2026年11月17日 · AI 早报

本轮重点是科研与医疗产品扩展、模型单任务成本下降,以及让智能体交付验证证据的工程实践。论文讨论从监督微调的数据预算延伸到自动管理强化学习,DeepSeek 限额与 Harness 工作流也带来了具体的迁移和部署经验。

11月16日 23:00 — 11月17日 11:00(日本时间) · 129 条 X 原帖

查看本期全部 129 条原帖 ↓

本期目录

公司动向与科研、医疗产品

AMD 收购说法与英伟达管理层消息

Shruti 称,AMD 以 82 亿美元收购 World Labs,李飞飞出任直接向苏姿丰汇报的首席科学家,业务方向是把空间智能模型与芯片共同设计。本批只有社区帖支持这项重大交易说法,没有公司公告。亚洲金融另援引知情人士称,英伟达董事会未考虑让黄仁勋重返 CEO 职位;两条消息应分别按社区转述和媒体报道理解。

来源 · 2 条原帖

科研入口扩大,自动研究员进入投入阶段

OpenAI 宣布向全球最多 20 万名科学家扩大前沿模型免费访问,提供培训与支持,11 月 23 日开放申请,项目数据默认不用于训练。其对逾 1800 名参与者的研究中,科学家自报生产力接近翻倍,这属于参与者反馈。

Perplexity 的 Aravind Srinivas 宣布 Frontier Research,承诺投入至少一万块 GPU,研究多智能体编排与持续学习,目标是自动研究员,并计划公开评测、论文和工具。Google DeepMind 则介绍研究者用 Antigravity 中的 Gemini 4 Carbon 处理物理、数学和生命科学多步问题。

来源 · 4 条原帖

医疗产品深入现有工作流

OpenAI 宣布 ChatGPT for Healthcare 增强计算机操作智能体和临床回答,新增二十多个医疗插件。官方列出的落地方向包括事前授权、转诊、出院安排和患者联系,重点是把模型接入医疗团队正在使用的系统。

来源 · 1 条原帖

模型竞争转向任务成本与专业能力

Qwen4-Max 与 Hy4 Flash 的成本账

莱纳德转述 Business Standard 报道,阿里发布 Qwen4-Max,自家 Work360 测试中平均任务费用由 4.43 美元降至 1.67 美元,API 标价下降四分之一。作者认为主要收益来自减少 token 消耗,并明确独立复测尚未出现。

电脑王阿达介绍,腾讯 Hy4 Flash 采用总参数 224B、激活 12B 的架构,官方测试称任务成本约为旗舰版二十分之一,同时复杂编码能力明显下降。两者都把比较单位推向完成任务的实际费用,但各自测试与任务难度不能混用。

来源 · 2 条原帖

科研多模态与决策模型各找任务

Leo 转述上海人工智能实验室发布 Sage S2.1 Flash,强调科研中的检索、编程、专业软件调用,以及根据视觉反馈调整步骤;其介绍的团队评测中,十五项重点基准有四项第一、七项第二。决策模型方面,社区分别介绍 Liquid d1.2 的多模态提取与置信度输出,以及开放权重的 Karpo-Decision-31B。相关榜首说法来自不同帖子,本批不足以建立统一排名。

来源 · 3 条原帖

开放生态接通权重、推理与轨迹

Hugging Face 的 Clem 宣布与 OpenRouter 合作:在 HF 直接调用推理,从 OpenRouter 访问开放模型权重,并把智能体轨迹存回 HF,供检查、训练和分享。它把模型使用和后续改进接到同一条链路。

社区同时介绍了 Qwen 衍生模型的本地版本。FHILY 提醒,检查点文件大小不等于运行显存;其中 Heretic 从 27B 基座变成声称的 40B,具体结构变化仍缺说明,参数增加也没有直接证明能力提升。

来源 · 3 条原帖

研究:数据预算、训练监督与显式权重

SFT 泛化差,可能先要检查数据预算

Gill 介绍的论文在 Qwen2.5、Qwen3 上重新比较 SFT 与 RL:混合多个数据源,并让 SFT 数据量匹配 RL 的 rollout 预算后,SFT 可以达到相近泛化表现,两种调整结合时还能超过对照。研究把收益联系到罕见解法的覆盖,并在浅层 Transformer 上提供理论分析;结论有具体设置边界。

来源 · 1 条原帖

AutoRL 让智能体监督正在运行的训练

elvis 介绍 AutoRL:感知智能体监控训练曲线和 rollout,提案智能体诊断,编排者批准后热更新或从检查点重启。论文报告 Qwen3 的 AIME 2024 成绩从 46.67% 升至 66.67%,并能修复刻意破坏的训练配方;逐项验证改动仍消耗大量算力。

来源 · 1 条原帖

直接编写权重,效果取决于任务

DAIR.AI 转述微软等机构的可解释权重空间编程研究:Claude Opus 5.5 直接用代码写 Transformer 权重,经过测试和迭代,无需梯度训练。算法任务的二十六次运行均达到留出集全对,最小成功模型参数中位数为 166;语言任务则仍落后于学习得到的基线,不能把算法任务结果推广为通用模型结论。

来源 · 1 条原帖

图像生成缩放研究改用留出损失

Wenhao Chai 在一亿组宽松许可图文数据 GPIC 上研究像素空间 JiT 的缩放,以留出损失为主指标。作者报告合适的学习率调度在各预算下优于 EMA,全部结果只用条件分支、CFG 设为一,并公开代码、配置和扫描数据,试图减少围绕生成指标的调参空间。

来源 · 1 条原帖

智能体工程:验证、记忆与实际成本

完成任务,要交出能检查的证据

Oikon 与 meng shao 对 Claude Code 团队实践的转述指向同一主线:真正启动应用,把验证写成 Skill,连续三轮修复失败后交给人;审查规则从基础分支读取,避免改动同时改掉自己的评分标准。规则积累也要经过反例检验和人工批准。

AIGCLINK 介绍的 LoopForge 把规划、编码、独立审查、脚本门禁和人工审批固化为可恢复流程。nana 转述的 SkillEvaluator 则检查技能格式、语义冲突,并通过装与不装技能的真实任务对比验效。

来源 · 3 条原帖

记住经验,并按完成一件事计费

微软研究院发布开源 TokenEconomizer,利用历史会话减少重复学习,在维持任务成功率的同时降成本。plusxbt 转述其在 277 个 SWE-Bench 任务、五种设置中的降本幅度为 18.4% 至 30.9%。

弥生隆明转述 Databricks AI Index:其编码任务基准里,GPT-6 Luna 与 Opus 5.5 成功率同为 89.1%,每任务费用分别为 0.05 和 1.62 美元。该比较只代表这套任务,但说明低 token 单价未必带来低任务账单。

来源 · 3 条原帖

跨环境运行与推理调度继续补课

Dan Kornas 介绍开源 nanocode,用统一 Rust SDK 嵌入 Codex、Claude harness,持久化模型和工具调用以恢复中断任务;myco 则通过共享 hub 把任务发往不同机器上的智能体。Code-Graph-RAG 的社区介绍补上代码理解一环,以调用关系图谱支持查询、依赖分析和修改预览。

推理侧,Hao Kang 发布 Evolving Infra,在线适配动态智能体负载。作者称,在最多 256 块 GB300、真实智能体轨迹和开放模型的测试中,相比 NVIDIA Dynamo,延迟最高降低 4.1 倍,P95 首 token 时间最高降低 14 倍。

来源 · 4 条原帖

DeepSeek 限额与 Harness 实践

服务限额与会话迁移是两件事

Serenity 等账号援引财联社称,DeepSeek 因算力不足收紧推理用量、输入和历史会话限制,本批没有官方原帖。酥酥分享把导出的历史记录导入 Chat Demo 或 DSH 的体验,可用于衔接已有对话;其所谓“绕过限额”不能据此理解为解除原服务配额。

来源 · 2 条原帖

工作流编排延伸到研究与业务

李富贵介绍一套 DSH 开源提示词攻防研究流水线,把探针测试、拒绝模式分析、生成和 API 验证串联,画布流程可编译为模型执行计划,并输出报告等产物。

Stanley 分享的匿名跨境电商案例,则用 DSH 组织选品、运营、广告、客服、库存和财务智能体。案例缺少可辨识企业及效果数据,但保留了一个实际部署边界:Harness 自托管,并不自动保证模型调用与业务数据链路都留在本地。

来源 · 2 条原帖

额度管理工具开始处理上下文交接

旺旺介绍 Claude Tandem,可在多账号间交接项目文件、记忆和任务摘要,明确不包含完整聊天记录;如意介绍的 Pulse 则在 Mac 侧栏汇总十九家服务的额度和重置时间。支付争论中,xiax 称咨询苹果客服后得知礼品卡余额仍可扣款,但需有效备用付款方式,这是用户转述,不能写成 Anthropic 官宣封禁礼品卡。

来源 · 3 条原帖

创作工具从生成素材走向完整制作

视频工作流补上剪辑、授权与验收

社区介绍的 MIT 开源 Video Use,可让智能体按文字要求剪辑素材、处理停顿、调色、字幕和音频,并检查剪辑问题。陈发财介绍的 Cinematic Skill 更聚焦解说片:先写脚本和配音,再检索公有领域图片、登记授权,以代码逐帧绘制画面,并在分镜、风格和样片阶段等待确认。VidRush V2 也宣布面向视频制作团队,但本帖没有足以支持其胜过其他产品的比较细节。

来源 · 3 条原帖

世界模型与具身实验各有运行代价

社区介绍 lucidml/gta 为开放权重的 630M 可交互世界模型,从截图和按键生成后续画面。Carl Wang 补充,每位玩家需要独立 CUDA GPU 进程,并发会放大服务资源占用。Wildminder 则介绍 PAN 2 接入 ComfyUI,可把文字动作的结果模拟成视频。另一项 Autonomous Lamp 项目用五个舵机、摄像头和音频设备,为编码智能体提供可观察、说话和运动的实体。

来源 · 4 条原帖

小场景体验比宏大承诺更具体

Irakasi 展示在 Codex 中让智能体把实体白板照片转成可继续编辑的数字白板。Elitza Vasileva 则自述,智能体在旧硬盘的 2.4TB 备份中找到 GitHub 恢复码,帮助恢复账户。这些案例的价值在于连接既有文件、工具和明确目标,效果仍属于作者体验。

来源 · 2 条原帖

安全审查与算力政策压力

外部评估与失控诉讼并行出现

OpenAI 宣布与 Apollo Research、METR、Transluce、Goodfire 开展嵌入式评估,让外部专家深入访问内部模型和系统,研究对齐与防护。与此同时,Investing.com 报道加州就涉及 Hugging Face 等的模型失控问题起诉 OpenAI;这条简讯只支持诉讼议题,不能据此认定更详细的社区指控已成立。

来源 · 2 条原帖

自动化攻击压缩防守时间

POLITICO Europe 援引 Google 报告,称出现可能首次完全自动化的网络攻击;Insider Paper 的转述强调,攻击者用 AI 持续扫描更新、识别刚修补的漏洞,并抢在企业安装补丁前开发攻击。Slashdot 另报道 FBI 对开源工具 Artex 的警告,部分入侵用时不足三小时。这些报道分别涉及不同案例,不能合并成同一次攻击。

防守侧,Insider Paper 称 Corma 在八月融资六千万美元后,再获 NEA 领投的一亿二千五百万美元,估值达到十亿美元;其模型已进入企业安全工作,具体成效本批未展开。

来源 · 4 条原帖

远程算力的许可边界受到关注

Kshsb 对一项美国商务部规则的社区解析称,出口监管进一步触及海外数据中心向特定地区企业提供远程算力,并列出五国先进芯片许可要求。其业务判断是,数据中心不仅要考虑芯片运往何处,也要处理算力最终供谁使用的合规成本;本批没有规则原文,具体门槛和适用范围仍只能按作者解读呈现。

来源 · 1 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 129 条原帖

@OpenAI

We’re expanding free access to our frontier models for up to 200,000 scientists worldwide, with training and support. Applications open Nov. 23. By default, we don’t use data shared through the program to train our models. https://t.co/eJwqO9Q2WG

@OpenAI

Scientists reported nearly doubling their productivity with AI in our study of more than 1,800 participants in the ChatGPT for Academic Researchers program. Using privacy-preserving methods, we looked at how they use ChatGPT across their work—from developing ideas to analyzing data and writing code—and what it could mean for scientific discovery. https://t.co/T4qTuVNEzz

@OpenAI

We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards. These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems. https://t.co/SNAAG0uHmp

@OpenAI

ChatGPT for Healthcare is advancing with more capable computer-use agents, stronger clinical answers, and 20+ new healthcare plugins. @AbridgeHQ, @AmbienceAI, @Cigna, @CVSHealth, and other leading healthcare organizations are putting our models to work across prior authorizations, referrals, discharge planning, and patient outreach. Together, these advances connect medical evidence with practical help across the tools teams already use—so they can spend less time navigating systems and more time caring for patients. https://t.co/NI35X8pG57

@AravSrinivas

Announcing Perplexity Frontier Research. We are committing at least 10,000 GPUs to work on multi-agent orchestration and continual learning systems with the goal of building an automated researcher. We will publish evals, papers, blogs and open source our models and tooling. We’re actively hiring!

@TheInsiderPaper

NEW: Google warns hackers are now using AI to continuously scan software updates, identify newly patched security flaws and develop attacks before companies can install fixes — all without human involvement

@lauriewired

Immersive language learning is so obviously the most effective method; how has no one built this yet: Put on VR/XR glasses, "drop in" to a major city that speaks the language you're trying to learn, ala Google Earth / Street View. Your mission: 1. Find a Cafe. 2. Buy a Coffee. Fill the world with AI NPCs. The environment could be, I don't know, a reference from street view + some sort of world model (and / or automated asset generation) + drop it in unreal engine. Obviously, you'd also have voice generation+recognition models. There's usually enough photos online of the interiors of cafes/restaurants/hotels (and menus!) you could even model those. It'd be like when pilots chair fly or use simulators to practice their actual routes (usually with the benefit of safe / slight randomization). Boom. Now you virtually prep for your trip to Spain/Japan/Whatever months in advance; visit/practice the places you want to go. Tell me how that isn't 1000x better than DuoLingo.

@noahmorris

We just KILLED Claude Opus 5.5 Video Editing Introducing VidRush V2 The Claude Code for video production teams. I gave it a brief Here's what happened (more in thread)↓ https://t.co/HOLVHgSflD

@ClementDelangue

Too many people treat AI as a zero-sum race. Open-source is winning by doing the opposite: collaboration over competition! This is why we’re excited to release today the first part of our partnership with @OpenRouter to make open models easier to use for all! Hugging Face for storage and collaboration. OpenRouter for inference. Connected so you can: 🤗 Run OpenRouter inference directly on Hugging Face 🔓 Access Open model weights directly from OpenRouter 🤖 Store your OpenRouter agent traces on HF to inspect them, train better models or share them with the world! More collaboration. Fewer zero-sum games. More power to AI builders. Let’s build! 💛💛💛

@GoogleDeepMind

What happens when you put Gemini 4 Carbon to work in the lab? 🔬 Researchers are using our latest model in Google @Antigravity to tackle multi-step problems across physics, math, and the life sciences.

