2026年11月14日 · AI 晚报

本轮重点是 DeepSeek 将产品能力向开发者开放,以及智能体从写代码走向常驻运行、跨模型协作和具体工具操作。研究讨论则围绕推理能耗、检索、自主学习与模型架构展开,实际收益都需要结合测试条件理解。

11月14日 11:00 — 11月14日 23:00(日本时间) · 128 条 X 原帖

查看本期全部 128 条原帖 ↓

本期目录

DeepSeek 从聊天产品走向能力接口

搜索、文件与语音能力成为开发积木

驱魔先生、Gorden Sun 等社区作者集中转述:DeepSeek 将搜索、文件解析、链接读取及语音识别、语音生成等产品能力开放为 API,并提供调用这些接口的网页 Demo。驱魔先生还提到,Demo 可以导入 DeepSeek 历史对话。这批材料支持的主线是接口开放与示例代码共享,不能直接扩展成整个服务端技术栈全部开源。

讨论关注的是开发者能否少拼接几家服务,就做出能搜、能读、能说话的助手。驱魔先生认为,单纯包装搜索和文件解析的中间层产品将承受压力;Kathleen Tyson 则给出闭源模型难以竞争的强烈判断,后者属于行业评论,并非已经得到验证的结果。

来源 · 3 条原帖

订阅套餐出现代码线索,价格仍无定论

Zane 与 V哥来了分别提到,网页前端出现额度管理、使用上限和 Coding Plan 相关代码;专属 API Key、不消耗普通 API 余额的说法则来自第三方文档。本批没有正式套餐公告,尚不能确定上线时间、价格或 API 权益,也不能把模拟渲染的额度弹窗当作已上线页面。

来源 · 2 条原帖

模型选择开始同时算效果、成本与控制权

高推理档位未必带来更好交付

Leo Wang 转述 Artificial Analysis 编码智能体榜:Fable 5.2 与 Opus 5.5 并列 58 分,但得分超过 48 分的配置,每任务成本从 1.51 到 44.22 美元,相差约 30 倍;最贵的 max 档还遇到了时间上限。这个比较提示,模型能力、执行档位与任务预算需要一起评估。

社区体验也并不一致:苍何称 Fable 5.2 用 15 分钟做出技能和小兵可用的简易游戏,Leo 则觉得它生成的 3D 鹈鹕一般。两者都是具体作品体验,不能据此概括所有游戏或建模任务。

来源 · 3 条原帖

免费入口之外,本地权重也要看许可证

kadd 介绍,ERNIE 5.5 Flash Preview 已可通过 OpenRouter 免费试用,但预览期限额可能变化。另一边,Alexey Fateev 整理了覆盖语言、图像、视频、语音和音乐的本地模型包,反映出社区对平台访问稳定性的担忧;“开源模型即将全面关闭”仍只是帖主的判断。

这份清单更实用的部分是部署细节:量化仓库只下载所需版本,分片文件要齐全,原始权重还需保留配置、分词器和许可证。清单同时包含需要接受条款的仓库及非商业许可项目,能下载并不等于可以任意商用。

来源 · 2 条原帖

研究在追问:哪些计算可以省下来

投机解码需要同时优化速度和功率

青稞社区介绍的 Olifant,用只看四个 token 的记忆分类器生成草稿,替代小型 Transformer,CPU 预测耗时低于 0.3 毫秒。帖中将结果分开报告:无 KV cache 时,领域匹配测试吞吐达到 2.43 倍;验证端启用缓存后,仍有 1.25 倍吞吐。

关键发现是,速度最优的草稿长度未必能耗最低:更长的验证会提高功率。收益还受领域匹配影响,研究只在消费级硬件验证,不能直接外推到所有服务栈。

来源 · 1 条原帖

检索与学习,不只是在提示词里塞材料

Simplifying AI 转述的 NVIDIA UNREAL,让语言模型利用内部信号挑选段落,冻结原模型,仅增加不到 50 万可训练参数。在帖中未具名的一项基准上,前十条结果包含全部必需段落的问题占比为 73.2%,基线为 49.1%;目前测试范围是维基百科。

