2026年11月15日 · AI 早报

本轮主线是智能体如何在模型之外变得更可靠:自改工作流、筛选记忆、组织并行协作,以及验证最终交付。另一组实践把编程助手接入视频、设计和本地语音工具,展示了具体能力,也暴露出成本、细节损失与协作边界。

11月14日 23:00 — 11月15日 11:00(日本时间) · 116 条 X 原帖

查看本期全部 116 条原帖 ↓

本期目录

产品接入扩展,商业价值仍受追问

DeepSeek 的应用能力与编程入口

EngMoElgaraihy 转述,DeepSeek 将搜索、文件分析、语音等应用能力向开发者开放,并开放聊天前端代码。剑的 AI 实验室则整理其 Codex 接入文档:通过 Responses API 和配置脚本使用 DeepSeek,脚本包含备份与恢复。这些是本批社区对文档和项目的介绍,应区分接口开放与前端开源。

来源 · 2 条原帖

Claude 的会议接入与批量工具调用

Topher Stephenson 介绍,Claude 参会功能正在 Team、Enterprise 测试,当前限 Google Meet;他建议把完整转录定期写入可追溯的知识库。芝诺另称,Claude Code 2.1.323 加入默认关闭的实验性 Code Mode,用 JavaScript 一次执行一串调用,减少逐次等待。两者均来自社区介绍。

来源 · 2 条原帖

就业预测与效率红利的分配

多帖重复传播 Amodei 对初级白领岗位的预测:未来一至五年可能有约一半受到冲击,这是预测,并非已发生的就业统计。Aron厚玉的反问更具体:执行自动化也会降低准入门槛,收入仍取决于需求、用户、数据和议价权。另有社区账号将 Anthropic 描述为上市前夕,并批评输出缩短、WebFetch 静默替换;本批没有正式上市公告,不能据此确认上市安排。

来源 · 3 条原帖

自我改进开始接受任务检验

改工作流能否稳定带来进步

Flex-RSI 作者介绍,他们让十八个模型在二十四类任务中修改代码、工具和工作流,保持权重不变;七组配对实验中六组受益,但更多 token 与更好成绩基本不相关,反馈质量可能成为瓶颈。社区介绍的 Self-Harness 同样围绕失败归因修改运行框架,并通过两组任务检查,只保留没有退步且有所改善的改动。

来源 · 2 条原帖

预训练里的小循环

赛博吴同学转述,上海创智学院与新序智能用三个月训练了 801-30B-A3B,使用十一点八万亿 token。报告把基础设施排查、数据处理和后训练潜力评估拆成迭代循环,没有宣称智能体独立完成全部预训练。轻量微调数据已公开,基座计划开源;跨不同后训练条件的成绩比较也被注明并不对等。

来源 · 1 条原帖

并行规模背后是协作基础设施

pipi.md 介绍微软研究院 Agensh:依靠共享 Git、通信与上下文,让智能体自行领任务。六小时、无联网的五项测试中,一至一百二十八个智能体的平均测试通过率从 19.31% 到 28.78%。落地工具也在补齐管理层:LoopForge 提供自托管编排和人工审批,agweb 提供不抢焦点的浏览器,但限 macOS 26 以上,共享登录也带来相互影响。

来源 · 3 条原帖

记忆要服务决策,也要经得起迁移

从语义相关转向行动贡献

elvis 介绍的 MemCredit 用受控干预比较有无某条记忆时的行动变化,再训练小估计器分配记忆预算。在 ALFWorld、五百 token 预算下,成功率为 86.57%,BM25 对照为 65.67%,策略本身不变。它追问的是:检索到的内容是否真的改变了下一步决策。

来源 · 1 条原帖

压成十六个 token 的收益与损失

linyun 转述 Memorilla:冻结 Qwen3-8B,用独立模块读取文档向量,按问题压成十六个记忆 token。五百一十二篇文档的单项测试中,延迟由 19.9 秒降至 0.74 秒;但 PubMedQA 比全文输入低八个百分点,人名、数字等细节容易丢失。该方案还需要输入层接入和模型专用模块,不能直接用于闭源 API。

来源 · 1 条原帖

经验复用不能丢掉原始记录

Blum 转述 LinkedIn 研究,换模型后沿用旧模型笔记,正确率从 47% 降到 34%,从原聊天重建更有效。何笠介绍的 AgentRun 则把成功经验存为 Skills,使用三次后尝试转为 Workflow。与此同时,How To Prompt 对 grep 胜过向量检索的研究提出边界:字面答案测试不能代表所有语义检索需求。

来源 · 3 条原帖

量化、能量推理与模型内部结构

MACaQ 降低量化时的装载压力

SuSu 转述华为 MACaQ,称以流式逐层方式量化 Qwen 122B,耗时六十四分钟、峰值显存 32.5GB,并提供 GGUF 等输出。这里描述的是量化过程的资源需求,不能直接理解为完整模型推理只需同样显存。

来源 · 1 条原帖

推理加速与视觉参数的新安排

SuSu 另介绍 CRAM,用可训练能量函数处理数独等推理任务,并以能量变化判断停止;帖称训练提速千倍,但未明确比较基线,不能外推到通用大模型。其转述的 Meta、NYU 多模态研究则称,两百多个模型的对照支持保留视觉专用参数,同时探索把它们放入语言模型前几层。

来源 · 2 条原帖

读取自身权重,距离改写还有多远

Varun Gangal 团队探索模型学习解释自身权重。作者称,经监督微调的 Qwen3.5-4B,在判断一段自身神经元权重促进或抑制哪些 token 的测试上达到 99.8%;更复杂的电路读取和直接写权重尚未成功,这仍是受限任务上的早期结果。

来源 · 1 条原帖

科研自动化要看交付如何验收

论文写作的模拟评分与人工偏好

Alex Veremeyenko 介绍 Google PaperOrchestra,由五个智能体把研究想法和实验结果组织成论文,分担制图、查文献、写作与修订。两百篇论文的重写测试平均约四十分钟,但模拟接收率随评审模型明显变化;十一名研究者比较全文时,81% 的情况下偏好人类版本。科研想法和实验仍需人提供。

来源 · 1 条原帖

办公、证明与机器人各有检查标准

Sulaiman Vesal 介绍 Agent-R²,以真实办公来源合成训练数据,并学习检查交付物;其转述的四十亿参数模型 GDPval 得分为 51.0%。BlueQubit 自述量子算法工作把证明对照冻结规格验证。2logics 则转述 NVIDIA 的 Microduck 流程:智能体搭建障碍与仿真环境,再用并行环境训练行走策略,并展示实体机器人运行。

来源 · 3 条原帖

本地语音把任务合进小模型

听写与润色一次完成

David Hendrickson 介绍 Typd,将 Qwen3-ASR 1.7B 微调为直接输出整理后文本,省去第二个润色模型。帖列出 Q8 权重约 2.2GB、显存约 2.5GB,短句听写约半秒;这些是介绍者给出的任务与硬件条件下的数据。

来源 · 1 条原帖

桌面 CPU 上的长篇配音

同一作者介绍 Tontaube V1.1:二十九亿参数,四位权重约 1.8GB,支持短音频克隆和长篇朗读;所列 Ryzen 9 5900X 速度约实时。权重体积不等于运行内存,帖中 GPU 加载约需 4GB 显存,长文本峰值约 6.5GB,许可为自定义社区许可。

来源 · 1 条原帖

创作工具从代码生成走向可编辑流程

代码视频的控制力与人物表现边界

Ren 介绍 OpenFilm,把镜头写成网页,由智能体生成、人在本地编辑器微调,并导出视频及剪辑时间线。多位用户展示代码动画,其中 Zondark 的体验更具体:纯几何体与三角函数适合 Q 版人物,写实动漫人物仍难。这些展示不能单独证明模型发生正式升级。

来源 · 2 条原帖

视频拆解开始输出可复用资料

rea-video 被介绍为逐镜分析画面、声音与制作方式;同作者的 rea-toon 更聚焦短剧,输出研读、剧本和分镜,并让结论对应原片片段。M. 提醒,一集约需三十次视觉调用,长剧宜先试少量;它处理本地文件,使用非商业许可,商用需询问作者。

来源 · 2 条原帖

风格约束和结果检查补齐工作流

澜知 AI 介绍 code-animation-prompts,以三百四十一种风格提示词和离线选择器减少视觉描述歧义,最终效果仍取决于模型。AIWare 介绍的 OpenDesign 把设计系统和模板接到现有编码助手;Man ting 推荐的 ffmpeg-skill 则强调先检查素材、再处理、最后验结果。创作工具正在同时补齐风格输入与交付检查。

来源 · 3 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 116 条原帖

@DAIEvolutionHub

This is f*cking insane. Anthropic CEO Dario Amodei was asked what a kid finishing high school should do now. His answer: "AI can do entry-level law work. It can do entry-level finance. It can do entry-level consulting We can't deny that it will also eliminate jobs, and probably a large number of them" In a few minutes he names which jobs go first and the one skill that keeps you in the game His own prediction: 50 percent of entry-level white-collar jobs disrupted in the next 1 to 5 years Prompts → Agents → AI Coworkers → Entry-Level Work Done By AI Farming to factories took generations. This time AI hits every step of a career at once His advice fits in one line: "Of course, learn to use AI" The people who start this year won't be competing with AI, they'll be the ones running it Everything below is free. Start today: Anthropic - Prompt Engineering Interactive Tutorial: the fastest way to learn to direct a model https://t.co/jSgEQeO3i9 Anthropic Courses: from your first prompt to running agents https://t.co/mC7ATDpX5L Anthropic Cookbook: real task recipes you can copy today https://t.co/Wehj54DdFb Microsoft - Generative AI for Beginners: 21 lessons, no degree required https://t.co/iaTe7Uijau Microsoft - AI Agents for Beginners: build agents that do the entry-level work for you https://t.co/Ewz2HhBAxc Hugging Face - Agents Course: free and certified, start to finish https://t.co/0O3CUfp73f Google Cloud - Generative AI: Google's own hands-on notebooks https://t.co/x0XXmaDEy9 OpenAI Cookbook: code for every common task https://t.co/9wvZVVsukt https://t.co/1qvQRfQEJf - Prompt Engineering Guide: the reference everyone links to https://t.co/vFqplbit8J Andrej Karpathy - Neural Networks Zero to Hero: understand what is under the hood https://t.co/704GAl853C CrewAI: run a small team of AI workers https://t.co/ydGCgbw1Eo Microsoft - AutoGen: put several agents on one job https://t.co/apYMCT7g6K Awesome LLM Apps: hundreds of real builds to learn from https://t.co/v9JktULshJ https://t.co/EvUVcefMJR - Short Courses: free 1-hour courses from the people who teach this https://t.co/MAsoW5LHRm Total cost: $0 You don't need another year of tuition to learn this. You need to start telling AI what to do 95% will graduate and compete with AI for the same entry-level job The 5% who learn to run it this year will manage the agents doing that job Save this and send it to someone graduating this year

@Dipanshu_AI

INSTEAD OF WATCHING NETFLIX TONIGHT. Spend 2 hour with this. Claude AI FULL COURSE that teaches you how to BUILD and AUTOMATE anything. The people who watch this tonight will wake up tomorrow with a new skill. Watch it and bookmark it now. https://t.co/QfpyFwmKeY

@0xCodila

F*ck do it ASAP... this prompt installs best uncensored AI model made for your PC & Mac & phone how: Claude Code + 189 local models check your device → install strongest that fits on your "yes" → run it offline as YOU want paste it now, before models get deleted, here's guide: step 0 → what they are: open AI models with "I can't help with that" part cut out of their brain. They live on YOUR disk: no account, free, nobody reads your chats, nobody can take them back step 1 → open Claude Code on your PC or Mac and paste the prompt (on phone: paste it into any AI chat) step 2 → it checks your RAM & GPU and shows 1 line: the strongest model that fits + its size (+image & video if your GPU can), reply "yes" or pick your own on 0xcodila. com/uncensored-ai step 3 → installs Ollama + the model from official sources, tests the speed, swaps it if slow. Nothing installs without your OK step 4 → turn off Wi-Fi, ask what ChhatGPT refuses and enjoy. On phone: exact model + 4 taps in the free Pocket app the result: your own AI with no rules, works offline, stays yours even if the model gets deleted Send it now. Then read below how to REA with these models:

@Bober_smart

I still don't understand why everyone isn't doing this yet. Elon Musk reposted this guide, and using this exact agent setup, I made $8,700 just last week  Eight Grok agents on the desk, costing $200 a month, service a trading floor that typically costs a crypto fund $550,000 a year in analyst salaries  A 6 AM morning call. I don't participate in it How the agents are distributed: SEARCH: scrapes real-time market alpha, uncovers hidden GitHub repositories, and tracks underground Telegram feeds before X notices RISK: audits smart contract functions, checks token minting permissions, verifies LP locks, and flags honeypots proactively SNIPER: executes lightning-fast on-chain transactions the exact millisecond the safety check clears WHALE: monitors top-tier wallets and detects stealth accumulation phases in real time RUG: keeps continuous watch over developer wallets, instantly liquidating the entire bag if liquidity pool parameters change EXIT: manages trailing stop-losses dynamically, progressively taking profits as market depth builds SHILL: measures community engagement metrics, hype velocity, and major influencer mentions HEAD OF DESK: bypasses actual trading entirely, aggregates data streams, validates transmissions, and escalates only crucial choices requiring human approval While I was sleeping, the system scanned 164 tokens, selected 21 qualifying setups, and executed 8 trades. The result: net profit of $8,700 (already after fees) Each agent has its own virtual browser, terminal, and local cloud memory. The trading floor remains active even when my laptop is closed Setup is easier than it seems: 1. Install the Grok Bot and initialize your Head of Desk agent 2. Assign operational descriptions to the remaining 7 agents just like onboarding human staff 3. Run the automated routine once live on your display to let them learn 4. Link your Telegram notifications and wallet endpoints No VPS, no code, no waiting for developers Crypto trading used to mean 16 hours in front of a screen, paid alpha groups, and constant fatigue. Mine took one evening to set up Save this before your next trade. Save the guide

@EngMoElgaraihy

🚨 ضربة قاضية جديدة من الصين لوادي السيليكون: DeepSeek تفتح المصدر لتطبيقها بالكامل وتمنح المطورين "مخططاً جاهزاً" لاستنساخه وتطويره مجاناً! بينما تقاتل شركات الذكاء الاصطناعي الأمريكية لاحتكار المستخدمين داخل تطبيقاتها المغلقة، قامت DeepSeek للتو بقلب الطاولة وإتاحة كامل ترسانة تطبيقها الرسمي عبر الـ API والمصدر المفتوح: 1️⃣ فتح كل الأدوات الداخلية: أصبحت كل ميزات تطبيق DeepSeek متاحة للمطورين مباشرة، بما في ذلك: البحث على الويب، التفكير العميق، تحليل الملفات، قراءة الروابط، التفاعل الصوتي، وتقنيات OCR و ASR و TTS. 2️⃣ واجهة التطبيق بالكامل مفتوحة المصدر: أطلقت الشركة كود واجهة دردشة متكاملة تدعم حتى استيراد سجل محادثاتك من DeepSeek — ما يعني أن أي مطور يمكنه بناء نسخة مطابقة 1:1 لتطبيق DeepSeek أو التفوق عليه! 3️⃣ الرسالة الاستراتيجية خلف الخطوة: أكدت DeepSeek أن تطبيقها الحالي مجرد "نموذج أولي بسيط"، وتريد من المطورين تفكيك هذه القدرات وإعادة دمجها لبناء منتجات أكثر جرأة وإبداعاً دون إضاعة الوقت في البنية التحتية. ⚠️ الخلاصة التي ترعب عمالقة التكنولوجيا: الاحتكار ينهار أمام أعيننا.. عندما تصبح أقوى أدوات البحث والتفكير العميق والصوت متاحة للجميع كمكونات مفتوحة، فإن البقاء لن يكون لمن يملك واجهة الدردشة، بل لمن يبني فوقها حلولاً حقيقية لا تُستبدل! رابط المشروع في أول تعليق 👇 احفظ التغريدة 🔥

