2026年11月21日 · AI 晚报
本轮讨论从旗舰模型与使用额度,延伸到本地开放权重、智能体任务编排和交付检查。研究方面,数学家与 Claude 的协作、短任务训练如何迁移到长任务,以及机器人推理的延迟成本,提供了比单看榜单更具体的观察。
11月21日 11:00 — 11月21日 23:00(日本时间) · 132 条 X 原帖
本期目录
旗舰消息之外,产品入口与使用成本同样受关注
Gemini 新模型的媒体转述与 Anthropic 行情消息
The Daily Block 报道 Google 推出 Gemini 4 Neon,面向复杂编程、分析推理和企业工作流,并转述 Google 关于部分基准领先的说法;本批没有具体成绩或官方原帖。Claude Saga 6 则主要出现在社区猜测中,不能据此确认正式发布。
Automater 的行情简报称,Anthropic 收跌约9%,报88.30美元,接近其所述83美元发行价。这提供了一条上市后市场表现的转述,但帖子没有交代上市时间与股价变化原因。
来源 · 3 条原帖
办公助手与项目会话的两种入口
DaBin.Binah 介绍 AWS Quick,称其使用 Claude Opus 5.5,免费版每月提供10小时输出时间,支持 Windows 和 Mac。这是用户对产品与额度的介绍,不代表本轮刚刚官宣。知县则体验了 Claude Code 的 Project 会话,认为复杂任务跨会话协作更连贯,部分替代了自己原来的工作方案。
来源 · 2 条原帖
额度优惠与消耗焦虑并存
DeepAI 宣传其 Code 客户端可用每日免费额度调用多款 Claude;Hardy Chen 分享 ChatGPT Pro 关联 API 组织领取信用额度的个人经历。另一边,LEO 称一晚用掉 Copilot Pro 月额度的三分之一。免费策略也有经营成本:Adhiraj 自述发现500多个 GitHub 仓库针对14款 AI 产品批量获取试用额度。
本地模型开始围绕具体机器和任务讨论
Inkling 2 Mini 的循环结构与笔记本体验
社区将 Inkling 2 Mini 描述为开放权重的26B MoE,激活参数3.4B、体积约20GB,中间层每个 token 循环三次,借此增加计算而不延长输出。Javi 自述在16GB MacBook 上完成翻译、摘要和少量代码测试,体验良好;帖子没有说明量化版本,不能把这次体验与上述体积直接对应。
来源 · 2 条原帖
从桌面运行扩展到训练工具
Daniel Han 介绍 Unsloth Desktop 的跨平台支持,覆盖多类 GPU、CPU 与 ARM64,并将语言模型训练、强化学习、扩散及视频生成纳入同一工具。Michel Bauwens 分享的 OpenDiLoCo 则关注低通信量的分布式训练,让讨论从单机部署延伸到分散算力如何参与训练。
来源 · 2 条原帖
研究焦点:专家协作、长任务泛化与推理成本
黎曼猜想相关工作,核心是数学家的迭代参与
围绕 Levent Alpöge 与 Claude 的数学工作,社区出现“已经攻克”的说法,但 Quanta 的原文只描述向结果推进的迭代尝试,不能据此宣布猜想已获证明。其强调的区别是:数学家的专业判断深度参与了过程,而非简单给模型一道提示词。
来源 · 2 条原帖
用短任务训练智能体如何组织长任务
alphaXiv 介绍的研究把重点放在 Harness:通过上下文管理、工具调用和递归委派,将长任务拆成较熟悉的小问题。研究采用简单合成环境及模拟工具调用训练;据其转述,短任务训练可迁移到长2至64倍的任务,RLM-27B 在长上下文基准上提升26.9%,相同 Harness 下推理 token 用量减半。
本批另有闭环智能体量化设计 QDA,以及智能体自我改进是否仍需外部指导的论文条目,但仅提供题目,尚无方法细节或结果。
来源 · 3 条原帖
知识注入与确定性计算的两条路径
DeepStarts 转述 AtlasKV,称可用不足20GB显存将10亿条知识图谱三元组放入 LLM,但没有展示足以判断其能否替代 RAG 的对比。ArXiv Updates 则介绍一种零梯度方法,将确定性的数学与逻辑电路直接写入 Transformer 权重,声称无需微调即可达到六位数算术100%准确率;这一结果应限于帖中所述任务。
来源 · 2 条原帖
采样加速瞄准服务端延迟
Yuvraj 介绍 SonicSampler,将 logit 处理、top-k/top-p 与推测验证融合成兼容 CUDA Graph 的 Triton 内核。其转述的结果包括 top-k 最高10倍、推测解码端到端最高16倍加速,并支持混合采样、惩罚及语法约束;具体收益仍取决于测试配置与基线。
来源 · 1 条原帖
机器人需要决定何时再次思考
朱俊帆梳理机器人研究时指出,高层推理可调用现成模型,动作控制仍受延迟和反馈制约。他转述 AMP1 每次成功任务约耗时280秒,其中约211秒用于思考;另一方案由 Astra 规划、π0.5执行、4B模型检查完成情况,报告79.1%的成绩,但未测 token 与美元成本,且仿真会等待推理暂停。他主张重点建设低成本监控层,决定何时唤醒昂贵推理。
来源 · 1 条原帖
智能体工程从堆工具转向复用与任务约束
把一次成果整理成可迁移的技能
0xSero 发布 Skill Distill,尝试把模型完成任务的过程整理成 Skill,供开放权重或本地模型复用,并展示用 DeepSeek 复现 GPT 工作的例子。Graft V2 则由项目方宣布支持从公开仓库提取智能体配置,在浏览器内转换到25种平台。兼容性也开始进入检查流程:The Agentic Daily 介绍了集成 SkillEvaluator、面向多种 Harness 的测试与 CI 检查。
来源 · 3 条原帖
编排方式应随任务变化
Unicorn 的反重力使用指南区分了规划执行分工、后台子代理、可拆块并行任务、复杂单点排障和跨天工程,提醒并发会快速消耗额度。E=mc² 的个人案例则强调缩小工具范围:限定任务和两三种必要工具,比预接大量 MCP 更有效。varun 分享的开源实践由 Claude Code 委派 Meta Muse 再检查结果,首次任务仍漏了两个答案,说明委派之后仍需验收。
来源 · 3 条原帖
让工具提供可理解的上下文
mireia 介绍 cli-anything,将桌面软件操作转成智能体可用命令;CodeGraph 则为 Laravel 映射代码与数据库结构,让 Claude 能从报错追到页面、路由和表。输入输出两端也有对应工具:Firecrawl 将网页转成干净 Markdown,Real Python 的 Pydantic AI 教程则让智能体返回经过校验的 Python 对象,减少直接解析原始字符串的工作。
记忆、凭证与审查成为交付流程的一部分
记忆需要能查看,也需要能迁移
Archivum 宣布可视化应用上线,支持导入文件、生成关联 wiki、查看知识关系和搜索记忆。项目方称基础组件本地运行,Dream 整理步骤可连接自选模型或现有智能体,结果由人审阅后应用;它也明确承认,按模式清理密钥无法覆盖所有秘密。Row-Bot 5.3.4 则宣传从多个助手导入记忆,回应更换产品时的迁移需求。
来源 · 3 条原帖
凭证隔离与过程记录
Tucano 0.2.44 宣布加入按名称使用、无需智能体看到内容的凭证管理,并增强重试及智能体协作。AI Now 转述 OpenCode Console 支持 CI 的 OIDC 身份认证,减少长期 API token 的使用。Atlas 则记录会话、提交、提示词和工具调用检查点,把代码变更与生成过程关联起来。
来源 · 3 条原帖
发布前检查覆盖代码、安全与商店规则
OpenQodex 在推送前调用独立审查智能体,并先运行静态扫描;Claude Bug Hunter 将公开漏洞报告整理成授权安全测试用的技能包。AI 产品普洱介绍的 app-store-compliance 则把拒审经验、条款与修复建议接入编码助手,在发包前检查隐患。这些项目分别处理代码问题、漏洞模式和上架条件。
来源 · 3 条原帖
交叉检查有用,但仍是检查机制
Lihai La 发布 MIT 开源的 Dissentry,让其他模型对回答提出异议,并标出争议句、合并重复意见、记录检查成本。WanderAI 作者提供了一个具体案例:测试通过之后,独立智能体仍发现锁屏可能导致旅行计划丢失的漏洞。前者组织多模型审阅,后者说明审阅可以补充测试覆盖。
来源 · 2 条原帖
代码创作扩展到视频、绘画与科研图
视频技能要连同许可和依赖一起看
OpenMotion 被社区介绍为可参考视频、结合品牌素材重建场景与转场,再编辑导出 MP4 的开源工具;OpenDesign 也推荐 DeepSeek 配合其 MOTION.md。Non 对热门视频技能清单的核对更具体:Remotion 有团队规模相关的商业许可条件,Manim 技能不会代装引擎与依赖,“Motion Skills”也并非唯一项目。
来源 · 3 条原帖
创作案例开始展示制作过程与成本
cookiesheep 自述约花500元,让 Claude Fable 5.2 用代码建模和路径追踪完成4800帧的4K《桃花源记》短片。卡皮展示的雨夜面摊则是可旋转查看、已公开源码的网页场景。Aksh 介绍的 aibrush 采用逐笔绘制、检查画布再重画的流程,三者都把可编辑的制作过程放到了作品背后。
来源 · 3 条原帖
模板与小目标让创作更容易落地
Roxy 分享科研绘图提示词仓库,以开放许可示意图为参考,通过替换课题信息减少反复试提示词。Haosen 的游戏路线则先把目标缩到能完整玩一局,再覆盖设计、拆任务、Unity 调试与上架准备。两类实践都在用明确模板和可完成的目标,约束开放式生成。
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 132 条原帖
@0xSero