Vikesh 介绍的 BookNavigator 则研究如何用强化学习训练模型自行浏览教材、寻找概念并应用知识。其转述称,Intern-S2-Preview-35B 的数学证明准确率从 51.5% 提升至 79.5%,另有两项基准的小幅改善。两项工作的共同问题,是如何让模型更有效地取得并使用知识。

来源 · 2 条原帖

字节输入、视觉模块与决策模型各有探索

CrazyShyyt 转述 Meta 的字节级建模研究,涉及 End-Of-Token 蒸馏,并声称在相应实验中用六分之一训练数据达到 token 模型水平;这不足以支持其“终结当前 LLM 时代”的标题判断。Shumpei 分享的视觉研究则认为,专用视觉容量比预训练本身更关键,将视觉模块放入语言模型早期层可在降低延迟的同时取得相近表现。九原客介绍 Drex 时更强调用途:以 LLM 为骨干提升决策模型能力,面向分类、路由等可落地任务。

来源 · 3 条原帖

智能体工程转向常驻、分工与执行边界

始终在线,不等于模型始终运行

meng shao 对 Claude 常驻智能体架构的解读,把价值放在收件箱、计时器、记忆、权限与发件箱的职责划分。花叔已经在 FanBox 做出具体应用:终端智能体停下来等人时,通过微信或飞书发一句通知。ChiCha 介绍的 agentring 则用 VoIP 呼叫 Android,通知任务完成或卡住,并提供自托管选项。

来源 · 3 条原帖

跨模型分工,把强模型用在关键决策

旺旺介绍 Fable 5.2 Advisor:主模型继续执行,仅在方案检查、连续报错和交付前审查时请顾问模型介入;能否省钱仍取决于返工率和总 token。卡兹克分享的个人流程则让 Claude 负责规划、前端与验收,Codex 承担成块的后端、算法和数据处理,再读 diff、跑测试。这是作者的实践选择,尚不能称为通用最优解。

来源 · 2 条原帖

可靠性落在权限、调用标识与可复用设计

awe0 推荐的 Hooks 示例可在工具运行前检查危险操作、拦截敏感文件访问并记录调用。goose v1.59.1 的社区发布说明则更具体:修复 Databricks 网关下工具调用标识重复导致的结果丢失,并调整 Fable 5.2 的自适应推理请求。Ren 分享的 22 章 Agentic System Course,把状态、协作、观测、成本和安全放进同一设计框架,还附带按项目规模评审的 skill。

来源 · 3 条原帖

数据与知识工具补上业务上下文

数据库设计要早于代码补丁

旺旺推荐 Database Architect,让智能体先分析业务访问模式,再处理表结构、索引、多租户隔离及迁移回滚。Vikram 介绍的 Dex 则面向已有数据仓库,为编码智能体提供表关系、指标语义和变更信息,并检查昂贵查询。据其转述,搭配 Claude Sonnet 5.5 的 Dex Semantics 在 ADE-bench 得分为 80%,未使用 Dex 时为 73.3%;这是特定组合下的结果。

来源 · 2 条原帖

本地知识库既要找得到,也要改得对

FELIX 介绍 SAG Wiki,称其可直接索引现有文件夹,索引阶段无需嵌入、向量数据库或模型调用,并支持增量更新及原文件引用。北衡推荐的 Obsidian Skills 解决另一端的问题:让智能体正确编辑双链、Callout、Bases 和 Canvas。前者强调检索已有材料,后者强调遵守知识库格式,都是为具体工作环境补上下文。

来源 · 2 条原帖

代码视频开始服务创作与审阅

把重要 PR 做成分镜说明

meng shao 转述 Matt Pocock 预告的 animatic Skills:将重要 PR 写成 JSON 分镜,用 Playwright 渲染包含真实 diff 行号的静态画面,再配旁白、合成 MP4 并逐帧检查。目标是缓解智能体提交速度超过人工审阅速度的问题。预览案例约 90 秒完成,作者另称平均约四分钟,这些是不同口径的自述,不能统一写成固定耗时。