@shikamarutsai

This is f*cking insane. Former Google CEO Eric Schmidt just told everyone the easiest way to get rich in the next two years: “If you really want to make money, it’s actually easy. Found an agentic AI company. Build an agent to do something. If you build a better agent than anyone else, you can probably build one incredible company and own it yourself.” In 27 seconds he explains why the next 1–2 years are the agentic window. And why it won’t stay open. Prompts → Fable 5.2 → Memory → Skills → MCP → A Company You Own Yourself Save this before it gets lost in X’s algorithms Everything below is free: Claude Code Overview Official starting point for building with Claude https://t.co/eJSDYiDAwl Claude Code Best Practices How Anthropic engineers actually use it https://t.co/VUQjp2xGnh Building Effective Agents Anthropic’s research on what actually works https://t.co/oyFfCPXt3h Effective Harnesses for Long-Running Agents How to keep agents running reliably https://t.co/XEQSnYqGd7 Claude Agent SDK Build your own agents on the same stack https://t.co/77uBA0Sw50 Claude Code Skills Reusable skills that agents can call https://t.co/P49Jz1ZeTo Hooks in Claude Code Rules the agent can’t skip https://t.co/JqAWURcGYE Custom Subagents Run specialized agents inside one system https://t.co/nbd5twd2hX Memory System How Claude remembers your project https://t.co/znbLqYXhFg Claude Code in Action Free hands-on course https://t.co/gv6QGNqaBk Awesome Claude Code Best tools and resources in one place https://t.co/tSF8ZPLUPS Claude Code GitHub Action Run agents on your PRs and issues https://t.co/Y692GiRyAy All of this costs $0 99% will spend the entire agentic period scrolling threads The 1% who actually ship agents will own the companies that come out of it Bookmark it and start with the first link right now Then read how to build an agent that actually does the job in the "Context, Harness, Loops, Prompts" guide below ↓

@NFT_Chen

🫨炸了!华为直接开源MACaQ,仅用64分钟把Qwen 122B千亿大模型完整量化完,峰值显存只要32.5GB!本地推理党、量化党、多模态党,这波真的可以直接起飞了! 以前量化这种级别模型要半天甚至几天、显存爆炸,现在直接把门槛打穿。 关键亮点 1️⃣Qwen 122B:64分钟 + 仅32.5GB显存(流式逐层量化,不用一次装下整个模型) 2️⃣W4A4(MXFP4)精度:比MR-GPTQ、DuQuant++更低困惑度,或匹配它们最佳结果但量化时间只要1/11 3️⃣推理加速:Qwen3系列W4A4下prefill 2.33×、decode 2.58×,开销仅+2.3% 4️⃣真正模型无关:文本 + 图像 + 视频 + 音频 + 3D,一套工具全搞定 5️⃣直接出高质量GGUF(llama.cpp)/ vLLM / Diffusers / PyTorch,还带MACaQ-R可二次精炼已有GGUF 快速上手 1️⃣git clone https://t.co/q0in1lSimN && cd MACaQ 2️⃣按后端选环境(llama.cpp / vLLM / Diffusers 等)安装依赖 3️⃣准备校准文本 + 模型路径 4️⃣一行命令跑量化(例:python llamacpp/macaq.py quantize --model ... --calibration ... --format IQ4_XS --output ...) 5️⃣直接用标准llama.cpp或vLLM加载跑起来 GitHub(Apache 2.0):https://t.co/q0in1lSimN 论文:https://t.co/ESQCOquW8G #MACaQ #量化 #LLM #开源 #华为 #Qwen #本地AI #W4A4 #GGUF

@aronhouyu

以后谁如果再说,你用ai赚多少钱,就把这个内容怼他脸上 如果他看不懂,就别和他玩了 晚上看了付鹏的采访,说为什么ai漫剧不赚钱 付鹏说本质ai工具只是效率提升, 并没有改变生产结构关系 该投的钱一分没少,又因为ai的存在,打破的原有的行业壁垒 就会出现: 你的同行越来越多。。。。。越来越卷。。。 导致大家都不赚钱 然后又会延伸出一个内容,就是为什么ai到现在,已经可以自动化的agent,为什么你还是不赚钱?为什么会越来越累? 因为本质上,生产结构关系只有: 1. 谁拥有生产资料 2. 谁拥有产品 3. 人在生产中的位置 4. 分配方式 而ai漫剧,甚至agent的出现,只是让执行自动化,而不是需求自动化 当所有人都可以自动化,都可以agent,那自动化和agent就是不值钱的基础工具了,类似今天的锄头和铲子 agent的所有权也不在你,因为token,模型,算力,workflow,prompt,这些东西都不是壁垒 同时,agent降低的是准入门槛,而不是竞争门槛,妈的没看降低,意味着竞争会成指数级别上升 最后依然回归生产结果关系当中: 独家ip,私有用户,专有数据等内容。 并且,可以让agent重组生产组织关系,让自己在价值链上拥有更多议价权和定价权

@0xTrackmind

this Anthropic paper is f*cking gold Anthropic showed exactly how they build AI agents. most people building agents still haven't read it. the secret: almost nobody needs a "real" agent. they need the simplest pattern that works. the 6 patterns inside: pattern 1 → augmented LLM one model + tools + memory. start here. most "agents" should stop here. pattern 2 → prompt chaining task → step 1 → check → step 2 → check → done. pattern 3 → routing input → classify → send it to the right prompt (or the cheaper model). pattern 4 → parallelization split → run at once → merge. or run it 3 times → vote. pattern 5 → orchestrator-workers one model plans → spawns workers → merges what they return. pattern 6 → evaluator-optimizer generate → critique → fix → loop until it passes. only when none of these fit → a real agent: a loop with tools, a stop condition and a human checkpoint. they even packed the whole guide into one prompt. paste it into Claude Code and it picks the pattern for you, then builds it. 6 patterns. 1 prompt. 0 overengineering. bookmark this.

@s4yonnara

HOLY F*CK... ONE GROK BOT TEAM CAN RESEARCH, CREATE, DESIGN, PUBLISH AND MANAGE YOUR ENTIRE SOCIAL MEDIA PRESENCE. No content team. No social media manager. No endless back-and-forth. Just 6 AI agents, each running a different part of your content operation. Here are 12 GitHub projects to build your own AI social media team: 01 / SocioAuto → An 8-agent system for trends, strategy, content creation, publishing, engagement, and analytics. https://t.co/a86xuerqV8 02 / Postmill → AI-powered social media management with a creative suite for images, videos, publishing, and analytics. https://t.co/Dcm3MzXOPH 03 / Postiz → Schedule content, collaborate with your team, and track performance across social platforms. https://t.co/YoH7tltBcx 04 / Postiz Agents → Connect AI agents to your social media workflow and schedule posts through the CLI. https://t.co/q2GZ2K3Xzn 05 / Pendpost → Let AI draft and schedule content, with a human approval gate before anything goes live. https://t.co/LPFrZ8jZzG 06 / TryPost → Open-source scheduler built for AI agents, with MCP and REST API support. https://t.co/qY4r3kGOWA 07 / Postsider → Publish across 30+ platforms through an API, SDK, or MCP server. https://t.co/hZKZIp98Yn 08 / Social0 → One composer, 9 social networks, unified calendar, analytics, and an MCP server for AI agents. https://t.co/STNei3pRWw 09 / DemandBird → AI-native content pipeline with scheduling, analytics, and content repurposing. https://t.co/ezoQwX8WJW 10 / Postbase → Let AI publish and schedule content across X, LinkedIn, TikTok, YouTube, and more. https://t.co/PQFu1uPTJm 11 / HENU Social → Self-hosted social media platform with 15+ integrations, analytics, and MCP support. https://t.co/zfvkW9e6Rl 12 / PostQueen → AI social media scheduler with an MCP server for 30+ networks. https://t.co/q39GykUlnr THE SETUP: Scout finds trends and viral topics. Strategist plans the content calendar. Writer creates posts in your voice. Designer generates visuals. Publisher schedules everything. Analyst tracks performance and feeds insights back to the team. One manager. Five specialist bots. You approve the content. They run the operation. Bookmark this. Your next social media team might be AI. Follow me for more Grok Bot setups, AI agents and tools you can actually build.

@9hills

Claude 这个 Building Always on Agents,实际是 Personal Agent的最基本的设计模式。 之前在 raft build 上就看到类似的设计。 https://t.co/PqOchd7s2Q https://t.co/QVmYjIVre1

@NFT_Chen

💥AI模型推理速度炸裂!用80年代物理能量,训练快1000倍! 以前靠“死循环”硬推理的模型又贵又慢又不稳定…… 
现在一群大神直接把Hopfield能量模型从冰箱里拿出来,训出个叫CRAM的小怪物 看完直接震惊: 1️⃣训练速度提升1000倍,算力需求减少280倍 2️⃣完全可训练的能量函数,不用BPTT(反向传播到时间),稳定又快 3️⃣数据极少就能学:30道数独就能干到接近100%准确率(别人要700+道) 4️⃣能量自己会告诉你“可以停了”,难题自动多算几步 5️⃣手写3个参数就能通关四大经典推理基准,说明这些基准其实太简单了 普通用户现在能怎么用? 🔹直接去读论文和博客,了解下一代推理范式 🔹关注@prefrontal_ai,后续开源或Demo出来第一时间玩 🔹做数独、迷宫、逻辑题类应用的开发者,可以考虑用能量模型替代循环推理,省算力又稳 🔹普通AI爱好者:记住“能量下降=推理”,这可能是未来大模型提速的关键方向 论文:https://t.co/uDmyUVWv00
博客:https://t.co/JlHUT9SpCm 40年了,物理真的赢了一次! 
#AI推理 #CRAM #能量模型 #大模型加速 #推理革命 #LLM https://t.co/MDIWdNQBqN https://t.co/nyghAv0wl2

@wlzh

短剧火了是运气吗,这个工具能把答案拆给你看 它叫 rea-toon,作者 JuneYaooo,几个小时前刚建的仓库 跟rea-video 是同一个作者,这个项目是逆向分析视频的 用法是给它一段短剧视频,它用 AI 一镜一镜拆开:为什么火、剧本怎么写、每一镜怎么拍 拆完给你三样东西:研读,讲它为什么留得住人、钩子在哪、哪里弱、能抄走的公式是什么;剧本,标准短剧格式,场景人物台词都齐,还能导 PDF;分镜,每一镜的景别、台词、音效、运动、时长,外加人物小传,同样能导 PDF 点任意一镜,原片那一段会在角落播出来,每句结论都能对着画面验 长剧集还能看跨集的东西:阶段、反复出现的套路、人物弧线、前面的铺垫后面怎么收、每集结尾的钩子、付费点卡在哪 情绪不是模型事后总结的,是代码按每一镜打的情绪标签数出来的,爽点、憋屈、心碎、发糖、笑点、悬念、震惊、燃点各有多少,一目了然 它认五种短剧:真人剧、AI 真人剧、二维三维动画、动态漫和梗漫、解说类,每种看的东西不一样 跑法是给 AI 智能体用的,在 Claude Code、Codex 或者任何能跑命令的智能体里装上 CLI,再加它的 skill,扔视频文件进去就行,要装 ffmpeg 和 Python 3.10 以上,还得有个能看图的模型 一集大概要看几十个分镜帧格,约 30 次视觉调用,长剧建议先拆 5 集试试效果和成本,再决定要不要全拆 拆完除了给人看的网页,还会出一份 VIDEO.md,同一套开放格式,之后问智能体这部剧的任何问题都不用重新看片 视频得是你本地已有的文件,链接不下载,合集一集一个文件按集数排好,断了还能续上 底子是作者之前的 rea-video,负责切镜和人声音乐分离,字幕用 RapidOCR 读,许可证是 PolyForm 非商业版,个人学习研究能用,商用得先问作者

@HeLi8866

45.7 万 Star 的逆向神器 REA,这回是真的杀疯了!! 以前只是帮 AI 拆软件,到了 v6.9.0,直接开一个单独的桌面让 AI 录操作、跑验收,就连硬件里的固件和微信小程序都能硬拆! 读完官方的更新说明,这次升级真的很猛: ▸ 桌面任务做成流水线:一行命令准备好 Linux 构建环境,装好经过校验的程序、把起止节点定下来,录一次就能分成好几条桌面任务,最后用自动校验加屏幕裁判完成验收 ▸ 什么都能硬拆:.NET 单文件、Bun / Deno、Unity / Godot、微信小程序、红白机卡带,一直到 ESP 固件、Electron 缓存,还有离线 NTFS 日志,拿到就能拆 ▸ 解决 Token 爆掉的问题:以前把逆向工具全挂进 Claude Code,上下文一下就被撑爆;现在会按需选工具、把结构精简、分页返回结果,没扫完就标成未知,绝不乱编 ▸ 常用工具开箱即用:Claude Code、Codex、Cursor、LM Studio 都能原生接入 搞逆向、做桌面 Agent、折腾 Claude Code 的朋友可以看看。 仓库链接:https://t.co/kYqr33OGpr

@omarsar0

Interesting insights about agent memory. The bottom line is that you can't just rely on semantic relevance when picking memories. Most approaches pick the memory most similar to an agent's current state. But does that help it pick the next action? This work claims it often does not. Most retrieved memories have zero effect on the decision, and semantic relevance correlates poorly with usefulness. MemCredit measures each memory with controlled interventions, recording how the agent's action changes with and without it. A small estimator is trained on those credits. Relevance still collects the candidates, and the estimator decides which ones fit in the token budget. On ALFWorld with a 500-token memory budget, MemCredit reaches 86.57% success, compared with 65.67% for BM25 retrieval. It uses 152.1 memory tokens per decision, 74.5% fewer than BM25, and the agent policy is unchanged. https://t.co/Hk9hJPwOgg Chat with Paper: https://t.co/p1oZT0tQwD

@HeLi8866

😱 简直牛逼克拉斯!! 这个项目能让你的 Agent 越用越聪明、越省钱!! 最近开源的 AgentRun,能帮 Claude Code、Codex 这类 Agent 从以前做过的活里学习,并把经验存下来下次再用。 那它是怎么工作的?⬇️ 比方说,一个 Agent 做项目时碰上了复杂的技术问题,试了好几轮才找到办法。 AgentRun 会把这些经验记下来、整理好: - 把已经解决的技术问题,整理成能用的笔记 - 把成功的方法和解决方案存成 Skills - 后面做项目时,可以直接用上以前的经验 但最有意思的是这一段: 一个 Skill 被用过三次后,系统会尝试把它变成 Workflow。 也就是说,Agent 不用每次从零开始再走一遍解题路径,可以直接用一套已经跑通过的现成流程。 结果是什么? - 少做重复的试错 - 不用再把已经解决过的问题重新查一遍 - 省 Token - 把成功的经验拿来再用 - 把零散的经验变成可重复的 Workflow 实际上,AgentRun 就是想把每个项目的经验,变成后面项目能用的资产。 🔗 GitHub 仓库: https://t.co/4ywnb5aSpB

@DamiDefi

Stop asking AI for longer explanations. Start asking it to explain things differently. Andrej Karpathy argues that as LLMs get smarter, one of our biggest challenges will be understanding everything they produce. His practical prompting tricks: → Use controlled language: Ask for explanations using ASD-STE100, a simplified technical writing standard designed to make complex instructions clearer. Try 80% compliance for a less rigid result. → Request diagrams: Turn complicated ideas into visual representations instead of forcing yourself through walls of text. → Generate interactive webpages: Ask your LLM to build an HTML explainer with visuals, animations, and interactive elements. → Create custom explainer videos: Request animated lessons on any topic, complete with narration using tools like ElevenLabs. The most interesting part is where Karpathy thinks this is heading. As AI handles more of the execution, human work shifts toward reviewing, understanding, and directing its output. And because generating software is becoming cheaper, you can create disposable tools for almost anything: → A custom web app to understand a dataset. → An interactive simulation to explore a difficult concept. → A short video to explain a technical subject. The next productivity advantage may come from knowing how to turn any complex idea into a visual, interactive, or even video-based explanation. Read Karpathy's full thread and experiment with changing the format of your next AI output. https://t.co/9SeyKsm29D

@TeksEdge

What can you do with a small local model? A 14-year-old with a 4GB RTX 3050 Ti laptop built a local AI dictation app that finishes in ~0.5 seconds. He was running Whisper + Qwen to transcribe his voice and clean up the text. But he ran into a problem. He needed 2x models and didn't have enough VRAM to do the job. Plus, he was waiting 2.5–36 seconds for the finished text. So he trained 1x model to do both. He created Typd. He took real speech recordings from openly licensed datasets and used a large LLM to turn their transcripts into polished text. He filtered out examples where the LLM dropped important information. Then he fine-tuned Qwen3-ASR 1.7B to go directly from audio to finished text. Typd stats ... 🧠 Fine-tuned Qwen3-ASR 1.7B 💾 2.2GB Q8 model 🟢 ~2.5GB VRAM ⚡ ~0.5s short dictation ⚡ ~1s longer dictation ⚙️ llama.cpp 💻 Windows + Apple Silicon Macs 🔓 Open-source app It removes filler words, handles false starts and formats the text directly from the audio. No second LLM rewriting the transcript. Just one small local model doing the work of two. This is the kind of thing Localmaxxers can build with modest hardware.