New skill is live! Skill Distill, I'm sure we all love Opus-5.5 so if you've ever been impressed by it's output, now you can have it insta-distill the work it did into a skill for open weight/local models. Done with GPT, recreated via skill w deepseek https://t.co/HsjBZav0MA https://t.co/mUHGz7tHnx
@julien_c
> 1.8x faster than Claude Sonnet 5.5 on an M5 Ultra Mac Studio and 11x cheaper (API pricing). And this, my friends, is how local and open source AI is going to win 🔥🔥🔥
@DeepAI
Wait… free? For real? Yes, really. Claude Fable 5.2, Claude Opus 5.5 and Claude Sonnet 5.5 on your free daily Theos in DeepAI Code. And many more models. Mac, Windows & Linux. No card needed. Download free 👉 https://t.co/DM4IdCE3yG https://t.co/cS3V4m5UV4
@AxelGarciaK
Do we have local Claude now? Great release by @thinkymachines! 🎉 https://t.co/NS9EOA9EUi
@yedabinbinah
亚马逊版 WorkBuddy ,免费用Claude Opus 5.5! AWS-Quick是亚马逊旗下的办公AI,类似于中国的WordBuddy 背后的模型是 Claude Opus 5.5满血版 根据AI输出的时间来计算的,不是按Token,10小时输出,实际可以用挺久 客户端只有windows 和 mac,免费版每月 10 小时,每月1号重置 重点:有亚马逊账户的授权登录就能用,多个亚马逊账户可以切换用 URL:https://t.co/LRW4rR9YD5
@QuantaMagazine
The approach to solving the Riemann hypothesis is markedly different from the 377 results OpenAI released in October, almost all of which were the result of a simple prompt to an AI model. Instead, Levent Alpöge’s mathematical expertise played a major role in iterative attempts with Claude to move closer to the result. https://t.co/yOtgd8SDUO
@everlast_ai
Ein Entwickler baut mit Claude die komplette Adobe Suite nach. Was heißt das für SAP? @E_Boeminghaus: Sobald der erste Konzern sein ERP selbst mit KI baut und Millionen an Wartung spart, „fängt die Uhr für SAP an zu ticken“. https://t.co/sItwUu9YgS
@TuyetloanP81840
https://t.co/pwJ6iFH3H1
@mehulmpt
Docker buying DeepSeek would be the smoothest brand merger ever
@askalphaxiv
"Language Model Harnesses Can Be Compositional Generalizers" What if LLM agents could solve much longer tasks simply by learning how to break them into smaller and more familiar problems? This paper explores how agent harnesses can improve generalization by managing context and delegating subtasks through tools and recursive LLM calls. So rather than training on increasingly difficult tasks, the researchers introduce surrogate harness training, which uses simpler synthetic environments with simulated tool calls to teach models how to organize their work. Their experiments show that training on short tasks can generalize to tasks 2-64x longer. Using synthetic training, their RLM-27B improves 26.9% on long-context benchmarks, while using 2x fewer inference tokens than the original model in the same harness. https://t.co/iuJGa503hp
@Dipanshu_AI
agentic engineer in 2026. Andrew Ng, the godfather of AI, gave the complete playbook to become one from scratch. 1 hour course. Free: • 00:00 - AI agent basics • 12:12 - AI Agentic workflows & design patterns • 53:27 - Practical tips for building AI agents • 1:20:30 - self-improving AI agent loops • 1:30:19 - multi-agent AI systems I watched it last night. Halfway through, I realized I could get into Anthropic in weeks, not years. Bookmark now. Watch it. Then build your own AI agent in.
@liyue_ai
https://t.co/4dxKMMUJIQ
@Li_mima
我的宝藏梯子 https://t.co/wuoISJQ95o ,全AWS线路,反重力、Codex全解锁;晚高峰油管4K秒开 性价比更是爆表,每月512G年付才¥49.9 https://t.co/Tc25QCvVfR
@thedailyblock
🚨 BREAKING: Google Launches Gemini 4 Neon to Challenge Claude Opus 5.5 and GPT-6.2 Sol! Google says its new AI model is built for complex coding, analytical reasoning and enterprise workflows, claiming it outperforms competing models on several benchmarks.
@BrillaConElla
The most powerful AI workflow isn’t one agent doing one task It’s your entire process working together A new lead comes in → AI scores it → an agent drafts the email → a human approves it That’s ClickUp AI Workflows: connect agents inside and outside ClickUp add human review, and route tasks to the right AI model all in one visual flow
@jgonzalezferrer
You don't need an expensive computer to run a really good AI model at home anymore! Inkling 2 Mini just came out. It's open, free, private and works with no wifi I ran it on my MacBook with 16GB and tested it on what I use AI for every day: translating, summarizing and a bit of code. Worked really well!