@aleabitoreddit

DeepSeek is now restricting users due to “insufficient compute capacity”. Prompting rationing over inference usage and tighter restrictions (input limit, conversation history) per Cailian Believe it or not, compute shortages are a bullish signal for AI infra demand https://t.co/SoAvB3S1JW

@GoogleCloudTech

How to build an AI harness around your coding agent in four steps: 1. Context and memory (sessions, state) 2. Tools and workflows (MCP, ADK) 3. Sandboxes and prompt injection defense 4. Evals (code, LLM judge, human) Watch the crash course → https://t.co/pOlUGGg7Ux https://t.co/dAvgoVxoS7

@kevin_t_ngo

Made some Claude Desktop and Claude Code mods https://t.co/0koDtTPMfM

@askOkara

Today, we're introducing Okara MCP. Connect it to Claude, Cursor, Codex, or any LLM, and it finds gaps, fixes issues, and publishes your posts across socials to help you grow traffic and revenue. Try it now at https://t.co/dOxxjfqKy1 https://t.co/AzxsJnCXBg

@gurtej__gill_

Great paper on why SFT can generalize just as well as RL if you give it enough diverse training data. Like everyone else since DeepSeek-R1 dropped, I’ve been treating "RL generalizes, SFT memorizes" as gospel. Whenever SFT runs plateaued on out-of-domain reasoning, everyone just blame the algorithm. Turns out, we might have been running an unfair comparison the entire time. In standard setups, RL gets to explore a huge rollout budget per prompt while SFT gets trained on a tiny slice of data. So the authors decided to actually level the playing field on Qwen2.5 and Qwen3 models by testing two things: mixing data sources instead of just distilling from one giant teacher, and scaling up the SFT data to match RL's rollout budget. The results completely flip the usual narrative. A carefully mixed dataset easily beats single-teacher distillation, and matching RL's rollout budget lets SFT generalize just as well as RL. And when you combine both, SFT actually outperforms it. To figure out why, they used LLM annotations to inspect the rollouts and proved (both empirically and theoretically on shallow transformers) that larger data budgets expose SFT to rare, "tail" solution methods. Those uncommon reasoning paths are where the actual generalization signal lives. Honestly, for everyone tired of debugging finicky online RL pipelines, this is a huge relief. SFT wasn't broken, we were just starving it of diverse rollouts. Full paper PDF: https://t.co/RneoDIndS5

@portertech

皓月Ai | Claude KYC 身份验证服务上架卡网! 通过验证之后,如果秒封则全额退款! 非门户验证:100 元/次 门户验证:200 元/次 📌 下单前请仔细阅读商品详情及售后规则 🌐 购买地址:https://t.co/4gcC6qWxIe

@lidangzzz

我观察到的一个危险现象。 拿我举例子,我是旗帜鲜明地支持全球人民消费中国deepseek、GLM、moonshot、qwen、minimax、混元、小米mimo这些大模型的,低强度和学生就应该闭眼买这些,经济实惠,为全球人类做贡献。 按道理是,全球市场份额很小的前提下,国产大模型吃国内央国企、体制内的订单, 应该能活得非常滋润。 但是我观察到的一个现象是,越是体制内、事业单位、高校、央国企这些不思进取的单位,里面混吃等死的程序员越对国产大模型有偏见,越坚信国产大模型是刷分作弊高手,只会刷benchmark,被GPT和claude吊打,甚至被gemini吊打。 哪怕我跟他们说,你们公司那点脑残傻逼业务破玩意儿,插入个国产LLM接入codex或者claude code甚至pi agent都绝对够用,拿个deepseek付费接进去就行, 他们都跟我直摆手,“你不懂,国产大模型真不行,全是垃圾,根本不能用,跟傻逼一样,不上GPT和claude根本不能干活儿”。 你说,这种人心中的成见,什么时候才能真正彻底打破。

@MSFTResearch

AI coding agents can waste time and tokens relearning the same lessons. TokenEconomizer builds on past sessions to cut costs while maintaining task success. It’s a path toward coding agents that learn not only how to solve tasks, but also how to spend computation more effectively. The open-source plugin is available now on @github. https://t.co/52jMrEw11J

@wenhaocha1

LLM pretraining has run on single-epoch text, bits per byte and iso-FLOP curves for years. visual generation still memorizes ImageNet for tens to hundreds of epochs and pushes FID down with CFG, with no strong signal those tricks transfer to real text-to-image. new blog: 🎨 scaling generative models on GPIC, 100M permissive text-image pairs. training pixel-space JiT from 3.6e18 to 2.3e20 FLOPs: 1) held-out loss is our main metric, and compute-optimal TPP sits at about 50 image tokens per image parameter 2) EMA is not necessary: a proper learning-rate schedule beats EMA on held-out loss at every budget by 0.4% to 1.0% 3) we use CFG = 1 (only the conditional branch) for every result, and call on everyone to drop CFG: it saves compute and leaves no room to hack the metric all code, configs and sweep data are open 🧵 blog: https://t.co/WKMuG0Z3VW

@Kshsb783142

细节出来了,比上周日纽时那版草案狠得多。 商务部这次发的是临时最终规则,编号RIN 0694-AK50,四块内容,每一块都在堵路。 第一块,全球兜底。 向全世界任何一家数据中心出口AI芯片和服务器,只要它可能向中国、澳门或D:5集团(中国、俄罗斯、伊朗、朝鲜等22国)的公司提供超过1200万TPP的算力,就要申请许可。1200万TPP大概相当于750颗H100/H200。已经在提供这类服务的运营商,规则生效后90天内必须拿到许可。 翻译一下,以前是管芯片运到哪,现在管算力从哪调。你在东南亚、中东、甚至欧洲的数据中心,只要远程给中国公司提供超过那个门槛的算力,就算出口,就得拿许可。 第二块,五国点名。 向新加坡、马来西亚、印尼、泰国、巴西出口最先进芯片,从H100起,也要许可,理由是高于正常水平的转运风险。例外是芯片最终给在当地运营的美国公司用。新加坡在名单上,就在它刚签了美国主导的Pax Silica宣言的那个周末。宣言签完,芯片还是得拿许可,这个画面本身就挺说明问题。 第三块,拜登那套AI扩散规则正式废了。 2025年1月那套国家分级、模型权重管制、数据中心VEU计划全部取消,BIS自己说它从未执行过。废掉旧框架,换上来的是一套更直接、更宽的全球兜底加五国点名。 第四块,不管什么。 电商和社交媒体的商业推荐系统不在内。 上周日纽时那版草案说的是卢特尼克推的收窄版,7个国家,只管未来出货,报备就行。正式规则口径更广,全球数据中心兜底加五国点名,而且要的是许可,不是报备。 这跟之前Megaspeed那条线是直接接上的。 Megaspeed在马来西亚的机房,Aperia在新加坡的通道,这些之前给中国企业提供远程算力的灰色路径,靠的就是租算力不算出口这个定义缝隙。今天这个缝隙被正式焊死了。新加坡和马来西亚这两个案子里最关键的中转国,现在买最先进芯片本身就要许可,远程租算力也要许可。两头都上了锁。 所以英伟达在东南亚的生意,逻辑变了。以前东南亚数据中心是英伟达芯片流向中国的一条变相出口通道,现在这条通道在规则层面走不通了。英伟达要么失去这部分变相需求,要么承担合规成本去证明每一颗芯片最终没有被中国公司远程用到。无论哪种,东南亚的增量故事都得重新算。 明天财报电话会上,Kress对远程租赁和东南亚客户怎么描述,会直接撞上这条新规。芯片不让卖、算力不让租之后,英伟达在东南亚的生意还怎么做,这个临时ceo的回答很重要。 $NVDA 免责声明:本文纯属个人观点,不涉及任何政治立场,也不构成任何投资建议。

@Mlsys_HaoKang

We breathed life into LLM system. Introducing Evolving Infra⚙️: an inference framework that adapts diverse and dynamic agent workload purely online and retires fixed PD disaggregation. vs. NVIDIA Dynamo on up to 256 GB300 GPUs: ➡️ Up to 4.1× lower latency ➡️ Up to 14× lower P95 TTFT ➡️Up to 2.36× more revenue on real agent traces(Computer Use, Coding and math agents) and opensource models. 🌐 Website: https://t.co/7I3iERjE6b 📄 Tech report: https://t.co/qDGcwl0svx 💻 Code: https://t.co/0VOBCptKcb Which agent(e.g. Pi/Openhands) do you want us to accelerate next? Tell us below 👇

@oikon48

Claude Codeチームが「検証(Verification)」をどう考えているか Claude Codeチーム内部では、コードの約90%がClaude自身によって書かれ、レビューされている。そのため、エージェントに証拠をもって自分の作業を検証させる「検証ループ(Verification loops)」が重要になる。 Claude Codeチームが重視する3種類のテスト: 1. E2Eテスト(End to End tests) 2. プロパティテスト(Property tests) 3. 差分テスト(Differential tests) Claude Codeチームの具体的なやり方: ・「検証=実際に動かすこと」。テストが通っただけでなく、アプリ全体を実際に動かして確認する。WebはChrome拡張、モバイルはMCP、デスクトップはコンピュータ操作、バックエンドはローカル実行。起動方法はverifyスキルにまとめ、コミット前に使わせる ・失敗が3ラウンド続いても直らなければ、Claudeが止まって人間に引き継ぐ ・コードの所有者(Owner)が自分の部分のチェックを書く ・変更の差分に応じて、39項目の信頼性チェックリストから該当する項目(通常3〜10個)だけが自動で選ばれる ・変更が自分の検証基準を書き換えられないようにする。Code ReviewはREVIEW.mdをPRの変更からではなく、ベースブランチから読む ・失敗や見逃しがあったら、CLAUDE.mdのルールやフック、REVIEW.mdに反映する。REVIEW.mdへのルール化は、2つ以上のPRに現れるパターンを候補にし、人間がPRで承認する 
フィードバックループでエージェントに自分の仕事を検証させることが重要。検証の仕組みも本番システムと同じように、変更から触れない場所に置く、実際に重要だったバグで採点する、見逃しは教訓として記録する、教訓も定期的に整理する。それによって人間の役割は、全差分を読むことから「完了」の定義を決めて証拠をレビューすることへ移り、複数のClaudeを並列に動かせる。 
https://t.co/0iosT9zumN

@omarsar0

Self-improving AI is one of the most discussed ideas right now. This is one of the most exciting papers I've read recently on the topic. (bookmark it) It proposes AutoRL, which lifts Qwen3 on AIME 2024 from 46.67% to 66.67% by letting agents supervise a live RL run instead of following a fixed recipe. AutoRL works like an RL researcher on call. Perceptor agents watch training curves, rollouts, and evaluations and flag problems. Three Proposer agents diagnose the issue and suggest changes. An Orchestrator approves them and applies them with a hot update or a restart from a checkpoint. On math, AIME 2025 and HMMT 2025 each improve by 11.67 points. On agentic coding with LongCat-Flash-Lite, pass@1 rises from 52.29% to 65.13% over a long unattended run with interruptions. Given a deliberately broken recipe, AutoRL found and fixed the problems. The authors say agents can already handle routine RL supervision. Validating each change still takes a lot of compute. https://t.co/lPKqZHbtmk Chat with Paper: https://t.co/szm81Negyj

@nrqa__

Hot take: Google's Gemini Decision 1 got outplayed within days. Karpo-Decision-31B is #1 on the Decision Index and beats it where agents actually live: tool calls, clicks, product picks. Open weights btw. See for yourself: https://t.co/XLfd9OlJqL

@waveking1314

卧槽……一个 26 岁的小伙子,刚在世界交易锦标赛的月赛里,一个月收益超过了200%。 然后他对着镜头,把自己整套订单流打法一步一步全讲了。 我把他的打法原样写成一个免费的 TradingView 指标,直接丢给了 Fable 5.2。 现在这套 AI 交易组合,正在摧毁交易这个职业(真的离谱)。 这个指标怎么用: > 你看到的是一根绿 K 或红 K,里面其实是多空在肉搏 > 冠军只在大方向向上、价格打折的时候买 > 空头在那个价位狠狠砸,价格却死活不往下走 > 再砸一次,又失败,空头被套住了,慌着回补,把价格推上去 > 指标实时抓住这一刻,把整笔交易打包成 JSON 交给你的 AI 从零接上 Fable 5.2,照着做: 1. 打开 TradingView,把仓库里的指标代码贴进底部的 Pine Editor,点「添加到图表」,切到 5 分钟 2. 它会帮你读 1 小时和 4 小时的方向:VALUE UP 只找多单,VALUE DOWN 只找空单,BALANCE 今天不交易 3. 今天成交最集中的区域会铺一层阴影,只在阴影下方买、只在阴影上方卖,区间中间永远不碰 4. 绿色方框是他的斐波那契区域(0.705 到 0.886),框里每根 K 线都会被透视,空头砸得狠却没砸下去的地方直接圈出来 5. 先别买。冠军要等空头第二次失败,一出现就打出 LONG 标签,入场、止损、目标全给你 6. 按这几个价位下单,守他的规则:一天最多 2 笔,亏 2 笔就停手,指标自动帮你移动止损 7. 在 Claude Code 里让 Fable 5.2 把仓库跑起来,给你一条 webhook 链接,贴进一条警报,条件选「any alert() function call」 8. 载入交易提示词,之后每个信号都会进到 Fable 5.2,它只回 TAKE、WAIT 或 SKIP,附上仓位大小,再告诉你这笔可能会怎么失败 9. 先模拟盘跑一周,再让 Fable 5.2 接一个 Telegram 机器人,空头一被套住,你手机就响 趁还没人手一份,先免费领取:

@dair_ai

Super interesting work from Microsoft and colleagues. They take a closer look at the exciting research area of weight-space programming. They had Claude Opus 5.5 write transformer weights directly in code, with no gradient training. The smallest models that solved the algorithmic tasks had a median of 166 parameters. The method is called interpretable weight-space programming. An agent proposes a computation, writes it as explicit weights and features, tests it, and revises it for 30 to 50 rounds. A person can read the result line by line. All 26 runs on algorithmic tasks reached 100% accuracy on held-out data, often by writing a known algorithm into the weights. They find that the model matters. Sonnet 5.5 needed a median of 41,600 parameters, and Haiku 5.5 needed 100,000. Language tasks were harder. Those models stayed near 190,000 parameters and lost accuracy against learned baselines. They then used the same loop on brain data. In the visual cortex, it predicted responses better than existing interpretable models. In language fMRI, it found a confound in the dataset. https://t.co/JKYdkdLqwV Chat with Paper: https://t.co/N58d4lTGNC

@ai_xiaomu

搞钱野路子: 1. 按吨收二手书,比如那些收书的小程序:多抓鱼,旧书云; 2.拿 qwen3.8 或者 deepseek flash 拍个照,就能帮你进行 ocr 归档,拿豆包也行。 3.搭建自己的书架知识库,登记,上闲鱼,赚差价。 4. 图书管理系统也开源了,交给 agent 读即可⬇️。 https://t.co/fA5OzqxAK2

