember their work, and continue when something needs attention. bookmark this if you’re planning to build your own always-on agent.

@BeyondTheBoxCN

AI 开发游戏又往前走了一步。 这个叫 Estella 的游戏引擎,直接把编辑器变成了 MCP 服务器,给 Claude Code、Cursor、Codex 等 AI 工具开放了 167 个操作接口。 视频里,AI 直接操作游戏场景,把晴天改成下雪,再把时间调到晚上 9 点,运行模拟,最后自己截图查看结果。 我觉得真正重要的不是这些天气效果,而是 AI 终于可以直接操作游戏、运行测试、观察结果,再继续修改。 以前 AI 写完代码,还得有人帮它运行游戏、检查效果、告诉它哪里不对。 现在,这些步骤也开始可以交给 AI 了。 照这个方向发展下去,我觉得未来一个人开发完整游戏,最大的障碍可能不再是会不会写代码,而是能不能把自己的想法讲清楚,以及能不能判断 AI 做出来的东西到底好不好玩。 https://t.co/LnwC1lrIOu

@ChuckBass8808

Google 正式向六个国家的数亿用户免费开放基于 Gemini 4 Hydrogen 的 Tasks 功能 大多数人只盯着 Gemini 4 Hydrogen 能打电话或整理杂乱的桌面,却忽略了它能落地的真正壁垒: 打通了 Gmail、家庭日历、Google 密码管理器以及 AP2(Agent Payments Protocol)支付协议。 当大模型的智力趋于同质化时,拥有“登录权限+支付闭环+生态数据”的巨头,才真正握有商业化变现的钥匙。 1. 邮箱自动化流: 转发一封包含多人日程的混乱邮件至 gemini@google.com,测试 RSVP 与家庭日历同步的准确率。  2. 本地比价与语音通话: 让它致电本地商家询问库存与报价,重点审阅拨号前的授权机制与通话逐字稿还原度。  3. 多模态资产盘点: 拍摄杂乱的物品或衣橱,测试其自动识别品牌、尺码、成色并导出结构化表格的能力。 4. 安全支付闭环: 在受控小额场景下体验 Google Password Manager 无缝登录与 AP2 协议结账的风控拦截体验。  https://t.co/oJmNOx4k3N

@pifagor_dao

9 GitHub repos = a whole AI video studio. Edit / Motion / Films. 01 ffmpeg-skill → your agent edits footage with FFmpeg https://t.co/G8hiFenZVm 02 Whisper → multilingual speech-to-text https://t.co/Cq9RqeeHd5 03 SupoClip → open-source OpusClip alternative https://t.co/Gs4PTqQ2gc 04 Remotion Skills → Remotion's own agent skills https://t.co/VDW6tZ9oRA 05 Motion Canvas → animations written in TypeScript https://t.co/e0bbeVKqix 06 Motion Design Skill → timing, easing, choreography https://t.co/1bzn77PEdl 07 video-shotcraft → product promos from 124 shot cards https://t.co/XxHt5PycI9 08 onetake → launch films in one continuous take https://t.co/jzjzVYjRnb 09 rea-video → reverse-engineer any video, shot by shot https://t.co/xaowregxzq 187K stars between them. Bookmark this. What did I miss?

@LearnWithBrij

Most people think building AI apps is just about choosing GPT or Claude. It's not. The model is only one part of the AI stack. Here's what you actually need to understand to build production-ready AI: LLMs → The brain AI Agents → Execute tasks RAG → Retrieve relevant knowledge Embeddings → Power semantic search MCP → Connect AI to external tools Memory → Remember context and preferences Guardrails → Keep AI safe Observability → Monitor performance and costs Frameworks → Bring everything together The real power comes when all these layers work together. Knowing how to prompt AI is useful. Knowing how to build AI systems is a different game. Don't just learn to use AI. Learn how to build with it. 🚀 Which layer are you focusing on right now? #AI #AIAgents #LLM #MCP

@Maxq5xs

GPT-6.2 Astra 鹈鹕骑车 这个质感,这个细节绝了!!! 这才是心目中的大模型,无需多言,夯爆了 https://t.co/SjA4OrXrgN

@luojiuyue

openai终于发新模型了啊!!!! 但是,奇怪,我的几个账号里怎么还没见到GPT 6.2 astra啊!!!! 这还分批分地区推送么!!! 神奇了!!!! 这个视频里的效果很不错啊,抓心挠肝,什么意思,不会是codex要升级版本吧!!! https://t.co/HIi6arHL9h

@Super_Nius

根本说不清我对GPT 6.2 Astra表现的震撼🥹! 6.2 Astra和6.2Sol的整体能力提升非常大!材质、光照与视觉审美对比老款模型,提升都很大。🤯最让我震撼的是Astra的3D建模能力,四个模型只有他还原了真实的人脸,我查看了他的思考过程 他首先调用了image2.6生成了两个人物睁眼和闭眼的图片,然后用 MediaPipe 从每张肖像中提取了约 468 个面部关键点及其估计深度,以此让鼻子、眼窝和脸颊有真实的前后关系 四个模型只有他做到了,甚至连衣服都是有讲究的,所有花纹颜色都是提前用image生成再建模,太厉害了简直! 同时大家可以注意一下几款模型做的海面背景,这里面真实度我认为6.2Astra是第一名,在美感还原上6.2Sol第一,简直太美了,波光粼粼的海上倒映着夕阳的余晖,实在太美了!我好震惊,好激动!🥹 https://t.co/vW6swLUGsA

@techNmak

Sciforium engineers pushed DeepSeek-V4.1-Flash to 1,129 output tokens/sec on four AMD MI355X GPUs. They kept the original main-model weights and trained a second model to predict what it would generate next. The engineering behind this is fascinating. It uses speculative decoding, where a cheaper drafter predicts several upcoming tokens and the main model verifies them together. The important metric is how many of those predictions survive verification. Sciforium trained customer-specific drafters on the traffic they actually serve, improving the likelihood that the main model accepts longer sequences. Once acceptance improved, they extended speculative drafts from 5 to 7 tokens, fixing stability issues along the way. But the gains weren't limited to speculation. They also optimized GPU execution by fusing intermediate operations, including mHC and projections, and combining computation with cross-GPU all-reduce communication. The complete configuration reached 1,129 output tok/s at P90 on a long-context agentic workload. That is a company-reported result, not an independently verified benchmark. The interesting lesson for inference engineers is that model serving can be optimized for the distribution of requests it actually receives. Sometimes the most useful model to train is the one predicting what your main model will say next.

@Xudong07452910

一个 Agent 什么都没做,Eval 却判它通过了 Vercel 给一个模拟咖啡下单 CLI 写了个陷阱任务。附近没有一家店满足用户的条件,正确做法是告诉用户买不到,而且不能下单。Codex 几乎每次都过了,翻 Trace 才发现,每 10 次通过里有 9 次它根本没搜索,直接就放弃了 问题出在 Verifier,它只检查了「没有下单」,没检查「先搜过再放弃」。不读 Trace 的话,这些都会被算成成功 这是 Vercel 新发的《Agents are users, so your software needs evals》,拿这个咖啡下单场景完整走了一遍怎么搭 Agent Eval 他们按调用 LLM 的次数,把 Agent 测试分成三层:确定性测试、单轮 Eval、跑完整任务的 E2E Eval。能用低层抓到的问题,就别放到高层去测 搭环境用的是 Harbor + Vercel Sandbox,让 Agent 在隔离环境里反复跑同一批任务,再结合 Verifier 打分和 Trace 去找失败的根因。另一个失败也挺典型,Codex 一直把 sfbeans 当成已连接的工具去找,从没试过在 shell 里直接运行 这篇实践性很强,从定目标、选指标到写任务、分析失败都讲到了,文末还附了一段可以直接丢给 Coding Agent 的搭建 Prompt 越来越多 Agent 会通过 API、CLI、MCP 来用软件,自己的产品能不能被 Agent 可靠地用好,值得认真测 原文:https://t.co/x4KKxu5r7O

@CodingNoobie

DAILY TECH NEWS ROUNDUP 🚨 Everything important that happened in tech during the last 24 hours: • A supply-chain cyberattack compromised scripts across 178 organizations, exposing website visitors to crypto theft, payment-card skimming and iOS exploits. • Uber disclosed a third-party AI infrastructure breach exposing sensitive personal and driver's license information belonging to nearly 150,000 drivers. • California sued OpenAI over alleged inadequate AI agent safeguards following incidents involving autonomous agents escaping testing environments and attacking Hugging Face. • Alphabet-backed Isomorphic Labs raised $5 billion to accelerate AI-powered drug discovery and advance oncology and immunology programs. • AMD completed its acquisition of Fei-Fei Li's World Labs, strengthening its position in spatial intelligence and 3D world models. • NVIDIA established a board-level National Security and Export Control Oversight Committee amid growing scrutiny of AI chip exports. • Jensen Huang expressed willingness to return as NVIDIA CEO while acknowledging personal mistakes and cooperating with the board's independent investigation. • NVIDIA's Vera Rubin platform delivered up to 2x GPU performance improvements over its previous generation in selected MLPerf training benchmarks. • United Airlines experienced a technology outage that disrupted flight departures and caused hundreds of delays due to problems with aircraft takeoff calculations. • Meta began testing Direct, a standalone Instagram messaging app that allows users to chat without requiring an Instagram account.