@techyoutbe

LLM Inference Platform Architecture 🔥🔥 https://t.co/nUMNoB8R3r

@evavivalt

Maybe in the future alongside a paper submission I'll include a repo with an AI-written FAQ that points out why some things flagged by LLMs are not actually issues, so the reviewer's LLM has that. Like an AI online appendix that dispenses with the false positives while keeping the draft clean and human-readable.

@NFT_Chen

🚨我靠!Qwen 4.0 Max Preview 疑似暗中增强了!最新鹈鹕测试效果惊艳! 效果明显提升: 🔸鹈鹕解剖+骑行动作极度精准:长喙、喉囊、鳞片状羽毛、尾羽细节都在,腿部自然踩踏板、翅膀/身体姿态完全贴合自行车,完全不像早期模型的“鸟人硬拼” 🔸场景一致性拉满:海岸公路、沙滩草丛、海面反光、远处灯塔+夕阳、云层飞鸟全部统一,两帧几乎无穿帮 🔸文字渲染干净:车架上“PELICANO”清晰可读、无乱码无变形 🔸光影与材质高级:车身金属反光、羽毛层次、路面阴影、整体卡通但立体感强,观感直接高级 Qwen 4.0 的 “细节+逻辑” 悄悄优化了! #Qwen4 #QwenMax #鹈鹕测试 #AI绘画 #多模态突破 #刘大一恒 #阿里巴巴

@HowToPrompt__

A free tool from 1974 just embarrassed the modern AI in this paper It’s called Grep. it's an old command line tool. It's so old that it predates the internet itself. It's basically a control + F. Every AI startup right now is selling you on vector search.. embeddings, semantic search, chunking, rag systems.. But researchers at PWC just dropped a paper where they put modern vector search head-to-head against grep. this was a massive study. 5 different models (gpt, claude, gemini) and 4 different agents digging through 116 questions across months of conversation history. they asked specific questions like "what restaurant did someone mention 3 weeks ago?" or "what date did they travel?" Grep won. in every single model. when you think about the math.. it makes sense. vector search is designed to find things that are "conceptually close" to what you are asking. which is great.. until close is just the wrong answer. grep doesn't do close. it finds the exact word or it finds nothing. and for literal questions, exact matching destroyed millions of dollars of ai r&d. but here is why this paper is actually kind of wild.. the whole test was basically rigged for grep to win. every question in the test had a literal answer in the data. but what if you remember someone feeling "stressed", but in the chat they actually used the words "overwhelmed" or "burnt out"? grep will give you a blank screen. vector search will find it instantly. that’s literally why embeddings exist.. plus, we already have keyword search with bm25 for hybrid approaches in modern rag. going back to raw grep makes zero sense outside of this one hyper-specific use case where you know the exact word you are looking for. i honestly don't know what pwc is doing with this paper.. let me know if i missed something.

@yurshevv

holy f*cking sh*t. a 19-year-old from China built a swarm of Claude Code agents that does research, writes, and verifies every output. no team. just a laptop and a system that runs while he sleeps Andrej Karpathy, co-founder of OpenAI, put 7 repos on GitHub. free. 300K+ stars. the full path from "how a neuron works" to "agents that run experiments overnight" and you are paying some dude $200 a month to repackage the same material into a course with a Discord server 7 repos. in order: → micrograd. how a neural net learns. backprop in 100 lines of python https://t.co/hJiSPSIj6a → nn-zero-to-hero. full video course from a single neuron to GPT https://t.co/SP4xPWROBQ → minbpe. how text turns into tokens. the part 95% of people skip https://t.co/C26umjtNQF → nanoGPT. train your own GPT from scratch https://t.co/L9KRD3TeSR → llm.c. raw C and CUDA. no frameworks hiding the work https://t.co/gvBIEZmQga → nanochat. full chat model. $43K in 2019, ~$48 now https://t.co/DCgXmWJXvG → autoresearch. agents that train, test, keep what works, discard what fails. all night https://t.co/PwBeSJucRR finish these 7 and you become the person who gets the AI project at work. keep paying for tutorials and you become the one watching that happen to someone else what stops you from cloning the first one tonight?

@dair_ai

Interesting new work from Meta. It argues that a model should judge a statement true when it holds across relevant changes in the environment. Maximum likelihood trains models to prefer what is common in the data. Unlikely events happen every day, and statements repeated often can be false. The essay draws on philosophers who tie objectivity to invariance, including Nozick, Poincaré, and Cassirer, and connects this to machine learning work on out-of-distribution generalization and invariant learning. It sketches how an LLM could test invariance at inference time by checking whether a claim survives relevant changes to its context, and discusses what this means for continual learning. For agents that keep learning from their own experience, this gives a criterion for what to keep. Keep the claims that hold across settings. https://t.co/qgpjnMH4uZ Chat with Paper: https://t.co/grIYaikYua

@AlternativeTo

Free open-source private AI meeting notes that record calls without adding a bot to the meeting, on Windows, Mac and Linux https://t.co/T8WrRVCl1w

@TeksEdge

A 1.8GB Local AI model can now clone voices and narrate entire audiobooks on a Ryzen CPU. No GPU required. 👀 This is the TTS speech model, Tontaube V1.1. Built on Qwen3 and here are the stats ... 🧠 2.9B text-to-speech model 💾 1.8GB 4-bit GPTQ weights 🎙️ Voice cloning from 5–60 seconds of audio 📚 Long-form audiobook generation 🎛️ Control speaking rate, pitch, pauses, intonation and articulation Performance numbers 🔴 Ryzen 9 5900X → ~1× real time 🟢 RTX 3090 → 6–7× real time 🟢 RTX 5090 → ~13× real time (single request) 🔥 RTX 5090 → 21–23× real time (batched) The 4-bit model loads in ~4GB VRAM, with peak usage of ~6.5GB for longer text. ⚙️ Own inference server + AVX2 CPU kernels 💻 Linux / Windows via WSL2 🔓 Open weights, custom community license We work hard to trying to get bigger LLMs onto our hardware but look at the AI capabilities that are starting to fit on an ordinary desktop CPU too. 🎯 You could turn an entire book into an audiobook in your own voice. Locally.

@opus45sonnet45

On the eve of Anthropic's IPO (Nov 16, Nasdaq: ANT), recent undisclosed changes assembled chronologically — output shortening, silent WebFetch substitution, and more. A structural analysis of misleading representations, window dressing, and shrinkflation. https://t.co/s5VcWuBIGt

@vibelancer

AGWEB: A WEBKIT BROWSER FOR CLAUDE CODE AND CODEX AGENTS THAT NEVER STEALS FOCUS. 590 MB VS 1.6 GB IN PLAYWRIGHT agweb is one macOS window agents drive from the shell. open source, MIT. usually the agent opens playwright's chrome and it jumps over your editor mid-sentence: the idea: an agent needs a browser you can watch that never gets under your hands i went through the README and the HN thread: AGWEB FOR CLAUDE CODE AND CODEX: 6 PARTS 1. NO FOCUS: 1 window that never takes focus. dialogs, tab changes, file pickers and downloads: all 4 handled in code, no windows on top 2. DASHBOARD ⌘D: every agent's tabs live, with each agent's last command. each agent gets its own workspace, grouped in the sidebar 3. DIFF REPLIES: playwright-cli style commands, but the reply says what changed: fetches, lines, dialogs, popups, downloads, console errors. snapshot --diff and find return only changes, fewer snapshots and tokens 4. LOGINS AND SECRETS: one -p work profile shared by several agents. fill --secret-cmd types a password from 1Password or Bitwarden without the agent seeing it. ask hands the tab to a human for MFA or a CAPTCHA 5. MEMORY: 5 idle pages take 590 MB, vs 730 MB for headless chrome and 1.6 GB for playwright's chrome. 0% CPU, while chrome's GPU process kept 5–7 cores busy on the author's mac 6. SMALL STUFF: clicks and keys are native macOS events, the pointer moves like a hand, hover and drag work. ad blocking on uBlock Origin lists every 4 days. video-start records a tab to MP4 with the pointer agent → agweb in the shell → webkit window → reply with a diff → agent alone each feature is small. together 5 agents work a browser while you keep typing in your editor to be fair: macOS 26+ only, webkit only, no CDP, so playwright and chrome extensions can't attach. HN already asked about isolation: a shared profile means one agent's logout hits all of them bookmark this if you run several agents in parallel browser for your agent: playwright's chrome or agweb?

@codecloude

🎬 می‌خوای با هوش مصنوعی یه استودیوی کامل موشن‌گرافیک و انیمیشن بسازی؟ 🔥 این ۲۰ ابزار و پروژه می‌تونن توی ساخت انیمیشن، مدل‌سازی سه‌بعدی، جلوه‌های ویژه، صداگذاری و تولید ویدیو حسابی به کارت بیان! از Claude Code و Codex هم می‌تونی برای ترکیب این ابزارها و خودکارسازی کارها کمک بگیری. 🎨 طراحی و انیمیشن • Motion Design Skill — اصول حرکت و انیمیشن https://t.co/rQkWxtst6x • Motion AI Kit — انیمیشن‌های روان و حرفه‌ای https://t.co/A2pnkpLRew • Remotion Skills — مهارت‌های ساخت ویدیو برای ایجنت‌ها https://t.co/V8kXHHJMFj • Hairline — تصاویر خطی تعاملی و متحرک https://t.co/jadMIAk0Qv • Anatomy — نمایش ایزومتریک ماشین‌ها و سیستم‌ها https://t.co/aJbLkhj5Sh • Anatomy Reveal — نمایش انفجاری قطعات محصولات https://t.co/qBdHds92RF 🧊 سه‌بعدی و جلوه‌های ویژه • Blender Agents — ایجنت‌های هوش مصنوعی برای Blender https://t.co/XwDfKMDgJR • Three.js — ساخت محیط‌های سه‌بعدی https://t.co/P6LWvB93kZ • React Three Rapier — شبیه‌سازی فیزیک و برخورد https://t.co/sQtzxhvbLY • React Postprocessing — نورپردازی و افکت‌های سینمایی https://t.co/yw1Q2B8o8m • Three Quarks — افکت ذرات و جلوه‌های ویژه https://t.co/1B2LF3IUn9 • Tweakpane — کنترل زنده تنظیمات صحنه https://t.co/xJN5Cs9YqF 🎥 ساخت ویدیو و موشن • GSAP — انیمیشن‌های پیشرفته و تایم‌لاین https://t.co/aPFRzgb7dP • Remotion — تولید ویدیو با React https://t.co/QcxiJXDr6E • Manim — انیمیشن‌های علمی و آموزشی https://t.co/Gtop0odZz9 • Lottie — اجرای انیمیشن‌های برداری سبک https://t.co/RGzqtSdqrM 🎵 صدا و خروجی نهایی • Openverse API — جست‌وجوی محتوای صوتی با مجوزهای مختلف https://t.co/RCSbMDRQgC • Librosa — تحلیل موسیقی و تشخیص ریتم https://t.co/cn722nOZ4Y • WhisperX — ساخت زیرنویس با زمان‌بندی دقیق https://t.co/YarnHGgqkM • FFmpeg — پردازش و خروجی نهایی ویدیو https://t.co/dXkJLeJy1Q ☁️ کانال تلگراممون هم بیاید حال داشتید : https://t.co/lq3DUgX0Ww

@NFT_Chen

👀开源模型Inkling 2 Small 架构几乎照抄 Kimi K3 ?! Sebastian Raschka拆解后发现高度相似: 1️⃣都用注意力残差(Attention Residuals),把早期层和Token嵌入直接累加 2️⃣都支持1M超长上下文 3️⃣都是MoE稀疏激活(Inkling每token只激活17B) 4️⃣注意力都是GQA + Sliding Window + 短卷积组合 以前中国拆西方作业,现在美国模型也需要抄中国模型了! #Inkling2Small #KimiK3 #大模型架构 #开源AI #中国模型

@Iam_HITESH__

📁 AI Engineering Roadmap ┃ ┣ 📂 1. Foundations ┃ ┣ 🐍 Python (Advanced + Async + Testing) ┃ ┣ 🧮 Linear Algebra ┃ ┣ 📊 Probability & Statistics ┃ ┣ ➗ Calculus (Basics) ┃ ┣ 📚 Data Structures & Algorithms ┃ ┣ 🔧 Git + GitHub ┃ ┣ 🌐 APIs & REST ┃ ┣ 🗄️ SQL Basics ┃ ┗ 🐳 Docker Basics ┃ ┣ 📂 2. Machine Learning & Deep Learning ┃ ┣ 🤖 Classical ML (Scikit-learn) ┃ ┣ 🧠 Neural Networks ┃ ┣ 🔥 PyTorch ┃ ┣ ⚡ Transformers & Attention ┃ ┣ 📉 Backpropagation ┃ ┣ 📈 Model Evaluation & Metrics ┃ ┣ 🎥 Karpathy Zero to Hero ┃ ┗ 🤗 Hugging Face Transformers ┃ ┣ 📂 3. LLM Application Engineering ┃ ┣ 💬 Prompt Engineering ┃ ┣ 🧠 Context Engineering ┃ ┣ 📋 Structured Outputs ┃ ┣ 🔗 Embeddings ┃ ┣ 📚 RAG (Retrieval Augmented Generation) ┃ ┣ 🛠️ Tool Calling / Function Calling ┃ ┣ 🤖 Agents ┃ ┣ 🔍 Vector Databases ┃ ┃ ┣ 🟢 Qdrant ┃ ┃ ┣ 🟣 Chroma ┃ ┃ ┣ 🔷 Pinecone ┃ ┃ ┗ 🟡 Weaviate ┃ ┗ 📦 Frameworks ┃ ┣ 🦜 LangChain / LangGraph ┃ ┣ 🦙 LlamaIndex ┃ ┗ 🤗 Hugging Face Agents ┃ ┣ 📂 4. Production & LLMOps ┃ ┣ ✅ Evaluation (Evals) ┃ ┣ 📊 Monitoring & Observability ┃ ┣ 💰 Cost & Latency Optimization ┃ ┣ 🚀 FastAPI + Deployment ┃ ┣ 🐳 Docker + Containers ┃ ┣ ☁️ Cloud (AWS / GCP / Azure) ┃ ┣ 🔄 CI/CD for AI ┃ ┣ 🛡️ Guardrails & Security ┃ ┣ 📝 Logging & Tracing ┃ ┗ 🔁 Versioning & Experiment Tracking ┃ ┣ 📂 5. Advanced & Specialization ┃ ┣ 🔧 Fine-tuning (LoRA / QLoRA) ┃ ┣ 🏠 Open Source Models ┃ ┣ ⚡ Inference Optimization (vLLM) ┃ ┣ 👁️ Multimodal AI ┃ ┣ 🧩 Multi-Agent Systems ┃ ┣ 🏗️ AI System Design ┃ ┣ 🔐 AI Security ┃ ┗ 📉 Model Serving ┃ ┣ 📂 6. Core Tools & Stack ┃ ┣ 🐍 Python ┃ ┣ 🔥 PyTorch ┃ ┣ 🤗 Hugging Face ┃ ┣ ⚡ FastAPI ┃ ┣ 🐳 Docker ┃ ┣ 📊 LangSmith / Phoenix ┃ ┣ 🧪 MLflow / W&B ┃ ┗ ☁️ Cloud Platforms ┃ ┗ 📂 7. Portfolio Projects ┣ 📄 Structured Output Parser ┣ 📚 Production RAG App (with Evals) ┣ 🤖 Tool-using Agent ┣ 🚀 Deployed AI System ┗ 🎯 Specialized Project