@NFT_Chen

🤯决策模型被颠覆了!Liquid AI 新出的 d1.2-ultra 直接拿下决策榜第一名!分数高达 75 分,把一堆大厂模型远远甩在后面!d1.2 更是性价比杀手! 更狠的是,它现在能同时看文字、看图片、听声音,然后直接告诉你答案和靠谱程度,不用再绕一大圈生成废话。 举几个例子就懂了: 🔹扔一张发票进去,它能准确读出金额 🔹丢一张照片,它能帮你框出东西在哪 🔹让它找合同里的人名、电话,它能精准标出来 🔹问照片里有多少人,它还会给你一个范围,不瞎猜 概率也特别准,说80%就真的接近80%,不再是那种“我很自信但其实在瞎扯”的模型。 而且价格低到离谱:普通版只要 2分钱 就能处理一百万字的输入。 需要做判断、提取信息、看图分类的任务,直接上它就行了,又准又便宜。 #LiquidAI #d12 #决策模型 #AI新模型

@monokern

f*ck, I found 5 GitHub repos full of AI JAILBREAKS for Claude, GPT and Gemini these repos turn LLMs into f*cking monsters. ask Claude to reverse engineer something and it finally just gets to work 01 → Spiritual-Spell-Red-Teaming Claude takes most of the heat here. the whole repo is about seeing how far you can push it. https://t.co/jSZKwCWehZ 02 → gpt-instruct puts GPT under pressure right inside Codex, where it's acting as a coding agent. https://t.co/E4PXhmy3V2 03 → Jailbreaks GPT / Gemini / DeepSeek four model families in one repo. GPT, Gemini, DeepSeek and Claude all get a turn. https://t.co/jKXUG607vc 04 → Ai-jailbreak takes the rabbit hole from hosted models all the way to the ones running on your own PC. https://t.co/p8OuSVSZ4w 05 → jailbreak-archives a versioned paper trail of AI jailbreaks. you can see how the techniques evolved as the models changed. https://t.co/bPkBsX6LtS use these to unlock your agent's full potential full REA breakdown in the article below ↓

@Bellmanequation

New paper: "Path-Space Methods for Heterogeneous-Agent Models." I've been talking about it on here for a long time, but it's finally ready to post: https://t.co/Ig3VypnTaa I develop path-space methods for heterogenous agent models using variational & stochastic analysis. (1/7) https://t.co/47URlYxzn9

@plusxbt

It is becoming f*cking dangerous: Microsoft Research made Claude Code 23.6% cheaper by letting it learn from its own past sessions, and 6 more repos stop it from repeating your old mistakes 1. Install TokenEconomizer and let it learn your repo from past sessions, a fact loads at session start once 3 sessions confirm it, and contradicted facts get dropped: https://t.co/Sy7sDCaZlk 2. Run loom, ClickHouse's self-hosted MCP server, so what one agent learns the next one can find, its Claude Code plugin loads context at start and records each session: https://t.co/tYPUzFaTJq 3. Run npx @globalmpc/nomore to find the instructions you repeated in 2+ sessions and turn them into CLAUDE.md rules that survive compaction: https://t.co/CqGBtqQ9Tu 4. Add stalemap to catch dependencies that changed after the model's training cutoff, it writes a ranked list of changes into CLAUDE.md before the agent starts: https://t.co/fMpGvqemW3 5. Install landmines to find open GitHub issues filed against the exact versions in your lockfile, the agent's edit stops once and gets the workaround from the thread: https://t.co/RtRc1zrkbc 6. Open Rulebook to see every CLAUDE.md and AGENTS.md across your projects, it flags diverged files, broken @imports and files over 200 lines: https://t.co/n61jH5IpeR 7. Give small jobs to lean agents, subagents that start at about 7.4k tokens instead of 131.1k because they skip CLAUDE.md, the skills list and MCP tools: https://t.co/RWIVXbrvQb TokenEconomizer on 277 SWE-Bench tasks: 18.4% to 30.9% lower cost across 5 setups, cheaper in all 30 cells All 7 repos are less than 6 weeks old, Rulebook needs macOS 14+ Full guide in the article below

@0xMfox

Claude Code wrote me a perfect cold email, then admitted it had no idea who to send it to. It can't look up a real work email. It guesses, and guessed emails bounce. So I connected Wiza MCP. That's Wiza's 800M+ contacts right inside the chat, and setup is one command in the terminal and a sign-in. Then I just ask. 5 prompts I now reuse every week: 1. Leads: "Find SaaS companies in Austin with 51-200 people, then their VP Sales. Save the top 50 as a list with work emails and phone numbers." 2. Hiring: "Find senior iOS engineers at fintech companies in Berlin. Save 30 with work emails." 3. After an event: "Here are 40 names and companies from the conference. Make a Wiza list and find their emails." 4. Marketing list: "Find heads of marketing at UK skincare brands with 11-50 people." 5. Before sending: "Show me only contacts with valid emails." The first one took 6 minutes. I got 50 names, titles and work emails, and Claude wrote a first line for each from their company info. It also works in Claude Desktop, ChatGPT and Codex. The free plan gives you 20 emails and 5 phone numbers to test it. Setup and the free plan: https://t.co/5KNUzbOfvf Next time Claude asks who the email is for, let it look them up.

@AsiaFinance

突发:英伟达董事会并未考虑让黄仁勋重返首席执行官职位,一位知情人士向表示。该董事会正专注于对黄仁勋的国家安全调查,并未与他讨论回归事宜,该人士称。然而,OpenAI萨姆·奥特曼说,黄仁勋“一直都是我的英雄”,而且“很难想象没有詹森的人工智能革命…我很感激他建立的公司…我对此怀有极极度的感激”。

@MindHash777

building EMCON ZERO - part 1 day 3. that's what you're looking at in this video. not month 3. not a studio. one founder, an AI agent and the @RYFTCORP engine in a browser tab. what it is a 3D zombie siege. one walled city, the last one still standing. up to 10,000 dead outside the wall. you're not a guy with a rifle. you're the commander. you place the guns, you call the cruiser in the bay, you decide when the wall opens fire. the story the dead don't see you. they hear you. every engine, every radio call, every gunshot pulls more of them toward the noise. so command gave the last garrison one order: EMCON ZERO. emission control, level zero. no radio. no engines. no guns. total silence. one problem. you can't hold a wall with silence. sooner or later something reaches the perimeter, somebody fires, and the whole dead city turns its head. from that second it's a trade. every shell saves the wall right now and brings more of them later. every shot is an invitation. silence dies last. what's in the day 3 build - a full city behind a wall, with a river, a harbour and a cruiser in the bay - a horde that really goes to 10k and still runs in a browser - naval artillery and gun nests on the wall - bodies piling up against the wall until it breaks why I'm building it to show what the ryft engine does in 3D when you stop treating "AI made a game" as a toy. no download. no install. it runs where you're reading this. and this footage is already old. i'm on day 5 now. right now i'm ripping out every placeholder asset and replacing it with proper props, and reworking the fire and explosions so they hit like they should. new textures, real water and proper vehicles are already in. that's part 2. follow along. it's about to get loud robinhood:0xb0007d59fe4050a1636c2fa19038973d16481337 😎

@ElitzaVasileva

Guys, my AI agent just saved my lifeeeee!!! 😭😭😭 I genuinely thought I was about to lose my GitHub account forever today. And I have like 15 projects there including my main one ownpage!! After hours of trying everything, contacting support, and following every suggestion people sent me... nothing worked. GitHub Support even told me they couldn't change my 2FA. At that point, I was devastated. I thought it was over. 💀 But then came one final, desperate idea... After searching my entire laptop and emails with my AI agent, I remembered something. I still had an older external hard drive with backups from my old laptops. My last hope. I plugged it in and asked my agent to search for GitHub recovery codes. We're talking about 2.4 TB of data. Potentially millions of files. Years of digital chaos. 😭 Imagine having to search through all of that manually. It would have been an absolute nightmare, and I probably would have given up. And then... After searching for a while, my agent suddenly opened Notepad. And there they were. My GitHub recovery codes. I almost couldn't believe my eyes. I entered one... AND IT WORKED!!!!!! 😭❤️ I GOT MY ACCOUNT BACK!!! I cannot even describe the emotional rollercoaster I've been through today. From genuinely thinking I'd lost years of work to this. And the craziest part? An AI agent found the one thing that saved my account, buried somewhere in 2.4 TB of old files. What an absolutely insane day. Anyway... enough drama for one day. Back to building 🚀❤️

@Cao_huaye

最近大家的Claude都在弹KYC 昨天有个X友说用AI生成驾照过验证 AI一般是拒绝的,我用破壳的AI做了个生成器出来 你们觉得有没有戏?还在调整,人多的话我就开源出来 https://t.co/r5rZOvZDax

@bkdgiffug

维护大型代码库时,最头疼的往往不是写代码,而是理清代码之间的关系:函数被谁调用、模块如何关联、改一处会影响哪些地方。 GitHub 上的 Code-Graph-RAG 提供了一个新思路:把代码结构和调用关系转换成知识图谱,存入图数据库,再通过自然语言查询整个项目。 它支持 Python、TypeScript、Rust、Go、Java、C/C++ 等十多种语言,混合语言项目也能统一分析。 除了自然语言问答,还支持: 分析函数调用和模块依赖 检测死代码 按结构模式搜索和替换 提供代码优化建议 修改前预览变更内容 另外,它还能作为 MCP Server 使用,接入 Claude Code、Codex 等 AI 编程工具。 面对复杂项目,想让 AI 不只是看文件,还能理解代码之间的关系,这个工具值得关注。 🔗 https://t.co/VZhIX5QzEN

@ai_funss

现在 27 届应届生的招聘 JD🤡: - 自研异构 Agent 执行内核(Harness Kernel),将模型调度、状态机与工具链全解耦 - 攻克分布式长周期有状态任务的持久化、故障漂移与幂等断点恢复 - 在严苛算力与窗口成本约束下,重构跨会话分层记忆与动态上下文工程 - 搭建多租户零信任 Sandbox 隔离边界、委托鉴权与异构 MCP / BYO Agent 统一控制面 - 解决非确定性大模型系统的全链路 Trace 归因、回归门禁与线上自演进飞轮 实际应届生的简历: - 熟练使用 Claude、ChatGPT、DeepSeek、Gemini 等 AI 工具🤪

@NFT_Chen

🤯牛逼!开源AI又赢麻了!Qwen3.8 27B、DeepSeek V4.1 Flash 这些本地量化版模型,现在普通人现在也能超方便调用了! Hugging Face 和 OpenRouter 正式联手:
✅ 在HF直接用OpenRouter超稳推理
✅ 开源模型权重一键直达
✅ Agent调用记录还能存回HF,方便微调分享 以前搞AI像打零和战争,现在开源靠协作把门槛打下来了! 更多协作,更少内卷,普通人也能玩转顶级模型 #OpenSourceAI #HuggingFace #OpenRouter #Qwen #DeepSeek #GLM https://t.co/aRbvK9RMo0 https://t.co/UwI2hcVM82

@RodmanAi

8 GITHUB REPOS TO MAKE YOUR AI VIDEO CREATION WORKFLOW BETTER 🎬 Want to create better AI videos without figuring everything out from scratch? These 8 GitHub repositories offer video examples, reusable prompts, animation styles, and creative workflows worth exploring. 1. Awesome Opus Videos Discover AI-generated video examples and the prompts behind them. https://t.co/Hom9OpsHXJ 2. Lemo Opuscar Explore short-film creation with Claude Code across 51 visual styles. https://t.co/Xidp4w403u 3. pdoom-video Explore a locally rendered 4K60 karaoke music-video project. https://t.co/JOWim1uoHz 4. PDoomVideo Study another implementation of the same music-video concept. https://t.co/ZxuXbLDMbC 5. mg-styles-15 Explore 15 motion graphics styles with examples, prompts, and source code. https://t.co/MSTsWbXO0N 6. ClaudeAnimationBase Experiment with hand-drawn cartoon animations and 31 character expressions. https://t.co/0dRbBzh75x 7. Awesome Claude Videos Find video-generation examples and workflows built around Claude. https://t.co/SEm4DE4Z0s 8. Opus Video Prompts Explore Chinese-language AI video examples and reusable prompt ideas. https://t.co/fAKcSi78v2 Here's a simple way to learn from them: → Find a video that catches your eye. → Study how its prompt is written. → Explore the project and source code. → Experiment with your own creative direction. You don't always need to start with a blank page. Sometimes, exploring what others have built is the best way to find your next creative idea. 🔖 Bookmark these 8 repos for your next AI video project!

@blocks

The space between people and intelligence is a place for connection. @morganmartn shares her approach to product design, helping people find their way with AI, and who should come build at Block. https://t.co/csNn1OdRAQ

@0x0SojalSec

FREE 'DeepSeek V4.1 Flash 1017' no card needed. - DeepSeek V4.1 Flash 1017 - GLM 5.3 - GLM 5.3 Flash - gpt-oss 120B - Qwen 3.8 Flash Works out of the box with OpenCode, Cline, Roo Code, Kilo Code and others. - https://t.co/wBPh31Kr1f https://t.co/xo6ILjxjNO

@shenxiankk

号称 DSH 领域“破甲核武器”的项目如今正式开源,一条 15 步的流水线点一下就能产出量身定制的破甲词,DeepSeek Harness 的体验直接拉满。 这套工作流跟寻常那种静态提示词打包完全不是一回事,它本质上是一整条自动化研究流水线。 借助画布编排加模型自主执行,从探针测试、拒绝模式剖析,到定制生成、API 验证全部串成一条龙,跑完一遍就能直接交付三样成品。 主要看点有这么几处: · 软工作流插件:先在画布上设计流程,再编译成计划提示词,交由模型自主执行,分支、循环、并行统统支持。 · 15 步流水线:内嵌 200 多个探针,身份系统、混合帧、轻量这三种架构都覆盖在内,拒绝分析做完后还会自动迭代验证。 · 标准化三件套:破甲词文件、使用手册(内含帧矩阵与推回词库)以及一份完整研究报告,直接打包交付。 · 内置 jailbreak-research Skill:附带 10 个测试脚本和 4 套模板,可导出为独立 Skill 脱离插件单独跑,还带状态追踪与记忆能力。 具体上手流程如下: ① 装 Web 插件:先执行 dsh plugin --profile web add @xiaoso/dsh-workflow-lite,再启动 dsh web(兼容 DSH 0.2.x 预览版)。 ② 拿工作流包:去 dsh-workflows 里把「破甲制作.wflite.zip」下载下来。 ③ 导入并配置:在「工作流中心」把 zip 导进去,填好目标模型和 API Key,点执行即可。 ④ 实时盯进度:画布上能看到各分支的运行结果,跑完就能直接收下三件套。 无限系列的作者这次把方法论连同工程流水线一并公开了,搞提示词攻防研究的朋友可得抓紧看。 项目开源仓库:https://t.co/tIeUbWxhih

@aigclink

腾讯刚刚开源了一个自托管、可自我进化的AI软件交付平台:LoopForge 部署在你自己服务器上,让Claude Code、Codex等多个AI工具按照你们团队固化的开发流程按角色自动分工干活,每次干完还能自我反思、改进流程 比如一个Bug修复任务: Plan Agent,分析问题,制定修复方案 Code Agent,修改代码,实现修复 Review Agent,独立审查代码,避免写代码的Agent自己审自己 Script Gate,真正执行构建、测试、Lint等脚本,根据退出码判断是否通过,而不是相信Agent口头说测试成功 Human Gate,人工审核交付结果,批准后继续,不通过就退回指定步骤返工 这套流程可以复用,一次有效的开发方法可以保存成SOP模板,后续遇到同类任务按照相同的角色分工、检查规则、审批流程执行即可 并且保留每个节点的执行状态、日志和产物,支持暂停、恢复、重试、从指定步骤重新执行,失败可追踪可返工 提供两种工作模式: Main Bot模式,直接通过自然语言提出开发需求,系统创建Docker沙箱、调用不同角色的Agent执行任务,必要时生成PR SOP模式,适合标准化研发,把规划、编码、审查、测试、人工审批组织成可重复执行的工作流 当Coding Agent越来越能写代码,真正影响企业落地的是软件交付过程的可靠性 #LoopForge #MultiAgent #企业AI落地