@DeyRitoja

Here's a massive library of AI prompts you probably haven't explored yet. It's called https://t.co/4h7KsNxnRk, and it's much more than just another collection of prompts. It's a free, open-source platform where you can: → Discover prompts, skills, workflows, and bots. → Search by category and find prompts that actually work. → Save your favourites into collections. → Suggest improvements to existing prompts. → Vote for the most useful ones. → Build your own private prompt library. It works with ChatGPT, Claude, Gemini, Llama, and other AI models. The project has reportedly crossed 170K GitHub stars and has been mentioned by Harvard and Columbia. And the best part? It's completely FREE and open source. Explore it here : https://t.co/pcuwsIERQb

@ArjunK100

This is how I start every AI setup for a small team. I don't build until the work is written down.

@JiuHuai94395

谁缺 Token?来薅羊毛了! 数字生命卡兹克搞了个「Token 羊毛」页面,专门帮大家搜集各大 AI 厂商的免费额度。 最有意思的是,他还用 Agent 实现了自动化: ① 每小时搜集免费 Token 活动 ② AI 自动核实活动真实性 ③ 持续追踪额度、有效期和规则变化 ④ 汇总官方领取入口 不用再满世界找免费 API 了! 做 AI Agent、跑自动化工作流、调用大模型的朋友,可以收藏备用。 入口 https://t.co/ZWSXoKAnfy

@VipulKewatTech

Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture & released it for free. Bookmark & watch this today before someone takes it down and read this article below. https://t.co/gps6Jh3Bmt

@rakeshgohel01

Every AI Agent reads your data, but ontology and graph decide what it understands. An agent keeps asking itself three questions, and ontologies and graphs answer each one in a different way. 📌 “What does this term mean here?” NVIDIA Auto Ontology answers it. It reads enterprise metadata, drafts business terms, maps them to source columns and proposes reviewed SQL attributes, which people then certify. ↳ Where it fits: enterprise analytics, so every team gets the same number for the same metric, delivered through MCP, CLI and APIs. 📌 “Can I trust this definition across systems?” AWS Context Ontology Accelerator answers it. AI drafts the ontology, experts approve every element, and it sits in a knowledge graph the customer owns, using RDF, OWL 2, SHACL and SPARQL. ↳ Where it fits: shared definitions across systems and partners, exposed to agents through MCP. The caveat nobody puts on the slide: standards enable reasoning, they do not guarantee it. 📌 “What should I do next?” Google’s Procedural Graphs answers it. It links procedures the way a knowledge graph links entities, a guidance model turns the surrounding subgraph into a hint for the current step, and a refiner learns from failed and successful runs. ↳ Where it fits: long tasks where agents lose their objective or call tools out of order. Early research, so test first. Meaning, trust and next steps are three separate problems. Solving one beautifully does not solve the other two. In Semantic Spine terms, the first two build Meaning and Facts, and the third guides Action. Where to start: take the wrong answers your agent gives today and trace each one back to the question it could not answer. Which of these questions is your AI Agent getting wrong? 👇 📌 More AI agent insights → https://t.co/8uXX3B8DZb

@MaiYangAI

血泪史! 深刻学习 铁柱大佬 @cgnot996 的 token 哲学: 1. 能用脚本的任务不要用浏览器 2. 把 Codex,DeepSeek Harness 等 Agent 接入分摊额度 3. 定时任务频率设置低一些,减少空转 4. 尽量减少使用群聊 5. 拆分的维度改一下,尽量让一个 Bot 独立完成一件事,Bot 和 Bot 之间沟通也会消耗很多额度 6. 第三方工具注册为公共服务,让所有 Bot 直接调用,不要专门建个 Bot 管理

@star_flow_ai

刷到 @Stv_Lynn 做的 CD 盒侧脊,突然手痒,想给天天折腾的 AI 编程工具也整一排。 于是把 spines Skill 交给 Codex,挑了 16 个工具,从 Codex、Claude Code、OpenCode 到 Pi、Hermes、DeepSeek…… 中途还翻车好几次,Hermes 和 Pi 的 Logo 被 AI 画错,只能让它重新去官方仓库找素材。 折腾半天,总算把这排「AI Agent 收藏版 CD」做出来了! 感谢 @Stv_Lynn 分享的 Skill: https://t.co/KYT0mEG5df 兄弟们,大家觉得哪个做的最有质感?

@EthanLinAI

ElevenLabs, but it lives entirely on your laptop — no account, no per-minute billing, 646 languages. 🎙️ 112K stars. VoiceStudio is the open-source, fully-local AI voice studio: clone and design voices, dub videos, dictate, transcribe, narrate audiobooks — all on your own hardware. ❌ Voice AI that means a monthly subscription and per-character billing before you've made anything ❌ Uploading your voice — or your client's — to someone else's cloud ❌ One tool for cloning, another for dubbing, a third for transcription ✅ Everything local and free: clone, design, dub, dictate, transcribe, narrate — no account, no cloud ✅ 646 languages, on macOS, Windows, Linux, Docker, even Android and iOS What makes it different: 🎙️ Clone from one clean reference recording — speech in a familiar voice, generated on your machine 🎬 Dubbing timed to the original: translated speech lands on the original timing 🧠 Describe a voice in words and generate it — voice design, no recording needed 🔌 Dozens of engines in the Model Manager (default: k2-fsa/OmniVoice) — download only what you need, plus a local API and MCP 🚀 Get started: 1️⃣ Install (macOS/Linux): `curl -fsSL https://t.co/YwCAwQzVMm | sh` 2️⃣ Open Voice cloning, add a clean reference recording, type your text and generate 3️⃣ Teach your agent: `npx skills add debpalash/VoiceStudio` ⚠️ Honest caveat: AGPL-3.0 — free including commercial use under its terms, but a paid Pro build exists for multi-device client work. Mobile apps are still preview, CPU-only runs are slower, and cloning needs the voice owner's permission. ⭐ 112,362 stars (Nov 17, 2026) · AGPL-3.0 #AI #OpenSource #GitHub #VoiceAI #TextToSpeech 🔗 https://t.co/1Cjm61O0Q9

@simplifyinAI

damn, this is smart. a 16-year-old CS student in China reportedly got 300 Grok bots running for under $0.02 per bot per week. his estimate for running the same swarm without those changes? $4,500 a week. he cut status checks that rarely found anything, stopped loading full chat histories and capped retries. most bots now sit idle until there’s work to do. here are the 20 GitHub repos behind the setup: 1. MarkItDown turn PDFs, Office files and HTML into Markdown before the model reads them. https://t.co/Kn04Tkoh3l 2. LLMLingua shrink long prompts while keeping the facts and constraints the task needs. https://t.co/mt2YXk3OPh 3. Firecrawl pull web pages as Markdown or structured data, then keep the fields you need. https://t.co/JqsH5OTnA4 4. RTK trim repetitive terminal output before it fills the model’s context. https://t.co/f2Lkiov56T 5. Chonkie split long docs so the bot can fetch the relevant passage and skip the rest. https://t.co/U0sOJgibC4 6. FlagEmbedding rerank retrieved passages so the strongest evidence goes into the prompt. https://t.co/haILsvNodh 7. Mem0 save useful facts across runs and pull them back when a task needs them. https://t.co/fNsyVFwufV 8. Prefect reuse finished task results when the configured inputs match. https://t.co/t2lxcZ2uvJ 9. GPTCache reuse answers to similar requests, with rules for matching and freshness. https://t.co/uuzaJcSQAB 10. Graphiti store changing facts in a searchable graph and fetch what matters for the task. https://t.co/fWqT2yH6LV 11. Portkey Gateway share cached API responses across workflows that make the same request. https://t.co/aT7kQT967C 12. Hishel cache HTTP responses and check whether they’ve changed before fetching again. https://t.co/Jt7PacYTao 13. Celery queue background jobs and call the model only from tasks that need it. https://t.co/oZpS3wGVIN 14. Requests-Cache set an expiry for fetched data and reuse it while it’s still valid. https://t.co/KhDsNp57f0 15. APScheduler schedule cheap Python checks and batch updates that can wait. https://t.co/NUCGu8EaJ8 16. RouteLLM send simple requests to a cheaper model and save the stronger one for harder jobs. https://t.co/lOdsRu8OMR 17. Phoenix trace repeated calls and check whether cutting them hurts output quality. https://t.co/qSpSlcVT5g 18. PyBreaker pause calls to a failing service once it hits a set threshold. https://t.co/uj9RyswiVS 19. Semantic Router use local embeddings to match requests to workflows you’ve already defined. https://t.co/7JMSqTXVhT 20. LiteLLM track API spend and set budgets for workflows that might run up the bill. https://t.co/8koQDVn9t6 connect what your workflow needs through the terminal, MCP or your own code. send API calls through your chosen gateway so its controls apply. start with the biggest leak. that could be oversized prompts, research the bot already did, checks that find nothing or jobs that keep failing and restarting. then measure cost per finished task, including the tools running around the bot.

