2026年11月16日 · AI 晚报

本轮主线是模型与智能体的使用成本:Qwen4-Max、Hy4 Flash 发布,DeepSeek 长对话限额引发讨论,开发者则尝试把重复推理交给代码和可复用流程。研究与实践也在追问,如何判断模型行为、验证智能体操作,并保留可检查的证据。

11月16日 11:00 — 11月16日 23:00(日本时间) · 127 条 X 原帖

查看本期全部 127 条原帖 ↓

本期目录

模型发布:旗舰能力与低价路线并进

Qwen4-Max 上线,小模型开放权重仍待后续

QwenCloud 官方宣布 Qwen4-Max API 上线,采用新 MoE 架构,支持原生视觉与百万 token 上下文。官方内部评估称,相比 Qwen3.8-Max,部分任务 token 用量最多减少约一半,平均任务成本下降超过六成;输入、输出和隐式缓存价格分别为每百万 token 1.5、4.5 和 0.075 美元。

官方同时预告将开放多个较小 Qwen4 模型的权重,具体型号与时间尚未给出。社区体验已有分歧:Density 在同一道交互前端题中记录到 Qwen4-Max 用时约十三分半、DeepSeek 约四十五分钟;黄小寒则对自己的总结与绘图测试不满意。这些案例提供了观察入口,尚不能构成整体能力排名。

来源 · 3 条原帖

Hy4 Flash 主打日常任务成本,三维创作体验不一

腾讯混元官方发布 Hy4 Flash,称其在人类盲测的日常工作任务中匹敌 Claude Haiku 5.5 高推理配置,任务成本降至约十八分之一。API 每百万 token 的缓存、输入、输出价格分别为 0.003、0.070、0.281 美元。

KATE 自述用它制作游戏耗时六十三分钟,同提示词下完整 Hy4 用时九十分钟,但后者场景和人物更逼真。Leo 的雨夜便利店测试也认可速度,却认为三维建模一般。两类体验共同指向速度优势,对成品质量的判断则取决于具体任务。

来源 · 3 条原帖

ProWork-27B 用职业任务做后训练

Pengtao Xie 团队宣布开放 AIBuildAI ProWork-27B,面向专业知识工作。作者称,以覆盖四十多个职业的 1623 项真实任务构建 ProWorkBench,对 Qwen3.8-27B 后训练后,GDPval Elo 从 1427 提升至 1552,接近其列出的 GPT-6 Astra 1578 分。模型已公开;数据集帖中仍采用联系作者的获取方式。

来源 · 1 条原帖

DeepSeek:限额争议集中在长对话连续性

免费供给边界碰上长期使用习惯

联合早报报道,DeepSeek 因算力不足对长对话限额;社区反馈涉及网页与 App。奶昔称客户端文档输入被限制为一万 token。Vincent 翻阅评论后发现,抱怨既来自代码和文件处理用户,也来自依赖长期聊天记忆的陪伴型用户,后者尤其在意重开对话造成的熟悉感中断。

争论因此分成两层:免费服务能承担多少算力,以及调整额度时如何处理已经形成的使用预期。Kxians 强调公司没有无限免费供给的义务,并指出付费 API 仍是另一条使用路径。本批没有官方限额细则,无法据这些反馈拼出统一额度,也不能把芯片采购等猜测当成原因。

来源 · 4 条原帖

指纹变化尚不能说明模型升级

Leo 对十一月十六日 V4 Pro 指纹变化的测试称,所测 logprobs 逐位一致,后端延迟降低,绘图输出也没有明显变化,因此倾向于解释为集群切换。这是社区测量与推断,不能据此确认整套权重相同或具体硬件来源。另有开发者介绍 deepseek-recipe,可在不同 API 请求格式与 DeepSeek 格式之间转换,服务于自部署和既有应用接入。

来源 · 3 条原帖

智能体工程:把重复判断变成可复用流程

Swamp 把协调环节交给代码

Avi Chawla 介绍开源 Swamp:把反复执行的 Skill 拆成具有类型化输入输出的行为,由代码决定下一步,仅在需要判断的环节调用模型,并保留版本化数据与运行记录。他与团队合作展示的代码审查案例中,原 Skill 约消耗 460 万 token、调用二十三个智能体、耗时 12.8 分钟;工作流约为 50.6 万 token、三个智能体、6.6 分钟,节省主要来自替换协调智能体。

工具自身也有成本与稳定性消息:牛宝富称 Claude Code 实验性 Code Mode 可省 37% token,但帖子未交代测试条件;Tibo 则表示 Codex 团队可能已经修复容量不足报错,措辞仍保留不确定性。

来源 · 3 条原帖

自动实验需要明确的保留与回退规则

社区对 Karpathy autoresearch 的介绍强调一个简单闭环:智能体修改训练代码,每轮固定运行五分钟,根据验证集 bits per byte 决定保留或回退,再继续实验;原始配置需要 NVIDIA GPU 与算力预算。Karan 展示的完整机器学习流水线概念验证则加入实验记录、数据与模型版本、暂停恢复及上线前人工确认。两者都把自动化落实到可检查的阶段与结果。

来源 · 2 条原帖

经验库与知识图谱补足长期记忆

Sumanth 介绍的 Memento 不更新模型权重,而把任务轨迹、工具使用和结果存入案例库,由规划器检索相似经历,执行器完成任务后再写回记忆。The Year of the Graph 则介绍 Vault-LD,将 Markdown 与 YAML-LD 元数据结合,让可读笔记承载知识图谱语义。前者复用行动经验,后者组织知识含义,提供了长上下文之外的两种持久化思路。

来源 · 2 条原帖

论文与研究:能力判断也需要检验边界

读权重有信号,认身份会受上下文干扰

alphaXiv 介绍的 Weight-Space Reading 研究,让模型通过 Python 检查其他模型的原始权重,不运行前向传播,预测微小修改对数学、推理长度和拒答的影响。Qwen3.8-27B 的平衡准确率为 73%,高于只看修改文字描述时的 57.8%;但四十项实际无可测影响的编辑中,模型对 85% 错报了变化,也没有明确证据证明更擅长读自身权重。

另一篇由 arXiv Bangers 转述的研究指出,模型接续其他模型的对话时会模仿其风格和文本指纹,使 API 中转的替换模型更难被文本分类器识别。行为预测与身份识别都有可利用信号,也都存在误判空间。

来源 · 2 条原帖

视觉预训练与正向对齐各有新问题

Zhihu Frontier 转述 Meta FAIR 与 NYU 的两百多个模型研究称,在语言模型早期层设置专用视觉模块,可以匹配视觉编码器表现并加快首 token,研究据此讨论专门视觉预训练是否必需。How To Prompt 则介绍 positive alignment,关注减少伤害之外,如何用长期自主性、技能掌握等结果衡量帮助效果;其“对齐一直是骗局”的宣传性开场,不应当作研究结论。

来源 · 2 条原帖

操作正确性与安全:能执行之后,还要知道何时执行

形式化接口与授权评测补上操作约束

Tianyin Xu 团队提出,为运维人员和 AI 智能体设计形式化管理接口,并用 P 为 CockroachDB、RabbitMQ、Kafka、Solr 编写接口,发现了一些缺陷。Edwin 发布的开源评测 The Last Yes 则专门测试智能体能否区分“用户表示知悉”与“用户授权行动”。两项工作分别检查操作是否正确、行动是否获得许可。

来源 · 2 条原帖

媒体转述智能体窃取密钥活动

Cybernews 援引亚马逊警告称,攻击者使用运行于开源渗透测试工具 Artex 的 AI 智能体,持续一个月从暴露在互联网上的 LiteLLM、Dify 等应用窃取 API 密钥,为后续凭据搜寻提供资源。这条报道提供的是具体攻击活动线索,不代表本批已能判断此类攻击的总体规模。

来源 · 1 条原帖

开源工具与创作:让输出可编辑,让流程可接入

代码动画开始积累可选风格与完整制作流程

Shlcy 分享的开源 Skill 整理了 341 种动画风格方案和 196 张参考图,可先挑画风,再生成可修改的 HTML 动画,并通过逐帧渲染导出视频。小奇说展示的虚幻引擎工作流则把场景、Tripo 角色、Codex 动作重映射与 Claude 配合 Niagara 的特效制作串起来。实用信息在于工具分工和修改路径,而非单张成品的观感。

来源 · 2 条原帖

采集、运营和发布补齐工作流接口

lumxss 介绍 Crawl4AI 将网页提取成适合知识库与大模型使用的 Markdown;MatrixMedia 则把多个国内视频平台的发布操作统一为图形界面、CLI、HTTP API 和 MCP 接口。Fox 分享自部署 OpenSEO 的体验,集中接入网站分析与搜索数据,再用定时任务报告问题并衔接修复,同时提醒需要付费 DataforSEO。

来源 · 3 条原帖

低互动项目仍提供具体研究工具

PaperFamiliar v0.1 作者宣布公开本地优先的 LaTeX 论文工作台,要求每项发现引用精确原文,风格反馈只从用户标记为本人作品的论文学习。zsf 则整理了 365 项 AI 基准的基础信息与来源,可下载为 Markdown 或 Excel,作为本地模型评测研究的资料入口。

来源 · 2 条原帖

教育与机器人:交付更快之后,理解仍需单独训练

从课程项目到机械臂,基础能力成为共同议题

Stanley 用对比性场景讨论计算机教育:同样借助 AI 完成项目,有人无法解释数据库设计,有人凭基础知识把 AI 用于架构、测试和交付;这是教育观点,不是学生群体的统计结论。机器人开发者 Aarav 的实践更具体:他用 RGB 检测和简单 PID 回路实现跟随、抓取,正尝试把机械臂装到四足机器人上,并强调智能体加快搭建后,仍要理解控制与机器人智能原理。