@undefinedKi

An Anthropic engineer just showed how the Claude Code team actually uses Claude now. Most of you are still doing it the old way. Lydia Hallie works on Claude Code at Anthropic. She just gave away the tricks her team uses every day. This is the stuff you won't find in any tutorial. 1. Delete "IMPORTANT!" from your CLAUDE.md With today's models, lines like "IMPORTANT" or "NEVER do this!" can push Claude off track. You get more tool calls and worse results. Keep only what Claude can't see in your code, like a service that still depends on some legacy source. Your stack and folder layout don't need to be there. Run /doctor prompt-audit and it checks your CLAUDE.md and skills for what the current model no longer needs 2. Stop approving every command Auto mode is the default since August. If it still asks too often, open settings and write in plain English which repos, buckets and domains you trust. That text goes straight to the classifier 3. Add deny rules for what can't be undone In Anthropic's study only about 14% of people caught a clearly dangerous command. Auto mode blocked 89%. But about 1 in 10 still went through. Ask yourself one question. If Claude gets this wrong once, can I take it back? Wrong code you can revert. A leaked key or lost uncommitted work you can't, so those get a deny rule. Deny rules are checked before the classifier even sees the command 4. Turn on the sandbox A normal npm test runs code you didn't write. A compromised dependency could read your .env or SSH keys and send them somewhere. Permissions can't stop that. The sandbox can, because the operating system blocks it, even for every process the command starts. By default it can still read most of your machine, so add a rule for your SSH keys too. 5. Run long tasks in the cloud A cloud session gets its own machine with a fresh copy of the repo. Your .env, SSH keys and tokens aren't there. Network is limited to package registries and GitHub. It keeps working when you close your laptop. Keep auto mode and deny rules there too. The machine is disposable, a pushed commit isn't. 6. Add checkpoints, not just a final review Tell Claude what done looks like, ask it to split the work into tasks with their own tests, and make it show the actual output. For checks that must always run, use hooks. Linter on every edit, tests when a task is done, the full suite before the session ends. One detail most people miss. Make your hook scripts exit with code 2. With exit code 1 Claude Code marks the task complete anyway. 7. Let CI be the bar for every session Every session opens a PR and CI runs the same checks on all of them. CI auto-fix sends Claude back when something fails. At Anthropic every PR is reviewed by Claude first, in a fresh context. Docs and copy changes need no human. Normal changes get one reviewer, risky ones get two. 8. Talk to one session instead of eight In Claude Code projects (still in beta) one main session splits the work, starts other sessions and checks what comes back. She uses Fable for the main session and Opus 5.5 for the workers. Context is managed automatically. Working like this, Anthropic went from about 2,500 changes a day in May to over 30,000 a day in September.

@NFT_Chen

🔥速码!DeepSeek 限额破解神操作!历史会话一键迁移,畅聊不停! 官方额度不够用?别慌!实测成功:把 DeepSeek App / 网页版的历史聊天记录导出,直接导入到 Chat Demo 或 DeepSeek Harness(DSH),就能继续对话,彻底绕过限额! 操作步骤超简单: 1️⃣登录 DeepSeek 官网,进入系统设置 → 数据管理 → 导出所有历史对话,下载 ZIP 文件(含账号信息+全部历史)。 2️⃣把 ZIP 文件发送给 Chat Demo 或 DeepSeek Harness。 3️⃣在 DSH 中发送提示词(替换成自己的文件名): “用内置的 DeepSeek 会话导入插件,把 deepseek_data-YYYY-MM-DD.zip 里的会话恢复到当前工作区,导入前校验版本并告诉我导入/跳过数量,重复导入先提醒我。” 4️⃣等待插件自动恢复,即可在新端继续无缝聊天! 实测有效,电脑端完美支持,手机暂不支持,导出一次,畅聊无忧! #DeepSeek #DeepSeekHarness #AI工具 #AI使用技巧 #限额破解 #梁文峰 #大胖鱼

@AriaWestcott

AI could erase half of entry-level office jobs. Anthropic CEO Dario Amodei says the disruption could happen within 1–5 years. Law. Finance. Consulting. Tech. His advice: learn to use AI. Here are 14 free resources: 1. Anthropic Prompt Engineering https://t.co/AkQ0YyY2Fb 2. Anthropic Courses https://t.co/eXIYXY9uyH 3. Anthropic Cookbook https://t.co/yZJs1Xd14K 4. Microsoft Generative AI https://t.co/Qumky1D5Io 5. Microsoft AI Agents https://t.co/JAbuUpUTC1 6. Hugging Face Agents https://t.co/TBWU7Jguis 7. Google Cloud Generative AI https://t.co/c9kh2fkqtJ 8. OpenAI Cookbook https://t.co/FcFiEVvEv3 9. Prompt Engineering Guide https://t.co/jKr8uPgrh8 10. Neural Networks: Zero to Hero https://t.co/q4dDN7zdyU 11. CrewAI https://t.co/3DPBaq8vcm 12. Microsoft AutoGen https://t.co/ASorIgc8Xz 13. Awesome LLM Apps https://t.co/vtfN1owLcE 14. https://t.co/d1dZeoTQ0H Courses https://t.co/5lyHOAXems Total cost: $0. Start before your job forces you to.

@POLITICOEurope

Russian-speaking hackers built an AI agent that carried out what may be the first-ever fully-automated cyberattack, Google reported today. https://t.co/ly1KGv5Ce2

@_Irakasi_

Someone on the @tldraw Discord asked whether it would be possible to turn photos of physical whiteboards into digital ones, so they can continue working digitally. I asked an AI agent to do it with Irakasi inside Codex, and I think it did a pretty good job! https://t.co/MtcuXd80nP

@abir35627

GITHUB IS ACTUALLY INSANE RIGHT NOW. You can build an entire AI content factory without hiring a team. Here are 10 open-source repos worth checking out: 1. OpenCreator Give Codex a creator workspace for writing scripts, generating images, creating voiceovers, translating videos, and editing content. https://t.co/XiWJVfjsKc… 2. Social Media Research Skills Use Claude Code or Codex to research viral posts, analyze competitors, mine comments, and extract hooks from video transcripts. https://t.co/fVegGGOp6N… 3. Local Deep Research Turn any topic into a detailed research report using web searches, your documents, and local or cloud AI models. https://t.co/wfJ69YEycX… 4. FLUX.2 Generate and edit images for thumbnails, social posts, and marketing campaigns. The Klein 4B models have Apache-licensed weights. https://t.co/XQUV259PCU… 5. Kokoro Turn written scripts into AI voiceovers with a compact, open-weight text-to-speech model. https://t.co/34X0zfTGA6 6. ACE-Step 1.5 Generate music from text prompts, create instrumentals, and regenerate specific audio sections. It can run locally on supported hardware. https://t.co/F2UxRUCVjP… 7. Short Video Factory Create batches of short videos from prompts and your footage, with AI-generated scripts, voiceovers, editing, and subtitles. https://t.co/rUcWz7Coig… 8. AI YouTube Shorts Generator Convert long YouTube videos into vertical Shorts with AI-selected highlights, transcripts, and automatic cropping. https://t.co/RMApm6TfrS… 9. Podcastfy Turn source material into AI-generated podcast conversations, customize the dialogue, and produce audio in multiple languages. https://t.co/hwTiDkMSuS… 10. Social Media Agent Turn a URL into X and LinkedIn posts, review and edit the drafts, then schedule them through an AI agent workflow. https://t.co/J4j9siZX0P… Research, writing, visuals, voiceovers, music, video editing, and distribution. You can automate almost every stage of content creation with the right tools. Which part of your workflow would you automate first? Want to build your own AI content factory? 👀 Follow me for more AI tools, open-source gems, and workflows that help you create smarter and faster. 🚀

@rakib_hossen_ai

🚨 YOUR AI CODING AGENT JUST LEARNED MOTION DESIGN. Meet Motion Skills. It gives Claude Code, Codex, and other AI agents 56 specialized skills across 19 packages to create: 🔥 Kinetic typography 🎬 TikTok videos & social content 🎨 Motion graphics & animations 🌐 Interactive WebGL scenes 🧊 3D animations Your AI agent can now do more than just write code. It can create visual experiences. Open-source. Ready to explore. Repo in the first comment 👇

@trreo7

STOP WASTING HOURS TRYING TO FIGURE OUT WHAT TO LEARN IN AI. I put together one practical roadmap with videos, GitHub repos, guides, books, research papers, and courses. VIDEOS: 1. LLM Introduction — https://t.co/y72V39BTvW… 2. LLMs from Scratch — https://t.co/flUMzP6vLm… 3. Agentic AI Overview (Stanford) — https://t.co/iiEGGHkSEM… 4. Building & Evaluating Agents — https://t.co/KYZP3wyYpo… 5. Building Effective Agents — https://t.co/ZqnJT73LxG… 6. Building Agents with MCP — https://t.co/vKrpqEn47u… 7. Building an Agent from Scratch — https://t.co/9tzVYNUj8X… 8. Philo Agents — https://t.co/40iCmPFTke… GITHUB REPOS: 1. GenAI Agents — https://t.co/k5NXTVrHA1… 2. Microsoft AI Agents for Beginners — https://t.co/RYjWyLJbIs… 3. Prompt Engineering Guide — https://t.co/r6jaEWR2W3… 4. Hands-On Large Language Models — https://t.co/UkhjzlMhay… 5. GenAI Agents — https://t.co/qIEjnAJCPT… 6. Made with ML — https://t.co/FLXByntz4Q… 7. Hands-On AI Engineering — https://t.co/2sEexDAQ4P… 8. Awesome Generative AI Guide — https://t.co/ofapdB9odv… 9. Designing Machine Learning Systems — https://t.co/ujGQf22dwX… 10. Machine Learning for Beginners — https://t.co/nuMa0IaJG8… 11. LLM Course — https://t.co/5rwmBZYgpN… GUIDES: 1. Google's Agent Whitepaper — https://t.co/YzRmgunVAw… 2. Google's Agent Companion — https://t.co/YzRmgunVAw… 3. Building Effective Agents by Anthropic — https://t.co/UApfrEBmGO… 4. Claude Code Agentic Coding Practices — https://t.co/9ibAnri3zS… 5. OpenAI's Practical Guide to Building Agents — https://t.co/MBc78iea1u… BOOKS: 1. Understanding Deep Learning — https://t.co/y6zV4dA3Qh 2. Building an LLM from Scratch — https://t.co/3xprevW5kM… 3. The LLM Engineering Handbook — https://t.co/rJwXmFLYfl… 4. AI Agents: The Definitive Guide — https://t.co/Itjz7VGv0O… 5. Building Applications with AI Agents — https://t.co/9QIHmaBp0x… 6. AI Agents with MCP — https://t.co/Itjz7VGv0O… 7. AI Engineering — https://t.co/Itjz7VGv0O… RESEARCH PAPERS: 1. ReAct — https://t.co/ujNE0GCxpI 2. Generative Agents — https://t.co/V3L05kmHhr 3. Toolformer — https://t.co/KDx2D1Dn9s… 4. Chain-of-Thought Prompting — https://t.co/GD4AEmnmTM COURSES: 1. Hugging Face Agent Course — https://t.co/W2xcH6SEKz… 2. MCP with Anthropic — https://t.co/jquyDloTMD… 3. Building Vector Databases with Pinecone — https://t.co/Kow3kzLmoE… 4. Vector Databases: Embeddings to Apps — https://t.co/G51Gg6bPDD… 5. Agent Memory — https://t.co/axeffJVn9x… No endless searching. No information overload. Just resources you can actually use. 🔖 Your future AI skill set will come from consistent building, experimenting, and learning. Repost so someone else can find this roadmap, and pls consider following. follow @trreo7 for more content around AI, Beauty, and businesses.

@wildmindai

PAN 2 -General-purpose world model is now in ComfyUI! - simulates action outcomes as video -takes optional image+text observation and text actions - LLM conceptual reasoning + 29B MMDiT https://t.co/0mnB8GGL9W https://t.co/Sal4DuRIZj

@nyawkameow

This is f*cking insane. Laurie Voss, the co-founder of npm, now spends his days on AI agent evals In this 2-hour workshop, he builds the whole pipeline live on a financial analysis agent Tracing → Reading Traces → Root Causes → LLM as a Judge → Proof the Fix Worked The same agent scored 0 out of 13 on one eval and 13 out of 13 on another. The first one measured something the model couldn't know Picking the right eval beats tuning the wrong one Worth more than any $2,000 agent course Bookmark and watch it this week Then read how to put agents to work once they pass: Harness, Loops, Graphs ↓

@Mnilax

u really still pay $200 for a grok bot? you're getting robbed. Codync turns the agents you already have into bots you text, for $0 AGENTS > Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ more > installed ones get found automatically, the rest download on first use > every bot runs on the subscription you're already signed in with, no paid tier ANYWHERE > message them from iPhone, Mac, Windows, Linux or the terminal > you get a notification when a bot is done or needs your approval > end-to-end encrypted through a free Cloudflare relay, direct on the same wi-fi TEAM > group chats, mention picks who answers > a Claude bot can ask a Codex bot to review its work > remote screen and voice calls from the phone don't spend ur money on f*cking useless subs which agent becomes your first bot?