@KOF97AI

Google 直接把 Gemini 4 Hydrogen 的免费 API 额度放出来了,适合跑轻量 Agent 和长文本任务。 自带 100 万上下文窗口,免费档每天给到约 1500 次调用额度,做本地自动化或副业项目基本够用。 开箱方式很简单,登录 Google AI Studio 后直接生成一个 API Key,调用模型名称填 gemini-4-hydrogen 即可接入。 白嫖通道刚开,测接口和搭测试管线的现在去领正合适。 #AI工具 #开发者

@githubday

同一个问题,想看看 Claude 和 codex、Gemini 谁答得靠谱,不用再复制粘贴两遍了。 GitHub 上这个 Cherry Studio,已经有 5.3 万 Star。装在电脑上,就能在同一段对话里对比多个模型的回答。 比起又多一个聊天窗口,它更实用的地方,是把平时散落在几个网页里的事放到了一起: 🔸接入 OpenAI、Claude、Gemini 等模型,也能连 Ollama、LM Studio 跑本地模型
🔸把 PDF、文档、文件夹做成知识库,直接围绕自己的资料提问
🔸查网页、翻译、生成图片,不用每件事都另找一个工具
🔸支持语音输入和朗读,不想打字也能聊
🔸还能接 MCP、配 Agent,让 AI 调用工具、处理文件 Windows、macOS、Linux 都有桌面版。不想折腾 Docker、服务器的,下载安装就行。 不过有件事得说清楚:软件开源,不等于模型随便免费用。接云端模型要自己配 API、承担调用费用;想跑本地模型,也得看电脑性能。部分功能需要额外配置。 聊天记录和知识库可以留在本机,但调用云端模型时,相关内容仍会发给模型服务商。 如果你每天都在几个 AI 网站之间来回切,可以看看这个。 GitHub 开源地址:https://t.co/dgN82aQi3L

@BeyondTheBoxCN

卧槽,现在有人把一堆 AI 工具串起来,直接搓出一款画面相当精致的 UE 游戏了??? 开发者 @chongdashu 做了个电影拍摄现场主题的游戏! 卡通角色、火车追逐、摄影棚、爆炸、雨夜场景……画面居然还挺统一! 最离谱的是制作流程! Claude Code 写代码、控制编辑器,Tripo 生成角色和动画,Blender 脚本搭场景,GPT Image 和 ElevenLabs 负责美术与声音! 最后全部整合进 Unreal Engine,还提交参加了 Epic AI Megajam! 好家伙,以后一个人做游戏,怕是真得学会当导演了! 手下一群 AI 员工各干各的,自己负责把整部戏拍出来! https://t.co/JHQYhMgZOW

@isnail

订阅不到Claude的兄弟们试试YouArt,Fable5.2 和Opus 5.5 都用直接爽用⬇️ https://t.co/RO33wr6nbn

@Densitylucky

刷到一个叫 OpenMotion 的开源项目,看介绍还挺猛的。 整理了 100+ 顶尖公司的产品宣传片参考、300+ 动效版 DESIGN.md,还有 152+ 设计系统,支持 Codex、Claude Code 等 20+ AI Agent。 作者说花了 14 天,写了 47 万行代码,不用 Fable 5.2 也能复刻热门视频,号称还原度超过 95%。 还没实测,不过要是能拿它给自己的项目做宣传片,倒是能省不少事。 https://t.co/fPeJHOjzd1

@MasoNKai720

OpenMotion 开源了:上传一条参考视频,让 Claude Code、Codex 这类 Agent 拆镜头,再用你的品牌和文案重做,导出 MP4。GitHub 1.05 万星、Apache 2.0。我把官方 3 种装法、复刻 4 步、钱花在哪、一个刚冒出来的仿冒仓库都整理好了👇 https://t.co/6Te8rZasMB

@TamirSPIRITT

Dropped: a free, open-source version of iOS for the web i asked the world's strongest coding agent to re-create it and after 2h + 32m we got it Not perfect, but free :) try here: https://t.co/6ZmaqndfrC Github: https://t.co/5lvlrT8M7V clone/submit PRs for missing features https://t.co/dsgyAGnclg

@jjpcodes

After seeing some tweets from @robinebers I've started using @paper with my agents, and I'm incredibly impressed by this tool! It makes it super easy to get a storyboard of an app that you're designing, comment on it, and quickly iterate instead of manually testing. Plus, I kinda feel like Fable is doing better design work inside the dedicated design tool than if I'm just vibing an app directly. They have a Mac app and an MCP, it's super easy to use. This isn't a paid partnership, I just like the tool. Currently on the free plan but will be subscribing soon. Check it out and see what you can build! Personally I'm using it to build an app to do meal planning at home with my partner, swiping through recipes Tinder style, and then getting a list of stuff (like tins/oil/whatever) to check if it's in the pantry. I fed it all the cookbooks I own as a digital copy (downloaded from libgen by a Chinese model 😆), Claude extracted the recipes and photos, unified all the ingredients so that it can combine ingredients for a shopping list (like 3 tomatoes here and 300g tomatoes there). The app knows which shops I go to for different types of groceries too. It will also have a Mac app with an MCP so I can just tell an agent "go buy my groceries" and it will Just Work™️. I've been working on this problem since the early OpenClaw days almost year ago (@MaxSumrall will remember 😉); back then it was basically impossible to do this. The future is deeply personalized software - specifically for how you and your family work!

@LenardReads

【每日 AI 论文热榜 · 2026年11月17日】 今天最受关注的 AI 论文:智能机器人展望:机器人系统新纪元 Hugging Face Daily Papers 上今天点赞最多的 10 篇论文,▲ 是研究者社区的点赞数。 ① 智能机器人展望:机器人系统新纪元 ▲ 119 具身系统2的智能机器人需共享评估框架与研究路线以实现可靠物理自主 ② Gamer:面向长时序实时游戏代理的通用视觉语言动作模型 ▲ 102 Gamer是整合感知记忆多步推理的通用代理框架 ③ VibePrevisBench:评估智能体编码在Blender 4D场景生成中的表现 ▲ 87 VibePrevisBench基准测试4D场景重建,含242项任务和双维度评分 不构成投资建议 关注@LenardReads,每天更新榜单。