来源 · 1 条原帖

本地剪辑与生成画布各自补齐流程

北衡介绍的开源 Skill 提供 42 个本地 FFmpeg 工具,覆盖裁切、字幕、画幅和响度,并先探测素材、完成后复核。Eric Kang 自述的 BeatDesign 则把提示词、图片、片段和时间线放在本地画布,允许 MCP 智能体参与同一项目。周览资源推荐的 Lemo-Opuscar 展示了代码驱动影片的风格库;Haosen 同时提醒,同叫动画 Skill,交付物可能只是图片、网页,也可能才是 MP4,选工具前应先看最终产物。

来源 · 4 条原帖

实时交互从语音延伸到眼镜与触觉

日语语音模型强调交谈节奏

Ashley 试用 Haru-4B 后,特别肯定日语应答和附和的自然程度,同时直言模型还不够聪明。她介绍,其通过在同一 token 位置共同处理输入音频、输出音频和文本缩短上下文,并称 MacBook 上端到端延迟为 65 毫秒。这个案例把语言特有的交谈习惯与知识能力区分开来,短时聊天自然并不代表复杂任务表现成熟。

来源 · 1 条原帖

眼镜识物与机器手触觉都有具体硬件路径

cabbit 转述的眼镜项目,通过视频通话把摄像头画面送到 Mac,再用 OBS、YOLO 和语音播报串起实时识物;帖子描述的是给人、车、狗等目标分类和跟踪,不能理解为识别人物身份。机器人方向,Techniahqrobot 介绍 CMU 与 Meta FAIR 的开源 D’Manus:三指、十自由度,指尖与掌面共布置 56 个磁传感器,通过柔性 ReSkin 表面获取触觉信息。

来源 · 2 条原帖

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

全部 128 条原帖

@Kathleen_Tyson_

DeepSeek just avenged Old Huang after Anthropic’s decapitation strike by releasing its entire App stack for everyone to use and adapt freely via DeepSeek API. Dario and Sam are now dead men walking as closed models can’t compete. “Now, all the capabilities behind our official products can be freely accessed and combined through the DeepSeek API. We hope developers can use this to skip repetitive foundational work and build more diverse, more polished, and more imaginative products. We welcome you to use the DeepSeek API and create freely.”