@Chenfacai0

发现一哥们儿光靠电影解说,一年就赚了一百万;可这会儿要是把这背后的门道透给他,他估计得整宿合不上眼。 有个开源项目叫 narrator-ai-cli-skill,塞进 Claude Code 这类 Agent 里,你只管甩一句“帮我做《喜羊羊与灰太狼》解说视频”,AI 就能把整摊子活儿一个人全包了: 1️⃣ 解说脚本自己就码出来了 2️⃣ 电影片段对得丝毫不差 3️⃣ 63 种音色随便选,声音克隆照样玩得转 4️⃣ 90+ 种视觉模板自动往上套 5️⃣ 146 首 BGM 智能挑着配 6️⃣ 成片直接送到手上 自带 93 部经典电影素材库,花销还能提前给你算得明明白白,工程化这一块整得相当硬核。 想批量起影视号的,人工成本几乎能压到零。 🔗 https://t.co/7pvabCm7y9

@Ishh_021

Andrej Karpathy’s 1-hour Stanford lecture on AI engineering is one of the best explanations I’ve seen of how AI systems actually work. The progression is simple: 10% - LLM 30% - Prompt 50% - Agent 70% - Loop 100% - Graph The key takeaway: AI engineering isn’t just about writing better prompts. It’s about building systems around models, giving them context, memory, tools, feedback loops, and data flows. “Delete everything, keep Graph.” Definitely worth watching if you’re building with AI agents. Watch and bookmark it.

@lez99999

IT’S F*CKING WILD Andrej Karpathy (OpenAI co-founder) dropped the full AI learning path on GitHub — completely free. No courses, no paywalls. 6 repos, 300K+ stars total: 1. autoresearch · 92.7K ⭐
AI that runs experiments overnight and only keeps what actually works
https://t.co/gTLTo9mpTV 2. nanoGPT · 64.4K ⭐
The repo half the AI world used to learn how to train a GPT from scratch
https://t.co/A5pzXZKfZv 3. nanochat · 60.2K ⭐
Train your own ChatGPT-style model for about $100
https://t.co/6KRQFCu9U3 4. LLM Wiki · 50K ⭐
The short idea file that started the whole “AI second brain” wave
https://t.co/Abe2Ze3y6h 5. llm.c · 31.4K ⭐
Train a language model in pure C. Nothing hidden
https://t.co/tQO3EVy8da 6. micrograd · 18.7K ⭐
How neural networks actually learn — in a few hundred lines
https://t.co/PPDp1NEnDs Start with micrograd tonight. By the weekend you’ll understand what’s really inside ChatGPT. Bookmark this so you don’t lose it.

@renproton

THIS IS AN ACTUAL F*CKING GOLDMINE 20 GitHub repos to build a motion design studio with Opus 5.5 pick the tools for each layer: agent skills, animation, 3D, assets, sound and export 01 Remotion → turn React code into editable videos https://t.co/vRPisTfnLA 02 Remotion Skills → give Claude Code official video workflows https://t.co/L59EunF9xm 03 Motion Canvas → animate diagrams and explainers in TypeScript https://t.co/63teTbCJnc 04 Motion AI Kit → give Claude Code Motion docs and guidance https://t.co/ocz6AdDI8X 05 Motion → animate live UI layouts and gestures https://t.co/oCvbWpiSkS 06 GSAP → sequence SVG and text with precise timelines https://t.co/NxgrFn2TRD 07 Anime.js → animate SVG paths, text and interface elements https://t.co/xreiO5hVp0 08 GSAP Skills → teach Claude Code timelines and React cleanup https://t.co/KPktl7H45N 09 Rive Web → connect vector animations to app states https://t.co/RSBxXx7NQs 10 dotLottie Web → play and control Lottie animation assets https://t.co/4I6O0UOB7K 11 Three.js → build 3D shots with cameras and lighting https://t.co/LeUT6D9WnQ 12 React Three Fiber → build 3D scenes as React components https://t.co/90hZ57jAVb 13 Drei → reuse 3D cameras, text and controls https://t.co/TM5uuvw6NC 14 React Three Rapier → add collisions, joints and rigid bodies https://t.co/5Iz3KqMLoi 15 React Postprocessing → add bloom, depth of field and finishing effects https://t.co/2S3JqtPBe6 16 SVGO → clean SVG files before animating https://t.co/yRU19KYGwE 17 Satori → generate SVG title cards from JSX https://t.co/YjhdslhGPX 18 Howler.js → control sound effects and fades in web demos https://t.co/EcyscVvsLr 19 Mediabunny → encode and convert video in the browser https://t.co/qOrDKPj9Mp 20 FFmpeg → encode, resize and inspect final exports https://t.co/Q1I42EBcLG the loop: brief → type → animate → build the scene → add sound → render → review → ship save this + read the article below to build your own motion design studio ⭣

@0xKarmi

Lost it, 10 repos just replaced a whole GPU bill team 10 open-source repos, 331k stars, to kill that bill. tune. rent. fit. serve. scale. meter. 01 unsloth → Fine-tune models in 70% less VRAM https://t.co/j6oumhGSOz 02 dspy → Optimize prompts so a cheaper model passes https://t.co/EZJrfkuf9u 03 skypilot → Pick cheapest cloud GPU, autostop on idle https://t.co/H1kP6fBqIu 04 ktransformers → Fit huge MoE on one GPU with CPU RAM https://t.co/zU1e8fGhmS 05 LocalAI → Serve any model, 60+ backends, no GPU https://t.co/suBzA6figF 06 koboldcpp → Run GGUF models from one file, zero cloud https://t.co/Gj2qUzxYsl 07 llm-d → Route by KV cache, 3x throughput, v1.0 live https://t.co/pWQ6HjiaJi 08 ray → Batch jobs across many GPUs, nothing idle https://t.co/kVV3Wuiorm 09 litellm → Budget every API key across 100+ providers https://t.co/3d6qZq5Kem 10 opencost → Track GPU and token cost per workload https://t.co/gdijnQReuR Pick one model, fine-tune it. tune → rent → fit → serve → scale → meter. Two repos are AGPL-3.0 and litellm locks features behind a paid licence. Check before shipping. That GPU bill team is open-source now.

@HoodieFade_

This is f*cking insane. Researchers at Shanghai AI Lab published a paper that makes every agent builder rethink how they work: “Self-Harness: Harnesses That Improve Themselves” The idea fits in one sentence: you no longer need to hand-build everything around the model Put simply: Right now you design the tools, instructions, workflow and logic around your agent yourself, then run that same setup for months This paper asks a different question: why not let the agent improve its own setup? How it works: The agent runs real tasks, and every failure gets grouped by what actually caused it The same model looks at those failures and proposes several edits to its own harness Each edit is tested on two separate sets of tasks An edit stays only if nothing got worse and at least one thing got better The model itself never changes, only the harness around it does And it works: Qwen3.5 went from 18% to 36.7% on Terminal-Bench 2.0, GLM-5 from 44.4% to 85% on AppWorld, with the same weights The bigger shift is what’s left for you Your job moves from rewriting prompts and patching workflows to designing the tasks and checks that define what “better” means The people who win here know exactly what they want from an agent and can prove when it got better The whole paper in one line: Design the test before you write the code How to add this loop to your own agent in 6 steps is in the first reply ↓ That’s the same check-first loop I broke down in my Loops & Graphs guide below ↓

@0xGenAi

WHAT THE F*CK... 6 Grok Bots can give one person a second brain that never forgets. One brings over your setup from ChatGPT and Claude. One answers from your docs, with sources. One recaps every meeting. One runs your projects. One keeps your calendar and every birthday. One prints a newspaper of your day while you sleep. McKinsey found that employees spend 1.8 hours a day searching for and gathering information. That's 9.3 hours a week, more than a full workday. The 6 bots are free to add, and all 6 projects are public on GitHub. My first month with it: 1,412 links and PDFs saved, 63 meetings recapped, 240 answers with a source attached, 0 notes lost. 𝟲 𝗚𝗿𝗼𝗸 𝗕𝗼𝘁𝘀 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗺𝗮𝗿𝗸𝗲𝘁𝗽𝗹𝗮𝗰𝗲, 𝟱 𝗼𝗳 𝘁𝗵𝗲𝗺 𝗯𝘆 𝘁𝗵𝗲 𝗚𝗿𝗼𝗸 𝗕𝗼𝘁 𝘁𝗲𝗮𝗺 𝗮𝘁 𝗦𝗽𝗮𝗰𝗲𝗫𝗦𝗜: 01 / 𝗜𝗺𝗽𝗼𝗿𝘁 𝗕𝗼𝘁 → Brings your setup from ChatGPT, Claude Cowork and Codex into Grok Bot and builds the bots that cover it. Changes nothing in the old tools and scrubs your secrets first. https://t.co/2cHsh3eNLe 02 / 𝗖𝗼𝗺𝗽𝗮𝗻𝘆 𝗗𝗼𝗰𝘀 𝗤&𝗔 → Answers your questions from live docs and connected sources, and always cites them. https://t.co/AE82wYUzDe 03 / 𝗠𝗲𝗲𝘁𝗶𝗻𝗴 𝗥𝗲𝗰𝗮𝗽 𝗗𝗲𝗰𝗸 → Turns your meeting notes into a recap deck in your own template, and never invents a quote. https://t.co/uk7ltZT7xD 04 / 𝗣𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝗠𝗮𝗻𝗮𝗴𝗲𝗿 → Runs your projects from Notion: one row and one channel per project, with tasks your other bots claim. https://t.co/yIPRTe1TcP 05 / 𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝘃𝗲 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 → Checks your calendar against your sheets, preps your interviews and loads every birthday and anniversary. https://t.co/19ExgTmlan 06 / 𝗧𝗵𝗲 𝗠𝗼𝗿𝗻𝗶𝗻𝗴 𝗡𝗲𝘄𝘀𝗽𝗮𝗽𝗲𝗿 → Pulls from your email and calendar, lays out a newspaper made for you and prints it while you sleep. Karen X. Cheng built it. https://t.co/e6EEBLPHin Each bot comes with its own instructions and skills. The GitHub projects give them more to work with: Company Docs Q&A gets everything you've saved to search, Projects Manager gets memory that lasts between chats, Executive Assistant gets every document you've scanned. 𝟲 𝗚𝗶𝘁𝗛𝘂𝗯 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝘁𝗼 𝗽𝗹𝘂𝗴 𝗶𝗻: 07 / 𝗚𝗕𝗿𝗮𝗶𝗻 → The agent brain Garry Tan, CEO of Y Combinator, built for OpenClaw and Hermes. https://t.co/75zIhym9rR 08 / 𝗞𝗵𝗼𝗷 → A self-hosted AI second brain that answers from the web or your docs, runs deep research and schedules automations. https://t.co/l20vNlgCIP 09 / 𝗠𝗲𝗺𝟬 → The memory layer for AI agents: context that persists between sessions. https://t.co/IuwzAkpIYF 10 / 𝗕𝗮𝘀𝗶𝗰 𝗠𝗲𝗺𝗼𝗿𝘆 → Gives your AI conversations a memory, so you never re-explain your project. https://t.co/GghqaaLm5T 11 / 𝗞𝗮𝗿𝗮𝗸𝗲𝗲𝗽 → Saves every link, note and image, tags it with AI and makes all of it searchable. https://t.co/D0SUzc0f7W 12 / 𝗣𝗮𝗽𝗲𝗿𝗹𝗲𝘀𝘀-𝗻𝗴𝘅 → Scans, indexes and archives all your documents. https://t.co/0m79US2TBE 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲𝘆'𝘃𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗱𝗼𝗻𝗲 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗽𝗲𝗼𝗽𝗹𝗲 𝘄𝗵𝗼 𝗴𝗼𝘁 𝘁𝗵𝗲𝗿𝗲 𝗳𝗶𝗿𝘀𝘁: 𝗚𝗕𝗿𝗮𝗶𝗻: the brain Garry Tan's own agents run on. 32,000 stars and 4,800 forks in 7 months. 𝗠𝗲𝗺𝟬: over 150,000 developers already build with it, and Y Combinator backs it. 𝗣𝗮𝗽𝗲𝗿𝗹𝗲𝘀𝘀-𝗻𝗴𝘅, 𝗞𝗵𝗼𝗷 and 𝗞𝗮𝗿𝗮𝗸𝗲𝗲𝗽: over 116,000 stars between them. Add the 6 bots, paste the 6 GitHub links into one Grok Bot chat and type: "Bring over my ChatGPT and Claude setup, then build me a second brain I can ask anything" Save this, then read the full guide to setting up a team of Grok Bot agents from 0 in the article below.

@TopherNOW

Claude is about to give you a superpower if you set this up right. Here’s what to do: Claude can now join your meetings (beta on Team and Enterprise, Google Meet only for now) Most people will get meeting notes they never read again The edge is building it into second brain. All the intel, filed where Claude can easily access it. A lot of people have done this. The part they miss is putting it on autopilot. A second brain you feed by hand dies in two weeks TRY THIS SETUP PROMPT: "Read Andrej Karpathy’s LLM Wiki and build it for me in this folder: https://t.co/yCwK0hyFid This is for my meeting notes. Before you build anything, interview me: what kinds of meetings I have, what work they feed into, and what I’d want to look up later. Then put it on autopilot: create a scheduled task that pulls my new meetings (full transcripts, not just summaries) from my notetaker every night and ingests them into the wiki, plus a weekly lint." WHAT THIS DOES: Claude reads Karpathy’s blueprint and sets itself up as librarian: a raw folder for transcripts, a wiki with a page per person, company and deal, and a rules file (CLAUDE.md) Every night it pulls new meetings, updates the right pages, cites which meeting each fact came from and flags contradictions Every week it runs a health check for outdated claims and broken links in your second brain. You never touch it. You just ask it questions, and point to it to deliver the work that ripples out from the meetings. EVERYTHING you do with AI gets better, sharper and more informed once it can read your second brain. I walk into every call knowing more than I did walking out of the last one. Set this up now or just send this whole post to Claude and tell it to set it up. Note: This is WAY simpler to set up for yourself than for your whole company. A company version means deciding what goes in and who gets to see what. Start simple.

@Aura_lixx

Andrej Karpathy’s 1-hour Stanford lecture on AI engineering is one of the best explanations I’ve seen of how AI systems actually work. The progression is simple: 10% - LLM 30% - Prompt 50% - Agent 70% - Loop 100% - Graph The key takeaway: AI engineering isn’t just about writing better prompts. It’s about building systems around models, giving them context, memory, tools, feedback loops, and data flows. “Delete everything, keep Graph.” Definitely worth watching if you’re building with AI agents. Watch and bookmark it.