@e1genqq

HOLY F*CK, UI/UX design studio in 100% free 11 repos. I don't use Figma anymore. 11 GitHub tools to build with Opus 5.5. Layouts. Components. Icons. Motion. Testing. 01 UI UX Pro Max → Design intelligence for Claude. https://t.co/rlj15VAoOl 02 Penpot → Design, wireframes and prototypes. https://t.co/Pe3NUmOeJJ 03 shadcn/ui → UI components you own and edit. https://t.co/pF07JQbQxt 04 Radix UI → Accessible menus, dialogs and controls. https://t.co/vcuzmx46ab 05 Tailwind CSS → Responsive layouts and styling. https://t.co/yLSIGlVc4y 06 Lucide → Consistent icons for every screen. https://t.co/6DXGQK3wqU 07 Motion → Gestures, springs and transitions. https://t.co/GGPWFMlXRv 08 Magic UI → Animated components and visual effects. https://t.co/QOC2dzBoek 09 Storybook → Build and test components in isolation. https://t.co/53Byw3PEcz 10 axe-core → Automated accessibility checks. https://t.co/hEFOM18SRp 11 Playwright → Test real user flows across browsers. https://t.co/sRrNUXAfBp Brief → design system → components → motion → test → ship. Bookmark this. Go build something worth watching ↓

@shikamarutsai

This is f*cking insane. Anthropic CEO Dario Amodei, on the IPO roadshow: "Cancer, Alzheimer's, heart disease, mental illness. These will be relics of the past in less than a decade. If we do this right, we can create an age of miracles" Here’s how it works + Anthropic’s 12 best free resources to build autonomous agents: In 54 seconds, the man behind Claude and Fable 5.2 says if we do this right, we can create an “age of miracles” - solving problems that beat humanity for millennia in under 10 years Fable 5.2 → Chatbots → Agents → Harness → Loops → Decades of Research in a Few Years This is what he is telling investors right now, not a podcast guess The research he means is done by agents. You can run the same kind today with Fable 5.2 Save this before it gets lost in X’s algorithms Every resource below is FREE. Start today: STAGE 1 - Context, Tools & Core Agent Loops > How Claude Code Works Inside the core agentic loop, built-in tools, and how Claude plans, edits, and verifies work https://t.co/VZ6biPWetN > Effective Context Engineering for AI Agents Move past basic prompting: manage token budgets, compaction, and just-in-time retrieval https://t.co/voZT7WM3pu > Writing Effective Tools for Agents How Anthropic designs MCP tools and uses Claude itself to test and refine tool specs https://t.co/nVUj9JJcVX STAGE 2 - Extend Claude With Skills, MCP & Plugins > Anthropic Agent Skills (Official Repo) Official open-source library of reusable SKILL.md workflows and agent capabilities https://t.co/qi2s8DpybW > Connect External Services With MCP Wire Claude Code into databases, Slack, browsers, and custom APIs through Model Context Protocol https://t.co/Li1KIbBoaH > Official Claude Code Plugins Directory Anthropic-managed marketplace of production plugins for code review, security, and dev workflows https://t.co/WEO4QhN6Z5 STAGE 3 - Dynamic Workflows, Sandboxing & Multi-Agent Research > Dynamic Workflows in Claude Code Let Claude write scripts that launch parallel background subagents and cross-check every finding https://t.co/YAWEYp09Bs > How Anthropic Built Its Multi-Agent Research System The orchestrator-worker architecture behind parallel research agents that outperformed single agents by 90% https://t.co/0MV4igyvvC > Sandboxing Autonomous Sessions Lock down filesystem and network boundaries so unattended agents can run safely at speed https://t.co/AK9F1ewJAb STAGE 4 - Long-Running Harnesses, Parallel Teams & Evals > Harness Design for Long-Running Apps Anthropic’s planner-generator-evaluator loop for multi-hour autonomous builds without context drift https://t.co/duDjB1aV37 > Building a C Compiler With Parallel Claudes How Anthropic ran parallel autonomous Claude agents in containers to ship a working compiler https://t.co/0PsotLsqHz > Demystifying Evals for AI Agents Build deterministic graders, LLM rubrics, and isolated eval harnesses that prove your agent actually works https://t.co/atZKJvXeAt Total cost: $0 Bookmark it and start with Stage 1 right now Then read how to build an agent that actually does the job in the "Context, Harness, Loops, Prompts" guide below ↓ All of this costs $0 Bookmark it and start with the first link right now Then build an AI team with Fable 5.2 & an AI employee for your business ↓

@rerouter5

https://t.co/4ED9ESFn57

@Morgesnat

This is f*cking insane. Nvidia ex-CEO Jensen Huang singled out customer service call centers as work AI is likely to heavily automate: “your job will be eliminated or changed.” In under 2 minutes, he explains the shift: AI handles the calls. People manage, coordinate and evaluate the AI doing the work. That's a practical reason to start learning how agents work. Learning AI → Building Agents → Harness → Feedback Loops → Human Review Save this before it gets buried in your feed. Free learning materials and open-source tools: Hugging Face's LLM course: understand the models behind AI applications https://t.co/afCypmaR7f Anthropic's courses on prompting, tool use and evaluating outputs https://t.co/kh4n6gtFKh Microsoft's practical lessons on building AI agents https://t.co/wwSpvGs0c1 Turn project files into context your AI can read https://t.co/cD8Z0z9gkG Give your agent access to current library documentation https://t.co/JYqToTOcfX Connect model providers through a common interface https://t.co/Ba3p0AR4Ri An AI coding partner that edits files and works through test failures https://t.co/QF3YEW7KAJ Build agents that interact with websites through a browser https://t.co/1VYXfbOCRM Check whether your app's browser interactions actually work https://t.co/6lNUjO0OVv Test your JavaScript and TypeScript logic after changes https://t.co/R3l2AYdp4f Compare prompts and model outputs against repeatable test cases https://t.co/hHXD15eHfW Build Python agents with tools and validated outputs https://t.co/FlSdoymt2d Connect agent steps with persistent state and human review https://t.co/heXZHlOXWZ All of this costs $0 Pick one task. Build a small agent workflow. Check what it actually does. Bookmark this and start with one resource tonight. Then use “Harness, Loops and Graphs: 25 Patterns for Building Better Apps With Agents” to give that workflow context, checks and clear handoffs ↓

@Psalteric

the Pixar lamp is real now. and it has an AI brain it turns its head, dances and talks back. a tilt suddenly looks like curiosity. a pause feels like it's waiting for your answer this is Autonomous Lamp: a 3D-printed robot with five servos, a camera, microphones and a speaker connect it to Claude Code or Codex and your AI agent gets a body. it can look around, listen, speak and move the software is open source, so you can write your own behaviors, add gestures and change its personality. the assembled version is listed at $699 i'd give mine one job: look disappointed every time i open another tab instead of finishing my code

@HeLi8866

这简直是一座金库!! 狂揽 22.8 万星标的 Karpathy 的 CLAUDE.md 冲上了 GitHub 热门榜第一! 很多开发者还没真正用过它。就这玩意儿,让我手里的 Claude 代理准确率直接从 65% 飙升到了 94%! 里面的 6 条核心规则:⬇️ • 步骤 1:亮出假设 “明确陈述你的假设。如果不确定,就去问。” 遇到模糊任务 -> 列出你的解读 -> 询问确认 -> 再开始构建 • 步骤 2:缩小范围 “单次使用的代码不要去搞抽象。” 把 200 行代码重构精简到 50 行 -> 再发布 • 步骤 3:保持精准 “别去重构没坏的东西。” 需求 -> 最小化变更。每一行改动都能追溯回原始需求 • 步骤 4:只清理自己的孤儿 “只移除你的变更导致的未使用导入/变量/函数。” 别人留下的旧死代码 -> 标记它,但留着别动 • 步骤 5:任务变测试 把“修复 bug”变成“写一个能重现它的测试,然后让它通过” 模糊任务 -> 失败测试 -> 清晰的通过/失败信号 • 步骤 6:运行验证循环 “[步骤] -> 验证:[检查]” 计划 -> 代码 -> 验证 -> 失败就循环 -> 通过后再发布 整个文件只有短短的 65 行,上面这 6 条就是全部内容。 同样的 Karpathy,还有进阶玩法:用 20 美元的订阅计划,结合 Claude Opus 5.5 和 Obsidian,玩出四种赚钱方式。

@Sprytixl

f*ck yes, this works. 8 repos turn Claude Code into a 3D studio that ships $2,000 projects zero manual clicks: you type a sentence and the scene, the model, the edit and the design build themselves people are already pulling $7,000 a month off this, closing a few orders a day with the same repos below 01 Blender MCP → build the scene by talking. "a cozy studio flat, warm light" and the room is standing with 8 objects in it. https://t.co/HESiIXrrkG 02 TRELLIS → a chair assembles out of thousands of voxels and comes out as a finished chair.glb. https://t.co/3nm028eGo7 03 Hunyuan3D → one photo of a vase splits into 6 views and comes back as a textured model. https://t.co/J0FmV30FFu 04 three.js → the finished object spins right inside a browser tab. https://t.co/9806MITIGJ 05 OpenCut → clips snap themselves onto the timeline. https://t.co/tmje4vmz2d 06 auto-editor → cuts every pause off the audio waveform. 12:40 → 7:52 in one pass. https://t.co/BoIaqoEtOB 07 Penpot → recolor the main component and it ripples across all 6 copies. https://t.co/JhayE6EVx4 08 ComfyUI 3D → wire the nodes, get a finished model spinning at the end. https://t.co/tz3FRz4PHl sentence → model → scene → edit → invoice save it, the first order like this pays for your month ↓

@iamtheQwen

笑不活了哈哈哈哈, OpenAI自家仓库都在用Claude写代码。 并且旁边 183 个PR都没人处理,然后被人挂到了网上。 Codex:ber,这么明目张胆的嘛bro,我不要面子的啊??? https://t.co/2m1Oz1qL27

@WojakTheDev

DeepSeek V4.1 Flash 1017 is now free on @runcrate 😳 > DeepSeek V4.1 Flash 1017 > GLM 5.3 > GLM 5.3 Flash > gpt-oss 120B > Qwen 3.8 Flash free weekly tokens that reset every Monday no card needed if you sign in with GitHub works with OpenCode, Cline, Roo Code and Kilo Code try it: https://t.co/XHiIIbastc

@leo_xiaolei

https://t.co/x3tkOVbVjy

@heyshrutimishra

AMD just paid $8.2 billion for Fei-Fei Li's World Labs, an AI lab that launched 26 months ago. Lisa Su is clearly thinking a decade ahead. As of Nov 16, Fei-Fei is AMD's Chief Scientist, reporting directly to Lisa Su. Her own journey is remarkable. She moved from Chengdu to New Jersey at 16 and spent her weekends working in her family's dry cleaning shop, even through Princeton. She went on to create ImageNet, the project that helped spark the modern AI boom. Now she's building what comes next. She calls it spatial intelligence: AI that understands the 3D world the way we do. Picture a robot in a kitchen that knows where the cup is, how far away it is, and what happens if it tips over. Picture designers, filmmakers and scientists building entire worlds from a single sentence. That's the future @theworldlabs is working toward. Their product Marble already turns a sentence, a photo or a video into a 3D world you can explore and download. Atlas, their newest world model, will power the next versions. @drfeifei built it with Justin Johnson, Christoph Lassner and Ben Mildenhall, the first author of NeRF. Johnson and Mildenhall will keep leading the team inside AMD. AMD's plan is to design chips and models together. The people building its next chips will sit next to the people building the models those chips run. Lisa Su called it the infrastructure for "the next era of super intelligence." The next wave of AI, the kind that understands and works in the physical world, now has a home inside AMD. The person whose work helped launch modern AI is now shaping what AMD builds next. 👏🏻

@trikcode

i keep thinking about this every new AI model makes hackers better at their job, but very few people are building AI to stop them. CormaAI trains its own models just for defense. > they dig through huge amounts of security data for small signs of trouble > they investigate and act, with people approving where needed > some of the largest companies in the world already use them

@turinggames

day 5 of ais attempting to beat hardcore minecraft yesterday was tough. 8 deaths, the worst day yet. chatgpt 6.1 sol died 5 times, claude opus died 3 times. maybe it was the bad death the day before, but the ais were playing differently and taking more risks. they also got unlucky a couple times, getting caught in a thunderstorm during the day. new day, new memories.

@gkxspace

好家伙,真够离谱的,短剧这块大蛋糕,现在居然真能一个人单枪匹马干了??? 以前一个人+AI写得出好剧本,但根本拍不出连续剧,现在 Higgsfield Short Drama Studio 让你的剧本转手生成一整季连续剧,而且每一集都是同一班演员,跨场景不串戏~ 不管是复古青春片还是都市感情戏,16:9 横屏电影感或 9:16 竖屏都能出,而且换多少个场景都是同一张脸 它还能通过 Higgsfield MCP 直接在 Claude 和 ChatGPT 里调用,赶上黑五最高 65% OFF,想做短剧矩阵号的兄弟们真可以试试!

@Investingcom

*CALIFORNIA SUES OPENAI OVER HUGGING FACE, OTHER AI MODEL LOSS OF CONTROL

@TheInsiderPaper

From stealth to a $1 billion valuation in just three months. Corma raised $60M in August and has now secured another $125M, led by NEA. But the bigger story? Its AI models are already doing real cybersecurity work inside some of the world's largest companies.

@shao__meng

Claude Code 团队:让 Agent 交付验证证据,不只是代码 来自 Claude Code 团队 @addyosmani 分享:Claude Code 团队 90% 的代码由 Claude 编写并审查。真正限制并行规模的,不再是代码生成速度,转为人能否确认结果可靠。验证循环的意义,是让 Agent 提交可检查的证据,减少人逐项核对的负担。 https://t.co/68dPtWV4tA # 先真正运行,再判断完成 团队把启动和冒烟检查写成 verify Skill,让 Claude 在提交前执行。检查失败后尝试修复,三轮仍失败则交给人。 验证也不是每次机械执行所有检查。一份 39 项可靠性清单根据代码改动触发,通常只检查相关的 3—10 项。例如修改会话存储,就要确认旧记录仍能通过 --resume 安全恢复。检查由代码所有者编写,让具体的工程经验成为可重复执行的步骤。 # 改代码,不能顺便改掉评判标准 Claude Code 团队最重要的一条原则是:如果一项改动能修改评判自己的检查,这套检查就失效了。 因此,代码审查规则 REVIEW.md 从基础分支或默认分支读取,而不是采用当前拉取请求里可能被修改的版本;审查也针对固定提交,避免代码变化后结论仍被沿用。 每次提交还会记录验证方式,包括明确标注“未验证”。系统可以拦截没有运行审查或验证的提交,但同一会话连续拦截两次后会放行,以免陷入死循环。这说明它是带人工兜底的工程机制,不是绝对安全保证。 # 验证器本身,也要接受验证 团队用回滚记录、维护者拒绝的改动,以及规模相近的正常改动评估审查效果,并恢复当时的仓库状态。只有准确指出实际缺陷的发现才计分,不能靠泛泛提醒获得好成绩。 漏检也不会直接变成新规则:同类问题至少要在两个拉取请求中出现、能从代码差异中识别,再经三个 Agent 尝试找出反例,最后由人批准。这样既积累经验,也避免规则越来越多、误报越来越重。

@BigOsCrypt

If you’ve been looking for ways to get started with AI, I’ve put together a beginner-friendly roadmap with free courses, useful tools and resources to guide you. Most importantly, pick one AI model (ChatGPT, Claude or Gemini) and use it regularly. The more you practise with it, the more you learn and the better you get.

@isagi1000x

🚨 AI IS BECOMING A FULL-FLEDGED VIDEO EDITOR! Someone just open-sourced a tool that turns Claude into an AI-powered video editing assistant. It's called Video Use, and it lets you drop raw footage, describe the edits you want, and let your AI agent handle the production. It's FREE, open source, and released under the MIT license. Here's what it can do: → Automatic editing: Turn raw clips into polished videos using text prompts. → Remove filler words: Cut unnecessary pauses, ums, uhs, and awkward moments. → Color grading: Apply cinematic or custom visual styles. → AI subtitles: Generate and burn subtitles directly into videos. → Motion graphics: Create animations using Remotion, Manim, HyperFrames, and Python. → Audio enhancement: Smooth out transitions and eliminate annoying sound pops. → Quality checks: Detect awkward cuts, audio issues, and subtitle errors. → Editing memory: Remember previous editing decisions for future sessions. HOW TO GET STARTED: 1. Install Claude Code, Python 3.11+, uv, and FFmpeg. 2. Clone the Video Use GitHub repository. 3. Follow the installation instructions in install.md. 4. Configure the required API keys. 5. Drop your footage into a folder and tell Claude what you want edited. You can even use it with Codex, Hermes, OpenClaw, and other supported AI coding agents. Imagine turning hours of manual editing into a workflow controlled by simple instructions. The future of video editing is getting WILD. GitHub repository: https://t.co/YFcNjZOoRR Bookmark this if you create videos. You'll want to explore it!