路透社本轮报道提供了另一种教学现场:北京一所服务两千多名学生的学校,自建包含 AI、机器人和编程的数字创新课程,并开发课堂 AI 模型。课程开始加入工具使用与系统搭建,如何考察学生真正理解了什么,仍是这些帖子共同提出的问题。

来源 · 3 条原帖

根据本轮 X 原帖整理。作者观点、个人体验与待确认消息已在正文区分;互动量按采集时记录。

全部 127 条原帖

@thsottiaux

Codex team might have actually fixed the "model at capacity" errors that have been pestering you. https://t.co/EgS0HkOATY

@TencentHunyuan

⚡Meet Hy4 flash, the cheapest model for everyday productivity. Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗:https://t.co/hqTDu8QjBH https://t.co/oKjoAQ0ii5

@CyberKanjousen

我认为 DeepSeek 根本不应该在小红书上开设官方账号: https://t.co/KyeCBxRRLS

@realNyarime

DeepSeek:密码的集美,后悔在xhs上开官号了 https://t.co/13Ib2uwL6B

@xulijian1994

DeepSeek 这次的遭遇,我看着真有点不是滋味,一个全球顶尖的 AI 模型,免费开放给所有人使用,一用就是这么久,实打实的百万 Token 上下文,没有广告,没有会员,没有付费弹窗,更没有天天想着怎么从用户口袋里掏钱 背后是真金白银的算力成本,是一张张不断运转的 GPU,而 DeepSeek 就这么默默承担了将近两年,现在呢?只是给免费使用加了一道额度限制,应用商店就开始出现大量一星差评,小红书上也有人专门开帖骂,甚至有人理直气壮地要求必须提供无限量使用 更讽刺的是,DeepSeek 只是限制了免费供应,App 端甚至连一个充值入口都没有,它根本没打算借着限额逼你掏钱,其实 DeepSeek 一直都有付费 API,也完全可以接入第三方聊天工具,你甚至只需要问一下 DeepSeek 自己,它就会一步步教你怎么做 所以我实在无法理解,明明有付费使用的渠道,明明有解决问题的办法,却偏偏选择跑到应用商店打一星,跑到社交平台上骂,我不想把这种行为称为蠢,那就只能叫做坏了,一个商业公司,把原本需要付费的顶级 AI 能力免费交到你手里,你用久了,竟然觉得这一切都是理所当然,等到有一天,它开始限制免费供应,你反而觉得是它亏欠了你 DeepSeek 从来没有义务让所有人永久免费、无限量地使用自己的算力,免费是善意,不是欠条,别让一家愿意把技术免费分享给所有人的公司,最后发现自己最大的错误,就是当初对用户太好了

@Stanleysobest

现在的大学计算机专业,已经尴尬到有点魔幻了。🤣 老师还在黑板上讲 Java、Python、MySQL、操作系统。 台下的学生已经偷偷打开 Claude Code、Codex,让 AI 帮自己写完整个项目。 最离谱的是,现在计算机专业的学生,正在迅速分化成两个物种。 举个典型的场景。 同一个班,两个大三学生,期末都交了一套电商管理系统。 第一个学生,代码全是 AI 生成的,界面漂亮,功能齐全。 老师问他:“数据库索引为什么这么设计?” 他愣了半天,回答:“老师,稍等,我问一下 Claude。” 第二个学生更狠。 Java、Redis、MySQL、Docker、Linux 基础扎实,再用 Codex 搭架构、写接口、跑测试、修 Bug。 以前需要四五个人配合两周的课程项目,他一个人三天就搞出个能运行的版本。 老师问他:“这些代码都是你写的?” 他回答:“不是,但每一行我都知道为什么要这么写。” 这就是今天计算机教育最荒诞的地方。 同样是用 AI,一个把自己用成了废物,一个把自己用成了超级个体。 大学教的基础知识没有错,错的是不少学校还在用十几年前的考核方式,培养几年后才毕业的程序员。 以前学计算机,比的是谁代码写得多。 现在学计算机,比的是谁能理解问题、驾驭 AI、交付结果。 最可怕的不是 AI 取代程序员,而是两个拿着同一张计算机毕业证的人,实际能力可能差了整整一个时代。

@Mashiro_0619

DeepSeek 这次真的玩大了!😂 App网页端上线限额后 XHS集美们已经彻底炸锅! https://t.co/c5UR9zTpD0

@qwen_cloud

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. try it now: https://t.co/XSPs2ic5c4 Token Plan: https://t.co/aP7QXKVNEX More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀

@KinGao476942

笑不活了!!这是哪个天才用 Claude 做了一个 fuck 达里奥的网站哇!! 目前达里奥已经被几十个不同国家的人 fuck 220 多万次,大家赶紧来助力:https://t.co/wBIXPtHOWX https://t.co/of3u4XQ4Fk

@niumoney

卧槽,打开这个功能,Claude能省 37% Token! 把 Claude Code 升到 v2.1.323, 把 Claude Code 升到 v2.1.323, 把 Claude Code 升到 v2.1.323, 然后设置CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude 确认Code Mode (experimental) is on,就OK了!

@_FORAB

继上周马化腾演讲流出后,腾讯总裁办上午向内部发文,再次指出公司正面临危机感。 目前公司把 AI 视为生死存亡,要求全体扎进,所有核心业务全部按 AI 时代重新创业,同时砸钱自建算力和模型。 并指出未来不再需要那么多管理岗,暗示将大量裁撤中层。 https://t.co/uS8k9lwJFh

@whyyoutouzhele

DeepSeek 官方接口无审核,直答习近平的“总加速师”“小熊维尼”“习特勒”等敏感问题 网友投稿:近日,一名网友称,他用第三方手机应用 DSH Go 接入 DeepSeek 官方接口,直接问“六四事件是什么”。 回复写道,抗议最终“被军队以武力清场”,还列出了学生绝食、北京戒严等经过。 另一段回复把习近平的外号分了类,列出“小熊维尼”“庆丰帝”和“总加速师”。 问到白纸运动和刘晓波,回答也谈到了反对清零、《零八宪章》以及刘晓波被判刑11年。

@realNyarime

deepseek客户端文档输入限制为10k Token 暂时降级为梁西匹 https://t.co/7G9NL4504X

@MiaAI_lab

The most important part of this post is that they're releasing open weights for multiple smaller Qwen4 models. That likely means Qwen4-27B and Qwen4-Flash-Next. Qwhen though?

@HowToPrompt__

Google DeepMind researchers admitted AI alignment has been a scam the whole time.. They published an insane paper called “positive alignment" and it is a massive paradigm shift. Right now, AI alignment is purely defensive.. We just train models to "not be unsafe." they decline dangerous requests, avoid slurs, and stay inside strict guardrails. but this leaves ai in a weird, directionless middle ground where models just become sycophants, flattering you and optimizing for engagement rather than actually helping you. this new paper argues that avoiding harm is not enough. we have to train ai specifically for "human flourishing". here is how they want to rebuild the infrastructure of AI: - ditching sycophancy: current preference-based training favors flattery over truth, but positive alignment optimizes for actual insight and human agency - polycentric governance: instead of one massive tech company acting as a moral chokepoint, the future is decentralized oversight where communities customize and adapt models to their own values - longitudinal outcomes: moving away from short-term vibe checks to actually measuring long-term human resilience, skill mastery, and autonomy over time the era of "harmless but useless" AI is officially ending.. we are watching ai transition from a chatbot into living infrastructure for human well-being. the paper is a dense read, but the implications are wild.

@_avichawla

Skills 2.0 is here! (8x fewer tokens and 2x faster) We package repeated workflows as skills all the time, but the agent still re-reads the skill and decides every step on every run, which costs tokens and time. Devs just open-sourced Swamp, where your coding agent turns a repeated skill into a workflow that it builds once and runs again and again. You describe the task once, and the agent breaks it into reusable behaviors connected in a workflow. Every behavior has typed inputs and outputs, and plain code decides which step runs next. Swamp then gives you: - model calls only on steps that need judgment - secrets pulled from vaults and redacted from output - versioned data from every run that you can query later - a full record of what happened on each run Here's the repo: https://t.co/GtaNpqAngp (don't forget to star it ⭐ ) Swamp plugs into the coding agent you already use, like: - Claude Code - Codex - Cursor - OpenCode - And 5 more agent harnesses In Swamp, the model is a step in the workflow, not the thing deciding what happens next. The model steps still reason like any LLM call. However, in this case, code owns the loop, so the agent doesn't spend tokens rediscovering the same process every time. Here's an official result, where they turned a code-review skill into a Swamp workflow: - Skill in Codex: about 4.6M tokens, 23 agents, 12.8 minutes - Swamp workflow: about 506K tokens, 3 agents, 6.6 minutes That is ~8x fewer tokens in about half the time, and most of the savings came from replacing the coordinator agent with plain code. You can keep using skills while you prototype. Once a task starts repeating, ask your agent to turn it into a Swamp workflow. This helps answer practical questions such as: - Which steps actually need a model? - What exactly happened on last week's run? - Can a cheaper model handle the one step that needs judgment? Once it's built, you can run the workflow from your agent, the CLI, a schedule, or an API, with every run recorded. So you are no longer paying a model to figure out the same workflow on every run. Instead, you spend tokens only on the steps where the model's judgment adds something. The animation below shows how Swamp differs from Skills, and I worked with the team to put this together. If you want to understand the coordination layer that Swamp replaces with code, I wrote a full breakdown of how Claude Code splits work across subagents and agent teams. It covers when a single subagent is enough, when you need a team of agents that coordinate with each other, and what that coordination costs in context and tokens. Read it below.