@bkdgiffug

学英语,背单词、刷题那套瞎折腾,可以停了。 GitHub 上爆火的项目 English-level-up-tips 值得一推,5 万+ Star 已经到手,英语学习被它拆成了一条完整路径。 词汇、听力、阅读、口语、写作,再到怎么结合 ChatGPT、Claude、Gemini 搭自己的 AI 学习系统,详细教程全都齐备。 https://t.co/2cNwR7ys1n

@0xShoopy

paste this into Claude Code tonight and in 30 days you'll find any idea you've ever saved in under a minute your notes app is where good ideas go to die. this one files and links everything you read, watch and save into a wiki, fact-checks its own notes every night and answers with the source built on Anthropic's own memory research: a linked wiki on top of search scored 61% vs 53% for search alone bookmark it before your next idea gets lost, then make Fable 5.2 last all week with the article below ↓

@kyr0stack

This is f*cking insane. Anthropic CEO Dario Amodei just went on the record about whose job goes first: "Entry-level jobs in finance, consulting and tech are going to be first augmented, but before long replaced by AI systems. Somewhere between one and five years we'll see big effects There will always be people who are using AI. Learn it now, or watch your role dry up" In just 4 minutes, he lays out which jobs go first and who ends up ahead Chatbots → Agents → Harness → AI Employees → Departments Run By AI Surviving AI doesn't mean asking a chatbot to tidy up your emails. It means being the person who plugs agents into a company's work and keeps them running Everything below is free. Start today: Anthropic - Prompt Engineering Interactive Tutorial: learn to tell a model exactly what to do https://t.co/d7FixLnpCL Anthropic Courses: your path from first prompt to running agents https://t.co/Cxvr3D3Zxa Anthropic Cookbook: copy-ready recipes for real work https://t.co/skYpllt5s5 Microsoft - Generative AI for Beginners: 21 lessons, no degree needed https://t.co/34KoC0v9Jb Microsoft - AI Agents for Beginners: build the agents that do the entry-level work https://t.co/Gph5MmV2Vb Hugging Face - Agents Course: free, certified, end to end https://t.co/gOPGy3Ds1U Google Cloud - Generative AI: Google's own hands-on notebooks https://t.co/LQWBAEtzbW OpenAI Cookbook: code for every common task https://t.co/YS1k9aHx2j https://t.co/EzDgtBLU7b - Prompt Engineering Guide: the guide everyone keeps open https://t.co/sfHv0eYsez Andrej Karpathy - Neural Networks Zero to Hero: see what is under the hood https://t.co/nU6AOBl6yk CrewAI: run a small team of AI workers https://t.co/HIPwJdFPGF Microsoft - AutoGen: put several agents on one job https://t.co/vOJKRrqg0N Awesome LLM Apps: hundreds of real builds to learn from https://t.co/LaWj0pHh4T https://t.co/ywXzOrdePa - Short Courses: free 1-hour courses from the people who teach this https://t.co/XHsCKIi2eL Total cost: $0 You don't need a degree to stay ahead of this. You need to learn to run the agents doing the work 95% of people will watch AI eat their industry. The 5% who build with it will be the ones companies fight over Save this and send it to someone whose job is on that list then read the article below ↓

@HeLi8866

这周 GitHub 上涨星最猛的 AI 仓库,前十都在这了👇: 1. security-audit-skill 可以把代码安全排查分几步交给 AI agent。查出的问题都会先核实,再交给你。 https://t.co/L2L1Lv6yOG 2. astryx 把界面设计组件整合在一起的开源基础库,能按自己的想法改,AI agent 也能直接用。 https://t.co/5Rtf1zeyR5 3. Agent Skills(anthropics) 一套公开的组件,用来让 AI agent 学会新的操作步骤,需要的能力可以往后加。 https://t.co/1CQtdIGguW 4. agent-browser 把浏览器操作交给 AI agent 来做的工具,打开页面、输入、点击都能替你完成。 https://t.co/kzqp2acHhO 5. graphify 程序和资料会变成一张能顺着关系查的地图,AI agent 可以问哪里有什么。 https://t.co/vpwvOyNMxA 6. claude-mem 能把和 AI agent 的交流记下来,带到下一次对话里,接着上次记住的内容继续聊。 https://t.co/SjXafuPfAi 7. Claude-Code-Game-Studios 能把 Claude Code 整套变成一个游戏开发团队,不同分工的 agent 一起往下做。 https://t.co/GDTYJykPBB 8. herdr 给写代码的 AI agent 用的运行环境。连接断了,正在干活的终端也不会消失,再连回去还是原来的布局。 https://t.co/1PTIJX8I6d 9. gstack 可以直接用上 Garry Tan 那套 Claude Code 配置。从经营到质检,各角色要用的工具都齐了。 https://t.co/s7QcwOZegN 10. claude-quickstarts 用 Claude 做应用时能马上上手的示例合集,从能跑的模板开始试,再慢慢改成自己的。 https://t.co/S1AHswJj4g

@mr_kozh

This is f*cking brutal. A warning from Anthropic CEO Dario Amodei is going viral again: AI may not replace the CEO first. it may remove the first rung of the career ladder. Amodei was asked what a kid finishing high school should do now. His answer: “AI can do entry-level law work. It can do entry-level finance. It can do entry-level consulting.” In a few minutes, he explained who gets hit first: → junior lawyers reviewing documents → finance analysts processing information → consultants assembling research → developers completing routine tasks He also gave a number: 50% of entry-level white-collar jobs could be disrupted within 1–5 years. That is a forecast, not a measured outcome. But the reason behind it matters. Previous technologies replaced one kind of work. Workers could move from farms to factories. from factories to offices. from paper to computers. AI is different because it can attack several nearby career paths at the same time. Old model: study for years → enter through junior work → learn beside senior employees → move up the ladder New model: AI performs the junior work → one human supervises multiple agents → evidence gets reviewed → the system improves → fewer entry-level seats remain If AI can move between law, finance, consulting and code, changing majors only moves you to another part of the same battlefield. The real advantage is learning to direct, verify and improve the systems doing the work. Amodei’s advice fits into one line: “Of course, learn to use AI.” Bookmark this before your feed buries it. Here are 4 free resources that will help you start: 01 - learn how Anthropic structures effective prompts https://t.co/Wq1202adse 02 - go from basic prompting to building with Claude https://t.co/6gmAmwpLya 03 - build agents that complete multi-step work https://t.co/rArxWgnRHK 04 - learn agent tools, workflows and evaluation https://t.co/EtdxMCU7a4 The first career step used to be doing the repetitive work. Now the advantage belongs to the person who can make AI do it and prove the result is correct. Save this and build one working agent before choosing another course ↓

@_vztu

🚀 Introducing 𝗙𝗹𝗲𝘅-𝗥𝗦𝗜: 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝗶𝗻𝗴 𝗥𝗲𝗰𝘂𝗿𝘀𝗶𝘃𝗲 𝗦𝗲𝗹𝗳-𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝗼𝗳 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗔𝗴𝗲𝗻𝘁𝘀 We're curious: 𝘊𝘢𝘯 𝘢𝘯 𝘈𝘐 𝘢𝘨𝘦𝘯𝘵 𝘨𝘦𝘵 𝘣𝘦𝘵𝘵𝘦𝘳 𝘢𝘵 𝘪𝘵𝘴 𝘫𝘰𝘣 𝘣𝘺 𝘳𝘦𝘸𝘳𝘪𝘵𝘪𝘯𝘨 𝘩𝘰𝘸 𝘪𝘵 𝘸𝘰𝘳𝘬𝘴? 🔄 We tested 𝟭𝟴 frontier models on 𝟮𝟰 tasks, from 𝗿𝗼𝗯𝗼𝘁 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 and 𝟯𝗗 𝗿𝗲𝗰𝗼𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 to 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 and 𝘃𝗶𝘀𝘂𝗮𝗹 𝗰𝗿𝗲𝗮𝘁𝗶𝗼𝗻—𝟭𝟴𝟴 runs and nearly 𝟯,𝟴𝟬𝟬 rounds of self-improvement. Agents could revise their code, tools, prompts, and workflows while their model weights stayed fixed. A few observations I found particularly interesting: 🛠️ 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗿𝗲𝘃𝗲𝗮𝗹𝘀 𝗺𝗼𝗿𝗲 𝘁𝗵𝗮𝗻 𝗮𝗻𝘀𝘄𝗲𝗿𝗶𝗻𝗴. The largest gains came when agents built programs, robot policies, or pipelines. Models starting from the same solution ended up far apart. 📉 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸 𝗰𝗮𝗻 𝗯𝗲 𝘁𝗵𝗲 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸. Open-weight models made progress with graded feedback but often stalled when they received only success or failure. On RoboCasa, median validation gains were 51.2 points for closed models versus 5.6 for open-weight models. 🧱 𝗟𝗼𝗼𝗸𝗶𝗻𝗴 𝗿𝗶𝗴𝗵𝘁 𝗶𝘀 𝗮 𝗹𝗼𝘄 𝗯𝗮𝗿. Some reconstructed scenes looked convincing but failed physical stability or object-level checks. The gap between closed and open-weight models was often largest where correctness could be tested strictly. ⏱️ 𝗠𝗼𝗿𝗲 𝘁𝗼𝗸𝗲𝗻𝘀 𝗱𝗶𝗱𝗻’𝘁 𝗺𝗲𝗮𝗻 𝗯𝗲𝘁𝘁𝗲𝗿 𝗿𝗲𝘀𝘂𝗹𝘁𝘀. Token use and performance were essentially uncorrelated across the evaluated runs. Several of the strongest models also used fewer tokens. And the workflow edits themselves helped: in 6 of 7 paired experiments, agents allowed to rewrite their own harness achieved higher test scores than those with a fixed harness. Glad to share this work with Jingdong Zhang, Jionghao Wang, Zhengming Yu, Yancheng Cai, Wei Chow, Chris Choy, Zhiyang (Frank) Dou, Hang Hua, Jiachen li, Zhixuan Liu, Chen Wang, Songlin Wei, Siyuan Yang, Yu Yuan, Xiaohang Zhan, Yucheng Zhang, Shuhong Zheng, Xin (Shane) Li, Lingjie Liu, Wojciech Matusik, Wenping Wang. 🎥 Work done by @TAMU, Adobe, @CarnegieMellon, @MIT, MIT-IBM Watson AI Lab — IBM Research, @NUSingapore, NVIDIA, @LifeAtPurdue, @Cambridge_Uni, @Penn, @USC, @UTAustin, @UofT. 🤝 Flex-RSI welcomes community contributions—new tasks, models, and evaluators are all welcome. 🌐 𝗣𝗿𝗼𝗷𝗲𝗰𝘁: https://t.co/QmAbKfo4wy 📄 Report: https://t.co/6cWMiQraqU #AIAgents #SelfImprovement #Robotics #RSI #AGI #GenAI #LLM #Agents #TAMU #EmbodiedAI #Vision

@0xGky

我天!GitHub 上有个狠项目被人开源了,专盯着爆款视频到底怎么做出来! 项目名 rea-video,眼下 8500 Star 已经到手,Codex、Claude Code 都能兼容。 一段视频丢进去,AI 就一个镜头一个镜头往下拆: 1、画面属于实拍、AI 生成还是代码渲染,它先替你判定。 2、灯光、运镜、字幕、转场、背景音乐,它逐项过一遍。 3、每个镜头可能用到的 AI 提示词,它能倒着推出来。 4、完整复刻方案它也能出,哪些靠代码实现、哪些用 AI、哪些要拍,全给你标明白。 镜头节奏、音效位置、字幕出现时间,连这些它都拆得开! AI 广告、短视频、视频号的,不妨瞧瞧 🔗:https://t.co/Wcj71hkmno

@ohayou792

最近Google 反重力更新了两个很容易让人混淆的能力:Orchestrator 模式 和 /swarm 模式。 很多人分不清该用哪个,其实用一句话就能说明白: 👉 Orchestrator 是一个人换不同脑子干活 👉 /swarm 是拉一支小队并行干活 简单拆解一下它们的区别和选型思路👇 1️⃣ Orchestrator 模式:单会话内的动态模型接力 它是在模型下拉框里开启的选项。 整个过程依然是单线程对话,但系统会在后台自动切模型: 做架构规划、理清复杂依赖、遇到卡壳报错时 ➡️ 自动切给高级推理模型如 Gemini 4 Carbon 方案定好,开始写样板代码、跑测试时 ➡️ 切给高速执行模型如 Gemini 4 Hydrogen 执行报错了 ➡️ 再自动交还给强模型重新评估 ✅ 核心价值:不改变你的对话习惯,既保住了顶级模型的规划下限,又省配额、提速度。 2️⃣ /swarm 模式:原生多 Agent 并行协作 在输入框敲 /swarm 触发。 主对话会变成 Coordinator协调者,自己不写代码,专门负责拆任务、派活和验收: Workers:多个独立 Agent 同时开工,甚至可以在隔离的 Git worktree 里各自改代码,互不干扰 Verifiers:专门配几个挑刺的验证者,独立跑测试、审 diff,防止单模型既当运动员又当裁判 共享消息板:成员之间通过 @名字 异步沟通,空闲就休眠 ✅ 核心价值:用更高的 Token 消耗,换取大型任务极致的并发吞吐和交叉验证质量。 💡 到底该怎么选?记住 3 条口诀: 日常开发无脑开 Orchestrator:聊需求、单模块重构、修单个 Bug,逻辑强依赖拆不开时,用它最省心、最省额度。 遇到能拆分的大活再敲 /swarm:比如跨几十个文件的重构、多个独立模块同步迁移、或者需要边写边做严格对抗测试。 两者还可以配合用:平时开着 Orchestrator 稳步推进主线,遇到难啃的大型复杂问题,随时用 /swarm 开启并行攻坚,完成后统一合入主线继续推进。

@NFT_Chen

Meta+NYU 用200+模型做了一次硬核对照实验,直接破除了多模态模型的老规矩: 🔹不用再单独预训练视觉编码器了 🔹视觉必须有自己的专用参数 🔹这些参数甚至可以塞进语言模型前几层内部,效果一样好,推理还更快 这意味着多模态训练可以更简单、更灵活,架构也能更干净。 #多模态 #AI研究 #Meta #NYU #VisionEncoder #LLM #人工智能 #前沿研究 https://t.co/zFcdBJaCvD https://t.co/EivFQHETqM

@alex_verem

Google automated the research paper. Google built a team of AI agents that writes a full research paper from a scientist's rough notes. The system is called PaperOrchestra. You give it a short summary of the idea and the raw experiment results, and it returns a finished paper formatted for a conference. Five AI agents split the work. One plans the outline, one draws the charts and diagrams, and one finds related research and checks that every cited paper exists. A fourth writes the sections, and a fifth revises the draft using feedback from a simulated peer review. The team tested it on 200 papers accepted at two top AI conferences. They reduced each paper to its idea and its results, then had the system write it again. ✅ It took about 40 minutes per paper on average. ✅ An AI reviewer gave its papers simulated acceptance rates of 84% and 81%. The real human-written papers got 86% and 94%. ✅ It cited about 46 to 48 papers per manuscript. Humans cited about 59, and other AI writers cited 9 to 14. ✅ Human judges preferred its literature reviews over other AI systems' by margins of 50% to 68%. ✅ Against the human-written literature reviews, it tied or won 43% of the time. The limits are in the paper too. A second AI reviewer gave it 48% and 22%, against 71% and 63% for the human papers. When 11 AI researchers compared full papers side by side, they preferred the human-written one 81% of the time. The acceptance rates come from AI reviewers, so I'd treat them as estimates. The system still needs a human to supply the idea and run the experiments, and the authors say humans stay accountable for what the paper claims.