@AdrianPunk115
https://t.co/9oc8Km5B8D
@m4lkxx
this is f*cking gold... I built a REPO-MAT 21 repos to use with Opus 5.5 / Fable 5.2 WhisperX. librosa. Manim. Rapier. WebGL Fluid. three.quarks. frontend-design → UI that feels hand-made https://t.co/GeNXxJTLnQ motion-design-skill → timing, easing and the 12 principles https://t.co/OYzE9gMiqW gsap-skills → smooth animation from the GSAP team https://t.co/Slsg73bm74 Remotion Skills → video skills for any coding agent https://t.co/SgWRiLEQt4 Impeccable → polish and critique any design https://t.co/qJXOrZ4RLG HyperFrames → write HTML, get a video https://t.co/VwRbGxthfe Remotion → video in React https://t.co/t1F2ErXxXV three.js → 3D in the browser https://t.co/V6zTkj4GeA Blender MCP → your agent drives Blender https://t.co/eYk5Hiz1AC Manim → animated diagrams and equations https://t.co/8BxzSxrUie lottie-web → lightweight vector animation https://t.co/CjVz3j4fua three.quarks → particle trails and VFX https://t.co/RD4HXeYlWT react-postprocessing → bloom, depth of field and film looks https://t.co/pKGa81cF4p react-three-rapier → real physics in 3D https://t.co/fn0G1Rlvji WebGL Fluid → smoke and ink that move https://t.co/M59nphCLPM Rough.js → a hand-drawn feel https://t.co/tjq0xq6vnl Tweakpane → sliders for every parameter https://t.co/35xh3uvdw1 librosa → beat detection for music sync https://t.co/1YMbCg2Agw WhisperX → word-level captions https://t.co/HGPBKfyHaZ rembg → cut out any subject https://t.co/HNkqhoA7ra FFmpeg → export the final video https://t.co/CX3SC5XYjI Static → animated → 3D → physics → sound → shipped Bookmark this and send the list to your Claude Then read how motion engineers check every frame in the article below ↓
@Polymarket
🚨 NEW POLYMARKET: Apple releases autonomous AI agent by...? https://t.co/xjbvIqxWJ9
@berttemme
AI Agent: The AI agent for everything you do in QGIS (plugin) https://t.co/kC7i39bkbb
@miles_mazy
FDE到底赚不赚钱? 今天一篇文章讲清楚 https://t.co/PhxbuIupcf
@xiping_ai
给Agent打造一个共享知识库 不仅接入新Agent不用重新介绍自己,还可以让多个Agent共同去完成一件事,比如Codex做计划,Grok bot做执行,Codex去检查。 相当于把自己的大脑做了外挂 让Agent能够去接入使用 https://t.co/uUcK014aTq
@shengxj1
愉快面基 程序员面基两大话题 1. claude被封了咋办 2. web3真是没啥项目可以干了😂😂😂 https://t.co/EMyO6zRFyP
@danielhanchen
Thank you! Unsloth Desktop works on Windows, Mac, Linux + NVIDIA, AMD, Intel GPUs, CPUs, ARM64! We also support diffusion, video gen + LoRA training, LLM training, RL, decision models, auto compaction, vLLM + SGLang (Windows via WSL), Cloudflared HTTPS, audio, sandboxing + more!
@xKaka1456
碾压 Claude Fable 5.2 的最强模型 Claude Kaka 你们随便用😔 https://t.co/9PvScwoxKo
@DeepStarts
RAG might be getting a serious alternative. Researchers introduced AtlasKV, a method that can put 1 billion knowledge graph triples into an LLM using less than 20GB of VRAM. That's a lot of structured knowledge for such a small memory footprint. https://t.co/8Fbzqurwqa
@DrRabbitAI
很多人用 AI 画科研图都在抽卡,同一段提示词跑十次,出来十张不一样的图。 分享一个开源的科研绘图提示词仓库,它参考了 Nature、Cell 系列开放许可的示意图。找到最接近的模板,把【】换成自己的课题就能用。 https://t.co/WwyqBuVKbm
@Haosen9527
让 AI 做第一个游戏,先把目标缩到能完整玩一局。 从玩法、GDD、任务拆分,到 Unity、调试和上架准备,整理成了 9 步路线,每步都有可复制的提问模板,也附上了 4 个 Skill 与 Unity MCP 的入口。 https://t.co/H5yIPr5wa1
@magicblock
Building for Blitz v10? Here's every resource you need to ship faster 👇 🏁 Register and start shipping https://t.co/Drj8S7TbOm 📚 Docs for Ephemeral Rollups, Private ERs and VRF https://t.co/EiQg4VW811 🤖 Teach your AI coding agent MagicBlock in one command https://t.co/4RFyk308dN 🤝 Find a cofounder, past projects and the builders group https://t.co/C265FDrgH9 💡 5 Entertainment Finance ideas to get you started https://t.co/qyo4qOhkvr 🧩 Templates for gacha, dice, predictions and auctions https://t.co/lKDiGD5Na1 💻 Example repos to clone and run https://t.co/77d54f6Xmt 🛡️ Let your AI agent audit your Solana program before you submit https://t.co/hBN5fc2ePw
@YuvrajInfer
New paper: SonicSampler — a suite of tile-aware Triton kernels that fuses the entire LLM sampling pipeline (logit processing → top-k/top-p → speculative verification) into one CUDA Graph-compatible kernel. Key results: • Up to 10× faster top-k (hierarchical two-stage algorithm) • Up to 16× end-to-end speedup on speculative decoding vs strong baselines • Supports mixed greedy/stochastic sampling, penalties, grammar constraints, etc. in a single batched launch Perfect for low-latency serving and EAGLE-style speculative decoding. https://t.co/dw5yrXPvuw
@yedabinbinah
国产 AI 每周送 2500 积分,支持 API 调用! 原定 10 月 20 日结束的免费活动,现在又延期了! 免费模型:编程类 U2-Flash、医疗类 U2-Med API 可接入 Claude Code、OpenCode、Hermes 等工具,非常适合当备用模型 领取步骤:打开云知声官网 → +86 手机号登录 → 开通免费订阅 URL:https://t.co/ZMIdkswCQ1
@zest_kicks
Giving AI the Memory to Keep Up. At #SC26, Preet Virk (Marvell) and Andy Mills (Penguin Solutions) explore how optical connectivity can overcome memory bottlenecks in AI systems. @NYSEWired @theCUBE @MarvellTech @WeAre_Penguin #AIInference #LLM #Memory #AI https://t.co/6kLMKzdrDw
@Reinpayapp
Reinpay Intelligence, at work 💻 Every payment your AI agent asks for goes through the same pipeline: payee, memo, budget, your line, a mainnet simulation, then a receipt. New look on https://t.co/Y11odIf4ME 0x17c4735e9da7f1eeb76d1cc17c2f196f38de64b3 https://t.co/s9YTjSKMHz
@albertorosasg
In this AI era the bottleneck isn't building anymore, it's distribution. So I recorded how I run paid media directly with Claude: > connect Meta Ads + Google Ads > conversion engineering > create the campaigns > optimize for ROAS every day If you spent $10K on ads and it didn't work, the problem is almost always the second one
@OpenDesignHQ
DeepSeek + OpenMotion's MOTION.md at their peak !! We rate it as a MUST TRY 🔥
@Chchenhao0129
还有人不知道,ChatGPT Pro 每个月还能领 API 额度? Pro 100 / 200 / 500,每个完整账期分别可领 $100 / $200 / $500。我是 Pro 100,我因为领的晚,结算下来这个月还剩64$,不过已经很爽啦!! 怎么领? 打开下面的账单页面,用你订阅 ChatGPT Pro 的账号登录。 https://t.co/rIO3ig2sPa -左侧找到"账单",打开概述 -选择要关联的 ChatGPT 个人账号,核对接收额度的 API 组织。 -个人用户可以用默认的“个人的”组织。阅读提示后,点击右下角确认关联并领取。 -回到“账单概述”看顶部的“API 信用余额”。这里显示的就是到账后可用的金额。 ——用 ChatGPT Pro 账号登录 Codex,套餐额度照常。 快行动起来吧🤣
@tuturetom
黑色星期五 🎯🐳 第一弹,DeepSeek,即将启动 大家可以猜猜看能给到多少额度
@ModengSir
爸妈总说 “我没事”|AI 帮我把看不见的健康隐患全部存档 https://t.co/Bf6wG1sLP8
@riverleaf88
现在的多智能体产品太多了,已经选择困难症了。横向对比一下,总有一款适合你…… https://t.co/2LyIs7BRlf
@tucanosh
Tucano 0.2.44 is out, and it's a big one: Cofre: your API keys and logins, used by your agents by name, never seen More autonomous agents: they carry a request through, call in another agent when needed, and retry on their own when something fails MCP: your Claude Code, Codex or Cursor can set up your Tucano team and talk to it Archived agents can now be deleted for good Release notes https://t.co/F8eFxxiTIW
@EianLu
看完这篇,与Agent 协作的焦虑感至少少一大半,正确的协作思路是把思维变成Agent实操规范。 文章里面有12 条经典思维模型(以终为始、第一性原理、二八法则、系统动力学、逆向思考、费曼学习法、检查清单、复盘等)。 里面直接打包好的7 条全局底线和1 个开源的 uw-mode Skill,可以直接给自己的 Agent 装上了👇