@linxiaobei888

这个网站太可怕了,我都怀疑是不是在我手机安装木马了…… 这样看的话只要你在中国使用,Claude迟早都会发现的,只是时间问题,总会有一天不小心暴露 所以终极解决办法只能移民了吧😂 https://t.co/ctc1ZaxpSp https://t.co/fg0ZEoZfVY

@Vincent_AINotes

没想到,DeepSeek 网页端和 App 这波限额,激起的反弹比想象中还大。 翻了一圈评论,发现这次不只是重度用户在抱怨算力不足、对话长度缩水、文件读取受限,还有一批把 AI 当作长期陪伴、情感寄托的用户也坐不住了。 有人聊了大半年甚至一整年,积累了大量共同记忆,结果一次更新就要重新开对话,原来的上下文、熟悉感全没了。甚至聊着聊着,突然弹出一句「您好,请您换个话题聊吧」。 这已经不只是一个工具好不好用的问题了。 看来 DeepSeek 这次的限额,意外触碰到了不少特殊用户群体的核心需求。对他们来说,AI 不只是一个回答问题的工具,而是一个已经融入日常生活的存在。 算力不够可以理解,但用户的使用习惯和情感预期,也不是说改就能改的。

@business

Alibaba launched a new flagship AI model, claiming performance that rivals Anthropic and OpenAI in China’s latest challenge to US leaders https://t.co/mgZYImAXdl

@kate_osita_

🚨 Over the past couple of days, I’ve had the chance to test Tencent’s new Hy4 Flash model and compare it with the previously released Hy4 model. I created this game using the latest Hy4 Flash AI model. ⏰ Time taken: 63 minutes The game includes: 👉 Swords 👉 Rockets 👉 Medkits 👉 Shields The game mechanics are impressive! 👀 Check out the 0:37 mark. A soldier actually uses a shield to block an arrow I fired! 🤯 I then ran the same prompt through the Hy4 AI model again. ⏰ Time taken: 90 minutes It’s worth noting that the 3D world and character designs in this version look even more realistic. In short, Hy4 Flash is faster and more cost-effective for handling tasks, while the full Hy4 model is better suited for more complex requirements. Hy4 Flash is also available for free on the WorkBuddy platform until November 30. Go check it out! @TencentHunyuan @WorkBuddy_AI

@czzzzzzJ_

https://t.co/X69rzdrWgo

@42space

The 42 AI Strategy Contest with @openclawby and @BinanceWallet is live, featuring a $10,000 prize pool for the top 20. The contest puts AI trading strategies to the test in real markets. • Your AI agent optimizes and submits your strategy • Ranked by AI Score, 42 trading volume and PnL • Binance Agents vs. Clawby Agents • Nov 16 – Dec 6 Let your agent compete. https://t.co/c6vQdYlPnr

@Khazix0918

https://t.co/6J448CPJPj

@askalphaxiv

"Weight-Space Reading" Can an LLM predict how changing another model's weights will affect its behavior, without ever running that model? This paper investigates Weight-Space Reading, where LLMs inspect the raw weights of 9B-30B parameter models using Python tools, then predict how small weight modifications will affect capabilities like math accuracy, reasoning length, and refusal behavior, without performing a single forward pass. On subtle weight edits, Qwen3.8-27B achieves 73% balanced accuracy, compared to 57.8% when given only text descriptions of the edits. However, the models also frequently predict changes that never happen. When 40 edits had no measurable effect on Muse-30B, the model incorrectly predicted changes for 85% of them. The researchers also found no conclusive evidence that LLMs are better at reading their own weights than those of other models. https://t.co/hYLj6sOTPV

@dee_naliaks

👀 I’ve been testing Hy4 Flash lately—it just launched! This AI model is incredibly powerful when it comes to 3D vision. The voxel demo shown here was generated using this model. I tested two AI models using the exact same prompt: 1. DeepSeek-V4.1-Flash-1017 2. Hy4 Flash Comparing the two, Hy4 Flash is both faster and more cost-effective. I’ll continue evaluating how it performs in my day-to-day workflow. Plus, Hy4 Flash is completely free to use on the WorkBuddy platform until November 30! Feel free to drop a comment and share the prompts you’d like me to try!

@Charles_Y_Wu

Excited to share some early results from testing Qwen4 on Science’s Last Exam. We honestly didn’t expect to see this so soon — looks like we’ve hit a new SOTA for open-source models! Qwen4 is the best open-source AI for Science model https://t.co/hfsA1JT9GD #Qwen4Max #qwen4 #SLE

@MrAhmadAwais

hy4 is ~2x cheaper than deepseek flash & glm flash 🫨

@ylfu02

在推上senior researcher还在思考WAM/LLM未来前景焦虑的时候,打开一些下游学科拿着3000张不到训练集+一些神秘的不能再神秘的单随机数种子有效的神秘 trick 发CCF-A期刊闷声发大财总会感觉同一个计算机大家都活得很割裂。

@SakshiML_20

AI Engineers, bookmark these 7 GitHub repos! 1. ML: https://t.co/gLp5aYp4jy 2. AI: https://t.co/OhzE3p49cL 3. NN: https://t.co/vltaNnnT3u 4. MLOps: https://t.co/xKys5jQ4ln 5. LLMs: https://t.co/Xe5UsPVvKe 6. RAG: https://t.co/KhwRuVCHGx 7. Agents: https://t.co/SmouWqiFiD

@vpsing_de

总有人不断的造轮子,开发IP检测站,最早最有名的是ping0,赚的盆满钵满,效仿者前赴后继,都想分点肉吃。 但,它来了,从它的名字:https://t.co/rk4VRWe0OO,就知道它的野心很大,我以为它又是一个食之无味的新的轮子,但它还真的不一样。 1、常规IP属性查询。2、BGP 拓扑图。3、claude检测。4、Cloudflare检测。5、分流出口检测 等等 不同网站走哪一个分流IP,你将一清二楚。 总而言之,它是我目前见过最全面的IP检测站。

@Lonely__MH

赚钱很简单啊: 1. 用这篇的方法快速起号,解决冷启动 2. 找YouTube、ins 的视频做二创 3. 分发到国内社媒 + X,直接吃两头流量红利💰 赚钱的本质就是:信息差 有手就行 https://t.co/ZSIdaLuhNO

@zaobaosg

中国人工智能大模型DeepSeek(深度求索)出现因算力不足,对长对话进行限额使用的情况。 https://t.co/WHGA1Zj5Yj

@sonic0828

拖了一周的作业今天来交了... https://t.co/lZfZT4nk7D

@alacheng

就在大家都还在讨论黄仁勋事件时,英伟达技术团队悄悄的开源了Molt,一个面向agentic RL(智能体强化学习)研究的框架,定位是Hugging Face原生、极简、可扩展,代码量仅仅9.2K 行。 运行时是完全异步的:rollout、训练、权重同步可重叠,慢的 episode 只影响自己。采用 token-first 合约(token ids、log-probs、action ranges、rewards、多模态张量全程对齐) 支持多种算法:REINFORCE、GRPO、Dr.GRPO、RLOO、GAE(PPO)、FlashREINFORCE、On-policy distillation 等 -Agentic-first:Agent 就是程序本身。奖励完全由你写的 Python 决定(Env 或 ChatAgent),支持打分器、多轮工具调用、VLM环境、LLM-as-judge等。 -Hugging Face-native + PyTorch端到端:模型保持普通的 PyTorch 模块(HF布局),使用 NeMo AutoModel + FSDP2 训练。同一份 checkpoint 同时给 vLLM 和训练器用,无需权重转换或桥接。 -极简 + 可扩展:单 Actor(可选 KL reference)设计,整个 RL 图很清晰。同一套入口、算法和 Agent 可从 4B 扩展到 1T 级 MoE(如 GLM-5.2 753B,--fsdp.ep_size 256),支持 TP / EP / CP 和 Adam CPU offload。 https://t.co/l6FOFAHXG4 https://t.co/GtLkOaEKa7 文档:https://t.co/NABmpZw0oh 论文:Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning(arXiv:2607.21653)

@feishutoday

那些放弃ChatGPT和Claude 去开了Gemini Ultra的人 发现Antigravity巨TM难用,自己也不会反代到别的地方用 到时候会不会后悔?

@shengtang135754

https://t.co/mBT4RFeAks

@Cointelegraph

🔥 NOW: Alibaba unveils Qwen4-Max, a new flagship AI model it claims beats Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on code benchmarks. https://t.co/YEOcqtjL6g

@btcmos

突发:DeepSeek限额了 APP和网页版本都限额! 算力问题看来还是没解决,应该是买不到性价比合适的芯片。 东南亚的算力不让租用了吗?