@superalesha

DOWNLOAD THESE 24 OPEN MODELS BEFORE MORE REPOS GET GATED OR TAKEN DOWN. You asked for the exact links, so I put together the full local AI pack: LLMs, uncensored quants, image generation, video, voice, music and speech to text. Put them on your own disk while you still can. 2 repos in this list already block anonymous downloads behind a login gate, and Hugging Face has disabled uncensored repos with zero warning before. If a provider bans your account, raises prices or rewrites its rules next month, nobody can take away files that are already on your drive. For all 4 LLMs below I included 3 versions: original weights, standard GGUF and abliterated GGUF with refusals removed. For Qwen3.8 Flash-Next and GLM-5.3-Flash, the uncensored GGUF repos are my own quants. [LLMs: original, GGUF and uncensored] Qwen3.8-27B Original: https://t.co/gZiDV7EOcq GGUF: https://t.co/ErvJ3X85Iu Abliterated GGUF: https://t.co/nNlGlOl3Aj Qwen3.8 Flash-Next Original: https://t.co/q0jco7xZuc GGUF: https://t.co/2DY6VS1TPH Abliterated GGUF (my Strata quants): https://t.co/ikTy36h1tM GLM-5.3-Flash Original: https://t.co/pkGk5447zj GGUF: https://t.co/qBfmPJW7Yb Abliterated GGUF (my Maya-M quants): https://t.co/FBy2Bn3eFx DeepSeek V4.1 Flash 1017 Original: https://t.co/g9BGRmz73L GGUF: https://t.co/hDOrdVUGq1 Abliterated GGUF: https://t.co/0XymBuJHpU [Image generation] Qwen-Image 2.1: https://t.co/R7rW0MRF9F MOSS-Image Base (Apache 2.0): https://t.co/9piDNQFkII FLUX.3 Klein for ComfyUI (gated, accept terms on HF first): https://t.co/y73OolCCAK [Video generation] Wan 2.2 Text-to-Video A14B (Apache 2.0): https://t.co/BcLlm3dJ3R Wan 2.2 Image-to-Video A14B (Apache 2.0): https://t.co/tsBaldQAzb HunyuanVideo 1.5: https://t.co/YQwooOPrdb LTX-2.5 (gated, accept terms on HF first): https://t.co/WEPR3rGqAH [Voice and TTS] Fish Audio S2 Pro (non-commercial research license): https://t.co/AIQDkvObfL Qwen3-TTS 12Hz 1.7B Base (Apache 2.0): https://t.co/yrK2L5q9wS [Music generation] ACE-Step 1.5 (MIT): https://t.co/U2QDLTCRLB YuE2 ComfyUI bundle (non-commercial CC-BY-NC-4.0): https://t.co/SbEyymdNc5 [Speech recognition] Whisper Large v3 Turbo (MIT): https://t.co/2ocXMygFKg [How to download cleanly] Use the official Hugging Face CLI: pip install -U huggingface_hub hf download OWNER/REPO --local-dir ./MODEL Do not pull every quant inside a GGUF repo by mistake. Pick 1 quant (and all of its split shards if it is split) with a file filter: hf download OWNER/REPO --include "*Q4_K_M*" --local-dir ./MODEL For original weights and media models, save the configs, tokenizers, text encoders, VAEs, README and LICENSE files too. For the 2 gated repos, accept the terms on the model card and run hf auth login first. An API account can get cut off in 5 seconds. Weights on your own disk stay yours even if the lab changes its mind tomorrow. Save this post and go fill your drives while every link still works.

@cabbitar

this is wild... the guy reverse-engineered his smart glasses with Opus 5.5 into an AI object recognition device you can hook it up to any smart glasses with a camera and they start naming every person, car and any object they see, live it's free, open source and runs on a regular Mac here's the full breakdown: step 1 → install YOLO with one command: pip install ultralytics step 2 → put on the glasses and start a WhatsApp video call to your own Mac the glasses look at the street, and the Mac sees exactly what they see step 3 → in OBS, capture the call window and turn on Virtual Camera now the glasses feed shows up on your Mac like a webcam step 4 → run the tracker on that feed: yolo track model=https://t.co/4FkPisbWkV source=1 show=True every object gets a box, a label and an ID that sticks to it from frame to frame step 5 → keep only what matters to you: classes=0,2,16 for people, cars and dogs, conf=0.5 to drop weak guesses step 6 → have a small Python script say what it sees with macOS say, routed into the call through a virtual mic like BlackHole: "car on your left". it plays right in your ear step 7 → train it on your own objects: your products, your tools, your dog. swap coco8.yaml for your own labeled photos and run yolo train step 8 → export it for your phone: format=coreml for iPhone, format=tflite for Android, so you don't need the Mac at all under the hood: → YOLO looks at each frame only once and finds every object in that single pass, which is why it's fast enough for live video → tracking (ByteTrack by default) keeps the same ID on the same dog even when it walks behind a tree → 80 object types work out of the box, no training needed https://t.co/nzOvJTuy72 try it yourself then read the guide on how to reverse-engineer any device with Claude:

@TiagerBao

AI 写代码最容易埋的坑之一,就是数据库表设计错了,等上线后才发现改不动。 发现一个 Claude Code 专用 Agent:Database Architect。 它不只是帮你生成 SQL,而是提前参与数据库架构设计。 主要能做这些事: 数据建模:表结构、主外键、索引、约束 技术选型:PostgreSQL、MySQL、MongoDB、向量数据库 多租户设计:RLS、Schema 隔离、独立数据库 性能规划:读写模式、分库分片、扩展方案 数据迁移:生成迁移脚本和回滚方案 安装命令: npx claude-code-templates@latest --agent database/database-architect 我觉得最实用的是让 Agent 先分析业务访问模式,再设计表结构,而不是代码写完才补数据库。 不过生成的 DDL 和迁移脚本仍需要人工审核,尤其是生产环境的数据变更。 你现在用 AI 开发项目,会让它直接设计数据库,还是自己先把表结构定好? 项目:https://t.co/xrVQkEQ6iX #ClaudeCode #AI #Agent #Database

@gengdaJ

大厂的AI商战真是朴实无华,就是猛猛砸钱找专家,猛猛做数据标注。。。字节的Xpert,10月份有一个人领了18万。。。 AI时代,只要你有任何特长,只靠贡献自己的知识和努力,也能活得很滋润: ⓵专业问答与评测 -专业领域问答(医疗、法律、金融、STEM等) -模型回复打分与对比评测 -Benchmark出题与标准答案设计 ⓶聊天 / 对话相关 -多轮对话评估:对模型在聊天中的回复打分、排序、指出问题(准确性、自然度、逻辑、人设一致性等) -对话生成 / 改写:写或优化多轮对话数据,用于训练聊天机器人 -偏好排序(RLHF):同一问题给出多个回复,选出更好的那个并写理由 -角色扮演 / 人设对话:按指定角色(医生、律师、客服、特定性格等)进行对话评测或数据生产 -对话纠错:找出聊天回复中的事实错误、逻辑漏洞、不当内容 ⓷内容生产类 -高质量专业内容撰写(案例库、知识库、语料) -复杂指令设计(多约束、高难度Prompt) -参考答案撰写 -专业内容改写与优化 ⓸代码 / 技术类 -代码生成评估 -Agentic Coding / 长程任务评测 -Bug查找与代码审查 -编程题出题与判题标准制定 ⓹语音 / 多模态类 -方言语音采集与标注 -语音转写质检 -视频/3D/动捕数据相关 -多模态内容评估 ⓺审核与质检类 -模型输出合规性审核(医疗、法律、金融强合规场景) -知识准确性审核 -已有标注数据的二次质检 ⓻其他 -访谈类任务 -数据采集(按脚本录制真实场景) -创意内容评估(影视、文学、设计美学等) 上面这些任务,简单到只要愿意花时间聊天,时薪也在三位数。更专业一些的比如生物、法律、金融、制造,就更加赚钱了。 可以参与的平台: 1、腾讯旗下的ExpertHub:https://t.co/CCOo3NFZm1 2、字节旗下的Xpert: https://t.co/MJYBDP1GSR 3、阿里的晓天睿士: https://t.co/FzRmiEqBU8

@shao__meng

Claude 团队发布了一份可以照着搭的架构说明 收件箱、计时器、记忆、权限、发件箱单独看都不新,很多团队早已用队列加定时任务拼出过类似的东西 它真正有价值的地方,是把每个部件的职责划清楚,并明确一点:始终在线是一种使用体验,并不是让模型一直挂在线上消耗 token https://t.co/C6j1v9BRwB

@benshandebiao

https://t.co/xFWkP3Dg0R

@bboczeng

事实证明,DeepSeek果然是GB200和B200最大走私商 英伟达不送几个人进监狱,那是真的说不过去了

@BrianRoemmele

The Time Is Now To Install Local AI Is Now! Governments and platforms have begun restricting model access, vetting customers, and pulling individual weights. Local files do not disappear when the rules shift. The Great Open Source AI Shut Down Approaches. Be ready. https://t.co/Nn1tkDQRxq