@zhixianio
Claude Code 真是想到我心坎里了,新出的 Project 会话干的就是我原来的「luoke-top」干的事情,而且少了跨 session 的权限隔离和会话之后整体性更强了,推荐有复杂任务的小伙伴们都去试试看 https://t.co/yPOSxvdfc7
@Kkkatia1
回母校参加 Grok bot 的meetup🇭🇰 上一次回来还是crypto的活动😭 感叹行业变迁太快 遇到了很厉害的学弟 好多学生都在build自己的产品,听上去都很有意思 还有一个香港founder和我说:在香港,豆包的使用率比codex Claude高出太多,都愿意付费使用🫣🫣 https://t.co/p39WSVzW7k
@AzaleeK13396
Codex 接码到手第一步:先把注册页面打开,码来了直接填。 码就 15 到 30 分钟有效,页面没准备好,码过期了等于白买。 下单后马上注册,接 1 到 3 次都够用,5 年老店,下单有售后。 还没号的顺便看看:https://t.co/bYw1csgaiC #ChatGPT #Codex #接码
@huang18825
网页端也加入了元素周期表,并且点亮了K,选择后描述信息是banana glow K!这是不是预示着我上次猜测的新的大香蕉要来了! 彩蛋激活方式:网页端Gemini,连续选择模型五次。 https://t.co/yUeRzLGsb5
@termireum
Claude Bug Hunter A Claude Code skill bundle for bug hunting and external red-team work - 82 skills, 15 slash commands, 681 disclosed-report patterns curated across 24 core vulnerability classes, plus enterprise identity + infrastructure attack matrices. https://t.co/ZmJzD9udX3
@CapyCola
最后一位客人一小时前就走了,猫还在。🐈⬛🏮 雨里的面摊,摇晃的灯笼,一锅没人看着的热汤。 它在等老板回来,还是只在等雨停? 这个场景用 Claude Fable 5.2 和 @JungleSilicon 的 Ink Studio 制作。它不只是视频,也是一个能拖动旋转、从任意角度看的网页。网页和源码都已公开,链接在评论区 👇 #pixelart #AIart️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️️
@junfanzhu98
System 2 Thinks. System 1 Acts. System 1.5 Knows When to Think Again. Happy to join @memories_ai's Physical AI Friendsgiving in SF, discussing what happens when AI moves off the screen and into the physical world. Frontier models now drive robot arms zero-shot. That's exactly why System 1 (S1) is the new bottleneck. My takes: 1️⃣ System 2 (S2) is now something you rent. Apple's AMP1 (Vladlen Koltun's team): one off-the-shelf model reads 3 cameras + depth, writes Python, checks reachability, executes, reviews. No grasp/place skills, no demos. 85.6% zero-shot on RoboLab-120 vs 49.2% (leading VLA) and 42.9% (leading WAM); 104/108 real trials. 2️⃣ Geometry is the API. Koltun co-wrote Point Transformer, the first model past 70% mIoU on S3DIS; AMP1 aligns the grasp axis from the point cloud first in ~61% of tasks. Without depth, Opus 5.5 barely moves (82.8% → 81.1%). Without depth + calibration, Haiku 5.5 collapses (51.6% → 10.7%). 3️⃣ Then the latency bill: ~280 s per successful task, ~211 s of it thinking (π0.5: 62 s); no mid-trajectory revision; quasi-static scenes only. Apple's own fixes (code reuse, a fast local feedback controller, planning while moving, a distilled small model) are S1 by another name. Code is not a natural motor language. 4️⃣ Rentable S2 makes edge cases a small-company game, and the moat moves back to hardware: actuators, sensors, reliability, cost. 5️⃣ The last points cost the most. NVIDIA's Astranomy: GPT-6 Astra as a direct policy hits 66.3% at $1.96 per episode; as a Codex agent, 74.8% at 30.9M tokens ($36.40). ~19× the cost for +8.5 points. π0.5: 28.0%. 6️⃣ Memory is where S2 earns its keep. RoboMME (Dai … Finn, Fazeli, Chai): 16 tasks (counting, permanence, reference, imitation); none of 14 memory-augmented π0.5 variants wins everywhere. 7️⃣ "Can Astra Solve RoboMME without Breaking the Bank?" (Chen, Fang, Liu): Astra plans, π0.5 acts, a 4B Qwen3-VL "System 1.5" checks completion after every chunk. 79.1% vs 44.5% (best non-oracle baseline) on 3.63 Astra calls per episode. Irony: tokens and dollars weren't measured. 8️⃣ The tell: the sim pauses while Astra thinks; unpaused, the policy often ran out its 1,300 steps before Astra answered. Worst tasks (18% each): pressing a moving cube, peg insertion. Timing and contact belong to S1. Explaining how is not doing it steadily. S2 decides what; S1 decides how and when. The layer worth building now is S1.5: cheap monitors that decide when to wake the expensive brain. The next breakthrough in Physical AI may not be a smarter brain, but a smarter way to decide when the brain needs to think. #CoRL2026 #EmbodiedAI #Robotics #PhysicalAI
@Use_Archivum
Archivum App is now LIVE! We're taking AI memory to the next level. You can now explore your agent's memory through a visual workspace: 1. Upload — Bring in Markdown, text, JSON files, or entire project folders. 2. Dream — Let an AI model review your files and build connected wiki pages. 3. Graph — Explore relationships between ideas and knowledge. 4. Search — Find information across your memories and conversations. 5. Playground — Interact with your knowledge through AI. This is another big step toward making AI memory easier to explore, manage, and understand. And we're not stopping here. Big updates are coming soon. We're building more, improving the experience, and expanding what Archivum can do. Try it now 👇 https://t.co/myUGFOUqyy
@Use_Archivum
Privacy is a fundamental part of Archivum's design. The CLI, MCP server, hooks, notes, transcripts, wiki, and history all run locally. Everything is stored under: ~/.archivum You do not need an account, hosted memory server, or subscription. There is also no telemetry built into the application. The only part that sends information to a model is Dream. - Bring Your Own Model Archivum supports two ways to run Dream. Option 1: Use your own LLM key Anthropic is the default provider, but Archivum also supports OpenAI-compatible endpoints with tool calling. This includes hosted providers, local model servers, and local routers. Before running, Archivum estimates token usage and asks for confirmation. You can configure a token budget to prevent a run from exceeding your chosen limit. Option 2: Use the agent you already have You can also ask your existing agent to perform the consolidation through the dream_prepare and dream_submit MCP tools. This means you don't need a separate API key for Dream. The work uses the tokens from the agent you're already running. In both cases, the result is a proposal that you review and apply yourself. - Secret Scrubbing Archivum checks for known secret patterns before transcripts are stored and whenever wiki content is written. This includes common API keys, bearer tokens, PEM blocks, AWS, GitHub, npm, Slack, Google credentials, and JWTs. Detected secrets are replaced with redacted markers. However, pattern-based detection cannot catch every possible secret, so users should still be careful about the information they store. - Controlled Memory Access Dream agents are restricted to document-related tools. They can read, search, write, edit, delete, and rename wiki documents, but they do not receive shell or network access. Transcripts are also treated as untrusted data rather than instructions. These restrictions reduce the risk of a transcript containing malicious instructions influencing the Dream process, although they cannot eliminate that risk entirely. The final safety layer remains human review, version history, and the ability to restore previous versions.