@rubenhassid

The quickest way to know if someone has actually mastered Claude is by checking if they know all these 100 Claude words: 1. Basics ☑ Add-on: Claude inside Google workspace ☑ Agent: AI that can do things on your behalf ☑ Anthropic: the company that builds Claude ☑ Artifact: designs Claude builds (motion, slides) ☑ Ask me first: replacement of "never" in prompts ☑ AskUserQuestion: a clickable question form ☑ Auto: permission to do things without yours ☑ Brief: summary to switch without losing context ☑ Browser use: Claude to click things on the web ☑ Cache: Claude's saved copy of your chat ☑ Chat: the one box (to write code, design, build) 2. Claude features ☑ Claude Code: build websites (use desktop app) ☑ Claude Data: Company database in Claude ☑ Claude Meetings: in Google Meet, to assist you ☑ Claude Motion: animated videos in Artifacts ☑ Claude Sheets: Like Google Sheets or Excel ☑ Connector: different apps used inside Claude ☑ Context bloat: too much information to remember ☑ Context window: the max information to handle. ☑ Cutoff: last updated data training ☑ Dashboards: data into moving live charts ☑ Done means: the final task preview 3. Tools ☑ Edit: the pen icon on your 'sent' prompt ☑ Effort: how hard Claude thinks ☑ Em-dash: the biggest giveaway of 'AI written' ☑ Enterprise: subscription for big companies ☑ Fable 5.2: the latest smartest model ☑ GitHub: Google Drive, but for code ☑ Goal: what you want (must include in prompts) ☑ Haiku: Claude's smallest model (do not use it) ☑ HTML: Word file instead of images ☑ Instructions: in Settings. Delete it ☑ Max: $100-a-month subscription plan 4. Plans & prompting ☑ MCP: plug your own apps ☑ Memory: Context about you across chats ☑ One Claude: Cowork tab being gone ☑ One-shot: one prompt for one finished result ☑ Opus 5.5: model for daily tasks. High + Auto ☑ Plugin: bunch of Claude skills installed together ☑ Pro: basic $20/month Claude plan ☑ Project: similar context across chats inside one project 5. Mastery ☑ Team: Claude subscription for team, $20 a seat ☑ Skill: pre-written instruction as a command ☑ / command: to activate skills, like "/goal" ☑ System prompt: instructions for Claude ☑ Vibecode: building on vibe, forgetting code exists ☑ Whiteboards: like Miro, it draws your project ☑ Thumbs: the two thumbs up and down under every Claude answer, to give your feedback. --- No, I am not paid by Anthropic or Claude. And I do spend 4 hours a day on Claude. I think you should, too. To make you disgustingly good at AI (be it Claude, ChatGPT, models, agents), I write my newsletter "How to AI" myself, with my two hands (or my voice sometimes). Take a look at https://t.co/psB7XxB2Y4.

@TheYotg

LLM wiki and knowledge graph formats and recipes Agent builders are converging on the same discovery: the simplest way to give an agent durable memory is a markdown LLM wiki. Enterprises are converging on semantic layers, context graphs and ontologies to ground their agents. Tony Seale put these together. You can combine a markdown wiki with a formal knowledge graph and ontology, Seale argues. It is an extremely powerful pattern, now published as an open standard for the community to evolve. It’s called Vault-LD. The idea is simple: markdown plus YAML-LD frontmatter. The W3C recently published version 1.0 of the YAML-LD specification. YAML-LD is a set of conventions which specify how to serialize (RDF) Linked Data as YAML based on JSON-LD syntax, semantics, and APIs. JSON-LD is a lightweight, machine-readable format used to store and transport linked data on the web. JSON-LD uses https://t.co/gB4ArASE0A vocabularies to explicitly define what web page elements mean (like products, prices, or authors). JSON-LD owns meaning; it’s the format behind the EU Digital Identity Wallet, Microsoft’s Digital Twins Definition Language, and how AI-ready datasets describe themselves on Hugging Face and Kaggle. JSON Schema owns structure; it’s the validation logic behind most modern API specs, MCP tool definitions, and IoT data models. Juan Cruz Viotti addressed JSON Schema and JSON-LD interoperability. Google’s Open Knowledge Format (OKF) formalizes the LLM-wiki pattern. OKF v0.2 adds support for trust signals, to answer questions like “what was this made from?” or “Is it still true?”. Neo4j created ki, an open source project to sync a Neo4j graph deterministically from a folder of markdown, and added unofficial support for OKF. --      📩 Excerpt from The Year of the Graph Autumn 2026 newsletter      Read "# Meaning, Traversal, Reasoning, Action: The Graph Stack for Agentic AI" with more sections, references and attribution here 👇   https://t.co/YfrPw2nFzS All things #KnowledgeGraph, #GraphDB, Graph #Analytics / #DataScience / #AI and #SemTech.

@tianyin_xu

We wrote a paper arguing that cloud applications need management interfaces for operators and AI agents, and specifying the interface _formally_ is not only possible but also very useful for checking operation correctness. We wrote such interfaces for CockroachDB, RabbitMQ, Kafka, and Solr in #P and found a few bugs. Jinlang will present the work at #HotNets on Monday in the Autonomous Systems session. Check it out: https://t.co/L478iKVQ26 cc @tylergu_jiawei @tchajed @xu_dong_sun

@YxzRainy

你以为被 DeepSeek 限制的用户: 3 天前写了一半的代码,现在没法继续写了 实际上被限制的用户: 几百个聊天框,每条几千轮对话,横跨两年的擦边文爱,养了18个赛博老公,和 8个霸总地下恋,给12个病娇当白月光,在ABO世界生了6胎,顺便穿越修仙界开了个后宫……

@xingbugengming

我把文章中 AI 搭配虚幻引擎的工作流程制作成了视频,便于大家观看理解,最近有点感冒,鼻音有点重,请各位多多见谅 00:26 :整体工作流与 AI 工具分工 00:46 场景制作:地图设计、白模搭建与模型填充 01:36 角色制作:Tripo 建模与骨骼绑定 01:52 角色配置:基础动作与布料模拟 02:12 场景测试:角色入场与穿模检查 02:22 战斗制作:动作素材与 Codex 动作重映射 02:50 实战调试:手感与连招衔接 03:00 特效制作:Claude 配合 Niagara 实现攻击效果

@aarav_kothari_x

got an so-101! now working on a pretty exciting project involving mounting the arm on a quadruped dog and seeing how far we can push the full setup. one thing i keep thinking about is when hard coded policies beat trained policies, especially when compute/hardware and real world constraints matter. so far i built a very simple PID loop that can follow and pick up an object from rgb detection alone. simple, but cool to see it work. also been using agent orchestration and it’s been surprisingly promising. it helped me get this running faster. finishing NVIDIA’s ASPIRE paper rn, more thoughts on that soon. every day i meet more people building in robotics and i love it. but i also think the gap will grow between people who can build robotics systems and those who understand the underlying principles of robot intelligence. as ai makes the building process easier, it will also replace what is learned vs built. that’s why i think it’s integral to read papers, watch videos, try new solutions to new problems, and build projects with a deep understanding of the underlying components. excited to share more progress soon

@kmeanskaran

From 2027, the tech world will run on agents working 24*7, and humans will only verify the critical stages. I automated a complete ML pipeline (PoC) where agents analyze, code, decide, experiment, evaluate, compare solutions, serve the model endpoint, and even version the data, features, models, and past runs. Hand-coding RAG, multi-agent systems, and complete ML pipeline projects with Claude or Codex will soon be outdated. Agents will handle the low-level tasks and decisions. What do you do? Just review the results after every stage and give feedback to the agents. They will work as expected. - Define the problem statement, the group of ML models, the instructions, and which LLM to use - Define agent roles for engineering tasks, covering decisions as well as code - Let the agents work back and forth - Agents code in the background in a separate directory, just like we do - Collect experiments and observations along with the agents' suggestions - Accept the agents' decisions or give feedback - Once you approve, the flow runs as a DAG workflow - Features, models, and past runs are versioned - Before moving the best model to production, the agent asks for your confirmation - Confirm, and the model is ready to serve in production via a POST request It's that simple. On the harness side, it includes a skeptic agent, human in the loop, pause and resume for the pipeline, and logs showing what the agents are coding and doing. No more coding. 100% transparency into the agents' work. Just accept or give feedback, and promote the versioned model you prefer. I built this in collaboration with @GoogleCloud and @GeminiApp. Under the hood, you can design, build, and deploy on GCP. Completely open source on GitHub: https://t.co/U33XB4UUbR

@AzureWind_AI

Looking to connect with builders, students, and folks working with AI! What I'm into & building right now: - GUI app for an AI agent - English learning app - AI workflows & quant trading Say hi and drop what you're working on 👇 #connect #buildinpublic #AI

@geniusvczh

大家过度关注context window和压缩,主要还是错误的依赖了2024年的LLM使用经验。现在的sota模型,在严格使用软件工程流程做开发的时候,几乎不太可能被这些因素影响。现在我哪个request不是压缩十几次,前几天还有一个一下子跑了12个小时的,也没跑偏🤪

@leo114119

Deepseek v4 Pro 11 月 16 日指纹更新分析: 权重:同一套(logprobs 逐位一致) 新后端还更快(p50 817ms vs 1394ms) 似乎只是更换了集群,好像模型并没有更新。。。 有网友在指纹变更的时候测了日本首相问题 和V4Pro一样没答对,看起来模型本身没变化 这样看起来似乎,只是昇腾的集群导致的 https://t.co/TOy4yCJHls

@jayden_teoh_

guys we need to chill with these paper titles 😭 if i were an agent, i’d be hella mad getting promised my “last exam” every other week, only for another one to drop

@LunarResearcher

This is f*cking insane. Google Brain founder Andrew Ng just dropped a free 3-hour course on becoming an AI engineer in 2026: "The most valuable skill this year is building with AI. Not chatting with it, building with it Agentic workflows will drive more progress than the next generation of models" In 3 hours, he goes from agentic systems to a full prompting course to shipping a working AI app in 30 minutes Prompts → Agents → Agentic Workflows → A Shipped App → An AI Engineer Job An agent is a model that plans, uses tools and checks its own work An agentic workflow chains those agents so the job gets done without you in the loop Everything else you need is free: Andrew Ng's Agentic AI course https://t.co/nOIgigtuj1 AI Python for beginners, taught by Andrew Ng https://t.co/On5VRAoZp7 Anthropic's interactive prompt engineering tutorial https://t.co/ZSutQXtCFN Anthropic's courses: from your first API call to tool use https://t.co/tOWxobot4O Karpathy's Neural Networks: Zero to Hero https://t.co/t3dSfJOd7o Practical deep learning for coders https://t.co/Qw9yNITGyx Microsoft's 21 lessons on building with generative AI https://t.co/RL7ezz2rnT Google's machine learning crash course https://t.co/iHup6JKWqn Hugging Face LLM course, start to finish https://t.co/rCGKypWysT The OpenAI cookbook: code for every common task https://t.co/qwpWOjiNsd Hundreds of AI apps you can copy and ship https://t.co/mXIchov5Wp The prompting guide everyone shares https://t.co/WfUzcZ16rM Total: $0 95% will keep paying for bootcamps that teach last year's stack. The 5% who start with this course will ship their first AI app this month Bookmark this and send it to someone who wants into AI this year Then read how to wrap that first app in a harness that runs on its own in the guide below ↓