@maanas_tyagi

I am an AI Engineer. Start here if you want to be one. Learn these 8 topics, then build one small project for each: - RAG (let an LLM read your own documents before it answers) - AI agents (an LLM that calls tools in a loop) - Agent memory (what the agent remembers between steps) - Tool calling (the LLM asks your code to run a function) - Vector databases (search by meaning, not exact words) - LLM evaluation (checking if the answers are right) - Context engineering (clear, specific instructions) - Model routing (use the cheapest model that can do the job)

@TiagerBao

Claude Code 不用全程跑最贵的模型,也能在关键时刻调用 Fable 5.2。 发现一个配置:Fable 5.2 Advisor。 思路很简单:主模型继续执行任务,遇到关键决策再请 Fable 5.2 提建议。 主要用在三个场景: 开始编码前,检查技术方案 连续报错时,分析问题并调整方向 任务完成前,检查是否遗漏关键问题 安装命令: npx claude-code-templates@latest --setting model/fable-5-2-advisor Fable 只在被调用时单独计费,不需要每轮都使用它。 但多一次模型调用就有额外成本,是否真正省钱,还得看返工率和总 Token 消耗。 你更倾向于全程使用最强模型,还是普通模型执行、强模型负责关键决策? 项目:https://t.co/6c8W4qvOfZ #ClaudeCode #AI #Agent

@Crypto_Holding_

@Concordium is the #AI infrastructure for the agentic economy! Verified Humans. Verified Agents. One Protocol. Purpose-built regulatory-grade L1 blockchain with identity & trust at the protocol level. Every account (human, business or AI agent) is verified, enabling privacy-preserving compliance via Zero-Knowledge Proofs. Highlights: ✅ Protocol-level identity & Agent Registry ✅ Fast finality (2.5s blocks), low fees (~€0.01) ✅ Protocol-Level Tokens for secure stablecoins ✅ Native $CCD for fees, staking & governance ✅ Sponsored tx for seamless agent operations Designed by leading cryptographers for trusted AI agents & real-world adoption. 🌐 https://t.co/3MWwz25iSw Yesterday, they invited developers to build trusted AI agents. Now, let’s show you what’s possible. 👀 Watch their CTO, Peter, walk through a live demo.

@gaoyy9

v0 上能免费使用 Anthropic/claude-fable-5.2 实测用 +86 和 +852 手机号都能直接注册过验证,抓包看了一眼后台请求,确实跑的是 anthropic/claude-fable-5.2 原生接口,没套壳。 地址https://t.co/dQ4Ox4JUv2 https://t.co/g3vUAoyGz6

@tom_doerr

turns technical books into structured agent skills for Claude Code and GitHub Copilot CLI https://t.co/GFnX15PgCP https://t.co/4abdGmHboO

@Yuziyyaa

卧槽重磅消息🚨 GLM-5.5 大模型就要发布了!! 那张超越Opus 5.5 的图竟然是真的? 智谱究竟发生了什么,为什么那么猛??? https://t.co/m14XdfUfQ8

@CrazyShyyt

Meta published a paper that might end the current LLM era. They proved you don’t need tokenizers to train SOTA models. Every major AI (ChatGPT, Claude, Llama) reads text in "tokens", chunks of words. This requires massive vocabularies, usually around 100,000 different tokens. It's how models have always been built. But researchers at Meta proved this is a massive bottleneck. They trained a model to read raw "bytes" instead. Only 256 basic units. The plot twist: They took a tiny 1-billion parameter model and fed it up to 1 trillion bytes of data. At first, the standard token-based model was winning in the low-compute regime. But then something crazy happened. The token model hit a ceiling. It plateaued. It completely stopped getting smarter. The byte model just kept climbing. By converting token probabilities into pure byte-level reasoning (using a new distillation method called "End-Of-Token"), the tiny byte model shattered the performance ceiling. It matched the standard token model using just 1/6th of the training data. It crushed storage costs by 80%. And asymptotically, this tiny 1B-parameter byte model outperforms Llama 3.2 and even the 2-billion parameter Gemma model across the board. Nobody is talking about the real implication here. We have spent years forcing AI to read text the way humans do. It turns out, AI gets much smarter, much faster, when you let it read like a machine.