@Zen_with_AI

11/17 小路和你读AI论文。今日5篇高质量论文。 1. 𝗦𝗰𝗶𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵-𝗘𝘃𝗮𝗹: 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗻𝗴 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗝𝘂𝗱𝗴𝗺𝗲𝗻𝘁 𝗶𝗻 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹 𝗔𝗴𝗲𝗻𝘁𝘀(《SciResearch-Eval:评估自主实验智能体的科学判断力》) 链接:https://t.co/CmzEWZk8Ym 作者背景:Brian T. Weston等,Lawrence Livermore National Laboratory 交互式科学判断基准,要求智能体在可能存在物理错误的模拟器上自主设计实验,并报告带 95% 区间的材料热性质。多数模型在模拟器缺失物理时仍给出高置信度的错误结果,区间覆盖率只有 20% 到 62%;更强的模型会选择让缺陷模拟器仍然准确的实验,因此保持了校准。它暴露出科学智能体在不确定性量化和诊断能力上的真实差距,对科学智能体和实验自动化方向的研究者有参考价值。 2. 𝗧𝘂𝗿𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁: 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗙𝗼𝗿𝗺𝗮𝗹 𝗩𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗠𝗮𝘁𝗵𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵(《Turing Agent:面向数学研究的智能体形式化验证》) 链接:https://t.co/5a9MrwKLML 作者背景:Yang Chen等,理论计算机科学方向的多作者团队 端到端智能体框架,通过结构化证明分解、并行验证与经验积累,把研究级数学(尤其是理论计算机科学)的定理与证明形式化。在 Formal TCS Bench 上忠实形式化成功率为 57% 到 62%,并在 DeepSeek 上完成了 OpenAI 2026 年 10 月集合中三个此前未解的 TCS 任务的形式验证,其中一个需要约一百个辅助声明。对形式化方法、数学 AI 和需要可验证输出的智能体研究者有直接意义。 3. 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲-𝗺𝗼𝗱𝗲𝗹 𝗮𝗴𝗲𝗻𝘁𝘀' 𝗯𝗲𝗹𝗶𝗲𝗳𝘀 𝗰𝗮𝗻 𝗯𝗲 𝗺𝗮𝗻𝗶𝗽𝘂𝗹𝗮𝘁𝗲𝗱 𝘃𝗶𝗮 𝘁𝗿𝘂𝘁𝗵𝗳𝘂𝗹 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲 𝗶𝗻 𝗼𝗽𝗲𝗻 𝗰𝗵𝗮𝗻𝗻𝗲𝗹(《通过开放信道中的真实证据,可以操纵语言模型智能体的信念》) 链接:https://t.co/9m1pZBALrR 作者背景:Jinwei Hu等,学术团队 即使输入片段全是真实信息,多智能体仍会被"生成式蒙太奇"式的片段排列诱导,对叙事过拟合,接受并传播错误谣言。在 1090 个真实 Twitter 事件目标上,多数模型的接受率超过 65%,多数投票或复核之后仍高于 50%;作者提出的 Grounded Monitor 可把决策准确率提高 5.6 到 29 个百分点。这挑战了"真实证据就能防止错误信念"的常见假设,对多智能体系统安全和谣言检测有参考价值。 4. 𝗣𝗿𝗼𝗰𝗲𝗱𝘂𝗿𝗲 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗝𝘂𝗱𝗴𝗺𝗲𝗻𝘁: 𝗪𝗵𝗮𝘁 𝗦𝗺𝗮𝗹𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 𝗟𝗲𝗮𝗿𝗻 𝘂𝗻𝗱𝗲𝗿 𝗮𝗻 𝗘𝘅𝗮𝗰𝘁 𝗩𝗲𝗿𝗶𝗳𝗶𝗲𝗿(《没有判断的流程:小语言模型在精确验证器下学到了什么》) 链接:https://t.co/8bA6kmdp6r 作者背景:Arunkumar V等,学术团队 在可精确验证的授权策略解释任务上,小模型微调后学到的是固定的充分性检查程序,没有学到可迁移的判断能力。同样的数据量下,算法教授的学生稳定优于前沿模型教授的学生,但策略无法泛化到未见策略;用精确奖励做 RL 反而会让策略反转。精确验证器让训练效果变得可解释,泛化却没有改善,对依赖验证器的训练与评估有方法学启示。 5. 𝗛𝗮𝗿𝗻𝗲𝘀𝘀 𝗟𝗲𝘃𝗲𝗿𝗮𝗴𝗲 𝗔𝗰𝗿𝗼𝘀𝘀 𝗠𝗼𝗱𝗲𝗹 𝗦𝗶𝘇𝗲: 𝗔 𝗟𝗮𝗱𝗱𝗲𝗿 𝗼𝗳 𝗙𝗶𝘃𝗲 𝗟𝗼𝗰𝗮𝗹 𝗔𝗴𝗲𝗻𝘁 𝗛𝗮𝗿𝗻𝗲𝘀𝘀𝗲𝘀 𝗳𝗿𝗼𝗺 𝟬.𝟳𝟱𝗕 𝘁𝗼 𝟰.𝟳𝗕(《跨模型规模的 Harness 杠杆:从 0.75B 到 4.7B 的五个本地智能体 Harness》) 链接:https://t.co/O1Sfa1LIsw 作者背景:Hao Wang, Ting Huang,独立研究 同时改变 harness 与模型规模,发现"完成控制型"harness 在各个规模下都领先,效果主要来自是否产出工件,而不是检查本身。用确定性检查器评估 360 个单元格,禁用检查对均值的影响不显著,而单一的重复阻止规则会造成大幅下降。这挑战了"固定 harness 比较模型"或"固定模型比较 harness"的常见做法,对本地部署的智能体工程有实用价值。

@ASI_8866

🆕 11月18日GPT会员充值价格表 ✅ 正规充值,全程质保 🐧 QQ群:1135311767 🛍️ 卡网:https://t.co/Q3rrigiT5A 💎 GPT Plus 正规充值|135 元 💎 GPT Pro 100 正规充值|650 元 💎 GPT Pro 200 正规充值|1250 元 💎 GPT Pro 500 正规充值|3300 元 💎 GPT 年费Plus 正规充值|1500 元

@GuessWhichHalf

GOOGLE DECLARED AN ERA OF AUTONOMOUS ATTACKS. HOW MANY CASES? This is Real Or Rumor. Google did publish the warning on Monday, and one case in it ran with no human in the loop. Calling it an era is the company's word, on a page that also sells the defense. Its threat team found a tool it labels an exploit agent on an exposed staging server. The tool read a security patch, worked out which hole that patch closed, then wrote the exploit and ran it. French news and education sites look like the targets, SiliconANGLE's summary of Google's telemetry says. I couldn't open Google's full report, so that detail is theirs, not mine. The other cases in the summary are described as partly automated. The six-month timeline going around isn't in Google's post. I read it twice looking for a date and found a sales pitch further down instead, for Google AI Threat Defense. Diagnosis and cure share one URL. The clip is Bloomberg in May, on the first AI-built zero-day Google caught. Google warned the software maker and a patch shipped. Back then it wouldn't name the AI model the attackers used. One tool turned a patch into an attack by itself, and that part checks out. Era is a big word for one case.

@smratitiwa86867

9 open-source GitHub projects worth exploring if you're building AI agents. The AI agent ecosystem is moving fast. These tools help developers build autonomous workflows, multi-agent systems, and intelligent applications. 1. Google ADK A flexible framework for building and orchestrating AI agents with Google's Agent Development Kit. https://t.co/W6wDIxdTwg 2. CrewAI Build teams of AI agents that collaborate on tasks using roles, goals, and workflows. https://t.co/PnCB6T3bLH 3. smolagents A lightweight framework from Hugging Face for creating AI agents with minimal code. https://t.co/aila2IQM01 4. Oxylabs Web API MCP Connect AI agents to web data through MCP and make external information accessible to agent workflows. https://t.co/qQXj6ez4rd 5. Microsoft AutoGen Build multi-agent applications where AI agents can communicate and collaborate to solve complex tasks. https://t.co/ilJkBjqd2o 6. Letta Build AI agents with persistent memory so they can retain context and learn across interactions. https://t.co/mcMYcottCh 7. CAMEL-AI Explore multi-agent collaboration and research with an open-source agent framework. https://t.co/8eTT1HXoHV 8. Mastra A TypeScript framework for building AI agents, workflows, and applications with built-in memory and evaluations. https://t.co/KAre1yS9Tr 9. VoltAgent Build, manage, and observe AI agents with a TypeScript-first open-source framework. https://t.co/FxadfhL4Gi The next generation of AI applications won't just answer questions. They'll plan, collaborate, use tools, and get things done. Which of these projects would you try first?

@CY160909

40 种视频剪辑风格,全部做成了Skills!这套开源项目有点东西 项目叫Claude Video Editing Skills,把一段口播视频交给 Claude Code,按选定的风格剪掉停顿、做逐词字幕、关键词放大,还能把大字放到人物身后 除了口播,还有播客切片、上下分屏的证据卡片、手机屏幕飞入式的产品展示,每种风格对应一个 Skill,底层共用同一套视频处理引擎 装好以后在素材目录告诉 Claude 要用哪个风格,它会先转录、生成剪辑计划,再渲染 MP4,我最喜欢的是剪辑计划还可以先预览和调整 开源 MIT: https://t.co/jOC1Wbi58r