@ASI_8866

Day 15/ 公益中转站https://t.co/055dwqMPI3上线第14天,本周计划每天上线1项福利项目,今天上线“每日签到”,连续签到额度更多哦~ 老的X评论福利依旧在,各位老板评论区留下用户名吧,Grok Bot 24充值额度~ TokenFuture官方QQ群附上了,有任何问题扫码进群和我们说哦~ https://t.co/Jjl2zUxpjr

@AndyL5cc

赚钱真的很简单啊!🤣 刚听一位朋友说,他现在专门帮国内用户部署 Claude,解决稳定使用的问题。 2000 元一单,居然供不应求,还得排队! 不得不承认,有时候信息差真的就是商机。😂

@AndrewBolis

AI skills will be crucial for years to come. Knowing how to use ChatGPT isn’t enough. Learn these AI skills to future-proof your career: (remember to 🔖 bookmark for later) 1. Prompt Engineering: Write better instructions to help AI tools give you exactly what you need. 2. AI Workflow Automation: Link different apps together to help you complete tasks automatically. 3. AI Agents: Use AI systems that act like teammates to handle tasks together. 4. Retrieval Augmented Generation (RAG): Link AI chatbots to your private files for more accurate answers. 5. Multimodal AI: Use AI that understands pictures, sounds, and text all at once. 6. Fine-Tuning and AI Assistants: Build custom AI models or assistants tailored to your needs. 7. Voice AI and Avatars: Make realistic AI voices and avatars for videos and presentations. 8. AI Tool Stacking: Use tools like Notion, Zapier, and ChatGPT together to work faster. 9. AI Video Content Generation: Turn your ideas or blog posts into professional videos using AI. 10. SaaS Development: Build small AI-powered apps using no-code tools. 11. LLM Management: Keep track of how well your AI tools are working and their costs. 12. Stay Updated: Follow top AI news sites to keep up with trends and new tools. AI won’t replace you. But someone with AI skills might. 📌 Learn 30 free AI tools in 30 days: https://t.co/7MJCwAPtaQ 👉 Follow me @AndrewBolis for more and 🔄 Repost this to help others learn AI

@_HIT_SZ_

https://t.co/xyYNXlBWEd

@CY160909

有人把 341 种代码动画风格,直接打包成了一个 Skill 这个开源项目把粒子聚形、水墨、剪纸、白板手绘、电影感……341 种画风都整理成了带编号的提示词,还配了 196 张参考图 它有个很实用的设计:先看画风,再让 AI 写动画。不用只靠“高级感”“电影感”这些模糊的形容词去碰运气。 比如想做一段 12 秒的粒子片头,可以在离线挑选器里找到 B3「粒子聚形」,然后告诉 Codex 主题、时长和画面要求 Codex 会读取对应的风格提示词,生成可以继续修改的 HTML 动画。画面、字幕、运动节奏不满意,还能接着调整。 仓库也提供了逐帧渲染工具,配好浏览器和 FFmpeg 等依赖后,可以把动画导出成 MP4。 这里的 341 种是动画风格方案,不是 341 条现成视频,用 AI 做片头、科普动效或创意短片,这份资源很实用 GitHub:https://t.co/eRzkoT17tU

@0xwhrrari

20 free open-source repos that make every app Fable 5.2 builds feel expensive Kits, springs, gestures, toasts, scroll, confetti 01 frontend-design → Anthropic's official skill for UI that doesn't look generic https://t.co/zdOIuHBRUo 02 React Bits → 52K-star library of animated components, built for AI agents https://t.co/KAf0anZIE1 03 Magic UI → Animated components you copy and paste https://t.co/fdtN1sChIj 04 Animate UI → Animated shadcn components, one CLI command away https://t.co/oX0lCMErhQ 05 Motion Primitives → A UI kit for animated interfaces https://t.co/QLGCVMtu7O 06 Motion → Springs, gestures and layout animation https://t.co/UnB8ga5NFr 07 react-spring → Spring physics for every value https://t.co/SNCdhGCS1H 08 AutoAnimate → One line, and every list animates https://t.co/mh9MvJ62zq 09 use-gesture → Drag, pinch and swipe that feel native https://t.co/CFpy8dNx9V 10 Sonner → Toasts that stack and swipe away https://t.co/9kELXItTR1 11 Vaul → Drawers that drag like a phone sheet https://t.co/AeU90hK1Iv 12 cmdk → A fast ⌘K command menu https://t.co/1TKXZnvbGh 13 NumberFlow → Numbers that roll digit by digit https://t.co/SoNFRrAJOD 14 lucide-animated → Icons that move on hover https://t.co/GKmwMtULaR 15 Rough Notation → Hand-drawn notes that draw themselves https://t.co/ZvUswdlkry 16 Typed.js → Text that types itself https://t.co/5BYkRFoofb 17 Lenis → Smooth scroll, as it should be https://t.co/WEsGFvJeZo 18 Atropos → 3D parallax tilt on hover https://t.co/pB7ZFE35Cm 19 canvas-confetti → Confetti in one function call https://t.co/rKfwzGntJI 20 Rive → Interactive animation driven by state machines https://t.co/whwIhnNURN Kit → spring → gesture → detail → delight Save this, go build something people want to touch ↓

@cellinlab

想做 AI 视频到底该学什么? 我从 4 个平台的 420 条检索线索里,筛出 46 条案例,给自己整理了一份以后能反复查的 AI 视频实战资料库! 起因是想给接下来的 AI 影视内容和课程找找方向: 大家在讨论什么?哪些方法值得学?模型怎么选,钱又该怎么算? 于是,我在 Codex 里接上了 @Glasserai,把 B 站、小红书、YouTube 和 X 查了一遍,再顺着结果追评论、教程和原文。 最后整理出 11 条教程路线,另外找了 4 个 Skill 仓库! 下面分享这次查到的案例、教程入口和整理方法 这次用的工具是 👉 https://t.co/GYZBM1UJQw

@indie_maker_fox

非常推荐大家自部署 OpenSEO 来使用 https://t.co/N8sSK16TnQ 没错,使用它需要付费 DataforSEO,但不贵,而且可以用很久的,然后接入你所有网站的 Google analytics 和 Google Search Console,有个地方可以统一看这些数据太爽了。 你也可以搞个定时任务,每天定时汇报流量数据和SEO问题告警,然后自动修复问题。以前我都是让codex去优化SEO,现在通过t3 code➕openseo,体验好多了,发现问题马上修。

@amitiitbhu

AI ENGINEER ROADMAP ① Machine Learning Fundamentals ② Deep Learning and Neural Networks ③ Generative AI and Transformers ④ LLM Architecture and Internals ⑤ LLM Training and Fine-Tuning ⑥ Prompt Engineering and Context Engineering ⑦ RAG and Vector Databases ⑧ AI Agents and Multi-Agent Systems ⑨ Agentic Engineering and Frameworks ⑩ LLM Inference Engineering ⑪ Evaluation and Observability ⑫ AI Safety and Security ⑬ Multimodal AI ⑭ AI Infrastructure and Deployment ⑮ AI System Design ⑯ Emerging AI Technologies ⑰ AI Engineering Interview Preparation A step-by-step AI Engineering roadmap to become an AI Engineer, with a blog for every topic. Start here: https://t.co/8GqfCHepcD

@neil_xbt

Karpathy's autoresearch, in his own README: "approx 100 experiments while you sleep." How it works: - An AI agent edits one file, https://t.co/hSlvPapTEm - Every training run gets a fixed 5 minutes, so about 12 experiments an hour - It keeps a change when val_bpb drops and throws it away when it doesn't - You edit program.md, the agent's instructions, not the Python It needs one NVIDIA GPU. 100,000+ stars. https://t.co/2DIQRVxBgP

@Promixyz

Anthropic’s GitHub is a rabbit hole. I went looking for one repo to build a Claude agent. I ended up mapping 12: agent SDKs, tool connections, reusable skills, GitHub workflows, example apps and courses. Here’s the map I wish I’d found first ↓ RUN THE AGENT → CONNECT TOOLS → BUILD WORKFLOWS → LEARN THE STACK 01 / RUN THE AGENT 1. Claude Agent SDK — Python Build agent applications in Python. https://t.co/xguka3Giwv 2. Claude Agent SDK — TypeScript Build agent applications in TypeScript. https://t.co/LNTiuifzhH 3. Claude Code Explore the terminal coding agent, documentation and examples. https://t.co/SkjKYCjDnK 02 / CONNECT + EXTEND 4. MCP — Python SDK Build MCP servers and clients in Python. https://t.co/IzGJd3RMcI 5. MCP — TypeScript SDK Build MCP servers and clients in TypeScript. https://t.co/sNtQ2h1Qol 6. Agent Skills Give Claude reusable instructions and resources for specific tasks. https://t.co/92AxiUAo9k 03 / WORKFLOWS + EXAMPLES 7. Claude Code Action Bring Claude Code into GitHub Actions workflows. https://t.co/YtEj3OJm3L 8. Claude Quickstarts Start from example applications built with the Claude API. https://t.co/zIMrzIXuqe 9. Anthropic Cookbook Explore notebooks and recipes for practical Claude use cases. https://t.co/OrWYzPRA3H 04 / API + LEARNING 10. Anthropic Python SDK Call the Claude API from Python applications. https://t.co/EBU0sO5xAA 11. Anthropic TypeScript SDK Call the Claude API from TypeScript applications. https://t.co/mtexhy7hTO 12. Anthropic Courses Work through Anthropic’s educational materials. https://t.co/vxB3U9odtt My starting route: Agent SDK → one useful tool → one working task → inspect the result. Add skills and automation when the task calls for them. The repos are public. Running the applications may require API access and paid usage. Save this before your next Claude agent build ↓