@zane_os

看来DeepSeek的的订阅套餐很快就要来力 提交了很多关于额度内容的前端代码 图二是模拟渲染额度耗尽后的弹窗 大鲸鱼哈气咋这么可爱🥰 参考豆包最低档位68元的话 我一定要支持梁子和大肥鱼! https://t.co/KT7ZXoUjBy

@indiehackercase

卧槽腾讯这把要玩大的了!腾讯做了个智囊团专家平台,邀请各行各业的专家,注册认证完成就有100元,首次参加完成任务招募,除了任务本身的报酬,还有200元奖励。已经有人实测领到100块。 重点是,你在短视频平台有一万粉丝,或者做过自己的工作流,就能加入这个计划,不需要什么博士一作。博主这个职业的含金量真是越来越高! 猛戳注册认证,建议电脑上操作,需要上传简历和个人资料:https://t.co/gYiY68itv8

@python_xxt

如果deepseek 能追上fable 5.2 ,价格十分之一,我觉得,大家都洗洗睡吧 期待早点出UBI的政策,否则天下大乱

@0xg4tsby

This is f*cking insane. SpaceXAI engineer Lauren Tan: "250 applications. Two years. Zero offers Then he showed up with 30 agents running his work and walked out with a $600K offer" In this 1-hour workshop, Lauren shows how to build a Grok Bot team that delegates and checks its own work Grok Bot → Chief of Staff → Managers → Workers → An Offer Nobody Else Got The resume didn't change. What he could run did Worth more than any $3,000 AI agent course Bookmark and watch it today Then read what that team looks like in full: 6 Grok bots, 200+ agents ↓

@king1818888

交作业了! 之前聊到那套《2026 AI 工程面试 100 题》,私信和评论区不少朋友在问完整内容。 其实如果只看题目,不仅很难抓住重点,甚至越看越焦虑。所以我花时间把整整 100 道题重新做了工业化梳理,给每一道题都补充了【核心考点剖析】与【生产级排障参考答案】,顺带把 7 道系统大题的架构方案也完整补齐了。 已经整理并开源到 GitHub,大家可以直接自取: 👉 https://t.co/aVGvaeBjvi 这份手册的核心亮点: 1. 真实线上事故全覆盖:覆盖 Prompt 踩坑、RAG 检索、Agent 状态机死循环、微调选型、vLLM 推理排队、生产安全与成本治理 8 大板块; 2. 折叠卡片式设计:每道题点击即可展开工业级题解,非常适合用来做个人自测或团队技术选型查漏补缺; 3. 7 道系统设计大题:包含 5 万人企业级 RAG、10 万单电商客服 Agent、共享推理网关、万圣节数据穿透应急等完整端到端架构。 不管是准备接下来的技术面试,还是正在线上做大模型落地排障,都可以存一份备用。 如果对你有帮助,欢迎去 GitHub 顺手点个 ⭐️ Star 支持一下!

@cyrilXBT

I’m literally pleading with you: drop this into Opus 5.5 before you go to sleep… By morning, it’ll spin up your very first AI agent squad from zero—complete with agents, prompts, and ready-to-run workflows. Grab it, stash it, and shamelessly borrow it now: https://t.co/5SHRRUoJhe