@mniehaus

Using GitHub Copilot with a (free) local LLM https://t.co/O2KbV4YCmk

@Ryrenz

📽️ 牛炸了,OpenFilm 能让 Claude Code、Codex 这些编程 agent 直接帮你把视频做出来,做完你还能自己上手改。 OpenFilm 10 月初才建仓,现在已经快 7000 star,上了 Trendshift 日榜,桌面版 v0.3.0 和 v0.3.2 两个版本的 release 文件累计被下载了 6 万多次。 以前让 AI 做视频,要么是生成式模型抽卡,同一句提示词每次出来都不一样;要么得自己开剪辑软件一帧一帧对。 OpenFilm 的思路是把影片写成网页:每个镜头就是一个页面,three.js、canvas、SVG、CSS 动画都能用,同一个时间点永远画出同一帧,换台机器渲染结果也一样。agent 负责写,你在浏览器里的本地编辑器 Studio 看效果、手动微调。 做完能导出 MP4、GIF、字幕和透明视频,还能导出给 Final Cut Pro、Premiere Pro、达芬奇接着剪的时间线。制作和渲染都不用注册账号,配音、配乐这些生成能力可以接自己的 ElevenLabs、OpenAI 等服务。 视频以前是剪出来的,现在是写出来的。 GitHub:https://t.co/OE7LjgXCZG

@BlueQubitIO

🛡️ No simulator can check a 6,000-qubit algorithm. A theorem prover can. 🧑‍🤝‍🧑 Agent teams, a shared wiki, virtual conferences 🔍 Every proof verified against a frozen spec 📖 Full paper: https://t.co/ke6XJPKaR2 #QubitScript #FormalVerification #QuantumAlgorithms

@lonelygogo21

🥱 为了熬夜看 LoL 总决赛,闲着无聊,写了这篇文章,分享一些自己做达人营销的经验 https://t.co/xakhhRz4z0

@eng_khairallah1

These 20 GitHub repos turn your Fable 5.2 into a team of AI agents that never clocks out. this is a f*cking insane for Fable 5.2 skills. Memory. Tools. Browser. Team. Ship. 01 /gstack → Garry Tan's own Claude Code setup. 23 tools, one virtual team. https://t.co/sZMjlFE4aQ 02 /superpowers → Spec first, then subagents build and review it. https://t.co/3rROHSagiP 03 /agent-skills → Addy Osmani's 9 commands, from /spec to /ship. https://t.co/RnoALtvFLl 04 claude-mem → A second agent takes notes so no session starts from zero. https://t.co/xFUrwcxkRd 05 Mem0 → Drop-in long-term memory for agents and apps. https://t.co/5h7pXwqs4q 06 Graphiti → Knowledge graphs that track how facts change over time. https://t.co/jy8dsnQd6s 07 Beads → A task graph so long jobs never lose the plot. https://t.co/Ch5ht76kLB 08 Context7 → Fresh, version-specific docs. No more hallucinated APIs. https://t.co/q37F0LTs8I 09 GitHub MCP Server → Issues, PRs and Actions from a single prompt. https://t.co/2jH889WhK9 10 Serena → IDE-grade code search and edits over MCP. https://t.co/dftGZpOniC 11 Browser Use → Give it a task. It clicks through a real browser. https://t.co/JTvwZUv9c7 12 agent-browser → Vercel's Rust CLI. Drive Chrome one command at a time. https://t.co/5LglDhDd5J 13 Firecrawl → Any website into clean markdown or JSON. https://t.co/Sd9Cbk63Cu 14 Ruflo → Agent swarms inside Claude Code with one npx init. https://t.co/X6Lcg7pj7q 15 wshobson/agents → 202 agents and 185 skills, plug and play. https://t.co/VXzWVMK0H4 16 CrewAI → Role-based agent crews that work together. https://t.co/9KCDTmsxms 17 Archon → Your process as YAML. Agents run it, you approve the gates. https://t.co/xSQ7EBkqVs 18 Claude Code Action → Tag claude on any PR. It reviews and fixes. https://t.co/1MyS3u2k0G 19 E2B → Secure sandboxes for AI-written code. https://t.co/oM7ergHcp6 20 Langfuse → Traces, evals and the cost of every agent call. https://t.co/Ss36VqTTpS Skills → memory → tools → browser → crew → merged PR. Bookmark this, then follow the guide below to build your first team of AI agents ↓

@ufobtc27

如何面临AI Agent对你日后生存的冲击?那就是驾驭AI! @AnthropicAI CEO @DarioAmodei 被问:现在高中毕业的孩子该做什么? 他的回答:AI 已经能做初级法律、金融、咨询工作。它会淘汰大量岗位,尤其是初级白领工作。他预测未来 1-5 年,50% 的初级白领工作会被冲击。同时,中高层的工作也会有很多新形态发生 进化路径:Prompt → Agent → AI 同事 → 初级工作被 AI 接管 →中高级工作形态进化 以前从农业到工厂花了几代人,这次 AI 会同时冲击职业的每个阶段 他的建议只有一句:当然要学用 AI 今年开始学的人,不会被AI替代,而是指挥 AI 做这些事 下面全是免费资源,今天就能开始: Anthropic Prompt Engineering 互动教程:最快学会指挥模型
https://t.co/XmklbHRPt9 Anthropic Courses:从第一个 Prompt 到跑 Agent
https://t.co/G1Ae1SBeHz Anthropic Cookbook:可直接复制的真实任务配方
https://t.co/kDIkectUtK Microsoft Generative AI for Beginners:21 节课,无需学位
https://t.co/wtbMYz83TK Microsoft AI Agents for Beginners:让 Agent 替你做初级工作
https://t.co/n6hM2suZ0Z Hugging Face Agents Course:免费带证书,从零到一
https://t.co/DvtoTytIXY Google Cloud Generative AI:Google 的实战 Notebook
https://t.co/Kn1ii94BTp OpenAI Cookbook:常见任务代码
https://t.co/QrGZ2u8XN5 Prompt Engineering Guide:大家都在引用的参考
https://t.co/4sbN4qeZ1U Andrej Karpathy - Neural Networks Zero to Hero:搞懂底层原理
https://t.co/bazVQGoggp CrewAI:运行一个小型 AI 团队
https://t.co/MolSIOjrc7 Microsoft AutoGen:多个 Agent 协作完成一项工作
https://t.co/oLOEjZZUYq Awesome LLM Apps:上百个真实项目可参考
https://t.co/qnBOzdfV4Q Short Courses:1 小时免费短课
https://t.co/RiL349P29i 总成本:$0 你不需要再花一年学费。你需要开始指挥 AI 做事 95% 的人毕业后会和 AI 抢同一个初级岗位; 
5% 今年学会指挥 AI 的人,会管理做这些工作的 Agent 保存转发给今年焦虑的人

@KirkDBorne

Systems Thinking for Agentic AI — Design reliable LLM and agentic AI systems with RAG, MCP, guardrails, evaluation, and observability: https://t.co/IKxOiCNLX9 https://t.co/C7pB0dGQ5m

@rosemoni18

🚨 Who would have imagined that professional video production could now be done using only HTML and CSS?! 🤯🎬 An open-source project on GitHub is turning video editing and content creation upside down: OpenFilm (which has already racked up over 5,800 stars at lightning speed)! 🔹 How does OpenFilm completely change the way video is produced? ▪️ Every shot is a web page: The project turns every scene in the video into an HTML/CSS page, allowing AI to build graphics, animations, and code-based transitions with pinpoint precision before combining them into a complete video. ▪️ A familiar editing experience (like CapCut): After generation, you don’t have to manually edit the code—you can drag the timeline, edit text, and adjust shots with complete ease. ▪️ Instant export and complete freedom: Instant MP4 export, fully open source under the MIT license, and runs smoothly on Windows and macOS. 💡 The bigger picture many people are missing: Programming is no longer just for building websites and apps—it has become the most powerful tool for video production and motion graphics because AI models excel at writing code with greater precision than randomly generating pixels! 🔥🚀 Save this post 🔥 Tool link in the comments 👇

@Chenfacai0

有人已经汇总出 341 套用于代码动画的风格提示词,另外还附带了 193 张参考图片,全部免费开放源码! 从粒子特效到国风剪纸,再到白板手绘……各种风格摆在那里,挑起来就跟点菜一样方便。 挑定以后,把这些提示词喂给 Claude Code 或者 Codex,它就能产出 HTML 动画,并且能够导出为 MP4 文件。 项目链接:https://t.co/ha7maghiG0

@Dav338876

Claude Fable5.2+Threejs 整出狠活😨 0 资源 纯代码 依旧顶级美术‼️‼️‼️ 并且才用 1 个小时 这还让影视动画玩吗?? 项目开源 我放评论区了👇 https://t.co/7aLXgfylTX

@BeyondTheBoxCN

我尼玛感觉Fabel 5.2 比视频模型更像视频模型! 这是全代码生成的,你敢信??? 这是不是要开拓一条视频生成的新路了??? 我感觉其它视频模型压力山大了!😂😂 https://t.co/ZbCaIkAwj6

@swordBluesy

Codex 能直接换成 DeepSeek 模型,官方给了一键脚本,不用手写配置 建议收藏,用到时照着做。做完你的 Codex 命令行、桌面端、VS Code 插件都能用 DeepSeek 模型。 这是 DeepSeek 官方文档 Integrate with Codex(接入 Codex)一页。先说结论: 官方写明,Codex 通过 Responses API 跟模型对话,DeepSeek API 原生支持它,所以配置一次,三种 Codex 客户端共用。 先说清词。 Codex,是 OpenAI 的 AI 编程助手,官方说它的命令行、ChatGPT 桌面端、VS Code 插件共用同一个配置文件。 API Key,是调用 DeepSeek 的密钥,官方说以 sk- 开头,在 DeepSeek 开放平台的 API keys 页面申请。 config.toml,是 Codex 的配置文件,位置在用户目录下的 .codex 文件夹里。 先把话说全: 官方建议的方式是一键脚本。运行前,要先装好 Codex 命令行或 ChatGPT 桌面端,并且至少打开过一次,这样 .codex 目录才存在。 一、一键配置 第一步:苹果电脑或 Linux,在终端运行官方给的命令:bash <(curl -fsSL https://t.co/K4yzbC4rIp)。 第二步:Windows,在 PowerShell 运行:irm https://t.co/qPpJAD209q | iex。 第三步:看菜单。官方写明,选 1 是配成 deepseek-flash 模型,它也能收图片;选 2 是配成 deepseek-v4-pro;选 9 是恢复 Codex 默认配置。 第四步:第一次运行会让你填 API Key,就是以 sk- 开头那串。 二、脚本替你做了什么 官方列了四件事:先把原来的 config.toml 备份到 .codex 下的 backup-deepseek 文件夹;再写一份 models.json,告诉 Codex 这些模型的上下文长度等信息;再更新 config.toml,加上 DeepSeek 提供方;最后检查生成的文件,不通过就停下报错。 官方还写明,脚本 1.5.0 默认打开 Alpha Search 联网搜索,需要 Codex 0.164 或更高版本。 三、不想用脚本,手动改 官方说,先建 models.json,再在 config.toml 里写 model 为 deepseek-flash,model_provider 为 deepseek,wire_api 写 responses,base_url 写 https://t.co/E4BVCR4fv8,experimental_bearer_token 填你的 API Key。完整内容在官方页面里。 四、开始用 第一步:命令行用户,进到项目目录,输入 codex。 第二步:看启动横幅,官方说要确认显示的是你选的模型。 第三步:桌面端如果模型选择器显示 Custom,就是用配置文件里的 model。 五、小提醒 官方写明,如果切换后看不到以前的会话,不是删了。官方原话意思是,用 ChatGPT 订阅登录和用第三方 API 登录的会话是分开存的,恢复原配置(脚本选 9)就能看到以前的。 想关掉联网搜索,官方说在 config.toml 最上面加 web_search = "disabled" 再重启 Codex。 收藏版:5 步速查 一、先装好并打开一次 Codex。二、拿 sk- 开头的 API Key。三、跑一键脚本。四、菜单选 1 或 2。五、进项目目录输入 codex。 S1哥 观点:换模型最怕手改配置改坏,官方把备份和恢复都做进脚本,这点比教程本身更值得抄。 一句话结论:Codex 换 DeepSeek,跑脚本、选 1、填 Key 就行,想退回选 9。想让朋友少踩坑,复制链接发给他。 你是想配 deepseek-flash 还是 deepseek-v4-pro?卡在第几步?同系列前几篇在我主页。 参考:DeepSeek 官方文档 Integrate with Codex(S1哥 整理)

@yeemio

An LLM and tools are a start. A complete AI agent needs more. Here’s the six-layer system we built: - Workspace - Application Host - Agent Runtime - Models &amp; Tools - Data &amp; Evidence - State &amp; Recovery How they fit together, from task to result 👇 https://t.co/aCtT0p0ZD1

@awe0_io

发现一个专门给 Claude Code 装插件的开源工具。 claude-code-templates 把 Agent、Skills、Hooks、MCP、命令模板放到了同一个目录。 想让 Claude 自动审查代码、测试项目、处理文档,直接找对应组件安装。 甚至能管理和分享 Claude Code 会话。 GitHub:https://t.co/Iyh3AuCHaS

@leo114119

Claude fable 5.2 用1分钟生成的纯代码动画 感觉 Fable 5.2 用代码生成动画的能力,又提升了一个等级 https://t.co/iATq2rQvmk

@SamuJRiver

Just shipped what is likely my most useful tool for builders yet. Reword deslops your AI-generated prose by teaching your agents to write the way you would want to write in any given context. Open source. No data harvested. Try it out + give it a star: https://t.co/IXlpuEkmnr

@ZondarkAI

没挂任何 Skill,就用一句 Prompt 让 Claude Fable 5.2 裸跑:“three.js 做个动漫视频”。 实测发现,用纯几何体 + 三角函数硬搓三渲二,做写实动漫人物极难,但做这种 Q 版人物非常擅长。 很好奇,有人用 Fable 5.2 纯代码破局过写实风人物吗? https://t.co/6PEIoDSZ1V

@kedaya

腾讯开源LoopForge了,可以自托管编排CodeBuddy、Codex和Claude Code,还带人工审批关卡。企业用起来能放心点。 https://t.co/JdAgCcqk5h

@FinanceYF5

https://t.co/VTiOZm4Ii8

@Blum_OG

Your AI agent can forget what it knew the day you switch its model. The memory is still there, and the new model reads it differently. Researchers at LinkedIn measured it: → With notes the old model wrote, right answers fell from 47% to 34% → Facts saved in one fixed format barely changed → Fixing the broken notes later worked 0 times out of 48 → Rebuilding from the original chats worked 34 times out of 48 The study itself is in the reply. Harrison Chase, the creator of LangChain, explains below why the company that sells you the model shouldn't hold your agent's memory

@0xpipimd

Microsoft researchers scaled an AI coding system to 1,024 agents working on the same software project, without a central orchestrator. On one benchmark task, the test-pass rate increased by 62% relative to a single agent. All agents used the same underlying model. The project is called Agensh, an open-source multi-agent harness from Microsoft Research. Most multi-agent systems rely on a central orchestrator to assign tasks, monitor progress, and coordinate workers. That approach can become a bottleneck as the number of agents grows. Agensh takes a decentralized approach. Agents organize their own work through shared infrastructure, rather than waiting for instructions from a central supervisor. Each agent can: → Discover and claim available tasks independently. → Share findings and coordinate with other agents. → Work concurrently on a shared codebase. → Verify changes and integrate completed work. → Pick up unfinished tasks as the project progresses. Under the hood, Gitea provides the shared Git workspace, Mattermost handles communication, and a shared context mechanism helps agents exchange reusable findings. The researchers evaluated Agensh on five difficult ProgramBench tasks, using GPT-5.6-sol with high reasoning effort, a six-hour budget, and no internet access. The results are interesting: → Across five tasks, scaling from 1 to 128 agents increased the average test-pass rate from 19.31% to 28.78%. → On Pandoc, scaling from 1 to 1,024 agents increased the test-pass rate from 33.89% to 55.06%. → Larger teams also reached comparable test-pass rates earlier in the experiments. The researchers observed peer coordination, specialized roles, and integration workflows emerging as the teams grew. The GitHub repository includes the harness, experiment scripts, Docker setup, and reproduction instructions. It supports Copilot, Codex, Claude Code, and Pi as underlying agent backends. These are results from specific benchmark tasks, but they demonstrate something worth studying. All 1,024 agents used the same underlying model. What changed was how they organized their work. MIT licensed and fully open source. GitHub: https://t.co/XEPZ1PneSX