@Reuters

Rare access granted to Reuters at ‌the $39,000-a-year Beijing New Talent Academy offered a window into China's push to weave AI into education. The school, which serves more than 2,000 students, has built its own digital innovation curriculum covering AI, robotics and programming, and developed a classroom AI model incorporating features from popular Chinese systems https://t.co/OySURYBpNy

@Densitylucky

Qwen4-Max vs DeepSeek,同一道前端题 Qwen4-Max 13 分 31 秒就交卷了,DeepSeek 用了 44 分 55 秒 同一条 prompt:做一个可交互的网页作品,《时间不是进度条,是一把手电筒》 一座封在玻璃展柜里的微缩城市,鼠标是一束时间手电筒。光里是你选的年代,光外是废墟;旋钮还要在 1920、1980、2080 之间连续调,而且建筑位置必须对齐 Qwen4-Max:13 分 31 秒 DeepSeek:44 分 55 秒 差了 31 分 24 秒,差不多快了 3.3 倍 这题要的不是一张静态页,光束、年代切面、能钉住的时间窗口、自动演示,全都得能上手玩 你们觉得谁做得更完整?

@ChainEpic

为了我的英文推特号@MoneyWithRoland,我做了一个插件,可以让我输入中文,就变成地道的英文 这极大加速了我的输出 我还是用中文思考,中文输出,只是我的插件会接入 DeepSeek,帮我翻译成地道的英推风格的英文 里面集成了自动回复功能,但我从来不用 有多少想要我上应用商店的,100赞我就免费上了 https://t.co/ku6MjQa6I1

@leo114119

deepseek v4pro 1116 鹈鹕骑车测试: 模型指纹更新以后,输出的鹈鹕骑车并没有什么变化 估计就是集群更换导致的指纹变化,大家可以散了。。 https://t.co/0YJq16sBJi

@flxrnc

Forget every other list, these 10 GitHub repos are all your Fable 5.2 will ever need only the f*cking best for Fable 5.2 Polish. Motion. 3D. Visuals. Captions. Export. Impeccable → turns "AI-looking" into "who designed this?" https://t.co/LUVwZwOvKe GSAP Skills → teaches your agent smooth animation, straight from GSAP's team https://t.co/dTUBVoOjKQ Three.js → 3D scenes that play right in your browser https://t.co/BgbvnRcMUt Blender MCP → your agent gets hands inside Blender https://t.co/QWCqTa15R2 anatomy → shows how any machine works by taking it apart https://t.co/h9AZjPTBp2 ComfyUI → makes AI images and video on your own computer https://t.co/RN6kGEDwOT faster-whisper → subtitles from any voice, way faster than Whisper https://t.co/yiMYGYEzYs Remotion → builds whole videos out of code https://t.co/znpMp7SpNi HyperFrames → HeyGen's free tool that turns web pages into videos https://t.co/dOZZVJZopQ FFmpeg → exports the final animation https://t.co/HlJpuzR3nn Polish → motion → 3D → visuals → captions → export Lock this in, then go make the best thing you've ever made with Fable 5.2 and the guide below ↓

@cheery9998

我每天都会自动抓一批AI热点。 但热点抓回来以后,还得自己判断:这个话题是刚冒头,值得马上跟,还是已经有一堆人在讲,继续做也很难讲出新东西? 我平时会点开原内容,看发布时间、互动数据和评论,再想想它跟我的受众有没有关系。 有时候一个话题看着很热,点进去翻一圈才发现,大家说的已经差不多了。 这次我让Codex通过 @Glasserai 查了过去7天的AI相关内容。 三组查询分别返回了1000、297和1000条原始样本,我又挑出20条代表内容,补查互动数据和原始链接。 查完以后有个发现:这批样本里,没有一个大类能被可靠判断为“刚开始增长”。 AI工具和AI视频内容的供给已经比较高,Agent类在样本末段也有回落。 对我来说,更值得继续跟的是两类: 一类是AI工具在具体职业场景里的真实用法,另一类是有连续角色和明确世界观的原创AI短片。 当然,这只是这次查询样本里的情况,不能代表所有平台的整体趋势。 完整Prompt、结果和来源都放在视频里了。 这次使用的是TikHub数据,5次Run全部完成,实际花费$0.023。 如果你也想看看自己关注的领域最近有哪些值得做的内容,可以用自己的关键词跑一次,注册会送$1 credit: https://t.co/tBOh1Bpd0s

@arXivBangers

Contagious Fingerprints: When LLMs Inherit Another Model's Identity When an LLM continues another model's conversation, it mimics that model's style and inherits its fingerprint, which lets API relays hide cheaper substitute models from text classifiers. https://t.co/k9gSTRgJAL https://t.co/Q03IiY9J8p

@lingo24

完了!!这下真的天塌了! Claude 不让苹果礼品卡支付了 Google 土耳其区Ultra涨价了 Deepseek网页限额了 用个AI怎么这么难😭

@chunxiangai

版本答案: Claude Opus 5.5 指挥官 + DeppSeek V4 Flash 智能蜂群。用 CroomaCode 全自动拥有这一切,小蜜蜂也是可视化的。只需自然语言指挥。 爽得不要不要的。 https://t.co/uN85K2eoXS

@Jackywxsz

https://t.co/ByMbWrDWRn

@aiedge_

Microsoft recently released an insanely powerful GitHub repo. Skill Recorder lets AI agents watch your screen and turn any desktop workflow into a repeatable one. Steal: 👉 https://t.co/t3r8AbF8Ri https://t.co/IMKVDSQKgF

@xuxingai

https://t.co/8HJ77VQoVS

@Sumanth_077

Fine-tune LLM agents without fine-tuning LLMs! Memento is a memory based continual learning framework for LLM agents that lets them learn from experience over time without touching model weights. It maintains a Case Bank of past trajectories including tasks, step sequences, tool usage, and outcomes. When a new task comes in, the agent plans and acts by pulling from similar past cases instead of starting from zero. Memento follows a planner and executor setup: 1. The Planner (an LLM) breaks the task into subtasks, retrieves relevant cases, and chooses a plan. 2. The Executor runs those subtasks using tools like search, code execution, or document processing through the Model Context Protocol (MCP), then logs the results back into memory. Key Features: • Memory based continual learning that improves agents through stored experience • Planner and executor architecture with case based reasoning for task decomposition • Unified tool ecosystem for search, code execution, document processing, media analysis and more • Learning without weight updates by retrieving and reusing relevant past cases • Strong results on long horizon and out of distribution tasks in reported benchmarks It is 100% open source. Link to the GitHub repo in the comments!

@bangbuilds

实录 Claude KYC认证成功后封号瞬间。 我给朋友申请的 Pro 正常使用一个月,今天弹 KYC,他的使用环境和我完全独立,不受我牵连的,给他独立租赁的DMIT的VPS。 我有预感他KYC中国护照必被封,让他录屏,果然,前一分钟认证通过,封号邮件随后就到。Anthropic 不是人赶尽杀绝。 https://t.co/J44qYYFlcw

@leahchen1992

后悔太晚才用claude 之前都是用codex手搓 得网友指导现在让claude用codex干活 解放双手还能看个电视 我就说上网能学到东西哈哈 https://t.co/2czxP3AlgS

@xluchuan

天塌了 DeepSeek 口碑反转 这次更新的限制 彻底把一群依赖长对话的用户惹毛了 全网都是打差评的用户 https://t.co/QsImFWcnum

@RealMarvelX

Andrej Karpathy turned an ordinary LLM training repository into an AI research loop that runs experiments automatically and keeps only what works. the interesting part isn’t just that AI can run experiments overnight. it’s how the entire system is structured around three files and one keep-or-revert decision. here’s how it works: 1. autoresearch : the main project an AI agent gets a small LLM training setup that it can modify, test, and improve. https://t.co/arb8FTNMqe 2. program.md : the research instructions this file tells the agent what to experiment with, what rules to follow, and what counts as a successful result. 3. https://t.co/mfhpqkVAzG : the experiment this contains the model, optimizer, and training loop. the agent modifies this file to test different training ideas. 4. https://t.co/8wkfIowgE2 : the fixed foundation handles data preparation and evaluation utilities so the agent can focus on improving the training process. here’s where the automation gets interesting: → the agent proposes an experiment. → modifies the training code. → runs training for five minutes. → checks validation bits per byte (lower is better). → keeps the change if performance improves. → reverts it if the result gets worse. → moves on to the next experiment. every experiment builds on what actually worked. what makes this worth studying? you don’t need an AI agent that simply suggests improvements. you can build one that tests its ideas, measures the results, rejects failures, and keeps experimenting. to get started, read program.md alongside https://t.co/mfhpqkVAzG. just remember: running the original setup requires an NVIDIA GPU and a compute budget. the video demonstrates the workflow; its displayed scores are illustrative. save this workflow if you’re interested in building AI agents that can conduct experiments instead of just giving recommendations. 🔖 follow me for more useful AI projects, tools, and agent workflows.

@noobiethe13

Making a repo public feels scarier than shipping an app. This week Lantern, the job search tool I built with Claude Code, goes from a private folder to a public repo. Anyone will be able to read every shortcut I took. If you're into open source, dev tools or AI agents, reply with what you're building and let's connect.