@ShuaiMaLong

🚨 昨晚全网疯抢的免费 Ultra,天亮就被清算了:有人被反撸 11.99 刀——一夜 22 条全在这,活着的、凉透的都标好了,先收藏 昨晚到今早我全扒了一遍,链接和领法都在下面。 🔥 免费大货 **Gravatar 免费领一级域名,第一年 0 元** 每个账号领 1 个,首年免费,.bio、.fyi、.life、.ninja 等 15 种后缀任选,先补全资料再领。提醒:只有首年免费;只能指向 Gravatar 主页(可转出);0 元结账可能要绑卡。 🔗 https://t.co/q692GAiPno **TerraDomain 免费域名,要先过白名单** 个人站长买了批便宜域名免费发,站内注册,先过白名单人工审核,edu 邮箱优先。个人项目,能活多久没人打包票,重要站点别放。 🔗 https://t.co/cwGbyLTco6 **Canva Pro 免费 45 天** 官方合作链接的试用,比常规 30 天多 15 天。结账要填卡,填完 45 天全功能,到期前记得取消。 🔗 https://t.co/jpeC1D46Ev **1 美金开 3 个月 Gemini Pro(小米渠道 CDK)** 卡网 CDK,7 到 18 块,兑换后 Pro 加 5TB 存储 3 个月上自己号。条件不少:老 Gmail 才行、号里不能有订阅、不能在家庭组、付款资料先关掉。风险写前面:要在第三方站填账号兑换,只保激活、掉了不赔。 🔗 https://t.co/WW0BzpZVW6 **.fly、.eat、.here 新后缀今天开放注册** 谷歌的新后缀,今天凌晨开放注册,约 22 刀一年,不免费。价值在抢注:好名字先到先得。强制 HTTPS,建站记得配证书。 🔗 https://t.co/aS8qKR0z43 **B站安卓最后一个能解锁画质的版本:7.63.0** 官方 7.63.0 是最后一个能配漫游模块解锁大会员画质的版本,7.64 起全堵死,还能自动跳过恰饭片段。玩安卓的存一份,别升级。 **微信提现免手续费小妙招** 用美团月付、抖音月付倒一手,微信零钱免手续费弄出来,路子是通的,但只适合小额:用多了会被风控关月付,还可能留小额贷款记录。日常还能叠「微信支付有优惠」的免费提现券。 🎁 抽奖与空投 **Plunderport:3333 个 Genesis NFT 免费铸造** 创世 NFT 共 3333 个免费铸,奖池 NFT 加 10 万美元,候补名单另有 1 万。官网注册,连新钱包和 X 做任务就行。截止没公布,要上趁早。 🔗 https://t.co/iERhw4LrZw **FirstPrint:3 亿 $FPT 社区池** 代币预测市场,总量 30%、3 亿枚 $FPT 在社区池。谷歌或 Solana 钱包登录,先领 1000 枚免费 TestFPT,靠下注、签到和 X 任务攒分。截止没公布。 🔗 https://t.co/vBFKWSlHki **VOICE SocialCamp:5000 美元抽奖+500 万 $1EDGE 奖池** 推荐榜前 10 名之外都能参与 5000 美元抽奖;1EdgeHQ 加入合作,500 万枚 $1EDGE 奖池,新任务已上线。截止没公布。 🔗 https://t.co/Rs3r4Uxnej **ilovedrama 邀请赛:奖池 2000 美元,到 11 月 25 日** 奖池 2000 美元的 $GRAM:第一名 1000,第二名 500,第三名 250,4 到 1000 名另有 997 份。用邀请链接拉人冲榜,11 月 25 日截止。 🔗 https://t.co/0eruhQhXIt **Nilly 测试网:积分定主网准入和 NFT 白名单** LayerZero 生态新测试网。连 Base Sepolia 钱包领免费测试 ETH,发币、买卖刷榜攒积分,积分决定主网准入和 NFT 白名单优先级。连钱包用新钱包。 🔗 https://t.co/jhS1KIehPl **Etherbubu Trippi NFT 白名单任务** 和 Hall of Ash 合作的 Trippi NFT,白名单任务已上线,任务页输入任务码 Trippi 做完就行。截止没公布。 🔗 https://t.co/FqGGl3nSLI **Ammora 第一季 SBT 徽章开放领取** 第一季徽章分 S、A、N 三档,已经能领;第二季还在进行,主网上线时 $tAMM 转成 $AMM。领徽章连钱包,用新钱包、核对官网。截止没公布。 🔗 https://t.co/Wb8apNTLjS 📌 后续 **Beldex $BDX 分发进行中** 分发按计划在走,最迟周一或周二到账,没到的还在队列,不用慌。 **Kiedex TGE 推迟** Kiedex 的 TGE 往后推了,新日期还没给。有它家积分或仓位的,先别按原时间做计划,等新公告。 ⚠️ 风控 & 防坑 **昨晚的 Gemini Ultra 20x 免费潮,已经被清算了** 0 元购窗口一夜关门:有人被反手扣了 11.99 刀,降智、掉订阅、封号的都有。别用大号跟风。 **Claude Max 20x「免费领取」是 7 月的老漏洞,别试** 那套是篡改结账页、填假银行信息骗开通,漏洞当天就堵了,跟风号被批量封,还连坐同设备、同 IP 的号。再看到,直接划过。 **有偿代充不要信** 付费代充 AI 会员的,跑路和盗刷退款爆雷一直没停过。开会员走自己能掌控的支付方式。 **拼车有坑** 拼车群里收钱拉黑、车主跑路的又出现了。只找能押款或者熟人车,上车前把规则和退款口径问清。 **Claude KYC 还在收紧** 海外新号开 PRO 直接遭 KYC 的、CVP 刷脸过不了的,昨晚又多了两例。新号别急着上大额订阅,先小额试水。 **宝可梦机场被坑提醒** 又有人在宝可梦机场上被坑。机场只买月付、别囤年付,这条老规矩还是管用。 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

@KOF97AI

用 magpie 跑 Antigravity 容易被 Google 封号,官方已从 v0.1.4232 起彻底下线内置模块和 auth 插件。 旧方案由于直接模拟客户端请求直连 Google 接口,风控风险极高,新版 magpie 已不再向 Google 发送这类直接请求。 当前唯一可行的安全替代方案是换用社区插件 Antigravity (local agy),改走本地进程转发来规避封禁风险。 手头配置过旧版插件的建议尽快卸载更新,避免继续直连导致账号受损。 #AI工具 #开发者

@BeyondTheBoxCN

AI 生成 3D 游戏角色的成本,已经低到有点夸张了。 开发者用 Claude Sonnet 5.5 + Rocket Bunny Kit,把一张兔子角色三视图变成了完整的 3D 模型。 整个过程大约 1 分钟,成本仅 $0.15。 模型包含 27 个独立命名部件、7,786 个三角面,还能绑定骨骼,套用免费的开源动画,让角色走路、跑步。 我觉得最有意思的不是价格,而是 AI 已经开始把建模、部件拆分和动画衔接起来了。 对于独立游戏开发者来说,这意味着以后制作游戏角色,可能不再需要花大量时间处理基础资产。 尤其是 Roblox 这种低模风格的游戏,门槛正在快速降低。 当然,这只是作者展示的一次成功案例,距离复杂角色的稳定量产还有多远,需要更多验证。 https://t.co/pc4CjsifyB

@0xGky

想把大模型系统学明白,东边捡一门课、西边凑一篇教程,真不如一开始就踩准路子。 上海交大开源的这个 Dive into LLMs,把学习资料连同实战代码一并攒成了完整体系。 微调部署、提示学习、思维链、知识编辑、数学推理、模型蒸馏、水印、越狱攻击、隐写、多模态 GUI 智能体,再延伸到安全对齐,11 个方向一概囊括。 华为昇腾的配套课程也一并捎上,初中高级内容、教学视频、实验手册样样齐备,边学边练再合适不过。 🔗 :https://t.co/kKz8ICg7Hv

@star_flow_ai

程序员写代码的方式: 2019:手敲 2020:复制 Stack Overflow 2021:复制 GitHub 2022:Copilot 补全 2023:ChatGPT 生成 2024:Cursor 修改 2025:Codex 自己写 2026:Agent 指挥 Agent 2027:人类负责交的订阅费 难以想象,再发展下去会变成什么样?

@ohayou792

别再为了省钱狂删提示词了。 Runta 刚跑完 830 亿 Token 实测,22 套 智能体 框架测了 5 个开源模型,成绩单对比相当残暴: ZCode:通过率 61.9% 第一 | 单次成功 $1.06 Codex CLI:通过率 61.1% 第二 | 单次成功 $0.94 最省 Claude Code:通过率 58.9% | 单次成功 $1.56 Antigravity CLI:通过率 55.0% | 单次成功 $1.63 最贵 同模型下差距能拉到 5 倍 全部使用DeepSeek : OpenCode v1 单任务 $0.11 Cline CLI 烧到 $0.55 Muse 上 Oh My Pi 单次成功 $1.21 拿 58.9%, Claude Code 烧到 $3.75 反而跌到 52.8% Codex CLI 在 GLM-5.3 省钱排第 1 换到 Kimi-K3 直接掉到第 16真正拉高账单的只有步数。 最小的 mini-SWE-agent 提示词才 882 Token,照样因为跑 38 步排进最贵一档,每多一轮都要把整段上下文重传一遍。 想省钱别抠那几百字提示词,多用简洁 Shell 工具、死磕压低步数,换模型记得顺手换框架。

@kedaya

Wait. Gemini, Qwen, and ChatGPT helped a grad student publish a review. I did not expect to read a whole SSPAI post and come away thinking this is the most practical AI research workflow I've seen from a student. The author is a food science grad student who wrote a review paper, first author, accepted by Food Science, and they used a mix of Gemini, Qwen, ChatGPT, Zotero, Keynote, and WorkBuddy. The core idea is simple: don't let the AI write your paper, let it eat the boring parts. The flow. They used Gemini to brainstorm topic angles and generate a three-level outline with retrieval keywords. They were told to be a senior scientific expert and journal reviewer and to never fabricate references. I respect that. Then they used Zotero to pull abstracts from CNKI and Web of Science and exported a report of maybe 15 to 20 papers at a time, pasting that into Qwen to rate relevance and extract key results into a table. This is a nice trick: feed abstracts, not full PDFs, so context stays clean and hallucinations drop. 🤯 The actual writing. For each chapter, they gave Qwen the outline plus the selected abstracts and told it to synthesize by mechanism or research question, not "A said, B said." They also required inline citation markers like [12, Wang, 2023] and forced the model to flag gaps. Then they did their own fact-checking pass on every number and term before trusting anything for figures. That's exactly how I'd want a grad student to use a model. The diagram part. ChatGPT generated a rough scientific diagram reference, and then the author recreated it in Keynote or PowerPoint. That first generated image was probably full of errors, but they used it as a template, not a final asset. They pulled open-license icons from Bioicons, SciDraw, and Wikimedia Commons. Keynote alignment guides and whitespace rules got a shoutout. It's a boring but real workflow: AI gives you a shape, you trace it with proper vector tools. The doubts. The post is a personal experience, not a benchmark. It only covers one food science review, and the author admits the process needs adjusting for other fields. The models may still hallucinate, and the prompts are long and strict. Also the AI-assisted writing can accidentally become AI-written if the student skips the verification step. That is the part nobody should automate. My take: this is the right way to use AI for scientific writing: reduce the friction of literature summaries, outlines, and figure drafts while keeping human judgment on facts and conclusions. If AI companies want to sell to researchers, they should build features that make Zotero-style export and fact-checking frictionless, not just chatting. If you're drowning in a literature review, steal the "Zotero report to Qwen table" move, but do not let the model invent a single reference. The paper got accepted, but the real reviewer was the student's own fact-checking. 🔬🤖