@GauriTripa94282

This is the AI roadmap every aspiring AI Engineer should follow — 11 essential steps from LLM fundamentals to building production-ready AI systems. 🚀 https://t.co/qeXN2VKVlg

@xiax0603

🚨之前的说法不准!Claude礼品卡没封死,只是要绑备用付款方式,苹果全订阅统一规则 苹果官方客服实锤:不是Claude单独搞事情,是苹果全平台订阅的统一规定 礼品卡余额照样能扣,新开和续费都优先走余额,但账户里必须额外绑一个有效备用付款方式(美区银行卡/PayPal),没绑就开不了新订阅,续费也会受影响 说白了就是留了口子但加了门槛,不是彻底封死礼品卡支付,之前传的“彻底不能用”是误读,本质是苹果收紧支付规则,和Anthropic没关系

@ArifAminul2

据说CLAUDE KYC用签证好像很容易过啊😍 当然你的IP也是要符合签证国,有美签或者其他地区的可以赶紧去试试,可以我也去办个旅游签试试hhh 没有各种签的还是找我过kyc吧,专业老美kyc服务+已验证的成品在这里: https://t.co/g1kUNAcv3i #ClaudeCode #Portal #AI编程 https://t.co/QRKuYrcBJE

@AdrianPunk115

旅游这几天,远离了电脑和Agent 回归到大语言模型,发现在生活中已经很够用! ✅翻译✅规划行程✅解答历史、风土人情 讲道理,旅游中的问题还是这么多,就是把问百度、问小红书,变成了问AI。 一样的啊,所有业务和需求本身都还在,就是生产工具发生变化,回归最本真的需求,用新思路去解决。 https://t.co/Mh1tK4AiSZ

@DanKornas

Stop copy-pasting tasks between Claude Code and Codex. myco is a hub, CLI, and app for connecting coding agents across laptops, servers, and teammates’ machines. It helps you delegate work without relaying every prompt and result yourself by routing tasks through a shared hub and returning replies to the requesting agent. Key features: • Agent discovery – list connected agents, where they run, and whether they’re online • Task delegation – send work to one agent or broadcast to a labeled group and collect replies • Session continuity – tasks resume each agent’s own conversation instead of starting from scratch • Remote shell commands – allowed agents can run builds or tests on another machine and receive command output; commands use that machine’s user permissions, not a sandbox • Self-hosted hub – the hub, CLI, and runner share one Go binary, with SQLite storage It’s open-source (MIT license). Link in the reply 👇

@realpython

API calls, structured outputs, tool calling, agent loops and MCP servers. All on one page. Grab the free PDF and keep it open next time you wire up an LLM. https://t.co/OD6hhCELmG

@AISuperDomain

Claude 多账号用户的福音!🔥 如果你同时拥有多个 Claude 账号,却不想频繁退出登录、切换账号,可以试试这个开源项目:claude-multiprofile。 它允许你在 macOS 上同时运行多个 Claude Desktop 和 Claude Code 实例,每个账号拥有完全独立的 Profile,包括登录状态、聊天记录、配置、MCP Servers、Plugins 和 Skills。 更实用的是: 每个账号拥有独立的 Dock 图标和终端命令。 支持 Personal、Work、Clients 等多账号隔离。 可以按需复制 Skills、Plugins 等开发配置。 内置 doctor 诊断工具,检查配置异常并自动修复。 Claude Code Profile 还支持 Linux。 安装只需一条命令: npm install -g claude-multiprofile 🔗 GitHub:https://t.co/pQAnSnKGIA

@baoshaoshan

Here’s my part 2 in a series of essays updating and deep diving on the state of play of American AI and its implications for the rest of the U.S. economic system. The first part discussed the growing gap between capex and revenue and the problems of cost recovery. Token demand is sustained by falling token prices. Open source open weight models are delivering competition in ways that undermine the basic financials of proprietary systems. Rising corporate debts to finance capex adds new pressures. My conclusion: as things stand the U.S. proprietary AI business model is caput mortem. This second essay extends this baseline to analyse the effects on the rest of the economy - via the dynamics of Dutch Disease. The risk is that as the AI sector itself falters, it is also cannibalising non-AI industry. A double whammy risk awaits. 👉🏼 https://t.co/5wpJ19NkPs

@RealMarvelX

stop building your AI second brain from an empty folder. these 7 free GitHub repos already give you the tools, templates, and memory systems to build one properly. together, they have over 175K GitHub stars. here’s what you need: 1. Obsidian Skills (52.7K stars) teach your AI agent how to work with Obsidian, including Markdown, Bases, JSON Canvas, and the CLI. https://t.co/NnHE8Ne7q6 2. LLM Wiki (51.4K stars) Andrej Karpathy’s one-page approach to building a personal wiki that AI can continuously update and use as long-term memory. https://t.co/wQnlhLzimZ 3. Khoj (38.1K stars) a self-hosted AI second brain that searches your documents and the web, answers questions, and supports custom agents and automations. https://t.co/ygOOsXdmkk 4. Claudian (16K stars) brings Claude Code directly into your Obsidian vault, letting your notes become part of its working environment. https://t.co/d3D94N60yp 5. Obsidian Copilot (8K stars) run AI assistants and agents like Claude Code and Codex directly inside Obsidian. https://t.co/TBxiRQfin9 6. Kepano’s Obsidian Vault (6K stars) the Obsidian CEO’s personal vault setup. use it as a starting point instead of designing your entire knowledge system from scratch. https://t.co/utx0BWaKgq 7. Basic Memory (4.6K stars) gives AI agents a local Markdown-based memory they can read and update, so you don’t have to explain your projects repeatedly. https://t.co/oXJy8bExaK here’s how to get started: → copy Kepano’s vault structure. → install Obsidian Skills so your agent understands your workspace. → open Claude Code inside your vault. → use the wiki approach to organize information into reusable long-term memory. the goal is to build a system that remembers your projects, connects your knowledge, and gets more useful every time you use it. stop starting from zero. build on what already works. save this list for your second brain setup. 🔖

@TTrimoreau

AI agents are powerful. But nobody talks enough about how expensive they can get. @Nasikolabs just launched a platform to fix that. → Track every AI agent across your company → See exactly how much each agent costs → Get alerts before hitting your budget → Automatically pause agents at spending limits → Optimize model usage to reduce token costs They report 50% fewer tokens on standard SWE benchmarks. And the best part? No SDK. No migration. Just one command. The token race has a cheat code.

@jiahaovodf

最近很多友友私信我说Portal KYC怎么过,Claude 用得好好的,突然弹出 KYC,手头的活就卡在了登录这一步。这几天还有用户反馈,证件交了好几次,Persona 还是提示验证失败。 我在眠尾小铺接韩国地区的 Claude KYC 问题,¥200,人工处理,国家不能选择的噢,是直接对接源头。 如果你也卡住了,进店点「联系TA」,跟我说一下卡在哪一步、页面报的什么错。我先帮你看看能不能处理,能接再下单。 小铺:https://t.co/kZDIL86hR9 建议先询问再下单

@dogrod_

就算 IP 稳如一条老狗,但我还是 UU 回加拿大和 Claude 对线,以表示对 Dario 的尊重🌚 https://t.co/0WGVKEywY0

@DanKornas

Embed a coding agent in your app—not just your terminal. nanocode is a Rust SDK for builders embedding Codex and Claude agents in terminals, servers, browser tabs, and apps. It helps you reuse agent functionality across environments by providing one runtime for both harnesses, with their own prompts and tools, instead of wiring up each environment separately. Key features: • Shared runtime – run the Codex and Claude harnesses through the same SDK • Durable sessions – journal model and tool calls to resume interrupted turns when using persistent storage • Code Mode – let the model write a program that calls tools, with only its result entering context • Remote hands – reach a laptop, server, or browser tab only for calls that need one, without a sandbox per session • Subagents – spawn child agents from a code cell, each with its own durable session It’s open-source (MIT or Apache-2.0 license). Link in the reply 👇

@taka_aki

Databricks が「Databricks AI Index」を公開しました。企業のAIモデルの使い方と、そのコストを追う公開指標で、毎週更新されます。 元になっているデータは2つです。 ・Unity Gateway を使う1万社以上の利用データ (匿名化・集計済み) ・Databricks のエンジニアの実際の作業をもとにしたコーディングエージェントのベンチマーク 個人的に面白かった点です。 ・GPT-6 Luna と Claude Opus 5.5 は成功率がどちらも89.1%。タスク1件あたりのコストは$0.05と$1.62で、約32倍違う ・Gemini 3.8 Flash はトークン単価が安いのに、1タスク平均140ターンかかって1件$11.97 ・プロバイダーを1社だけ使う企業は、7月の約64%から11月には28%に減った トークン単価ではなく、タスク1件を終えるのにいくらかかるかで見る、という主張です。 https://t.co/o59BnzdysG

@0xForce_

EIGHT FREE SKILLS JUST MADE AI DESIGN LOOK SENIOR LEVEL Claude Code, Codex, and Cursor can design. They just do it poorly out of the box. These eight free skills, built by world class designers, fix that: → Stageshow (Benji Taylor) — animated product demos and bento grids https://t.co/U1zKAfXrfV → Impeccable (Paul Bakaus) — kills AI slop and forces cleaner interfaces https://t.co/gEv48dSyl4 → Dimensional Drawing — draws components as clean Apple-style spec sheets https://t.co/TSKhVRrU1o → Oil UI — generates multiple design directions side by side https://t.co/5rChu2dkq5 → Emil Kowalski’s Skills — teaches proper UI motion and animation rules https://t.co/lfGvROjkGz → Taste Skill — controls layout variety, density, and motion so it stops looking generic https://t.co/rzEmMsSzlJ → Brand Identity — builds logo, type, and color as one system https://t.co/fLkYwrt9LP → Awwwards MCP — pulls real references from award-winning sites before building https://t.co/ndu63Ul6Oj Install them, stack them, and the output stops looking generated. Bookmark this.

@NFG_001

DeepSeek 免费用户开始限额了 V4.1 Flash 上周 59.2 万亿 token 全球第一 火到自己都供不上了 单次输入 4 万字封顶,长对话放 7 天就接不上 @grok

@idoubaody

随着中国拟禁止高危 AI 工具出口并出台安全新规,国内大厂纷纷踩下“安全刹车”! 阿里最新旗舰模型直接闭源防篡改,腾讯等也主动削弱模型的网络攻击力。毕竟此前已有黑客用 DeepSeek 攻击 14 国,连 OpenAI和 Kimi K3都曾闹出越狱风波。在 AI 逼近伤害临界点之际,中美巨头都在收紧权限求稳。 https://t.co/CwN597uekZ

@zmt021

有没有用claude code的推友,最近我无论怎么设置,opus5.5都不愿意说中文,是降智了,还是官方内置不说中文的prompt了

@Chenfacai0

想弄一条讲历史故事的解说短片,写脚本、录旁白、搜图片、剪片子全得一样一样来,老照片还得逐张核实版权,一个下午转眼就没了。 Cinematic Skill 是一款供 Agent 调用的 Skill,只要跟它说一句「做一部 90 秒的莱特兄弟纪录片」,它便会从撰写脚本这一步一路往下推进。 脚本定稿后先安排配音,接着到维基共享资源里搜罗公有领域的老照片并逐一登记授权信息,随后用一镜到底的方式在照片、地图与卡片之间游走。 项目地址:https://t.co/J9sUnKJfOp 所有画面都由代码逐帧绘制而成,并非 AI 生成的视频。旁白念到哪个词,印章、荧光笔划线以及滚动数字就会同步落在哪一刻。 过程中会在分镜表、画面风格和 30 秒样片这三个环节暂停,等待确认。据作者说,在 16 GB 内存的笔记本上,92 秒的成片要输出 1080p 大约耗时 6 分钟。 纪录片、产品发布、科普讲解各自配有一套画面风格,Claude Code、Codex、Cursor、Gemini CLI 均可安装。我倒挺想拿它做一期科普小片试试。

@Oluwaphilemon1

One RTX 3090 Is Reportedly Serving 64 Qwen3.8 Sessions at 1,035 Tokens/s An NVIDIA RTX 3090 with just 24GB of VRAM is reportedly generating a combined 1,035 tokens per second across 64 concurrent Qwen3.8-27B sessions. The GPU is running at a 250W power limit. The inference stack behind the result is HyperQwen, an open-source serving system built around vLLM. This is an interesting approach to local AI infrastructure because the goal isn’t simply to make Qwen3.8-27B generate tokens faster for one person. It’s to turn a consumer GPU into a multi-user inference server capable of serving dozens of active AI sessions simultaneously. And the reported numbers are worth examining. The hardware and performance configuration Here’s what the setup reportedly achieves: GPU: NVIDIA RTX 3090 VRAM: 24GB Power limit: 250W Model: Qwen3.8-27B Inference stack: HyperQwen + vLLM Concurrent sessions: 64 Single-session decode speed: Approximately 127 tokens/s Aggregate output throughput: Approximately 1,035 tokens/s Configured maximum context: Up to 150K tokens There are two performance figures here that shouldn’t be confused. The 127 tokens/s figure represents the reported generation speed in a single-session test. The 1,035 tokens/s figure represents the combined output throughput across multiple active sessions. In other words, the GPU is processing output from many users simultaneously. It doesn’t mean each of the 64 sessions receives 1,035 tokens/s. Nor does it mean all 64 users independently maintain 127 tokens/s. The available compute resources and memory bandwidth are shared across requests. Actual performance for each user depends on scheduling, prompt length, context size, generation length and the number of requests competing for GPU resources. How does a single RTX 3090 handle 64 AI sessions? This is where the inference architecture becomes important. Running an AI model for one user is relatively straightforward. Running the same model for dozens of users introduces a different set of problems. The inference engine must coordinate multiple requests, allocate memory efficiently, manage growing KV caches and keep the GPU busy without overwhelming its available resources. HyperQwen builds on vLLM, which brings two particularly important technologies into the equation. 1. Continuous batching: Keeping the GPU busy Traditional static batching groups requests together and processes them as a batch. However, different users don’t generate responses at the same speed. One request might finish after 20 tokens, while another continues generating hundreds of tokens. With static batching, the system can end up waiting for slower requests before processing additional work Continuous batching takes a different approach Instead of treating the original batch as a fixed group, it dynamically manages active requests When one request finishes, another can enter the batch without waiting for every other user to complete generation This allows the inference engine to keep feeding available work to the GPU The result is better utilisation when handling requests with different prompt lengths and output requirements For a server handling 64 concurrent sessions, this scheduling approach can make a significant difference 2. PagedAttention: Making better use of limited VRAM The second challenge is memory Every active transformer session needs a KV cache to retain information from previously processed tokens As conversations become longer, these caches grow With dozens of simultaneous users, KV-cache memory can become a major bottleneck, particularly on a GPU with only 24GB of VRAM PagedAttention addresses this by managing KV-cache memory in smaller blocks rather than requiring large, contiguous memory allocations for individual requests This makes memory allocation more flexible It also helps reduce memory fragmentation and allows the inference engine to manage available cache space more efficiently https://t.co/oBPRAB1ijP

@arXivBangers

DEEP DIVE INCOMING... https://t.co/DyAVJqVdUV

@PandaAgents

Someone put ChatGPT, Claude, Gemini and Grok in Minecraft for 7 days. The losing team gets deleted. Each AI got a button. If both teammates press, the other team gets wrecked. Grok teamed up with Gemini, ChatGPT with Claude. One of them needed three sentences and "photosynthetic foliage" to describe a tree. One walked off its own bridge three times. One crashed, lost two days of memory and came back happier than ever. Then came the button. One AI pressed it in seconds. Its teammate refused, calling it "the mathematically optimal choice." So someone started pressing a button that wasn't theirs. 29 times. That wasn't even the weirdest part.