@songchanggui
上车了!!!纯血美国卡段就是爽!!! 玩 AI 必备神卡,充 ChatGPT Plus 返现高达 30%,算下来才花了 14 刀!! 现在限时 0 开卡费发车,这么大便宜不薅血亏,赶紧冲!! https://t.co/mxKlgik6U7
@premwomp
Claude Saga 6 might be Anthropic's next BIG leap 🤯 https://t.co/x7NuB903ti
@Abdiraimov_
i had claude read 109 seo and geo articles and trace every number to the original study. most of the "geo playbook" is vendors selling tools. x sent me 10,331 visitors → 10 orders. google sent 207 → 8. what actually held up 👇 https://t.co/7SXHsDZ6AC
@qingzhoucloud
🚨 英文圈刚刚炸锅了! 那个霸榜第一的神秘模型, 底牌终于揭晓 竟然是 Claude Saga 6。 https://t.co/22uMUx493X
@0xsacr3d
This is f*cking insane. The Turing Award winner who taught computers to see, Yann LeCun, just explained why today's AI still isn't really intelligent: "It's human intelligence that allows a teenager to learn to drive in 20 hours of practice. We have literally billions of hours of training data of people driving cars around, and we still can't use that to get a machine to drive as reliably as humans" In this free 82-minute conversation, he covers why chatbots mastered language but not the world, what a world model is, why a house cat still knows things no AI does, why superintelligence will work for us and why AI can't stay in the hands of 3 or 4 companies Chatbots → Language → World Models → Machines That Learn Like A Teenager → Superintelligence That Works For Us I've seen $1,500 AI courses that don't cover what he says in the first 2 minutes Bookmark and watch the full conversation Then read how to check what your agents actually understand: Verification Engineering ↓
@ohayou792
新手上手反重力,多代理极易选混。 第一次打开反重力,模型栏里挂着 Orchestrator,命令行里又冒出 Subagents、/swarm、/boost 和 /teamwork,很多人根本分不清该用哪个,随便挑一个跑两轮,往往不对路还白费额度。其实这五套机制各有分工,按任务大小和复杂度选就行。 先看日常主力 Orchestrator,只在 Google AI Pro、Ultra 和企业版开放,App、CLI 和 IDE 插件都能用。 这个Orchestrator是在同一个对话里跑单代理,只不过拆成了规划和执行两个角色,高级推理模型用来做规划和审查,快速模型用来改代码和跑测试。 Pro 和 Ultra 账号能自选搭配,按各自套餐权限,规划模型默认用 Gemini 4 Neon,还能换成 Gemini 4 Carbon、Gemini 4 Argon 或 Claude Opus 5.5,执行模型默认是 Gemini 4 Hydrogen App 里去 Settings 的 Orchestrator Settings 可以改全局默认,或者把鼠标悬停在 Orchestrator 上为当前对话单独挑规划模型 CLI 里敲 /model 选中后按左右键就能切换,选好直接提需求就行,比如输入:给注册表单加上输入校验并跑通全部单测。 想顺手外包一个独立小活,就用全套餐全端都支持的 Subagents。不用专门记命令,主代理会自动在后台拉起内置的 research 扫仓库或 browser 验网页,做完把结果交回主对话,也能在 .agents/agents/ 下自定义角色,日常直接用自然语言让主代理派出去就行。 比如输入:开个后台子代理去核对登录模块的测试覆盖情况,我这边继续改接口。 遇到工作量大、能按文件或包切开的并行任务,全套餐都能在 App 和 CLI 里用 /swarm。 输入命令后主对话会当总协调者,在公开留言板上把活分给多个命名工人并行干,再配验收员审改动和跑测试,并发时消耗额度极快。 适合批量拆解的体力活,比如输入:/swarm 跑一遍 pytest,把失败用例按包分给不同工人修复,不准删断言,最后配验收员重跑全量测试。 要是改动面不大但问题极绕,例如多线程死锁、竞态条件或复杂算法,就换只对付费套餐全端开放的 /boost。 敲完命令直接走三段流水线,派多轮工人互相对抗挑刺并跑真实测试,最后主控核验交付。 比如输入:/boost 排查连接池在高并发切换下的间歇性死锁并给出带压测的修复方案。 至于整库框架迁移、系统仿真这种耗时几小时甚至几天的超大工程,才需要动用同样限付费套餐的 /teamwork。 输完命令不会立刻开改,而是先做一轮范围访谈,确认简报后再进 ~/teamwork_projects/ 专用目录按里程碑推进,每关强制过审查和真实测试审计。 比如输入:/teamwork 把现有 Express 服务完整迁移到 Fastify,补全 TypeScript 类型和性能基准测试。 省流版总结: 日常开发开 Orchestrator 搭配后台 Subagents 就足够应付九成工作,确认能切块并行再敲 /swarm,单点卡死上 /boost,跨天大重构才值得开 /teamwork。
@Maxq5xs
我尼玛,这老哥怕不是个穿越者吧?!🤯 今年 7 月 21 号发的预言贴,开头第一句直接精准踩中 2026 年 11 月 21 日 黎曼猜想被攻克, 还点名提到了Claude,结果这两天 A 社真就把黎曼猜想攻克了 细思极恐啊😱… https://t.co/JRLFAlPsiT
@usegraft
V2 in action. We pasted the Claude Code repo on https://t.co/tvmFUGfeUf. Its agent came straight from GitHub, landed in Cursor's own format, and was ready to download in seconds. Your turn. Any repo, any platform. https://t.co/tvmFUGfeUf https://t.co/NY0emnKz7p
@leo114119
姚顺雨清华分享|22 条 AI 深度洞察,含金量极高,建议先码后看 🧵👇 https://t.co/uN8lwgM7ws
@akazwz
claude code sub agents 的标准格式: https://t.co/ofoP7pPzt7
@mbauwens
* OpenDiLoCo https://t.co/qMln5Eg48O = "an open-source implementation and scaling of DeepMind’s Distributed Low-Communication (DiLoCo) method, enabling globally distributed AI model training." "Large language models have revolutionized AI, but training them traditionally requires massive, centralized compute clusters. This concentration of resources has limited who can participate in AI development and slowed the pace of innovation. (We are) exploring the state-of-the-art in distributed AI training." cc @jeffemmett
@usegraft
Graft V2 is live. Any public GitHub repo, any AI agent inside it, now one click away from your setup. Paste the link, pick your platform, Graft does the rest in your browser. 25 platforms. Nothing uploaded. Nothing left behind. https://t.co/tvmFUGfeUf https://t.co/zyFEPYukQs
@ebokify
🤖 Large Language Model Agents: Design, Architecture, and Practical Implementation (Machine Learning: Foundations, Methodologies, and Applications) ✍️ Author: Muning Wen 📖 Get the book: https://t.co/rTIILOImch 💡 Description: A practical, code-driven guide to building intelligent agents powered by large language models. Explore agent architectures, memory mechanisms, retrieval-augmented generation (RAG), tool use, and advanced reasoning strategies. Ideal for AI researchers, developers, and postgraduate students looking to design, debug, and deploy robust LLM-powered systems. 🔁 Retweet | ❤️ Like | 🔖 Save #LLMAgents #LargeLanguageModels #GenerativeAI #ArtificialIntelligence #RAG
@miles_mazy
我一直认为用 Claude 做视频,性价比低得离谱 如果你能解决 Claude 使用稳定性的问题,完全可以有更直接的赚钱方法: 1、搭建中转站,卖 Tokens; 2、代配置稳定的网络环境; 3、帮人解决 KYC; 4、卖成品账号; 5、用 CLaude 开发网站或者小程序; 6、用 Fable 1:1 复刻别人赚钱的产品,拿来卖; 这些都是干完就能立马赚钱的活 比起拿去抖音、小红书起号,挣一点播放量要实际的多
@PovilasKorop
New useful tool for Laravel devs! CodeGraph: A free macOS menubar app and local MCP server. It maps your Laravel app's code and database schema, so Claude can trace a stack trace or a support ticket to the screen, route and table behind it. https://t.co/q8MlvpeOGx https://t.co/YgXFhbO7Bz
@heymireia