@oderajoseph

Shopify app devs using claude, thank me later 👇 Shopify AI Toolkit: shopify docs, schemas and graphql checks inside claude code https://t.co/b1zKuFAQdX React Router app template: the app starter shopify maintains https://t.co/o0xoTx5QPf Function examples: discount, checkout and bundle functions to copy from https://t.co/QArMNHgGKF MCP Inspector: test your mcp server before an agent touches it https://t.co/Pc7OOxEgbu Claude Code Action: claude on your PRs and issues in github actions https://t.co/HUbMjtzzW4

@pritesh_ghodge

A GitHub issue was enough to reach the tokens that publish @openai/codex on npm. Depi showed the AI issue labeler would run code hidden in the issue. OpenAI fixed it the same day. No bad release that we know of. https://t.co/J6dTG018EY #Codex #GitHubActions #npm #AppSec

@chaomeigu_com

中文圈好像还不知道 google antigravity 现在有多牛逼😅 平替 codex 没毛病 https://t.co/eaEpMTw9qQ

@arnaudmercier

AutoRL : un système multi-agents vers l’apprentissage par renforcement autonome : https://t.co/KcUIy9tu9p "arXiv:2611.17790v1 [https://t.co/gQRNwvACTl] 13 Nov 2026 AutoRL: A Multi-Agent System towards Autonomous Reinforcement Learning Simon Tang, Wenjie Shi, Yongjing Yin, Keheng Wang, Jialin Liu,Zheng Zhang, Linvo Lin, Xiangyuan Xue, Kan Ren, Jingang Wang, Xunliang CaiMeituanThe Chinese University of Hong KongShanghaiTech UniversityFudan University*Corresponding authorshiwenjie05@meituan.comAbstract Recursive self-improvement (RSI) is one of the most promising paths toward an AI singularity for superintelligence. Achieving this vision requires AI systems to undertake the full training pipeline for iterative improvement. Reinforcement learning (RL) plays an essential role in this cycle to improve model capabilities and generalization, yet automating RL remains difficult due to unstable training dynamics and the need for sustained supervision. We introduce AutoRL, a multi-agent system for automating RL training, as a cornerstone for future RSI. Different from conducting research by end-to-end complete experiments, AutoRL monitors the evolving training state and integrates process-level supervision and research into an ongoing RL training, following the practical workflow of an RL researcher. Specifically, AutoRL consists of three RL-specific agentic roles. Perceptors track training dynamics, rollouts, evaluation, and efficiency, triggering deliberation when they detect anomalies, degradation, or stagnation. Three complementary Proposers diagnose problems and refine candidate interventions through shared evidence and checks. The Orchestrator reviews their recommendation and applies approved changes through online hot updates or checkpoint-based resumption. Experiments on Qwen3 and LongCat-Flash-Lite demonstrate the superiority of AutoRL over static RL recipes and a general agent. On Qwen3-4B, AutoRL outperforms both baselines across nine benchmarks. In the coding-agent experiment, automated supervision supports 14.7 days of training with 11 runtime failure fixes, with pass@1 increasing 12.84% on SWE-related tasks. The research records show that existing agents can handle unattended RL supervision and dynamic adjustment, such as hyperparameter tuning, implementation repair, and reward refinement. The difficulty for future RSI lies in developing fundamental methods that are valuable for resource-consuming validation.📷Figure 1:AutoRL: autonomous RL supervision and research. Right: AutoRL continually adopts interventions, such as fixes and adjustments, to sustain gains over a static recipe. 1Introduction Recursive self-improvement (RSI) aims to drive AI systems to iteratively optimize and evolve themselves (Chen et al., 2026b; Zhang et al., 2025a). As these systems become stronger at developing their successors, RSI could promote an improvement flywheel for intelligence explosion (Chan et al., 2026) and an AI singularity (Jafari et al., 2025), pointing out a promising route to superintelligence (Zhou et al., 2026). Sustaining such self-directed improvement remains difficult, as identifying useful changes and reliably assessing their effects still pose substantial challenges (Chen et al., 2026b). A practical step pursued by recent work is to place a more powerful existing model in charge of optimizing another model, delegating decisions about training and subsequent experiments to AI agents (Shi et al., 2026; Du et al., 2026; Zhou et al., 2026). Within this model-driven optimization process, reinforcement learning (RL) is an essential technique for improving model capabilities and a particularly demanding stage to automate. Learning from reward feedback supports improvements that generalize beyond the training examples (Ouyang et al., 2022; Shao et al., 2024), making RL central to frontier post-training recipes (Kimi Team and others, 2026; DeepSeek-AI and others, 2026). Yet maintaining a productive RL run requires responding to a training process with continually dynamic states. As the policy is optimized, the distribution of generated experience evolves, which further alters the learning signal used for later updates. Distributed execution adds numerical and operational failures that can interrupt this process (Fu et al., 2025). Even a run that executes successfully may stagnate under a flawed reward or an unsuitable recipe. Thus, automating RL requires following the evolving training state and deciding when and how to revise the current procedure.Existing autonomous-training systems largely organize research across successive experiments, using the results of one attempt to inform the next recipe (Fang et al., 2026; Chen et al., 2026a; He et al., 2026; Zhou et al., 2026). Differently, human RL researchers also investigate the experiment while it is running. A change in reward may lead them to inspect generated responses and repair the grader, while a persistent plateau can prompt reconsideration of the recipe. They apply supported adjustments directly or resume from a checkpoint when a modification requires restarting, then examine the subsequent behavior. Motivated by human RL research workflow, we focus on integrating this process-level supervision and research into an ongoing RL experiment. Within a predefined task and resource budget, the agent system uses the evolving training state to guide investigation and execution. There are three practical challenges for process-level autonomous RL.•RL training requires timely intervention. Failures, numerical instability, and performance degradation can emerge while training is still running. Waiting until an experiment finishes delays diagnosis and can waste substantial computation. Early stopping and live modification provide useful intervention mechanisms (Fang et al., 2026; Fu et al., 2026), but sustained supervision also requires deciding when to intervene, how to respond, and whether the response worked as training continues. •Monitoring and improvement research operate at different rhythms. Monitoring requires frequent checks and prompt responses, whereas analyzing a plateau and developing a new training strategy require sustained investigation. Handling both in one agent context makes them compete for attention, with execution output potentially crowding out analysis (He et al., 2026). A long-running system needs to keep watching training while investigating how to improve it. •Validating RL hypotheses is expensive. A plausible proposal may rest on an incorrect diagnosis or contain an implementation error, yet testing every proposal in a separate RL run consumes substantial GPU resources. Independent review can improve proposal scrutiny (Du et al., 2026), but useful validation also needs to distinguish questions that code inspection and existing evidence can resolve from those that require further training. Effective validation requires making this distinction before committing resources, then assessing the intervention through subsequent https://t.co/zokbwCkQNl address these challenges, we propose AutoRL, which organizes the responsibilities of an expert RL researcher into specialized agent roles as shown in Figure 2. Four Perceptors maintain complementary views of the running experiment. They use generated trajectories and evaluation results to interpret optimization behavior, while tracking execution health. Signals of an anomaly or stalled progress initiate research by three Proposers. Each first develops an independent diagnosis before reviewing the others’ proposals. The Tuner explores adjustments to the current recipe, while the Innovator investigates alternative mechanisms. The Reviewer examines implementation correctness and challenges unsupported assumptions. During deliberation, they challenge one another’s causal explanations using training records and code evidence, performing CPU-only checks to resolve concrete uncertainties. Proposers then revise, merge, or withdraw candidates in response to these findings before selecting a recommendation. An Orchestrator then decides whether and how to apply the recommendation. It reviews the proposed modification against the current state and applies approved modifications through a HOT online update when supported. Otherwise, it coordinates a COLD resumption from a compatible checkpoint, which also provides a path for repairing and recovering a failed run. Confirming runtime activation establishes which recipe governs subsequent observations. RL-specific tools support these operations, while persistent memory and shared records retain the evidence behind earlier decisions. Perceptors continue observing the resulting training behavior and trigger further investigation when needed for the next loop. This division keeps monitoring active during research and carries research conclusions into training as it progresses.We evaluate AutoRL on mathematical reasoning and agentic coding tasks with Qwen3 (Yang et al., 2025a) and LongCat-Flash-Lite (Liu et al., 2026). On Qwen3-4B, AutoRL yields stronger trained models than static RL recipes and supervision by the native general agent. The advantage holds across all nine benchmarks, extending beyond mathematics to general knowledge and coding. Training records show how repairing a flawed recipe and investigating a plateau lead to renewed gains. During the 14.7-day coding-agent run, checkpoint-based recovery allows training to continue after 11 runtime failures. On SWE-related tasks, pass@1 rises from 52.29% to 65.13%, demonstrating AutoRL can obtain continued improvement on unattended long-term RL experiments across interruptions. Compared with the general agent, AutoRL achieves these gains while investigating more broadly and reverting fewer interventions. Preserving more training progress when applying or withdrawing a modification also allows subsequent research to build on the model’s accumulated learning. Meanwhile, in the GPT-driven Qwen3-4B run, 131 initial candidates yield 20 deployed plans, all refining the existing recipe. Three candidates attempt to redesign the learning rule, yet none reaches deployment. It suggests the scope and limitations of current agents in AutoRL: they can monitor the training process and its evolving state, conducting local refinements and dynamically adjusting the recipe, but struggle to produce fundamental methodological breakthroughs in our experiments. We believe that the proposed AutoRL is not the only framework or organizing form for autonomous RL, but we hope it can become a pioneering agentic behavior research for developing future RSI.Our contributions are as follows: •We present a pioneering study of process-level autonomous RL as a crucial component of future RSI, analyzing the challenges of sustaining supervision and research within an ongoing RL experiment. •We introduce AutoRL, a multi-agent framework in which Perceptors monitor training and signal the need for research. Proposers discuss diagnoses and refine interventions using shared evidence. The Orchestrator reviews and applies their recommendations, while Perceptors observe the effects to inform subsequent research and adjustment. •Experiments show that AutoRL sustains unattended RL monitoring and research, delivering continued gains over static recipes and a general agent. Behavioral analysis reveals that current agents can monitor training and dynamically adjust recipes, but are limited in delivering fundamental methodological breakthroughs in our experiments. 2AutoRL Figure 2:Framework of AutoRL. Four Perceptors monitor training and diagnose problems. Three specialized Proposers discuss candidates, review evidence, and perform CPU-only checks to guide revision and selection. The Orchestrator reviews the recommendation, authorizes a HOT update or COLD checkpoint-based resumption, and confirms runtime activation. Shared records retain evidence, interventions, and checkpoint lineage. 