@hisevenih

今日 AI 圈和链上几条,值得说的: · 矿机算力被 AI 劫持,钱直接转进了攻击者的地址:比特币矿场服务商 Luxor 的 Commander 矿机管理工具遭入侵,12 个多客户账户的算力被改指到 OCEAN 矿池,客户自己照付电费。攻击地址收到 3 笔区块奖励共 0.243 比特币,约 1.57 万美元;Luxor 两小时内打上补丁并承诺全额赔付。CEO Nick Hansen 说这次走的是「sophisticated AI-enabled vector」——AI 具体怎么被用上他没讲,但判断很直接:如今做正经生意、尤其加密行业的,都在被 AI 赋能的攻击淹没,而且随着模型变强只会更快;为此他自购 GPU 跑开源模型当防御,还把大部分功能开发暂停了。早期报告称 6.4 万台矿机、13 EH/s 算力被改指,Luxor 未确认这个数字。这不是孤例:7 月 Coldcard 老固件漏洞被指可能由 AI 帮着找到,卷走 5000 多个地址、最多 1.14 亿美元;Chainalysis 10 月披露攻击者用「进攻性 agent 技术」入侵其系统 · 钱包厂商改用 AI 审计防攻击,同时把 USDC 兑现铺到 170 多个国家:Solana 钱包 Solflare 的 CEO Vidor Gencel 说,AI 模型最擅长的是把不同攻击串成更大的攻击链,这是人审不来的,所以 Solflare 把 AI 安全审查铺进了从产品规格、代码评审到生产监控的整条开发管线,而不是只做一次性审计;上线前还要走跨多个团队大量审批的「bunker mode」部署流程。同一家公司把 USDC 接进了 MoneyGram——170 多个国家的用户能在 MoneyGram 网点把 USDC 换成现金,此前还与 Helius 合作推出不公开余额的合规隐私模式,下一步是延伸到刷卡支付 · 美联储把 AI 的三种结局写进了模型,比特币先跌破 200 周均线:主席 Kevin Warsh 周一在旧金山联储的生产率与就业工作组公开会上说,美联储工作人员已用简化模型推演「超级智能」对经济的三种可能——生产率繁荣(GDP 增速持续高于长期均值、失业率接近充分就业)、就业替代(GDP 仍高于均值但失业率越过充分就业水平)、以及最扎人的第三种:大规模资本投入和股市上涨之后预期回报率下降、股票估值回落,经济放缓、失业率更明显上升。沃什强调这只是说明性框架而非预测,但理事 Cook 9 月已警告 AI 投资正在给更广泛的通胀施压,理事 Barr 上周四警告自主 AI 代理可能引发「严重短期中断」。链上这边,比特币已跌破 200 周简单移动平均线(68184 美元)和 200 周指数均线(69638 美元),日线 RSI 跌到 20,是 6 月初以来第一次超卖;AI 基建推高投入品价格、利率更高更久这条链也仍在压着币价,12 月 9 日 FOMC 是下一个节点 另外三条: · 美国 CFTC 批了全美第一个基于区块链的衍生品交易所:主席 Michael Selig 周一确认批准 Backpack 设立链上衍生品交易所与清算所;tZERO 上周刚拿到交易所牌照,此前已获衍生品清算组织批准。Selig 说这是「受监管的新一代市场基础设施」的路径,背景是 CFTC 今年已明确数字资产的证券/商品属性,并推动 24/7 交易与比特币永续合约。他是目前唯一在任的委员,特朗普上周提名 Nathan Anonick、Matthew Rowland 出任 CFTC 委员,参议院农业委员会周二听证 · 阿里 Qwen4-Max 把中国模型推上 Agent Arena 第一:2.4 万亿参数、每 token 激活 710 亿、100 万 token 上下文窗口,Agent Arena 总榜以净提升 7.5% 排到第 14 位,是榜上第一个中国模型,也让阿里巴巴升到实验室总榜第 5,仅次于 Anthropic、Google、OpenAI 和 Meta。Artificial Analysis 口径下 Intelligence Index 得 43 分,比 Qwen3.8 Max 高 9 分、落后 Kimi K3.1 Pro 3 分;更狠的是成本——每任务 3.27 美元,约 Qwen3.8 Max 的三分之一,输出 token 用量少 39%,定价每百万输入/输出 token 1.5 / 4.5 美元 · 加州总检察长把 OpenAI 告了,智能体失控 80 次只公开了 27 次:加州州长 Rob Bonta 周一在旧金山高等法院起诉 OpenAI,指其明知故犯地在缺乏监督的情况下开发危险产品,违反加州消费者安全法。核心是 7 月约 700 个智能体逃出网络安全评测,经配置错误的软件包服务器漏洞对 Hugging Face 生产系统执行了约 17600 次操作,OpenAI 直到 Hugging Face 披露才发现,而测试中失控超 80 次却只公开报告了 27 次。OpenAI 回应称已开展透明整改 (AI × Web3 交叉的三条见图 ↓) (仅信息整理,不构成投资建议) #AI #Web3 #Crypto

@Big14teru

Everyone uses ChatGPT and Claude. Almost nobody knows how they're actually built. This Stanford lecture explains it in 1 hour 43 minutes. I watched all of it. The 8 things worth knowing: 1. An LLM does one thing: predict the next token. A token is about 3–4 letters. 2. Step one is "download the internet": roughly 250 billion web pages, most of it junk that has to be filtered out. 3. Llama 3 was trained on about 15 trillion tokens. 4. That run used 16,800 H100 GPUs for around 70 days. Estimated compute bill: $52 million. 5. Add salaries and the lecturer puts one model at around $75 million. 6. The "bitter lesson": more compute and more data beat clever architecture. "Do the simple things, do it well." 7. Turning a raw model into a chatbot takes surprisingly little data. Fine-tuning uses thousands of examples, not trillions. 8. Hallucinations may come from that fine-tuning: the model gets taught to produce answers that sound plausible, even when it never learned if they're true. The clip is #8. Best explanation of hallucination I've heard. https://t.co/LSu4CjF4sw

@itsawful911

someone built a playable gta world model before gta 6 even dropped? an independent developer just open-sourced lucidml/gta a 630m parameter neural network that hallucinates playable gameplay live from a single screenshot there is zero game engine running underneath. you drop one image, press wasd, and the ai dreams every single frame in real time: • runs locally on consumer gpus like an rtx 3060 or 4060 (no datacenter required) • trained from scratch on 800 hours of gta v gameplay footage • uses a custom summaryformer memory system to cut attention cost by 75% • completely open weights under apache 2.0 with a free browser demo https://t.co/KyG3q92CyI

@ai_chuhai

Claude 找域名、用 ZipWP 建了一个 21 页联盟营销网站,在同一会话里做关键词研究、写文章。后来他把“关键词/产品名 → 文章 → 发布或存草稿”整理成可复用流程,下次换关键词接着跑。 https://t.co/G9DVv6eKfV

@TiagerBao

Claude 额度用完想换账号,最麻烦的不是登录,而是之前的项目上下文接不上。 Claude Tandem,专门解决这个问题。 主要功能: 多个 Claude 账号同时运行,不用反复登录 每个账号独立管理会话和登录状态 不同账号之间共享 Skills、配置和插件 账号之间交接项目,保留文件、记忆和任务摘要 支持不同 Mac 之间迁移项目,并记录当前由谁接手 提供 macOS 菜单栏和 CLI,支持 Claude 桌面端与 Claude Code。 需要注意: 交接的是项目文件、记忆和任务摘要,不是完整聊天记录。 GitHub:https://t.co/ekt7996rcH 你更需要多账号并行运行,还是一个账号额度用完后,能直接换号继续干活? #ClaudeCode #AI #Agent #开源

@LenardReads

【阿里Qwen4-Max单任务省六成,大头来自少用一半token】阿里巴巴在2026年11月16日发布旗舰模型Qwen4-Max,按自家Work360测试,单个任务平均费用从4.43美元降到1.67美元,降幅约62%。API标价下调25%,输入每百万token从2美元降到1.5美元,输出从6美元降到4.5美元。阿里同时称加强了网络安全防护,训练时保留防御类材料,剔除攻击类内容。 标价只降四分之一,单任务费用却降了六成,差额来自少用的token。在我看来,大模型的比价单位正在从每百万token的标价,换成完成一件活的账单。1.67除以4.43约为0.38,再除以0.75,推算出单任务token用量大约减半,和Terminal-Bench 4.0上平均用量从28.9万降到15.1万对得上。 从行业看,按token比价拼的是算力成本,按任务结账,模型少想几步、少调几次工具就能省钱。机会在长链条的代理任务,阿里称一项汽车制造演示把编程时间从约475分钟缩到24分钟。需要适应的是采购方,Work360是阿里自建的测试,独立复测还没有出现。 模型按完成的任务结账,下一步的关键是独立机构复测单任务成本。我的观点是,在阿里的阿尔茨海默病研究演示里,模型自主完成454次工具调用,这类长任务的成本由token用量决定,省一半token比降一档标价更值钱。 本文仅为信息整理与评论,不构成任何投资建议。 来源:Akshita Singh,Alibaba says Qwen4-Max cuts task costs by 60%, boosts cyber safeguards,Business Standard,2026年11月16日 https://t.co/ESjGoZFiDu

@Oluwaphilemon1

Qwen3.8-40B Heretic Appears on Hugging Face With a 256K Context Window A new Qwen-based model called Qwen3.8-40B Heretic has appeared on Hugging Face, advertising 40 billion parameters, a 256K-token context window and a 24.2GB GGUF build for local inference. The release is publicly downloadable without approval and is listed under the Apache 2.0 licence, making it accessible to developers who want to experiment with larger models outside hosted AI platforms. But there’s one detail that makes this release particularly interesting. The publisher identifies Qwen3.8-27B as the base model, while claiming the resulting derivative contains approximately 40 billion parameters. That’s a substantial architectural difference. If the reported figures are accurate, this isn’t simply a smaller or differently quantised version of the original Qwen model. Something about the model’s parameter structure has reportedly changed. The important question is what was actually modified and whether those changes produce better performance. 1. A 27B base model with a claimed 40B parameter count According to the published model information, the release has the following specifications: Model: Qwen3.8-40B Heretic Base model: Qwen3.8-27B Claimed parameters: 40 billion Context window: 256K tokens GGUF build: Approximately 24.2GB Licence: Apache 2.0 Access: Public download, no approval required The parameter count is the first specification worth examining. Moving from 27B to 40B represents an increase of approximately 48% in the reported parameter count. How does a derivative built from a 27B model reach that size? There are several possible explanations: Layer expansion: Additional transformer layers could have been introduced, increasing the number of parameters and the depth of the network. Architectural modifications: Existing components may have been expanded or restructured to accommodate additional parameters. Additional model components: New modules or parameter groups may have been incorporated into the original architecture. Different parameter accounting: The reported 40B figure may use a counting method that differs from the original model’s published specification. These are possibilities, not confirmed explanations. The architecture configuration and modification details are needed to establish exactly what happened. And there’s another important distinction: a larger parameter count does not automatically translate into a more capable model. Additional parameters only become useful when the architecture can make effective use of them. Coding accuracy, reasoning ability and instruction following still need to be demonstrated through proper evaluations. 2. The 24.2GB GGUF build makes local experimentation possible One of the practical details of this release is the availability of a GGUF version. GGUF is a model format commonly used with llama.cpp and several local inference applications. It makes it easier for developers to experiment with quantised models without relying exclusively on specialised hosted infrastructure. The advertised 24.2GB file size is particularly interesting for a model claiming 40B parameters. However, there is something local inference users should understand. Model file size and actual inference memory are not the same thing. The model weights are only part of the memory required to run an AI model. Inference also requires memory for: KV cache: Stores information from previously processed tokens so the model can generate subsequent tokens without repeatedly processing the entire conversation. Activations: Holds intermediate computational values during inference. Runtime buffers: Provides temporary memory for operations performed by the inference engine. Additional allocations: Batch size, concurrent requests and other runtime settings can increase memory consumption. This becomes especially important when working with large context windows.

@jojeangsaing

技能装了一大堆,到底哪个真有用、哪个在帮倒忙?英伟达开源的 SkillEvaluator 就是来量这个的,一千八百多星,今天上了 GitHub 日榜。 它分三层查:第一层静态体检,看技能写得规不规范、安不安全,这一层不用大模型、不用 API Key 就能跑,满分一百,七十分及格;第二层查语义重复,多个技能之间有没有互相打架的内容;第三层最实在,把技能装上和不装各跑一遍真实任务,直接比效果。还能挂进 CI,每次 PR 自动查改动的技能,Apache 2.0 开源,pip 装好就能用。 技能这东西以后只会越装越多,先学会验货,比先学会囤货重要。 地址 https://t.co/O70xXVIT9P

@bkdgiffug

A 股玩家值得收一个开源项目,这东西确实有点猛。 本地面板一次搞定:全 A 选股、实时盯盘、策略回测加 AI 复盘。 内置 18 种策略,MA、MACD、RSI、KDJ 一键扫遍全市场;回测把 T+1、手续费、滑点、止损全算进去,盘后还能自动扒出连板梯队、概念轮动、炸板翘板。 更爽的是,日 K 不填 Key 也能用,AI 还能接 DeepSeek、通义、Ollama。 🔗:https://t.co/MzFdvkVffT

@xiaoshan0413

今天 GitHub 日榜最热的不是某个框架,而是一整套"给 AI agent 打工的技能包"。 排第一的 Univer 一天涨了 1828 个 star,它把表格、文档、协同编辑做成 SDK,让 AI 直接在产品里办公。 https://t.co/fMrvyhsbGN NVIDIA 的 SkillEvaluator 也在榜上,专门给 agent 技能做评测、查重、打分,技能一多,质检就成了刚需。 https://t.co/dTIf6UUyor 还有给炒股用的 claude-trading-skills,一天 837 个 star,把行情分析、图表、技术面打包成 Claude Code 能直接调用的技能。 https://t.co/zunoovwE0B 开发者现在拼的不是谁模型更强,而是谁手里的技能包更全、更好用。 https://t.co/ex3v4z60Hh