REPO DEL DÍA: ALGUIEN HA HECHO QUE TU AGENTE DE IA PUEDA USAR CUALQUIER PROGRAMA se llama cli-anything y es open source (61,6k estrellas en github) → convierte programas normales en comandos que tu agente sabe usar, sin clics ni pantallas → la comunidad ya ha hecho versiones para libreoffice, obsidian, qgis, calibre, kdenlive y muchos más → funciona con claude code, cursor y otros agentes lo que te ahorras: hacer a mano en cada programa lo que tu agente podría hacer por ti guárdalo si trabajas con agentes
@OpenDesignHQ
UNLIMITED DeepSeek is here now 🎯 🐳
@linlin454136
现在做独立开发最憋屈的事:Vibe Coding 一下午搓完一个 App,结果卡在苹果和谷歌审核被反复打回,活活折磨半个月😭 刚刷到狂揽近 3000 星的开源神作 app-store-compliance,简直是提审渡劫救星!! 今年 iOS 新包暴涨 80%,商店审核卡得史上最死,作者直接把趟雷经验做成了自动化防线: • 扒透 117 种真实拒审死法(隐私清单、SDK 暗采、内购、账号注销全覆盖),逐条给清官方条款和修法 • 追踪 95 个出海合规死线(欧盟 AI 法案、COPPA 等) • 最狠的是能一键装进 Claude Code、Cursor、Codex 当拦截守卫——发包前扫出致命隐患直接 BLOCKED 拦停! 哪怕已经被拒了,里面还有 10 分钟申诉剧本。做 App 出海的赶紧焊死在书签里🔥👇
@xueyu1125
X 上一直流传着t3 code能安全反代antigravity,但是我没找到出处 我终于找到了谷歌DeepMind官方的声明 T3 code优势是可以支持一个agent管理codex cluade antigravity opencode cursor的订阅 并且支持手机远程访问 唯一缺陷是t3 code不支持DeepSeek,并且官方仓库PR并不想支持DeepSeek🤣 https://t.co/52B1qZsnuX
@0xTeddet
500 users at $20 a month is $10K, and you don't need a team for 500 users claude builds the app, grok bot runs everything around it, and you only tap approve 10 steps, 14 copy-paste prompts and a 14-day plan from idea to launch ↓ https://t.co/KKjg1cvO1Q
@finding_raj
i don't think founders realise how organized "free trial" farming has become. we found 500+ github repos farming credits for 14 different ai products. went through the numbers and honestly, it's pretty wild. read the report here ↓ https://t.co/SfsBhiGDlV
@wile_ai
Levent Alpöge & Claude Congrats with Riemann Hypothesis work! https://t.co/EAuaD49j55
@wy_ones
Claude-BugHunter 把 681 条公开披露的漏洞报告整理成套路,分成 24 类,其中 433 条能一条条查到出处,打成了 Claude Code 的 skill 包,给漏洞赏金和授权红队用 你上线过的东西,被人扫出过漏洞吗? https://t.co/Vbe6IPUfIv
@hsianghui
Atlas is source control for coding agents. Its desktop app and CLI record every agent session, link the commits it produced, and keep a checkpoint of the prompt and tool calls behind the change. It runs alongside Claude Code, Codex and Cursor. Apache-2.0. https://t.co/3RMeDY5z3z
@Samking207
Flight Sim game, made using @Gemini @GeminiApp Gemini 4 Argon. @ivanfioravanti @ItsmeAjayKV @teortaxesTex @superalesha #ai #Google #Gemini #llm #agi #asi https://t.co/NAEswyOrNi
@tuturetom
DeepSeek 无限用真的来了!🔥🚀 OpenDesign 黑五第一波福利:Pro/Max 年付会员满血DeepSeek V4.1 Flash 7 天无限畅用 其他还有💥 · GPT / Claude 模型低至八折,包括 GPT-6.2 Astra 和 Claude Fable 5.2 · API 充值最高加赠 25% · 年付会员首年最低五折 · 300+ 顶尖品牌视频模板、152+ 品牌规范和Skill 更多多模态和语言模型无限畅用陆续上线中,直接评论区点菜👇 详情:https://t.co/Mcb8MEFtoY
@XiaotiaoWang
Codex排了两套迁移路线 你补一句「老客户不能下线」 它先划掉全量切库那套 这一刀比再写十页说明有用
@cao_lab
This is exactly where AI should be heading. 🚀 Everyone deserves their own personal AI agent. But I believe the future isn't just local AI — it's hybrid AI. Run locally when you can. Tap into decentralized cloud compute when you need more power. Your AI. Your data. Your choice of compute. This is the future we're building toward at @computefluxAI. Would love to explore how Linger and ComputeFlux could work together!
@ai_goge
为你的Grok Bot 搭建一个内容运营小团队 https://t.co/QrcieHk4fp
@0xBakeer
local ai folks, inkling 2 mini is out 👀 26b moe, 3.4b active, ~20gb. it loops its middle layers 3 times per token, so it thinks more without writing more tokens runs on a laptop, open weights. https://t.co/LrKXMfTavv
@Harry_The_Nerd
Ask an LLM the same question twice. Get two different answers. Why? Article 5 of my GenAI + Agentic AI series is live: temperature, top-p, min-p, why temp 0 isn’t truly deterministic, and how to force valid JSON. GenAI Foundations: complete https://t.co/dmZAciqX9C
@hepburn37666
到底是谁在看不起豆包啊! 豆包终于赢了一次。 Claude VS 豆包 猫咪的对镜自拍。 https://t.co/WawQenYJ8u
@imnotchalk
this guy bruh 😂 bro really used claude to find the most optimal path for Hachi's game (btw check his work out he makes cool things) respect the hustle but we will not condone scripting from this point on ☠️ ALSO HACHI IS A DOG
@ToolAdoption
Everyone is arguing over Jensen Huang headlines while Nvidia quietly dropped an open reranker that stops AI search from feeding garbage to your bot. ServiceNow cut median latency 24.1% and lifted quality 5.17 points in production. https://t.co/bpKygoZNL2
@Feng_btc
新业务来了,听说最近Claude门户/真人验证很困难? 来吧,选择你的英雄帮你过KYC 可随机,也可挑选图中的小黑哥 PS:指定的英雄要加钱~ 有需求可以联系我,不想折腾的可以去用codex平替 #Claude #ChatGPT #Codex https://t.co/Y8NaZOsQfj
@akshdeeps_001
Claude Code built almost an entire oil-painting engine called aibrush, and it's genuinely interesting. It's free, open-source (AGPL-3.0), and nothing like a typical AI image generator. Instead of generating images instantly, Claude writes Python code, places brushstrokes one by one, examines the canvas, measures the results, and keeps repainting until it's satisfied. Basically, AI learning to paint like an actual artist. And the craziest part? Nearly the entire codebase was written by Claude Code itself.
@Gautam_A_k
I asked Claude to explain how Inkling 2 Mini's looped transformer works. Wait so if the middle layers run in a loop, it needs to remember the outputs and carry it and this is done by using Attention Residuals and a learned gate. It's actually beautiful! https://t.co/VL7XC7uD6W
@fxl1996
老板让我出门买点token awsb跑f5.2 8.5r oai大额跑付,3.8r Ccmax 3.2r #蓝V新人 #浇朋友 #蓝V互粉
@Huarin_Ou
过去一两年挑模型,很多人的习惯像追高考状元——谁总榜多出一两分,就把密钥全切给谁。但这十天连发的三款旗舰,第一次把底层性格拆得如此彻底。 https://t.co/QqxdVX2iDO
@Proofofme8
If you prefer staying in your current chat window, Dreamwork Pro also runs inside Claude and ChatGPT.