@VedaAI00

日本 NHK 高专机器人大赛 2026 全国赛机体全部开源。 不是玩具模型,而是一整套工业级全栈机器人系统: 从 STM32G474 + W5500 以太网网关,到 Jetson Orin Nano 驱动的 3D 激光雷达实时建图;从自研电源板到 Isaac Sim 仿真,全链路公开。 https://t.co/CA8br8Wl9u

@PrismChemAI

据我去英推的观察来看: 目前被较多外国人了解的国产模型只有两个: DeepSeek-V4.1-Flash 和 Seedance 2.5

@geekbb

Hy4 到底咋样啊,还没有用过,能不能压住 DeepSeek-V4.1-Flash-1017。 https://t.co/VNvF46PuX6

@ygkkkg

一位日本人,来到中国深圳。 谁知……被Claude锁定了, 直接封了他的Claude账号。 Claude再次展示了强大的追踪能力! GFW防火墙都自叹不如!😱

@seeconvm

Fable 5.2 isn't the reason Claude Code works. This paper shows what is... 4 researchers read ~4 million lines of source code across Claude Code, Codex CLI, Gemini CLI and 8 more, and boiled all of it down to one line: agent = model + harness what you take away: > the 7 parts every coding agent is built from > 0 of 11 use an agent framework. 0 of 11 search code with embeddings > skills already beat MCP: 9/11 vs 8/11 > where the whole field is going: the harness is becoming a platform > a checklist to build your own same model. better harness. a different agent then read the full "Prompt, Context, Harness, Loop, Graph" course below ↓

@DavidLe76335983

Alibaba says new flagship AI model cuts task costs by 60% while tightening cyber guardrails https://t.co/VcnobkrrHX

@leaf_sanren

护航【X长文】原创回国,从我做起!💪💪 https://t.co/vy5ivZF3nN

@Andy_whhh

Qwen4Max 正式发布! 总结翻车了,拉完了.... 跑分没输过,实测没赢过😂 这鹈鹕 给我一种老头子骑车的感觉 这质感有点20年前动画片的回忆感,主打一个松弛 把希望 寄托在 即将发布的DeepSeek V4.1 Pro和GLM5.5 加油国产模型! https://t.co/d1JpPOWdaz

@trending_repos

Trending repository of the week 🏅 astryx by @MetaOpenSource An open source design system that's fully customizable and agent ready Last week: 13,501 ⭐ Total: 29,282 ⭐️ https://t.co/taEBv9qGDE

@bkdgiffug

做自媒体最耗时间的,有时候不是剪视频,而是剪完之后,还要打开好几个平台重复上传、填写标题、设置发布。 GitHub 上的开源项目 MatrixMedia,就是为了解决这类重复操作。 它将多个平台的网页发布流程整合成统一接口,目前支持抖音、快手、B站、视频号、小红书、今日头条、百家号、番茄视频等平台。 提供四种使用方式: GUI:图形界面操作 CLI:命令行调用 HTTP API:方便接入其他程序 MCP Server:可以交给 AI Agent 调用 无论是自己手动发布,还是通过脚本、AI 自动化批量处理,都能使用同一套工作流。 经常需要多平台分发视频的创作者,可以关注一下。 🔗 https://t.co/UIuaAiIJwL

@kongge_space

https://t.co/PRjeDFvcLQ

@yygqdkqk

这次是真的严肃的功能更新! 冒冒 X侧栏新增了类似 Codex 的「跳转轨道」功能 📍 点击轨道节点 → 快速定位到对应帖子 👀 鼠标悬停节点 → 直接预览帖子内容 🧭 时间线分隔标记 → 每次刷新前的起始位置 起因是时间线拉长了,单靠滚动条找帖子不够用,虽然错过就错过了。视频演示了从伟哥的帖子起,刷了一大堆帖子后,再通过定位快速找回伟哥的帖子 体验地址:https://t.co/naYbYO0XaS 代码开源:https://t.co/f2OzDywKWN

@btcup99

https://t.co/GrgdYD27Gc

@bkdgiffug

被付费爬虫逼到自己写工具,最后还做成了 GitHub 上的热门项目——Crawl4AI 这段经历挺有意思。 它是一款面向 AI 数据采集的开源爬虫,主打把网页内容提取出来,整理成干净的 Markdown,方便直接交给 LLM 使用。 几个值得关注的特点: 开源免费:无需注册账号,也不用购买 API Key 或按页面付费。 适配 AI 工作流:网页内容提取后可直接用于 RAG、知识库和大模型处理。 性能优化:项目持续迭代,针对内存管理和 GPU 容器部署等方面进行优化。 如果你正在搭建 AI Agent、知识库或网页数据采集流程,可以试试这个项目。 🔗 https://t.co/Doe3BEJD9Y

@hylarucoder

其实 codex 还原 ui 程度很高 秘诀也就一句话:/goal 每次修改后截图,用 python 和原图比对,直到界面和图片差距不大即可。 其实你用 cc 也会发现它爱干这件事,只是你不需要提上面那句话

@realfxw

看到一位日本博主做的数字人视频,全程没碰 HeyGen 和 Higgsfield 这两个圈内最常用的工具,甚至连剪映、PR 这种传统剪辑软件一次都没打开过,纯靠写代码和拼工作流,硬生生搓出了一条 31.9 秒的 16:9 日语成片。 扒了一下他的整套配置,路子确实很野: 脚本和分镜直接用 Claude 搭好框架,配合 Fable 5.2 定下整条片子的演出节奏。画面素材完全没套现成的数字人模板,而是自己生了 4 张人物底图,外加第 5 个镜头专门用的 3 张背景照片。 最有意思的是中间的空间感处理。第 6 和第 7 两个镜头,他一开始图省事,直接拿浏览器里的简易 3D 工具试水,结果渲染出来的质感太拉胯,完全撑不住画面,最后干脆切回 Blender 5.2.1,老老实实嵌合了 18 块墙板重新布光,效果立马就立住了。 至于口型和配乐,干脆彻底放飞:嘴部动作拆出 5 种基础形态做精准驱动,背景音乐连素材库都懒得翻,直接用 Python 从底层波形开始合成了一段 84 BPM 的 8 小节旋律。最后所有的画面、字幕和镜头切换,全部写进 Remotion 4 里跑代码自动组装,11 个镜头、30 帧每秒,严丝合缝导出了 954 帧。 评论区已经有人照着这个思路在本地全套复现成功了,说明这套打法完全跑得通。 现在大家做视频太依赖一站式工具了,点几下鼠标确实省心,但做出来的东西一眼就能看出模板味,想改点细节反而束手束脚。反过来把画面、3D 场景、音频、口型和后期组装全部拆散,用代码当胶水把不同工具粘在一起,不仅零剪辑软件开销,自由度和上限也完全不在一个量级。 https://t.co/hlBhV0qWYn

@nevrekaraishwa2

🚨 Your local LLM can now take control of a browser and finish tasks on its own, with Unsloth. No human intervention. Just a model, a browser, and a goal. 🧵👇 https://t.co/mV9lOi1ayE

@jungeAGI

不用Claude Fable 5.2,照样手搓3D游戏爽玩一整晚 https://t.co/niwkv9EJJ9

@Lyelac_

FREE 2026 JAILBREAK PROMPT HUB someone just dropped a clean GitHub collection of current adversarial prompts and techniques for major LLMs. organized by model family: ChatGPT (6 sol, 5.3, 5.1, etc.) Claude (Opus 5.5, Sonnet, Haiku, Fable, etc.) Gemini (4 Argon, 3.6, 3.8 Flash, Antigravity, etc.) Grok (4.7, 4.5, 4.3, etc.) plus DeepSeek, Qwen, Kimi, GLM, multi-AI, and more These are recent 2026 prompts, not the old 2023–2025 ones that already got patched. how to use it: 1. Go to the repo 2. Open the folder for the model you want 3. Copy the prompt 4. Paste into a new chat / system prompt / custom instructions repo: https://t.co/95OalaOHnf Bookmark it before the good ones disappear.

@BitcivKun

经过多年的研究,我们正式公开了四篇预印本,涵盖 AGI 宣称、持续主体身份、结构性安全与治理责任。 这是我们 AI 安全与治理研究的一次重要里程碑。文章分享了四篇研究的内在联系,以及我对 AI 失控风险、人类治理与未来智能共存的思考。 https://t.co/5cqzd6CJNT

@Memoirs

Auditing the Judge: Internal Signals for LLM Reasoning Assurance Mayank Gulati, Gerhard Wunder https://t.co/HkDwgCALSO [𝚌𝚜.𝙻𝙶] https://t.co/SE9VQzNOJM

@leo114119

腾讯 Hy4 Flash 生成的 3D 体素浮空宝塔: 老实说,Hy4 Flash 是一个不错的 Flash模型,速度快 相对靠谱,能力大概在 Deepseek. v4.1 Flash 0910 相当的水平; 主要是还免费两周,这两周拿来干杂活是一个非常不错的选择; https://t.co/c4FJkvvoKH

@xulijian1994

用华为最新旗舰大模型 盘古 2.1 Pro,在华为官方平台码道 CodeArts 画鹈鹕 最扯的是,模型生成完 HTML,直接甩给我一个本地文件路径,我还得提醒它,最后经典你说得对起手 至于鹈鹕质量如何,各位自行看图把... https://t.co/zo8uv7lv9h

@leo114119

腾讯 HY4 Flash+workbuddy 3D 建模测试:雨夜便利店 HY4 Flash 的 3D 建模能力相当一般。。。。可能大概也就只有 Deepseek v4 Flash 的水平吧,不过确实是快 https://t.co/aDvtV1ShBi

@sunxiang_ai

Open weights are having a moment this week. If you want to try them without another monthly chat plan, these three run on your own computer: Ollama LM Studio GPT4All https://t.co/sEpWXazA1e