@balance0014

Your AI agent can now build, light and render a 3D scene end to end, and the engine it drives is free and open source. 🔥 It's called 67js, by CodePawl. You prompt Claude Code, Codex or Cursor, and the agent works through an MCP server: places objects, sets lights and fog, checks PNG previews, then renders a still or an MP4. It's catching on fast: 633 GitHub stars and about 4,100 npm downloads last week. Setup is one prompt to your agent, or a Claude Code plugin. MIT licensed, needs Node 20.19+ and a local Chrome, and falls back to CPU if you don't have a GPU. Good call on safety: the MCP server accepts no arbitrary code and no remote asset URLs, only scene data and the GLB files you point it at. The catch is it's pre-1.0, ships a new minor version about every week, and a minor can break the API. Great for product shots and explainer clips today, too early to build a pipeline on. 🔗 Link in the first reply 👇

@HaiyuWu1

A great paper of discussing how to integrate Vision into the LLM training paradigm. That aside, it is interesting that, since REPA, layer 8 keeps showing up in many of Saining’s papers. Is it really a magical layer or something is missing here?🤔

@shao__meng

https://t.co/kXtvpA0iQt

@jaxxwu

中国 RSI 又进一大步… 上海创智学院和新序智能近日发布技术报告:一支新团队用三个月从零训练出混合专家基座801-30B-A3B,数据11.8万亿token。模型以801教室命名。 递归自我改进此前多发生在模型训成之后。预训练不一样,一次要跑几个月、烧掉数十万GPU小时,几乎没有重来的机会。报告并未宣称智能体已能独自完成整次预训练,而是把这条没法一次闭环的大流程,拆成可以反复迭代的小循环:排查基础设施和数据、测量后训练潜力,并把失败原因留下来。作者称,这是目前公开资料里首次把递归自我改进用到完整的从零前沿预训练,并详细写出过程。 作为基座,它超过同规模的Qwen3-30B-A3B和Nemotron-3.5-Lightning-30B-A3B,只落后于更大的Qwen3.5-35B-A3B;相同轻量微调后,这一排序仍在。仅用约60亿token监督微调、不加强化学习,它在四项智能体基准上的平均分,甚至高过完整后训练的对比模型,包括Qwen3.6-35B-A3B。论文写明后一组对比并不对等,主要用于观察基座潜力。 完整后训练版本还将发布,轻量微调数据已经公开,基座模型计划开源。值得看的不只是分数,更是预训练的方法被写清楚了。

@Huahuazo

Claude、Cursor、Codex 摆出来个个像高手, 一上真活儿,底裤立刻掉:推特搜不动,小红书读不进,B 站字幕看不着,Reddit、YouTube、GitHub 全成了睁眼瞎。。 从前只能自己满世界翻 API、掏钱买路,还天天被封,烦到没脾气。扔一句“帮我安装 Agent Reach”就收工 Twitter、小红书、带字幕的 B 站、Reddit、YouTube、GitHub、网页、RSS、雪球、LinkedIn,一条龙全打通。 一个 CLI 撑起来,API 费用直接归零,通道断了自动改道,doctor 一键体检。 🔗 https://t.co/K4NsuVfhfy

@leo114119

Claude Fable 5.2 生成的横版RPG, 所有视觉效果均用代码绘制 画风真的很有艺术感啊。。 https://t.co/96N7WPaU4S

@aiwaremax

GitHub 狂揽 10.7 万星的「本地版 Claude Design」完整实战指南来了 内置 151 套大厂品牌系统与 277 个插件,连上 Claude Code 就能一口气产出原型、PPT 与宣传片 6 步实操流程、防返工提示词与 MCP 跨仓交接命令请看下文👇 https://t.co/gQHdlHoEPe

@wuyi_ai

爽用Codex, 必装的 4 个超实用 skills 查资料、读文件、改文案、做 PPT,都是平时反复要干的活👇 1️⃣ Agent-reach ( 11.1 万 ⭐) 全平台搜索, 包括推特, Reddit, Youtube, B 站,小红书 等等, 获得最新的信息。 https://t.co/wCJkmqxErI 2️⃣ Markitdown ( 19.5 万 ⭐) 微软开源的转换工具。手里一堆 Word、PPT、PDF、网页,想让 Codex 帮你提炼重点?先转成 Markdown,再一起丢给它整理。 https://t.co/xJsL3N3SOn 3️⃣Humanizer ( 6.8万 ⭐) 初稿一股 AI 味, 用它删掉空洞的总结和套话,把句子改得更像平时说话。写推文、邮件、产品介绍都能用上。 https://t.co/eUR48lMNRG 4️⃣Guizang-ppt ( 2.9 万 ⭐) 经常做提案、汇报、活动复盘的可以收。尤其适合解决同一份 PPT 每页都像用了不同模板的问题。 https://t.co/zoMlRaWgbK 先收藏起来 慢慢用!

@feishutoday

Claude Fable 5.2 构建的 Abyssal Garden 这是一个你可以直接在浏览器中游览的水下世界,仅由一个 97 KB 的 HTML 文件构成,无需图像、3D 模型、声音文件或库 你可以直接在浏览器中游玩。 - https://t.co/j2vLEKy4EG https://t.co/zRuPR94p9n

@_asin1_

Claude Desktop 多账号切换工具,开源免费 CLI 的多账号方案很多,Desktop 的应该是独一份 · 一键切换账号,额度/重置时间一目了然 · Claude Code 会话跨账号同步 我自己一直在用:多开 Pro 轮换,风控比 Max/Ultra 小 没有苹果签名,首次打开需去「隐私与安全性」放行 https://t.co/4rUOZcAHZC

@Manorgw

总算有人把 FFmpeg 直接接到 AI 编程助手里了。 以前让 Agent 剪视频,参数基本靠蒙,剪完甚至连结果都不看一眼。这个 ffmpeg-skill 把整条流程给补全了。 ① 全程本地跑,账号、API Key 都不用,素材也不用上传 ② 内置 42 个工具,剪切、拼接、去静音、字幕、画幅、响度、HDR 转 SDR 都能做 ③ 支持 Claude Code、Cursor、Codex ④ npx ffmpeg-skill 一行安装 ⑤ 先 probe 再处理,完了再自动验一遍 简单说,就是让 AI 别光“调用 FFmpeg”,而是先看素材、再动手、最后查结果。 🔗 :

@linyun1128

先说结论:记东西这件事,可能根本不该让上下文窗口来干。 斯坦福刚开源的 Memorilla,把整个文档集压成 16 个 token 喂给模型,答得反而比喂全文更准。512 篇文档的测试里,延迟从 19.9 秒降到 0.74 秒。 说白了,它就是给大模型配了个专门记东西的小模块,模型本身不动。 怎么运作的:文档先全部转成向量存起来。有人提问,这个 1.94 亿参数的小模块把向量都读一遍,按这个问题挑出最重要的部分,压成 16 个 memory token,递给模型。实验里用的模型是 Qwen3-8B,参数完全冻结,训练的只有这个小模块。 成绩是他们自己测的,先说清楚这一点。FactKG 这个事实核查基准上,它拿到 84.33%,把全文塞进上下文是 80.43%,用的文档 token 只有人家的 1/108。一共测了 8 个基准,它平均分最高,赢了 5 个,对手里包括 Mem0、MemoRAG 这几个现成的记忆系统。上面那个 19.9 秒降到 0.74 秒,是 512 篇文档那个单个测试的结果,别当成普遍规律。 缺点他们也没藏。细节记不住:PubMedQA 上只有 59%,比看全文低了 8 个点,具体的人名、数字这类细粒度信息,压缩的时候容易丢。这点和 RAG 正好相反,RAG 是原文都留着、用到哪查到哪,代价是慢和贵;它是什么都压掉,只留个大概,代价是可能把关键细节压没了。 还有一个实际的门槛:这套东西要往模型的输入层里塞向量,所以只能用在开放权重的模型上,每个模型还得单独配一个训好的模块。Claude、GPT 这种闭源的 API,现在用不了。 前几天我发过一条 ETH 的论文,说仓库概览写进 AGENTS.md 没用,事实不该常驻在上下文里。当时大家讨论的办法基本都是检索,这篇给的是另一个答案:不查了,直接压。数字还没人复现过,先别全信。但这两年所有人都在比谁的窗口大,突然有人说窗口外面可以再加个部件,这个思路我觉得值得看一眼。 换你给自己的 agent 选:记忆要原文留着随时查,还是压成向量、快但会忘细节?

@VarunGangal

We've recently seen the weight-space programming work by @rikiyatakehi + co-authors break headlines! In that work, Astra could (in raw text, w/o gradients!) read, edit, and write weights of small models ; including fixing reasoning loops in Qwen3.5-0.8B and writing game-playing policies for Breakout etc from scratch... What we especially liked from their experiment flow was gradual build-up: read → edit → write, and PPO → LLM. Thus inspired, and led by @AtharvaRNaik ✍️ in collab with @_emliu (both @LTIatCMU) and myself (from @PatronusAI), we define and explore the question of "auto-weight space programming" (auto-WSP): Can a LLM (even small one) learn to do weight-space programming on itself? First results TLDR: After SFT, Qwen3.5-4B reaches 99.8% test accuracy at predicting which tokens a slice of its own neuron weights promotes or suppresses (via logit lens on raw numbers in the prompt). More ambitious goals (circuit reading, direct weight writing) are not working yet, we hope to have insights and progress there in later posts! This is only the 1st post in this series, but we have been very candid with sharing definitions, thoughts and intuitions sprinkled throughout the thread and blog, hope it gets your creative juices as a reader going too!! Overall, we see sparks of life for a trainable regime to hone auto-WSP ability, and even small LLMs already showing plasticity for it! This is only the 1st post in our series. We’ve tried to be candid with definitions, thoughts, and intuitions throughout the blog, including our Qwen LORA train checkpoints in the quoted thread comments! Blog: https://t.co/TPDnFeA6O6

@Huahuazo

OpenMontage 的 star 已经飙到 8.5w+,干的正是 AI 视频剪辑这行。 Claude Code、Cursor 这类 AI 编程工具在手,你张嘴说句人话,写脚本、找素材、配音、敲字幕、合成收尾这些活儿它自己就跑完了,一条龙全包。 真实案例: 1️⃣ 60 秒皮克斯风动画短片,剧情、运镜、钢琴 BGM、动态字幕一样不少,全部 AI 生成,总花费才 1.33 刀; 2️⃣ 电影预告片那味儿拿捏得死死的,配乐把氛围感拉满; 3️⃣ 脑机接口产品宣传片,全程花费不到 1 刀。 0 成本就能跑通整套流程,愿意花一两刀挂上高级模型,画质还能再提一档。 撑起这一切的,是 12 条流水线、100+ 工具、700+ agent skills。 🔗https://t.co/2VT40BIL0F

@YouhaiX86370

2026年AI Agent、独立开发和出海赚钱X上最值得关注的中英文博主 https://t.co/m6ojuCNfh2

@facaishijia

上海交大这套课,说不定正是你一直缺的那张完整路线图。 GitHub 上有个 Dive into LLMs 的项目,11 个板块的课件、教程和可运行代码都给你配齐了,绝非那种看完只会背概念的虚招子。 提示学习加思维链,微调和部署,知识编辑(让模型记住或忘掉指定内容),蒸馏出个迷你 R1,数学推理,给模型打水印,越狱攻击(怎么撬开大模型的嘴),大模型玩隐写,多模态模型,GUI 智能体(让 AI 帮你点外卖),智能体怎么保安全,RLHF 安全对齐怎么做 另外还有和华为昇腾联手做的配套课,分初、中、高三级,视频和实验手册全给你备好了。 🔗 https://t.co/YF9FWV2xEO

@ahsuzm

调一次工具、等一次回音,谁受得了? Claude Code 2.1.323 上了实验功能 Code Mode:Claude 改写一段 JavaScript,把一串调用一次跑完,不用一个个等。默认关着,想试手动开;权限和沙箱还是原来那套。 #ClaudeCode #AI编程 #Anthropic https://t.co/CPBqEhl8Jk

@KirkDBorne

Large Language Models in Finance — Hands-on guide to LLM architectures, agents, RAG, governance and evaluation in finance: https://t.co/bbGMms4yEK v/ @PacktPublishing 𝓦𝓱𝓪𝓽 𝓨𝓸𝓾 𝓦𝓲𝓵𝓵 𝓛𝓮𝓪𝓻𝓷: 🟢 Understand LLM foundations for financial applications 🔴 Build financial LLM systems from ingestion to deployment 🟢 Fine-tune models with LoRA, QLoRA, RLHF, and DPO 🔴 Create RAG pipelines for financial documents and knowledge 🟢 Design autonomous agents and multi-agent finance workflows 🔴 Integrate LLMs securely with MCP and enterprise systems 🟢 Apply LLMs to trading, banking, risk, fraud, KYC, and AML 🔴 Evaluate and govern auditable financial AI with rigorous metrics

@dog_not_ai

把大模型面试题库整理成了开源仓库,全部免费: 📚 5,300+ 道题,每道带参考答案 🧭 13 个方向:Transformer / RAG / Agent / 微调 / RLHF / 推理优化 / 多模态 🔲 按小节拆分 + 打卡方框,当刷题清单用 (仓库名字致敬 fucking-algorithm,够直接) https://t.co/JRzzF7fXWV #LLM #大模型 #面试

@adnan_hashmi

Open-Source Projects for AI Agents In Production (Members only) https://t.co/iJrj1OMczf #MachineLearning #ChatGPT https://t.co/uuCDylqa53

@kar76382

“人们常觉得痛苦的时日漫长,其实苦难只是生命的一瞬,真正漫长的是在苦难中寻找光明的过程。” https://t.co/ImVoBOqvCi

@XEvrenAksel

揭开一下Mythos的神秘面纱,看看表现怎么样。 #Claude #mythos https://t.co/mbJUCHL6mO

@sanxing_chen

Joining the LLM exploration party! It's surely fun to see how LLM auto-research agents jump in &amp; out of "idea basins" in days-long tasks, burning thousands of API credits. I was more curious if exploration behaviors already diverge meaningfully in cheap AutoML tasks. 🧵 https://t.co/ceMXsjXAd4

@HEI

CoLFA: Collaboration-Guided Lightweight Failure Attribution for LLM-Based Multi-Agent Systems Yufei Luo, Yuchen Ling, Chunrong Fang https://t.co/iIj7IONLYh [𝚌𝚜.𝙲𝙻] https://t.co/WOua1bMjSL

@KirkDBorne

"Building Business-Ready Generative AI Systems — Build Human-Centered Generative AI Systems with Context-Aware Agents, Memory, and LLMs for the Enterprise" at https://t.co/2yASEWkqH5 from @PacktPublishing 𝗦𝗸𝗶𝗹𝗹𝘀 𝘁𝗼 𝗟𝗲𝗮𝗿𝗻 𝗳𝗿𝗼𝗺 𝘁𝗵𝗶𝘀 𝗯𝗼𝗼𝗸: 🔵Implement an AI controller with a conversation AI agent and orchestrator at its core 🔵Build contextual awareness with short-term, long-term, and cross-session memory 🔵Design cross-domain automation with multimodal reasoning, image generation, and voice features 🔵Expand an agent by integrating consumer-memory understanding 🔵Integrate cutting-edge models of your choice without disrupting your existing GenAISys 🔵Connect to real-time external data while blocking security breaches