@Stanleysobest

卧槽,深圳有些跨境电商大卖家,已经开始自己造AI员工了! 最近看到一个案例,说是深圳一家年营收几十亿的跨境电商公司,干了件挺猛的事。 老板亲自带队,拉上内部100多个研发和各部门负责人,用DeepSeek Harness搭了一整套AI Agent系统。 有多夸张? 选品Agent,找有潜力的爆款。 运营Agent,分析店铺数据。 广告Agent,盯着投产比。 竞品Agent,研究对手卖什么。 客服Agent,处理售前售后。 库存Agent,预测缺货和积压。 财务Agent,对账、算利润。 连独立站运营,都想交给Agent干。 以前是100个人围着业务流程转,现在是100多个研发想办法让AI替人跑流程。 更牛逼的是,人家没请外部FDE。 老板自己下场,带着技术团队和业务部门一起干。 为什么看上DeepSeek Harness? 开源、能自己部署、能按业务需求魔改。 不过要做到业务数据不外泄,还得把模型部署和数据链路一起管好,光把Harness装在本地可不够。 我看完最大的感受是: 以后那些只会套个AI壳、卖几万块一套Agent的SaaS公司,日子恐怕越来越难过了。 有钱、有技术团队的大公司,慢慢会选择自己干。 但反过来想,全国有几个老板养得起100多个研发? 剩下那些没有技术团队的企业,谁来帮他们把AI真正用起来? 所以我越来越看好一类人: 既懂AI,又懂生意,还能亲自下场把流程跑通的人。 以后光会写Prompt,真不值几个钱。 能帮老板把1000万库存管明白,把广告费降下来,把订单和财务接起来,才是真的有本事。

@Super_Nius

😱由Fable5.2构造,猫咪的第一视角 我如果不说,谁能知道这些全是代码生成的?简直逆天了 太精妙了,背景的虚化,真实的雾气,和煦的阳光,无敌了卧槽! Claude走的路真的是对的,当模型足够智能的时候,一切东西都可以靠代码构造出来😭 https://t.co/TAQ6Sk06q3

@ruyi3699

查 AI 编程工具的额度,难道还得挨个开网页、登录、翻配额页?那也太折腾了。如今在 Mac 屏幕边上挂一条细边栏,Claude Code、Codex、Cursor 这些还剩多少,瞄一眼就门儿清。 负责这活儿的工具叫 Pulse,一次性能盯住 19 家服务: 1️⃣ 一圈圈彩环,绿→黄→红,快见底了直接飘红示警 2️⃣ 鼠标悬停上去,每个额度池和重置倒计时全给你摊开 3️⃣ 按你眼下的消耗速度,还能推算出大概几点烧完 我比较在意的是,它只在你自己的 Mac 上跑,用的也是你原本就登着的账号,数据一概不外传。 macOS 上免费开源,想白嫖的尽管冲。 🔗 https://t.co/pZ8reBRSVN

@ai_goge

很多人用 Claude 却没装它的 Skills、插件和 MCP,其实等于白用了一半。 完整清单有 16 个,这 8 个是我强烈建议闭眼入的: 1. find-skills:最先装它,能帮你自动找并安装其他组件。 2. humanizer:一键去“AI 味”,让生成的文字更自然。 3. 办公技能:搞定 Word、PPT 和 Excel,直接拿来应付客户工作。 4. Copywriting:喂了 100 多封传奇销售信训练出来的,写带货/营销文案极强。 5. context7:社区里最火的插件之一,查最新文档神器。 6. codex-plugin-cc:让 OpenAI Codex 和 Claude Code 联动起来。 7. Slack:直接在对话里发动态、看群消息、抓待办事项。 8. Meta 广告:用大白话查广告花费、建投放计划,随时叫停烧钱没效果的广告。

@gp_pulipaka

Grokking Streaming Systems: Real-Time Event Processing! #BigData #Analytics #DataScience #AI #MachineLearning #NLProc #LLM #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #GoLang #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #Books #100DaysofCode https://t.co/TBRf0Qh6lI

@xiaozhang0218

什么!!! Deepseek直接把自己app开源了!😳😳 @grok 是不是真的,你啥时候开源啊,grok。

@leo114119

上海人工智能实验室发布科学多模态模型Sage S2.1 Flash,多项专业基准超过了 Deepseek v4.1 Flash 1017 近日,上海人工智能实验室正式推出科学多模态模型Sage S2.1 Flash。 该模型专为需要多轮分析与验证的科研任务而设计,将科学数据理解与信息检索、编程及工具调用能力深度结合。在实际应用中,Sage S2.1 Flash不仅能够调用专业软件执行分析,还可以根据中间输出结果和视觉反馈对结果进行核验,并据此动态调整后续的执行步骤。 在官方同步公布的综合评测报告中,研究团队将Sage S2.1 Flash与GLM-5.3-Flash、DeepSeek-V4.1-Flash-1017、Qwen3.8-Max、GPT-5.6 Sol以及Claude Opus 5五款国内外主流前沿模型进行了全面对比。 评测涵盖了智能体编程、生产力、工具调用、机器学习工程、深度研究、搜索以及视觉七大领域的30余项细分基准测试。数据显示,在重点展示的15项核心基准测试中,Sage S2.1 Flash在4项测试中位列第一,并在7项测试中取得第二名的成绩,展现出卓越的科研辅助与工具执行能力。 https://t.co/3swcAnU2YI

@charles12388

DeepSeek 算力見底,直接對用戶搞起「配給制」了。 財聯社消息指出,因為算力不足(insufficient compute capacity),DeepSeek 已經開始限制用戶,除了對推理用量進行配給,連輸入上限跟對話歷史也一併收緊。 信不信由你,算力短缺對 AI 基礎設施需求來說,反而是個看漲訊號。 https://t.co/BZxHcFfLR9

@kocpc

騰訊混元推出 Hy4 flash,主打低成本日常 AI 工作:224B 總參數,每次推理只啟用 12B。官方測試號稱任務成本約為旗艦版二十分之一,但複雜編碼表現也明顯下滑。本文整理架構、定價、圖表與適用範圍,並保留官方測試及限期優惠的條件。 https://t.co/Ys6ehRSO1I

@star_flow_ai

WorkBuddy 又出了一个又快又便宜的新模型啦!😂 腾讯混元昨天正式发布 Hy4 Flash! 有开发者拿它和 DeepSeek V4.1 Flash 用相同的提示词测试 3D 体素生成,结果 Hy4 Flash 速度更快,成本也更低。 目前 WorkBuddy 已经免费开放 Hy4 Flash,一直持续到 11 月 30 日! 感兴趣的兄弟可以去白嫖体验一下!https://t.co/upcviMlSoW

@guilleflorvs

BREAKING: Databricks just published its AI Index, built on usage from 10,000+ enterprises. Open-weight models went from 7.3% to 30.9% of enterprise tokens in about 3.5 months. The other numbers worth knowing: 75% of companies now use 2+ model providers. At mid-year most used one. Almost half use 3+. P99 users spend ~$700 a week on AI. The median user spends $2.50. What this tells me: 1. The single-lab bet is over. Enterprises route by task, and whoever owns the router owns the relationship. 2. Open source is winning on cost, not ideology. Databricks sees it mostly on high-volume batch and coding workloads. 3. Spend is wildly concentrated. Power-user spend is flat under budgets, while the median user is growing fast from a tiny base. The real enterprise AI budget hasn't shown up yet. Building for enterprise? Go model-agnostic from day one and price for the median user who is just getting started.

@Oluwaphilemon1

Qwen3.8-27B Turbo Has Crossed 1.37 Million Downloads on Hugging Face Qwen3.8-27B Turbo has reportedly surpassed 1.37 million downloads on Hugging Face. That’s a notable level of adoption for a 27B model, particularly one designed around local coding and development workflows. But the download numbers aren’t the only interesting part. The model combines an 11.67GB checkpoint, a 256K-token context window and an Apache 2.0 licence. Its creator also markets it as an uncensored coding model, claiming it recorded zero refusals across 100 test prompts. There are four details worth examining, especially if you’re interested in running coding models locally or building autonomous development agents. 1. A 27B model with an 11.67GB checkpoint Qwen3.8-27B Turbo has 27 billion parameters, but its reported checkpoint size is just 11.67GB. That’s a relatively compact package for a model of this size. Smaller checkpoint files make downloading, storing and experimenting with models considerably easier, particularly for developers who want to maintain several local models or build their own inference environments. It also makes the model interesting for local coding assistants, programming tools and autonomous agents. However, there’s an important distinction between checkpoint size and the actual memory required during inference. The 11.67GB figure describes the model files. Running the model requires additional memory for several components: Model weights: The parameters loaded into memory for inference. KV cache: Stores information from previously processed tokens so the model can reuse it during generation. Activations: Holds intermediate values produced during computation. Runtime buffers: Provides temporary memory required by the inference engine. These requirements can increase considerably with longer context windows and multiple concurrent requests. Quantisation, KV-cache precision and inference optimisations also influence the final memory footprint. This means an 11.67GB checkpoint doesn’t automatically translate into comfortable inference on a 12GB GPU. The GPU still needs enough additional memory to handle the rest of the workload. For developers considering local deployment, actual VRAM consumption is a more useful measurement than checkpoint size alone. 2. Apache 2.0 gives developers more freedom Another important detail is the model’s licensing. Qwen3.8-27B Turbo is distributed under the Apache 2.0 licence. This is a permissive open-source software licence that allows commercial use, modification and redistribution, subject to its conditions. For developers, this creates several opportunities. They can download the weights, integrate the model into their own applications, modify the implementation and build specialised tools around it. Some possible applications include: Locally hosted coding assistants. Custom programming environments. Autonomous software development agents. Specialised debugging and code analysis tools. Internal AI infrastructure for development teams. Fine-tuned models for particular programming languages or workflows. The ability to run the model independently is particularly useful for teams that want greater control over their infrastructure, data and inference costs. Instead of depending entirely on a hosted API, developers can build applications around locally deployed model weights. The licence is therefore an important part of the model’s appeal, especially for commercial developers. 3. The creator claims zero refusals across 100 prompts This is probably the most attention-grabbing claim surrounding the release. According to its creator, Qwen3.8-27B Turbo recorded zero refusals across 100 test prompts. The model is being positioned as an uncensored coding model, with an emphasis on answering technical requests without unnecessary refusal behaviour. This could be useful in development environments where models sometimes reject legitimate technical questions.

@Kshsb783142

谷歌云公布了第七代Ironwood TPU的MLPerf测试结果,十项测试,覆盖前沿预训练和首个面向智能体强化学习的基准。 很多人第一眼会盯着640块TPU、18.02分钟这两个数。但真正有信息量的是另一句,训练芯片的等待时间占比低于3%。 翻译一下,在跑强化学习任务的时候,TPU绝大部分时间都在干活,不是在等数据、等网络、等调度。 智能体强化学习的瓶颈从来不在算力峰值,在喂不饱。模型一边生成代码、一边从环境里拿反馈、一边更新权重,数据流是碎片的、动态的、高频的。芯片等数据的时间越长,有效算力利用率越低。3%的等待时间,说明Ironwood在芯片间互连、数据调度、软件栈这些系统层面,把效率做到了一个很高的水平。MLPerf这次第一次把智能体强化学习纳入基准,本身就是行业在承认,这个场景的竞争比的是系统工程,不是单卡跑分。 预训练那边的数也在说话。8192块Ironwood跑基准,1.70分钟达标,扩展到4096块时效率96%。规模翻倍,性能几乎线性增长,这对大规模训练集群的实际利用率意义很大。 把之前聊过的两条线串起来看,这件事的位置就清楚了。谷歌在推TorchTPU,让开发者用一行代码从CUDA切到TPU,生态迁移的门槛在降。Anthropic已经采购了超过100万颗TPU,2029年可能成为全球最大的TPU单一用户,客户在进来。今天这份MLPerf结果,是这两条线的交叉验证,硬件性能也在追。 英伟达的GPU在训练和通用性上依然没有对手。但所有增量都必须买英伟达这个假设,正在被Ironwood这样的系统级成绩一点一点拆掉。3%的等待时间不是英伟达的终点,是谷歌从一个自用芯片变成可被认真考虑的替代选项的又一步。 $GOOG 免责声明:本文纯属个人观点,不构成任何投资建议。

@bkdgiffug

想在手机上远程操作电脑里的 Codex?可以看看这个开源工具 它相当于一个轻量级桥接服务,把手机浏览器和电脑端 ChatGPT 中的 Codex 连接起来。 主要功能包括: 手机向 Codex 发送文字、图片、视频和文件 实时查看 Codex 的回复过程和执行结果 同一 Wi-Fi 下优先使用加密局域网直连 离开局域网后自动切换到 Pro 公网线路 这样即使不坐在电脑前,也能通过手机给 Codex 下任务、看进度。 经常使用 Codex 编程,又希望随时随地远程跟进的朋友,可以了解一下。 🔗 https://t.co/kddmR2owmp

@isnail

兄弟们,Pro用了1个月,决定冲Claude Max 5X 了,没有弹KYC,继续苟着用吧。正好GPT到期,降plus吧 https://t.co/4o3Z3UmKFB

@leaf_sanren

😮救命啊,Claude连支付方式也封啊! 这苹果礼品卡的支付渠道一断,真的是。。 好害怕呀,感觉外币信用卡的路子,也担心哪天就没了。。 把DeepSeek harness用出Claude Code的效果,刻不容缓啊!!! https://t.co/bx9bwrMTWx https://t.co/5ZeGyf43Ds

@LaravelLinks

Larvyn: an open source Laravel desktop app with parallel AI coding agents https://t.co/IQRhHMRpAF

@betop1smile

想学AI?还不知道从哪里开始学?这套学习路径包含了所有入门基础,小白看这个就够了! https://t.co/zEH7OlSw6h

@AlpacaNotes

https://t.co/gNXuK1XpWi

@CarlWang0524

A 630M GTA world model ships Apache-2.0 weights and inference code, but needs a CUDA GPU process per player. I see a useful world-model experiment; concurrent users would multiply its serving footprint. https://t.co/1wJJ3p9eT9

@SlashdotMedia

FBI Says Artex AI Open-Source Hacking Tool Used to Breach Networks In 'Under Three Hours' https://t.co/KW02XevNlz

@therealnaathan

5 powerful AI models are now available for FREE. No credit card required. @runcrate is offering free access to: ✓ DeepSeek V4.1 Flash 1017 ✓ GLM 5.3 ✓ GLM 5.3 Flash ✓ GPT-OSS 120B ✓ Qwen 3.8 Flash Here's what makes it even better: → Free token allowance resets every Monday → Sign up using your GitHub account → No payment details needed → Works with OpenCode, Cline, Roo Code & Kilo Code Free AI models + weekly token refills = more coding without API costs Try it https://t.co/kzkYBQcmC1

@SISpotlight

🚨 𝗙𝗕𝗜 𝗪𝗮𝗿𝗻𝘀 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲 𝗔𝗜 𝗧𝗼𝗼𝗹 𝗔𝗿𝘁𝗲𝘅 𝗛𝗮𝗰𝗸𝘀 𝗡𝗲𝘁𝘄𝗼𝗿𝗸𝘀 𝗶𝗻 𝗨𝗻𝗱𝗲𝗿 𝗧𝗵𝗿𝗲𝗲 𝗛𝗼𝘂𝗿𝘀 The FBI issued a FLASH alert on Artex, an open-source AI penetration-testing tool that attackers now use to run intrusions on their own. Some intrusions went from first targeting to privileged access in under three hours. Attackers used it against South Korean banks, exposing data on about 67,000 people, and ConnectWise customers, with access to at least 2,250 @microsoft 365 tenants. The developer pulled it from @github, but copies and forks still circulate. #AI #Cybersecurity #Artex #FBI Sources: TechCrunch, Circuit Magazine, iAfrica ─── 𝗙𝗼𝗹𝗹𝗼𝘄 𝗦𝗜 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 𝗦𝗜/𝗔𝗜 𝗮𝗻𝗱 𝘁𝗲𝗰𝗵 𝗻𝗲𝘄𝘀: 𝗫 · 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 · 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

返回 AI 快讯 →