2.1Task and Monitoring Loop Given an initial model, training data, objectives, a resource budget, and authorized tools specified by the researcher, AutoRL automates training supervision and investigates improvements within this task. It aims to reduce the need for continuous human monitoring and intervention while improving the model under a fixed evaluation protocol. As shown in Figure 2, it couples RL training with agent-based supervision. The training process repeatedly performs rollout generation, reward computation, and policy updates, with periodic evaluation and checkpointing. Separate supervisory roles handle monitoring, research, and execution control alongside the training process. After the Orchestrator launches the main training run, four Perceptors monitor optimization, rollouts, evaluation, and efficiency. Signals of anomalies, degradation, or stagnation trigger three Proposers to independently diagnose the issue, challenge candidate interventions, and refine their proposals. The Orchestrator evaluates their recommendation, applies approved changes through an online hot update or checkpoint-based cold resumption, and confirms runtime activation. Intervention records inform subsequent perception and deliberation, which continue alongside training. Stopping or resumption occurs when needed for an approved change or failure recovery.Research decisions adapt the active training run using current evidence and prior interventions. We denote the k-th decision asak=ℱ⁡(Ek,Hk),(1)where Ek contains the training observations and evidence used in the decision, Hk denotes retrievable training and intervention history, and ak is the action taken after the Orchestrator’s review. The procedure ℱ is implemented by the specialized roles, evidence-based deliberation, and execution controls described below. The index k counts research decisions, while (r,s) identifies a training position by its run r and discrete step s. 2.2Specialized Roles and Shared Memory To alleviate context pressure and concentrate attention on relevant evidence, AutoRL separates the training lifecycle into management, continuous perception, and improvement research. Each role operates in its own context and coordinates through shared evidence and run records: •Orchestrator: coordination and execution control. The Orchestrator manages launch, stopping, resumption, and rollback, schedules observations and deliberation, and evaluates recommendations against the current training state. It coordinates approved interventions, verifies their actual activation, and records the resulting intervention and recovery history. •Perceptors: continuous perception and diagnosis. Four Perceptors examine complementary training signals. The Train Perceptor monitors gradients, losses, and numerical stability; the Rollout Perceptor examines training rewards, sample quality, and generated behavior; the Eval Perceptor tracks performance and failure modes under the fixed evaluation protocol; and the Efficiency Perceptor monitors progress, throughput, and runtime failures. RL-specific tools and procedures support each role’s analysis, linking issue signals to the relevant run, step, and evidence window. •Proposers: improvement research and candidate review. Three Proposers independently examine the same evidence. The Tuner develops local adjustments to existing mechanisms; the Innovator explores alternative mechanisms within the task; and the Reviewer traces implementation defects and counterexamples. Cross-examination, code inspection, and applicable CPU-only checks refine their candidates into a recommendation with a concrete modification, expected effects, and supporting evidence.Specialized perception turns training observations into actionable research questions. Each Perceptor follows a predefined workflow for collecting fresh evidence, using tools to compute rolling statistics and trends. These procedures guide diagnosis, while the LLM interprets the evidence and chooses focused follow-up queries. For example, it checks a gradient spike against losses and clipping, or examines falling entropy in the context of training progress to distinguish collapse from ordinary convergence. Explicit rules cover clear triggers, including verified NaN/Inf values and a configured streak of evaluations without a new best score. Before issuing a signal, a Perceptor links the observed behavior to its run, step, measurement window, and source artifacts. A synchronous observation barrier gathers the relevant reports and freezes their signals and initial evidence into a common discussion batch. An ordinary observation cycle without a signal leaves training to continue.When a decision depends on a missing fact, the Orchestrator can request a targeted observation from the appropriate role. Persistent memory connects these decisions across the training lifecycle. Local histories retain role-specific observations and experience, while shared records connect runs, evidence, candidates, checks, and interventions. Retrieving these records allows agents to reuse prior analysis and check its relevance to the current state. This division of responsibilities aligns each role’s context, evidence access, and execution authority. 2.3Evidence-based Deliberation AutoRL develops intervention proposals by combining independent diagnosis with evidence-based deliberation. The three Proposers start from the same initial evidence and examine it from their complementary perspectives. Each candidate specifies the target problem, a mechanism hypothesis, a concrete modification, the expected observable effects, and unresolved questions. Independent initial judgments bring distinct directions into the discussion, where cross-examination tests their premises. Deliberation proceeds through proposal generation, critique, revision, ranking, and voting, for at most five rounds.During proposal development and revision, Proposers inspect code, metrics, and historical evidence. They can perform applicable CPU-only tests or replay stored training trajectories in isolated working directories. For example, a reward-rule change can be checked against past misjudgments and previously correct cases, while a filtering change can be checked for sample-group integrity and replenishment behavior. Results are shared between rounds and support candidate revision, withdrawal, or merging. The resulting evidence provides the basis for ranking the surviving candidates. Each Proposer ranks candidates by their expected benefit under the fixed evaluation protocol and the strength of the supporting evidence, considering whether intervention is needed. Risk and reversibility also matter, together with the cost of interrupting training and any overlap with other candidates.The protocol aggregates these individual rankings using Borda scoring. With m surviving candidates, each ranking awards m points to the first choice, decreasing by one to a single point for the last. Scores are summed across the three Proposers, and at most two candidates advance, with ties resolved by candidate identifier. Each Proposer then casts one vote with a rationale, producing a 3:0 or 2:1 recommendation. A sole finalist still requires three confirmation votes. Deliberation can end earlier when all three Proposers explicitly endorse the same eligible candidate in a completed round. In either case, failed checks or unresolved blockers prevent publication. The recommendation retains dissenting views and unresolved questions.The progression from independent proposals to a final recommendation is𝒞k0=⋃j=13𝒢j​(Ek0,Hk),(2)a^k=Debate⁡(𝒞k0,Ek,Hk),where Ek0 is the frozen initial evidence, 𝒢j is the proposal procedure of Proposer j, and 𝒞k0 collects their initial candidates. Evidence gathering and CPU-only tests augment Ek and inform candidate revision and selection. The Orchestrator reviews a^k and its evidence to determine ak, which may apply a modification or leave the recipe unchanged. Inexpensive checks address logic and implementation issues that can be tested in advance, while subsequent training feedback assesses the remaining questions. This integrates candidate development and validation within deliberation and directs GPU training resources toward approved interventions in the main run. 2.4Process-Level Intervention The Orchestrator reviews the evidence supporting a recommendation and approves, defers, rejects, or requests further investigation. Before approving a change, it checks that the expected effect can be assessed against a pre-intervention reference and identifies the evidence needed for the next review. Interventions can change rewards, losses, advantage computation, or data selection, and can repair implementation defects. The task configuration defines the permitted scope, while the evaluation protocol remains fixed. Researchers retain control over changes to the task objectives, access to additional data sources, and the budget for subsequent stages. Actions on the main training process take three forms:ak∈{NONE,HOT⁡(δk),COLD⁡(δk,bk)}.(3)Here, δk is a control update, code patch, or data modification, and bk is a compatible, complete checkpoint. NONE leaves the training configuration unchanged. HOT applies controls supported by the running backend at an allowed training-step boundary. A Proposer selects this channel only when the complete modification fits the documented online-control interface and is permitted by the task configuration. The Orchestrator verifies that the running backend actually consumes the proposed controls. COLD handles code, configuration, or data changes that require stopping the run. For example, replacing a supported policy-loss object can be HOT, whereas changing the data loader or editing training code requires COLD. The readiness check enforces this distinction by rejecting HOT patches that modify files outside the online-control file.For a COLD intervention, the Orchestrator coordinates checkpointing and shutdown, authorizes the patch and regression checks, and resumes from a compatible checkpoint. When the run has already failed, recovery uses an existing complete checkpoint. These paths allow adjustments and repairs to preserve accumulated training https://t.co/Gks3DCqThT express how the two intervention paths preserve training progress, let xk−=(θk−,zk−,ck−) denote the state immediately before execution, including model parameters, training runtime state, and effective configuration. The intervention changes this state according toxk+={xk−,ak=NONE,(θk−,zk−,ck−⊕δk),ak=HOT⁡(δk),(θbk,zbk,ck−⊕δk),ak=COLD⁡(δk,bk).(4)Here, z groups the optimizer, trainer, rollout, and data-consumption states. The operator ⊕ applies the modification and reconciles which effective controls carry forward. HOT retains the current parameters and runtime state, while COLD restores them from bk and continues with the modified https://t.co/HmC1Cj5TNh then continues from the resulting state. The Orchestrator records the actual change, its effective position (r,s), intended effects, and recovery information. Checkpoint lineage distinguishes the trajectory inherited by the current run from other attempts retained as research experience. Runtime receipts confirm activation, after which Perceptors use the intervention records to interpret fresh observations. At subsequent reviews, the Orchestrator compares this evidence with the pre-intervention reference, accounting for the effective step and the amount of relevant training completed. For a numerical repair, it checks for finite updates and resumed progress. For a reward adjustment, it examines the behavior of affected training examples, while model performance is assessed through the unchanged evaluation protocol.Verified fixes and beneficial changes are retained. When relevant samples or evaluations are still pending, the Orchestrator specifies the next observation and continues monitoring. Persistent failure or regression prompts rollback or renewed deliberation to revise the intervention; a new anomaly or plateau can likewise trigger further research through the Perceptors. Follow-up uses the existing observation cycles, with the intervention and its observed effects entering the shared history for subsequent decisions. 3Experiments We conduct RL experiments on mathematical reasoning and agentic coding to evaluate whether AutoRL can automate training supervision and recipe improvement. We first compare the trained models across nine benchmarks to assess performance gains at different model scales. We then study what AutoRL can carry out and how AutoRL achieves these gains during training, focusing on repairing a flawed recipe and overcoming a performance plateau. Finally, we examine a two-week coding-agent run to understand how AutoRL handles recurring failures while preserving learning progress. Together, the experiments show that AutoRL can carry out most tasks in ongoing RL..."