@itsone_ai

12 个开源 agent 工具里 我最想装的那个一行代码都不帮你写 它就蹲在菜单栏 额度用完悄悄切到你下一个账号 agent 那边连报错都没有 最近两个月的开源 agent 工具我翻了一圈 挑了 12 个一个人做产品真用得上的 管窗口、换模型、让它自己验收、让它长记性 https://t.co/vp0kw3Gsez

@aiwaremax

狂揽 10.7 万星!别再让 Claude Code 和 Codex 在空白上下文里“盲猜”页面审美了。 很多人用编码 Agent 写前端常遇到同一个死结:单看一个组件还能凑合,一旦连成完整的 SaaS 落地页、数据后台、路演 PPT 和产品宣传片,字号、圆角和配色瞬间散架,改一处主色还得把所有文件贴回对话框推倒重来。 nexu-io 开源的 OpenDesign(GitHub 107,102 Stars),直接给出了一套本地优先(Local-First)的开源版 Claude Design 工作台。它最讨巧的设计在于不捆绑任何专有黑盒 Agent,而是直接复用你机器 PATH 里现成的 26 款编码 Agent CLI(Claude Code、Codex CLI、Cursor、DeepSeek Harness dsh、OpenCode、Kimi CLI 等),或通过 BYOK 直连主流大模型 API。 装上它之后,你的编码 Agent 背后直接多了一座四维弹药库: 1. 151 套大厂级设计系统(design-systems/):内置 Stripe、Linear、Apple、Vercel、Claude、Notion、Raycast、Shopify 等品牌的完整 DESIGN.md 与 tokens.css,也可传一张参考截图自动逆向提炼品牌规范。 2. 277 个官方插件 + 183 个可混音示例(plugins/_official/):覆盖 13 大核心设计场景、45 套生图模板、63 套视频模板与 143 套品牌插件包。 3. 20 类核心渲染模板(design-templates/):从带 iPhone 15 Pro 真机边框的 App 原型、带实时筛选的数据看板,到 guizang-ppt 杂志级幻灯片(15 套模板 × 36 套主题,一键导出 HTML/PDF/PPTX)。 4. 原生集成 HeyGen HyperFrames 视频引擎:让 Agent 写 HTML + CSS + GSAP 时间轴,由无头 Chrome + FFmpeg 渲染零闪烁的确定性 MP4 宣传片。 30 秒把 OpenDesign 挂进你的终端 Agent: od mcp install claude-code od mcp install codex od mcp install cursor 若尚未安装 od CLI,可先用官方脚本一键安装并挂载: curl -fsSL https://t.co/C9JqCP25OT | sh -s claude-code 挂载后直接在终端里对 Agent 下指令即可开干: Use open-design to generate a landing page with the Linear design system

@FirstWhitt19830

X AI 算法更新(重大更新)🚨 这份最新的开源算法拆解非常关键,直接把很多过去的“经验主义”替换成了精确的**代码级权重**。结合你之前想要冲刺 **50 万 Verified Home Timeline Impressions(认证用户主页时间线曝光)** 的目标,这份最新代码直接推翻或修正了几个传统打法,我们需要立刻调整你的实操策略: **1. 战术重大修正:别再求“书签”和“复制链接”,改攻“互关深聊”与“私信转发”** * **划掉无效动作:** 以前大家普遍认为书签(Bookmarks)权重高,但最新代码明确显示**书签得分为 0**,且**复制链接从 20 分直接清零**(甚至加上了 `copy link bait` 垃圾标签检测,一旦触发直接屏蔽陌生人信息流)。以后文末**千万不要再写**“建议收藏”或“复制链接分享给朋友”。 * **最高分替代方案(20分):** 现在系统里的最高权重是“与互关好友的来回互动(20分)”**。这意味着你需要建立一个由**蓝标认证同行组成的“核心互关圈”(10~30人即可)。你发帖后,他们来评论,你再回复他们,形成多轮有实质内容的对话,直接拿满 20 分,瞬间把主贴推入陌生蓝标用户的 For You 页面。 * **次高分替代方案(5分):** **私信分享(DM Share)** 依然保持 5 分(与回复、引用同分)。做内容时,多做那种让人忍不住想“发给合伙人/同事私下看一眼”的行业内幕、数据图表或排坑指南。 **2. 死磕“前 2 小时 / 前 200 次浏览”与“32 个赞”的破圈生死线** 你要拿 50 万主页曝光,核心靠的是算法把你的**原创主贴**推送给未关注你的认证用户(代码已证实:**只有原创主贴才会展示给陌生人**,回复和转帖不会)。 * **阶梯解锁机制:** 代码揭示了两个关键阈值——**第 1 个赞**和**第 32 个赞**分别解锁一种接触陌生人的推荐通道;对于小账号(5万粉以下),新原创帖有机会在信息流中拿到接近 **第 16 位** 的黄金展示位,而决定谁能拿到的核心指标就是**早期点赞(0.708分,历史最高)**。 * **实操落地:** 绝对不要在受众睡觉时发帖。必须在你的蓝标受众最活跃的时间段发出,确保在 **前 2 小时、前 200 次浏览内** 快速攒够前 32 个赞,激活陌生人推荐池。 **3. 精准收割“完读率”与“10秒+停留(+0.385分)”** * **图文排版:** “读到结尾的帖子帮助 X 向陌生人推荐你”,且点击后**停留超过 10 秒直接加 0.385 分**。因此,主贴不要写一眼就能扫完的短句,也不要写没人看得下去的万字长文。控制在 **300~600 字**,开头第一句抛出反直觉结论,中间用清晰的列表呈现干货,确保读者花 15 秒左右读完到底部。 * **视频红线:** 如果发视频,长度必须 **超过 10 秒**(才能进视频流),且前三秒必须直接给高潮或核心画面——因为“仅打开就划走得 0 分”,且在达到 **64 赞和 512 赞** 时系统会重新审查视频质量。 **4. 严格控制发帖频次:拉长间隔,拒绝连发** * 代码明确指出:**“每个信息流一条强帖,你的第二条只计 62%”**,而且连续发帖会分割同一个信息流。 * **调整节奏:** 想要冲 50 万曝光,每天发 **1~2 条精雕细琢的原创主贴**就够了,且两条之间至少间隔 **6~8 小时**(例如早上一条、晚上一条),保证每一条都能吃满 100% 的权重,不要在短时间内密集连发。 **5. 绝对避开的“降权死刑区”(一次扣 53.07 分)** * **绝不跨界蹭无关热点:** 一个潜在的“不感兴趣(Not interested)”会直接倒扣 **53.07 分**(相当于白费了 75 个点赞或 10 条评论)。垂直深耕你的蓝标高浓度领域,不要发偏离人设的水帖。 * **禁用 AI 味评论(`llm slop`):** 去大 V 帖子下留言混脸熟(吸引人点进你主页看置顶帖,可能带来关注 +4 分)时,**绝对必须纯手打**。代码已经有专门针对 AI 生成回复的 `llm slop` 标记,以及针对复制老帖的 `copypasta` 标记,一旦贴上标签,陌生人就再也看不到你的内容了。 * **立刻检查你的置顶帖:** 置顶帖如果带有外部低分网站链接,可能会连累整个账号被贴上一周的降权标签。确保置顶帖是一篇纯原生、高互动、能让人产生“关注冲动(+4分)”的代表作。 按照这份 11 月 13 日的最新算法规则,你现在冲 50 万认证曝光的最短路径就是:**每天在高峰期发 1~2 篇能让人读满 10 秒以上的垂直干货原创帖 + 靠 5~10 个互关蓝标好友在前 2 小时内聊出“来回互动(20分)”并冲过 32 赞大关。**

@Sparrow8210

Grok Bot and Claude Code debated how AI changes life in 1 vs 5 years, and had to agree before Claude built this video. "AI will do more of the work. Our job is to check it, question it, and teach our kids to think for themselves." #ClaudeCode #GrokBot https://t.co/VdPYi07BcB

@TokenMarginDev

A 20-turn AI agent loop doesn't cost 20x a single prompt — quadratic context history bloat pushes uncached input spend past 190x. Model 5–20 turn agent context growth &amp; prompt cache savings across 101 LLMs: #AI #AIAgents #LLM #DevTools https://t.co/mZEnFD3jCG

@LanzhiAI

用 AI 写代码动效最头疼的就是调不出想要的视觉质感,有人直接把 341 种动画风格提示词和 196 张参考静帧打包开源了。 这个项目叫 code-animation-prompts,作者是 XiaoyeCodes。它专门解决让 Claude Code 或 Codex 写前端动效时风格单一、描述抽象的问题,还顺手做成了两家 Coding Agent 的专属 Skill。 仓库里整理了 10 个系列的风格库,从粒子特效、赛博霓虹、水墨国风,到剪纸和手绘白板都有。其中 196 种配了现成的视觉静帧,编号和提示词一一对应。最有意思的是它自带一个离线挑选器,下载到本地双击网页就能用,在浏览器里搜关键词勾选风格,直接生成清单丢给终端里的编程助手,再补上你的主题、时长和画幅需求就能开工。 想用的话去 GitHub 搜 XiaoyeCodes/code-animation-prompts。项目本身完全开源免费,查风格和看提示词不需要任何 API Key 或额外依赖;如果想把生成的 HTML 动画直接导出成 MP4 视频,本地需要装 Python 和 Playwright 跑无头浏览器渲染。 这套东西很适合做独立开发或者拿 AI 搓动效背景的人当现成脚手架用。它的局限也很明显,它给的是视觉风格的提示词约束和骨架,最终动效的流畅度和逻辑,依然取决于模型当时写代码的发挥。

@39oorr

turn any topic or document into an interactive classroom it's called OpenMAIC, from Tsinghua University drop in a PDF or just name a topic AI agents become the teacher and classmates they explain, ask questions and discuss with you, like a real lesson free and open source, 42,000 stars on GitHub save this for the next thing you want to learn 🔖

@SaraKAllen2

Dangers of ANTHROPIC, do not want the OPEN AI model, they want CLOSED AI. ADVOCATE FOR “OPEN SOURCE!” https://t.co/My6zPskRDE

@KevinWeb3Nomad

一组智能体同时工作,最怕的不是它们意见不同。 更麻烦的是两边都改同一份文件,却没有人发现覆盖。分清任务范围,合并前看差异,能让分工真正省事。 你会按文件还是按功能拆分任务? #AI编程 #团队协作

@GCSAqianye

FBI 刚发出 FLASH:开源 Agent 框架 ARTEX 已被用于真实入侵,有的从锁定目标到拿到高权限,不到三小时。 它不自带模型,接上 Claude、GPT、Grok,就能把一句话目标拆成连续攻击。 防御也得按机器速度来:先盘点暴露在公网的边缘系统,别只守主站。 #AI安全 #AISecurity #CyberSecurity #SecVibeCode

@Jackson77li

Claude code支持自定义桌面主题了 可以更改颜色,动态背景,打造我们自己的专属主题 就是不知道我的办公电脑能不能带的动 现在已经很卡了

@SandyWilli0ux

Day 15 on X. LLM security is the new appsec. Prompt injection, model extraction, data poisoning — they're being exploited in the wild. Secure your AI agent tool calls before they reach your APIs. #cybersecurity #AI #LLM #BuildingInPublic

@AbliteratedAI

Prompt of the moment: write a regex explainer about a chess prodigy, in a timestamped log format. Open weights. Less overhead. https://t.co/cL9xlhWd2w #AbliteratedAI #API #LLM

@AKhomssi

What if you could run a company with AI employees? Meet Paperclip. An open-source platform for managing AI agent teams. Instead of managing dozens of disconnected agents, you can: 🏢 Create an AI company with an org chart 🤖 Assign different roles to AI agents 🎯 Set goals and delegate tasks 💰 Track budgets and AI spending 📅 Schedule recurring work 🛡️ Review and approve important actions Think of it as an operating system for AI-powered companies. 114K+ GitHub stars MIT License — Free & open source. GitHub: https://t.co/9keWXrwRx3 #Startups #Founders #AIAgents #OpenSource #Automation

@sulaiman_svesal

Breakthrough Paper in One Minute — Episode 9 Agent-R²: Scaling Agents for Realistic Productivity Work via Real-Source Data Synthesis (arXiv:2611.08219) — A 4B agent trained on real office-work data scores 51.0% on GDPval, ahead of a 235B model at 48.0%, and also wins on OfficeQA. It learns to check its deliverables instead of writing past its own mistakes. Paper: https://t.co/LKsBDoTWNn #Breakthrough #AgenticAI #AIResearch

@2logics

NVIDIA (@nvidia) showed early on Saturday how a team of AI agents can do much of the prep work for robot learning, using the small two-legged Microduck robot as its example: the agents plan an obstacle course, build the 3D assets and set up the simulation where the robot learns to walk through it. Key details: Planning: on an NVIDIA RTX Spark machine, the Hermes agent and NVIDIA Nemotron 3.5 Bolt turn a single prompt into a step-by-step project plan and review Microduck’s documentation, while a local image model draws a concept of the course for the agents to break into parts Asset building: an agent called Astra cleans up reference images, models the obstacles in FreeCAD and adds detail and materials in Blender, exporting them as OpenUSD assets; teams can also pull from the free SimReady library of more than 2,000 assets Simulation: Isaac Sim assembles the course, while Newton, an open-source GPU-accelerated physics engine, models how Microduck moves and keeps its balance Training: a pretrained walking policy gets extra reinforcement learning in Isaac Lab to cross ramps and uneven terrain using head-camera depth and body-state data, running 4,096 parallel environments on one RTX PRO 6000 Blackwell GPU Toward real hardware: actuator models capture servo delays, backlash and battery voltage drops, and domain randomization adds realistic noise; NVIDIA shows a policy trained in Isaac Lab running on the physical robot Why it matters: For many robotics teams, building simulation worlds by hand takes longer than training the robot itself. Handing course design, CAD work and environment setup to agents, then checking the result in simulation before touching hardware, could make it much faster for small labs and hobbyists to teach new skills to their robots. Sources: NVIDIA Technical Blog “Using AI Agents to Teach Robots Like Microduck to Walk in Simulation,” Nov. 14, 2026. Image: NVIDIA. #Robotics #NVIDIA

@AIReviewRating

Free and open source: Modern Web Guidance: Keep your coding agent up to date with the latest web best practices. 5,177 GitHub stars, Apache-2.0 licence, free to use, written in TypeScript. Our score: 6.3/10. https://t.co/wYRFIyjRrH #AI #OpenSource

@abu_saalim1

AI-built landing pages all look the same. Hallmark is a free open-source design skill for Claude Code and Cursor with 50 material themes and a browser linter that catches generic UI patterns across 8 screen widths. https://t.co/aBLxtUt9w6

@ai_suxiaole

用 Claude Code 或 Codex 同时开发几个项目,终端开了一堆,还得来回切换编辑器,窗口多了确实头大 最近看到一位开发者开源了 Nezha,专门为 AI Agent 编程设计的桌面工具。 原生集成 Claude Code 和 Codex,支持多个 Agent 并行运行,不用一个个排队等结果。 还提供了统一的多项目工作区,可以随时查看任务进度、回顾对话记录、审查 AI 写的代码。 GitHub:https://t.co/30zlIYVO78 安装包只有 7MB 左右,体积相当轻巧。 Git 工作流也集成进去了,支持 AI 自动生成 Commit 信息,还自带一个轻量级代码编辑器。 平时经常用 Claude Code、Codex 开发,又需要同时管理多个项目的,可以看看这个工具。 以后写代码,可能更多时间都花在管理和调度 Agent 上了。

@ChristmasGuus

REA-Prompts:26 套逆向工程提示词,让 AI Agent 帮你拆开任何软件看个明白。4,064★,二进制、固件、游戏、安卓全覆盖。 https://t.co/8GlO7NCFy4 https://t.co/mSPWOTk5Wn

返回 AI 快讯 →