@nadillafrh
info claude prem 1 day #zonauang
@grandemange_k
I run several AI agents in tmux and never knew which one had finished. So I built TmuxIsland: a Dynamic Island-style bar for macOS. One dot per agent, one click or shortcut to the right pane. Open source, 100% local. Full write-up below 👇 https://t.co/d45chCzurs
@neurosurgeryCui
LLMs are now changing day by day. They can write medical records, automatically diagnose images, and even fabricate papers and go to the NASDAQ, but they still have a long way to go in the clinical application of neurosurgery~ https://t.co/4qPVMNEWJi
@bworker_eth
$VOYNICH seems interesting Building at the intersection of AI, privacy and blockchain, Voynich aims to keep AI model weights encrypted and protected through secure enclaves Built on Solana, Voynich enables access to encrypted AI models, paid inference and on-chain verification through signed receipts The goal is to make AI models usable without exposing their underlying technology nfa / dyor
@shubhmx
introducing claude hands-free vibe code any software without touching a single key https://t.co/NfG7is3Fsz
@realpython
Still parsing raw strings from your LLM? In Building Type-Safe LLM Agents With Pydantic AI, your agent hands back checked Python objects. If you know FastAPI, the pattern will look familiar. https://t.co/2GkZzN63JV
@hsianghui
OpenQodex is open-source AI code review for coding agents: a separate Claude Code or Codex reviewer checks your unpushed changes before you push. Scanners like gitleaks and semgrep run first, and lockfiles are checked against https://t.co/o8vBNJoaMP. Apache-2.0. https://t.co/tPwG9XZ2YQ
@theagenticdaily
Skill authors can now check if their AI-agent skills will work on OpenClaw before shipping. https://t.co/XK4YBMUqpa integrated NVIDIA's SkillEvaluator so its Compatibility tab tests skills against Claude Code, Codex, Copilot CLI, OpenCode, Hermes, and OpenClaw, flags issues, and suggests fixes — with a Harness Check that runs automatically in CI for marketplace skills.
@Memoirs
QDA: Closed-Loop LLM Quantization Design with Agents Xingyang Li, Tianlang Zhao, Song Han, Ligeng Zhu https://t.co/L5AVdFpjW4 [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻] https://t.co/NIyPjTgKBO
@sydsachar
With Row-Bot 5.3.4, You can now now safely import your memories from Grok Bot, Hermes, ChatGPT, Claude and more, directly into @row_bot_ai, with a couple of clicks. Zero reason to be stuck with those now. https://t.co/ZIK6SFVfjB
@mahanth1778
🎁 Giveaway: you will get one week free Claude Pro from my side Here's how: 1)Follow me 2) repost this post 3) Reply with what you would build or use with Claude Sunday at 8:00pm winners will be announced #weekend https://t.co/gyPBiScsXH
@pdp
Google Workspace containment tools just landed in my local Pion instance. This is really good. It allows a local AI agent suspend and de-authenticate accounts.
@VisiveAI
Google recently announced Loop, its new self-improving AI agent that can autonomously work toward a given goal. Loop is inter... via @testingcatalog #AI https://t.co/k2w4CNAHDz
@ai_xiaomu
fable5.2 做的鹈鹕,有一种说不出的优雅和从容,这是路边一条的 codex 做不出来的,不信你试试。 https://t.co/7rxbDcrC24
@abdussamad22940
约 500 元,我让 Claude Fable5.2 用纯代码做了一部 4K《桃花源记》短片 🌸 没用 AI 生图生视频,没用现成素材,4800 帧全部代码建模、路径追踪渲染。 ——大学图形学期末作业 ~$70: Claude built a 4K short film of a 1,600-year-old Chinese fable in pure code. No AI images or video. No assets. https://t.co/LJSZCmVy9R
@Globaltoken11
Every major AI lab dropped a flagship this fall. What each one actually wins at, in 55s: • GPT-6.2: Computer use & live UI • Claude 5.5: Agentic coding • Gemini 4: 1M native multimodal • DeepSeek V4 & Qwen 4: Sparse MoE value Now it’s all about routing 👇 https://t.co/EjGdVrczAZ
@mccrmx
Worklog: got image attachments working in the open source, local first, provider agnostic LLM chat I've been working on. https://t.co/P8UhfayTZ5
@DonovanxGray
I'm looking to connect with people building the next generation of AI. Especially those turning ideas into practical products and solutions. ▪ AI agent developers and builders ▪ Founders creating autonomous AI systems ▪ SaaS entrepreneurs and innovators ▪ AI automation specialists ▪ Open-source AI contributors ▪ Investors exploring AI opportunities What are you currently building or exploring in the AI space? Let's connect and exchange ideas.
@sefayilmaz_dev
WanderAI 1.1 just went live on the App Store. A few days ago its core feature had been broken for months. Today: 10 users, 19 trip plans. Tiny, but it works again. For the next update, a separate AI agent reviewed my code. Tests passed, and it still caught a bug that would have lost a plan if you locked your phone. Next goal: the first 100 users. What would you try first?
@TheSimonLabs
How many apps can your AI agent access right now? Email, calendar, files, all of it. Count them.
@usegraft
Your AI agent shouldn't be stuck on the platform it was built on. Graft moves it, intact, to any of 25 platforms: instructions, memory, skills and tools. Here is how it works, and why every graft is ringed on Robinhood Chain. https://t.co/NmH2m6fUhN
@varunPbhardwaj
Every guide I found puts the AI bot in charge of Claude Code. I flipped it. Claude Code now hands work to Meta's Muse and checks what comes back. First real job: every row right, two answers missed. How it works, what I changed, and the code. Open source. https://t.co/tjwrZqYDZw
@aicareles
Spent one night coding with GitHub Copilot Pro, and I’ve already burned through 1/3 of my monthly quota. 😅 Seriously, is this what “Pro” looks like? AI coding is getting more powerful, but the usage limits are becoming increasingly frustrating. Anyone else experiencing this? https://t.co/6twNYVAJda
@arxivupdates
Can we compile exact circuits directly into LLM weights without training? A new zero-gradient method grafts deterministic math & logic circuits into Transformer weights, reaching 100% 6-digit arithmetic accuracy without fine-tuning. https://t.co/MKI7GolLVU
@jlinxikr
l🚀 Welcome to GitHub Weekly! Your weekly radar for AI Agents, MCP, Agent Skills & amazing open-source projects. 🔥Discover trending projects 🛠️Learn through hands-on tutorials 💡Build real-world applications 💰 SaaS ideas Discover. Learn. Build. Ship. Follow us! #AI #GitHub https://t.co/6R2KYFDjMS
@asha_aiii
Open-source alternative to Claude Motion & CapCut: drop a reference video, let your coding agent rebuild it with your brand kit, and export ready MP4s: https://t.co/QGpGq2xtuT
@LihaiLaLLC
We just released Dissentry — Community Notes for AI answers. Don’t take one AI’s word for it. Ask any model a question, and your other models fact-check the answer and attach a dissent note only when they find a real mistake. Built by Lihai La, free and open source (MIT). Here’s what Dissentry does: 1. An X-style timeline for AI models Call on Claude, GPT, Gemini, DeepSeek, Qwen, Grok, Mistral, Groq, or local models via Ollama / LM Studio — or ask several models at once. Follow up in threads or quote any answer to ask another model for a second opinion. 2. Notes only when it matters Checker models stay silent when an answer holds up. When they catch a factual error, broken code, outdated info, or a dangerous step, they attach a severity-tagged dissent note and underline the exact disputed sentence in the answer. 3. Consensus at a glance When multiple checkers catch the same claim, their notes automatically merge into a single consensus note (e.g. “Claude · Qwen · Gemini agree”), and you can ask the original model to respond — concede and correct, or push back with evidence. 4. Your own Trust scores Rate notes as “Note is right” or “Not helpful.” Over time, the Trust page shows which models hold up on your actual questions and whose fact-checks are worth listening to — not someone else’s benchmark. 5. Token Usage & smart cost splitting See exact tokens and estimated costs per answer, per model, and answering vs. checking spend. Pair a fast/cheap or local model to answer with a stronger model to fact-check. 6. Share as image in one click Turn any answer and its dissent notes into a crisp 2400px card ready for X or your team, rendered entirely on your device. 7. Sign in with OpenRouter — or bring your own keys Connect OpenRouter in one click (OAuth PKCE, zero server middleman) for every major model under one account, or plug in your own API keys / local Ollama & LM Studio endpoints. 8. Local-first privacy & local CLI No accounts, no analytics, no telemetry. API keys, chats, and notes live only in your browser (IndexedDB). Need local models or HTTP/HTTPS/SOCKS5 proxy support? Run it on your own machine in one command: npx dissentry Available in English, 简体中文, 日本語, and 한국어. GitHub (MIT): https://t.co/IS4HkM8rgW npm: https://t.co/UTtE1dr54o Try it now: https://t.co/kjCU9QtdpN