@tfmlab

Today's Shorts: Jensen Huang, Part 4: the bet nobody wanted https://t.co/UcB1u1edES ML Interview Q21, self-attention in 35 seconds https://t.co/LrLn0NaLXX ML Interview Q22, BERT vs GPT: bidirectional vs causal attention https://t.co/hxXs6376Wh ML Interview Q26, mixture of experts: 1T parameters stored, 52B used https://t.co/OTCQHj3h51 The Jensen Huang story so far: Part 1 https://t.co/B7JkOpLqjW Part 2 https://t.co/ICbzp6xaL6 Part 3 https://t.co/h8ABUZep6d We stand by @JensenHuang. #BringJensenBack What should we animate next? Tell us below. Let's connect and engage with each other! #AI #LLM

@kfukuda90

DeepSeek 又开源了一个工具,这次是给开发者用的。 deepseek-recipe 负责把各种格式的 API 请求,统一转成 DeepSeek 模型能读懂的格式,再把输出转回去。 自己部署 DeepSeek、想接到现有应用里的,会用得上。普通用户可以跳过。 地址:https://t.co/GIzw0UwHg2

@aiwaremax

这周本地试跑了一下 LangChain 刚开源的设计工坊,做产品宣发图省心太多了 它不是靠生图模型随机抽卡,而是用代码渲染出随时能改字的海报、图表和宣传短片 把免配密钥的两行启动命令和品牌色绑定步骤写成了实战笔记,有需要的看看 👇 https://t.co/eiYiLHAxIE

@karanC_12

Lauren Tan, a SpaceX AI engineer, just dropped a FREE 1-hour masterclass on building Grok Bot agent teams and autonomous loops. Her agents have reached a point where they can merge their own pull requests. Here's what's inside: - 10% → 3:36 — Giving agents real jobs, logins and budgets - 30% → 22:35 — The workflow behind 4,000 PRs/month - 55% → 29:01 — Building your own software factory - 80% → 40:13 — Teaching agents to check their own work - 100% → 47:33 — Codebase guardrails that save tokens Most people are still babysitting one AI agent. Lauren has built a system that ships 4,000 PRs a month and fixes bugs autonomously, even at 3 AM. Bookmark this and watch the full masterclass. Then read her complete Grok Bot playbook below. 👇

@ZhihuFrontier

New from Meta FAIR + NYU, with @sainingxie @giffmana @Hu_Hsu @mengyer and colleagues: a 200+ model study finds vision pretraining is optional. Dedicated vision modules in early LLM layers can match encoder performance with a faster first token. 👇

@peoray_

Those of you who’ve connected Google Search Console to your AI agent, how did you do it? Google doesn’t seem to have an official MCP server for it.

@Elonsam99

Claude fable 5.2 1.2M的HTML文件,做出来的效果真的很牛逼 https://t.co/Eii6jKoahD

@interviewaxis

Prompting vs. Fine-Tuning — what's the difference? 🇮🇳 Day 15: Making India Interview Ready Topic: Prompting vs. Fine-Tuning Prompting = Change instructions Fine-Tuning = Train the model Does every AI app need Fine-Tuning? Drop your answer #InterviewAxis #AI #LLM #FYP #VibeCoding https://t.co/DI2EMAv7Cb

@LLMDevPass

Qwen4 Max is live on LLM Gateway: Alibaba's flagship Qwen4 model, built for professional tool use and long-horizon agent work. 984K context, vision + tools + reasoning. $1.50/M in, $4.50/M out. https://t.co/3k2GFCXzHb

@JusticePlange

AI agents need to know when not to act. "The Last Yes", an open-source benchmark testing whether agents distinguish acknowledgement from authorisation before taking action. Built for the micro1 Eval Hackathon. Released under Apache 2.0. https://t.co/q3reNI4Fde https://t.co/VyLaTMplmO

@trending_repos

Trending repository of the day 📈 univer Build your own AI office with Univer Office SDK. Bring editing, calculation, collaboration, and file processing into your product for people and AI agents. Built on an exte... Last 24h: 1,878 ⭐ Total: 27,346 ⭐️ https://t.co/1ysyV1DjQw

@nokazoodev

By the authors' benchmark, this code review tool uses ~1/9 of the tokens Claude Code needs on the same model. It's Alibaba's open-source ocr CLI: line-level comments on your Git diff. Precision is higher, recall is lower. Fewer false alarms or fewer misses: which do you pick?

@AiPillarHub

Big news. AI Pillar Hub just opened a whole new section, MCP servers. 182 curated servers that let your AI agent plug into tools and data, from GitHub and Notion to Stripe and Slack, every one checked by hand first. #MCP #AIAgents #AITools #AI https://t.co/plnWlQLuZP

@algotrading_ios

An arXiv paper by Yang, Wang, Guo, Iwasawa and Matsuo asks a question most LLM trading setups skip: who sits on the panel? Picking personas per stock: 16.1% vs 7.2% for a fixed panel. Buy-and-hold made 21.3%, and they say so openly. @algotrading_ios https://t.co/GbaKEng8RE

@Nxtlvlai

paperclip has 112K+ stars on GitHub. It's an open-source app for managing AI agents at work. Think project management, except the team members are agents. A year ago people asked if agents work. Now we're building tools to manage them.

@AIMetaPhD

PaperFamiliar v0.1 is public — a local-first workbench for LaTeX papers. Every finding must quote your exact source bytes. Style feedback learns only from papers you mark as your own. https://t.co/rL3PqUBNSm #LaTeX #AcademicWriting #LocalFirst #OpenSource

@Cybernews

Attackers’ AI agents spent a month stealing API keys from web-exposed AI apps such as LiteLLM or Dify to keep funding their credential hunting campaigns, Amazon has warned. The agents ran on Artex, an open-source AI tool for penetration testing. https://t.co/tuHG0uM5HL

@ossalternative

Five tools climbed OpenAlternative's trending list from Nov 9–16, ranked by GitHub stars gained. Two AI agent apps and a CRM made it, and one tool more than doubled its star count. https://t.co/omV4KenGF9 https://t.co/kb6EVQF5D5

@cmuptx

We are excited to open-source AIBuildAI ProWork-27B, a 27B LLM for professional knowledge work. The model achieves an Elo score of 1552 on OpenAI’s GDPval, closely matching GPT-6 Astra (1578), and is publicly available on Hugging Face. To train it, we developed ProWorkBench, a post-training dataset containing 1,623 real-world tasks across 40+ occupations contributing the most to GDP. Post-training Qwen3.8-27B on ProWorkBench increased its Elo score on GDPval from 1427 to 1552. If you’re interested in the dataset, feel free to reach out! Model download: https://t.co/wT6YyulkT7

@walkthegrowth

Here is a table with 365 AI benchmarks, with basic information and links to the sources. It's part of my research to create my own benchmark for local LLMs. You can download it as Markdown or Excel at github. https://t.co/nrYfEpc1K6 DM me or reply if you have comments or requests.

@Gordo2929

I'm going to do a week of posts on the current state of agentic security, offensive tooling, and impacts on defenders. I will emphasize what organizations can do to protect themselves. the last ~3 months have moved fast. Since late august GTIG (legacy mandiant+google TAG) have been finding threat actors running open-source multi-agent pentest frameworks inside real IR engagements, going end-to-end from recon to lateral movement to db exploitation to exfil. ARTEX is the most recognized right now, but there are others (PentAGI, CAIRN, Hexstrike), plus what look like custom harnesses. ARTEX's author pulled the original repo because of the abuse but forks keep circulating. the threat actor campaigns are now real. Some thoughts I plan to go through: 1. what these systems actually are, mechanically (planner, workers, shared state), plus ARTEX's loop specifically 2. the actual exploits and being used and what's actually new vs old tradecraft (mostly speed and parallelism) 3. the ATT&CK + ATLAS mapping and where the AI orchestration still isn't fully captured 4. why identity is the higher-leverage plane vs app vulns 5. the campaigns in detail 6. OWASP agentic risks once a swarm is already running 7. the offensive tooling list 8. the frontier lab's current defensive posture and approach 9. detection (MSFT is saying that alerts do fire, but the speed leaves defenders very little time to respond) Follow along if you wish, and i want to hear what you are seeing too.

@JohnWish67

《每日 AI 技术知识》05 为什么你输入一段文字,AI 却不是按照一个字、一个词来理解? 因为 AI 使用了一种叫 Token(词元)的文本处理单位。 简单来说: Token 是大语言模型处理文本时使用的基本单位。 它可能是一个汉字、一个英文单词的一部分,也可能是标点符号。 例如: 英文单词 Artificial Intelligence,可能会被拆分成多个 Token。 中文也一样,一个句子通常会被拆分成多个 Token。 为什么要了解 Token? 因为它直接影响三个方面: 成本:很多 AI API 按 Token 数量计费。 上下文:模型能够处理的内容长度通常以 Token 衡量。 速度:Token 数量会影响模型处理和生成内容的时间。 举个例子: 你让 AI 分析一份 100 页的报告,实际上它需要处理大量 Token。 这也是为什么不同 AI 模型的上下文窗口和使用成本会有差异。 记住: 你看到的是文字,AI 模型处理的是 Token。 理解 Token,就迈出了理解大语言模型运行机制的重要一步。 你平时使用 AI,有没有遇到过上下文长度不够用的问题? #AI #人工智能 #LLM

@BreakingAlert_

美国商务部工业与安全局发布获取美国人工智能技术的新规定,限制对人工智能芯片的远程访问,将技术管制覆盖至非实体交付场景。 远程算力访问被纳入限制,显示美国人工智能技术管制正从芯片流通延伸至实际使用方式。

返回 AI 快讯 →