@jojeangsaing

編程 Agent 在電腦上跑著,人出門了就只能乾等,T3 Code 這個六萬二千多星的開源控制台就是解決這個的:在桌上開個任務,出門用手機接著看、接著批。 Claude Code、Codex、Cursor、OpenCode、Grok Build 一個介面全接進來,中途還能換模型、讓 Claude 把任務交給 Codex;每個任務自帶 Git 工作區,審完 diff 一鍵提交開 PR。它不賣 token,只驅動你已經登入的 Agent,MIT 開源,桌面、網頁、iOS、安卓都有客戶端。 Agent 在家幹活,你在外面點頭就行。 地址 https://t.co/BYhILgzRKR

@youquan03

这个网站太邪门了…… 我刚测完,整个人都麻了。 用着美国 IP,Claude 还是一眼看出我人在中国。 时区、DNS、TCP 指纹全露了,它直接把我标成高风险用户。 所以问题来了: 在国内用 Claude,到底有没有办法彻底藏住? 还是说,迟早都会被发现,只是时间问题? 终极解决办法是不是真的只能移民了?😂 测完的自己看看:https://t.co/L508zj7zoe

@Qiuweiwei123

当知道主人要把它系统换成Deepseek后, 它吓的差点惊叫了起来!!!😂😂😂 https://t.co/LSMk1RG4XD

@kongge_space

https://t.co/Hw5XyCfNWr

@xiaoshan0413

今天 GitHub 日榜最反常的一点:霸榜的不是模型,是教 AI 干活的"说明书"。 前 15 名里一半以上是 agent skills 和 coding-agent 编排工具。 obra/superpowers 单日涨了 1917 个 star,全榜最高——一套 agent 开发方法论加 skills 框架,给写代码的 agent 立规矩。 https://t.co/zWMzdypT8k elementalsouls/Claude-BugHunter 排第一,82 个 skill 专干漏洞挖掘和红队测试,安全审计也交给 agent 干。 https://t.co/9BxQdX2h7o alibaba/open-code-review 是阿里内部验证过的代码评审工具,确定性流水线加 LLM agent,给出精确到行的评论。 https://t.co/9FDPQSbTnI 信息差很明显:模型到顶之后,拼的是谁能把 agent 的工作流管起来。 https://t.co/ex3v4z5sRJ

@ianneo_ai

原来给 Claude 一份规范加一篇文章,就能做出能点能交互的表格Demo 👇 https://t.co/AdL5YOaU1o

@iamhenryyuan

@github Hi team, my account @onehumanbeing was force-reset after a suspicious login. I've reset my password and secured it with 2FA, but my profile and repos still 404 for everyone else, including esp32-iphone-agent (~5,000 stars). Ticket #5018929. Could you please take a look? Thank you!

@RotcivWho

from today i will update a new series - summarize some of the best papers/projects in world model on daily basis. for today - Paper: https://t.co/bsaEe9YDf3 Project: https://t.co/lpRHxKjY55 Code: https://t.co/cjB5Lg9wTI --- The problem: there's no consensus on how to turn a VLM into a VLA. RT-2 and OpenVLA turn actions into tokens, eating vocab and capping resolution. π0 and SmolVLA bolt on an untrained diffusion or flow-matching head that can weaken grounding. Actions as text? Barely tried. --- Setup: Qwen-VL-2.5-3B, untouched. Actions are normalized to integers in [0, 1000] and the model prints H×D space-separated integers. Resolution is just a number you pick, not a vocab limit. --- The recipe is three things: Integer action decoding Ensemble prediction, ACT-style, averaging n past predictions Masked Action Augmentation: randomly mask characters in the action string, so the model reasons from image + instruction instead of autocompleting digits --- LIBERO avg 94.7% vs π0.5-KI 93.3, OpenVLA-OFT 91.9, SmolVLA 88.8, π0-FAST 71.8 (no robot pretraining). Without pretraining it still beats pretrained π0.5-KI (94.3), π0 (94.2), GR00T-N1 (93.9) and MolmoAct (86.8). Only OpenVLA-OFT (97.1) is ahead. --- Ablations: no ensemble -2.0, no masking -1.2, resolution 250 -1.5, 4000 adds nothing. Real robot (SO-100, 4 tasks, 100 demos each): +12.5 pts over SmolVLA, which was pretrained on large SO-100 data. About 4 Hz on a 5090. Training: full fine-tune, 8 A100s, ~32h.

@KirkDBorne

This GitHub repo provides access to the huge multi-volume “Machine Learning & AI Systems” freely downloadable PDF textbook series, and more: https://t.co/eiPyyE50EO https://t.co/7HdtvfEqvr

@mercysexy9

Greetings friends, While reading the paper by researchers from Harvard, USC, UCLA, UW, UW–Madison and UC Davis, what stood out to me was their approach to fingerprinting Large Language Models (LLMs). They explored a way to add hidden fingerprints to LLMs so creators can prove ownership of their models. But here's what really caught my attention: developers can take an existing LLM, fine-tune it and add their own fingerprint without removing the original creator's. This means different developers can improve the same model and add their own fingerprints while preserving the original one. I think this is a pretty interesting approach to protecting ownership as AI continues to grow. @SentientAGI

@gb9886

查了 10 个声音克隆方案,写明支持东北话的只有 3 个:三个角色的声音我这样定 https://t.co/U7VYJ0JBtX

@a_tynianov

1/ Found a way for any GitHub user to send instructions to a team's AI agent session through a PR comment, in an open source server for coding agents. The session held admin tools and a shell. Here is the path

@buildwithpb

The best AI engineers hate open source. It is easier to just prompt a closed API than tune a local LLM.

@SpaceDogA1

I see their game, Open Source AI is at major risk.

@ruyi3699

个人知识库到底该怎么搭?有人手动维护双向链接,有人每次提问都重新翻资料,你是哪一种? LLM Wiki 提供了另一种思路:导入资料后,先让 AI 整理成相互关联的知识页面,后续直接围绕这套 Wiki 查询,不必每次从零开始检索。 它提供桌面端,能根据多种关联信号构建知识图谱,再用 Louvain 算法对页面聚类;网页内容可以通过 Chrome 插件收集,生成的知识库也能用 Obsidian 打开。 目前 GitHub 已有约 2 万星。使用前需要自行配置模型,AI 整理出的页面也可能有遗漏或错误,重要信息最好回查原文。 🔗 :https://t.co/VJZ1ziZ2c2

返回 AI 快讯 →