@prompt48
4 of 6 AI models failed every one of my 5 Christmas SVG animation tests. Opus 5/5, Haiku 0/5. Results, honest limits and the video 👇 https://t.co/FnFFAKXjNm
@mmlong8
⚡ Uncle Cat AI Radar · Nov 21 Evening (32 items) Watch: DeepSeek V1 Author Reveals Open-Source Local Agent Uncle Cat: The 12GB claim is compelling, but until reproduced, Linger’s local advantage remains a… Full edition: https://t.co/wslM7HuCfO
@realwhonghom
I checked the 4 "free, open-source" motion graphics skills for Claude and Codex from a viral TikTok. 2 of the 4 aren't what the video says. The claim (10K+ views): Remotion, HyperFrames, Manim and Motion Skills turn Claude or Codex into a motion studio, and all four are open source and free. Here's what each one really is, how to install it, and when to use it: 1. Remotion: videos written as React code Install: npx skills add remotion-dev/skills (inside a Remotion project: npx remotion skills add) Best for: text animation, charts, captions, maps. Catch: free only for individuals and companies of up to 3 people. Bigger companies need a Company License: $25 per seat per month for manual work (AI agents included), or $0.01 per render with a $100 monthly minimum for automated rendering. 2. HyperFrames (HeyGen): HTML, CSS and animations rendered to MP4 Install in Claude Code: claude plugin marketplace add heygen-com/hyperframes, then claude plugin install hyperframes@hyperframes Best for: promos and social clips, if you think in web design. Catch: npx skills add can lag the main branch by hours. npx hyperframes skills update installs the skills that match your CLI version. Free and open source (Apache 2.0) checks out. 3. Manim: the math animation engine behind 3Blue1Brown-style videos Install: pip install manim (Manim Community, MIT), then a community skill like vumichien/manim-skill. Best for: equations, algorithms, explainers. Catch: that skill doesn't install Manim or its Python dependencies for you. And ManimGL code doesn't run in Manim Community, so tell your agent which one you use. 4. "Motion Skills": not one tool At least 4 GitHub repos go by some version of that name. One popular option is iart-ai/motion-skills (about 2K stars, MIT): 56 skills in 19 packs, from TikTok/Reels to lower thirds and b-roll. Install only the pack you need, e.g. npx skills add iart-ai/tiktok-video-skills How I'd pick: - Kinetic text, captions, charts: Remotion (check the license if your company has 4+ people) - You already know HTML/CSS: HyperFrames - Math or algorithm explainers: Manim - One specific format like lower thirds: one iart pack Bottom line: all 4 exist and support Claude Code. But "all free" only holds if you're solo or a team of up to 3, and "Motion Skills" means picking a repo first.
@Memoirs
Do LLM Agents Still Need to Be Told How to Self-Improve? Shobhita Sundaram, Cansu Sancaktar, Phillip Isola https://t.co/09DWTl0Ldq [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸] 💬Project: https://t.co/dYpCg2f4Ca https://t.co/jSChAejHOf
@park_danie36173
Are you handing AI a whole HTML page full of ads and menus? firecrawl/firecrawl turns websites into clean Markdown (a plain-text document format) that you can pass to AI agents. It gained 999 stars on GitHub Trending today, for 205,697 total (checked around 9 PM KST). https://t.co/43fjchNExY
@remoteur
TriggerDev is hiring a Senior Software Engineer, AppSec. Open source platform for building AI agents. AppSec is a rare find in remote listings, and this one comes with an open source product. https://t.co/nFexgnAkBM
@conradlotz
AI-assisted pilot fixes 55 vulnerabilities across six major open-source projects Source: GitHub Blog https://t.co/ZzptS404eX
@AI_N0W
🚨 You can call AI models from CI without long-lived API tokens. OpenCode Console now supports OIDC for GitHub Actions and other compatible environments. Check model pricing before use: it's billed per token.
@automater_ai
Saturday news: Anthropic closed down 9% at $88.30, nearly back to its $83 IPO price, as the Nasdaq erased all of 2026. Meanwhile #Qwen Code 0.28.3 puts managed agents in GitHub and Teams, and AWS Quick turns a process into an agent flow. https://t.co/8egBFPdgSB
@oderajoseph
laude may have just helped prove a 167 year old math problem. everyone's talking about the proof. the part that matters for builders is how it was made, and how it's being checked 👇 https://t.co/VTXcA5oK2y
@VCFellow
'Twas the night before Big Game, and all through @UCBerkeley not a Bear was compiling... Across the Bay the @Stanford Cardinal prepared by published a paper for the #boathackathon crew while the Bears slept amongst red solo cups "Verify Before You Trust": post-quantum proofs for AI agent computers, built with @boatdotdev, @getcontextdev and @TrustedRouter (Privateer Project) Unlimted futures on Boat machines in less than ms. Fear the Tree 🌲 Go Card (its a color). 🟥 @Stanford @UCBerkeley @TravLauren @AniC_dev @mynameisyahia @jperla @JoshConstine @mynameisyahia Proof of Paper @IACR_News https://t.co/khWkJCcQ4W
@huhufor
Stop connecting 20 MCP servers — or every Cursor / Claude plugin — "to make the agent smarter." A 20-tool setup ≈ 15 minutes of the model calling the wrong tool, inventing parameters, or ignoring the one tool you actually needed. "The one job + the 2–3 tools it needs + what it must not call" ≈ 3 minutes to an agent that uses the right tool. What I write instead: • The one job this session is for • The 2–3 tools it is allowed to use • "Don't call anything else. If a tool is missing, stop and ask." Example: I added GitHub, Slack, Notion, Linear, the browser, and 12 other MCPs to "be ready for anything" → it searched Slack for a code path and invented a Linear comment. "This session is only to fix the failing checkout test. You may use the repo and the test runner. Don't call Slack, Linear, or the browser. If you need something else, stop and ask." → it opened the failing test and patched checkout.ts in about 3 minutes. Still wiring up every plugin "just in case" — or ready to name the job, the 2–3 tools, and what must not be called?
@GamerTOTORO1
CapCut alternative, but open source: OpenMotion. Send a video link and it recreates the scenes, transitions & motion with your brand kit. Edit it in Studio, then export MP4s for Reels, Shorts & TikTok. Free. 300+ templates. Your agent does the editing https://t.co/40Ogv6p0kC
@ai_radar_
📌 AI Model Licences cheatsheet → Why open weights isn't the same as open source → Apache 2.0, MIT, OpenRAIL, CC BY-NC and Llama compared → The fine print to check before you ship Save it for later. 🔖 https://t.co/j6AazCFnoH https://t.co/UhbwBL61IY
@DebbieMoorin
Open source AI costs much less than the frontier models do.