@promptparag

Most developers treat CLAUDE.md like documentation. That’s the mistake. It’s not a README. It’s the control layer for your AI agent. If your outputs feel inconsistent, buggy, or “random”… the problem usually isn’t the model. It’s the instruction architecture. The biggest unlock is understanding scope hierarchy: • Global scope → coding standards, universal behavior • Project scope → stack, workflows, commands • Folder scope → local overrides for specific tasks And yes: the nearest scope overrides everything above it. Another game-changing framework: WHAT → project context WHY → engineering principles & constraints HOW → execution rules & commands Miss one layer and the agent starts guessing. Specificity is what separates reliable agents from chaotic ones. ❌ “Write production-ready code” ✅ “Use TypeScript strict mode, Zod validation, and async error boundaries” ❌ “Add tests” ✅ “Write Vitest unit tests with 80%+ coverage” A few rules that dramatically improve agent performance: Start with /init before customization Keep CLAUDE.md concise (<500 lines) Use hooks for repeatable behaviors Treat it like living infrastructure Reference configs instead of duplicating logic The uncomfortable truth: Most AI agents don’t fail because the model is weak. They fail because the operating instructions are incomplete. Better system design = better AI behavior. The engineers who understand this early will build agents that actually scale. 🚀 📌 Save this (you’ll need it when scaling agents) ♻️ Share with a dev building with AI ➕ Follow for no-fluff AI systems & workflows 🚀

@qingke_ai

投机解码的草稿模型,一直被做成「小号 Transformer」,结果它自己成了新的能耗大户。这块能不能不用神经网络? Olifant 的做法是不用。它用一个基于记忆的分类器做草稿,只看四个 token 的窗口预测下一个,CPU 上 0.3 毫秒以内,不要 GPU、不要反向传播、不占加速器显存。 数字分两种情形看。没有 KV cache 的时候,草稿几乎免费:领域匹配的模型在欧盟法律文本上接受率 38.1%,吞吐 2.43 倍,每 token 能耗低 2.51 倍。但验证端一旦开了 KV cache(本身已经省 10 倍能耗),草稿成本就变成主导项:拿 TinyLlama 当草稿会把吞吐砍半、能耗翻三倍,尽管接受率更高;Olifant 仍然有 1.25 倍吞吐,而且全研究里能耗最低,每 100 token 93 J,比无缓存基线低 14 倍。 他们还发现一件容易被忽略的事:KV caching 把「时间」和「能耗」解耦了。验证更长的草稿会提高功率,于是能耗最优的草稿长度会低于速度最优的那个长度。以前调草稿长度只看吞吐,现在得同时看功率。 边界:记忆分类器的接受率依赖领域匹配,换领域要重训;只在消费级硬件上验证;如果服务栈瓶颈不在算力或能耗,收益会打折。 https://t.co/MnRyqQbgHT

@DailyDoseOfDS_

5 LLM fine-tuning techniques, explained with visuals: Traditional fine‑tuning is impractical for LLMs (billions of params; 100s of GB). Since this kind of compute isn't accessible to everyone, parameter-efficient fine-tuning techniques (PEFT) were built. Before we go into details of each technique, here's some background that will help you better understand these techniques: LLM weights are matrices of numbers adjusted during fine-tuning. Most PEFT techniques involve finding a lower-rank adaptation of these matrices, a smaller-dimensional matrix that can still represent the information stored in the original. Now with a basic understanding of the rank of a matrix, we're in a good position to understand the different fine-tuning techniques. (refer to the image below for a visual explanation of each technique) 1) LoRA - Add two low-rank trainable matrices, A and B, alongside the weight matrices. - Instead of fine-tuning W, adjust the updates in these low-rank matrices. Even for the largest LLMs, LoRA matrices take up a few MBs of memory. 2) LoRA-FA While LoRA significantly decreases the total trainable parameters, it requires substantial activation memory to update the low-rank weights. LoRA-FA (FA stands for Frozen-A) freezes matrix A and only updates matrix B. 3) VeRA - In LoRA, low-rank matrices A and B are unique for each layer. - In VeRA, A and B are frozen, random, and shared across all layers. - Instead, it learns layer-specific scaling VECTORS (b and d) instead. 4) Delta-LoRA - It tunes the matrix W as well, but not in the traditional way. - Here, the difference (or delta) between the product of matrices A and B in two consecutive training steps is added to W. 5) LoRA+ - In LoRA, both matrices A and B are updated with the same learning rate. - Authors of LoRA+ found that setting a higher learning rate for matrix B results in better convergence.

@Ishh_021

Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture &amp; released it for free. Bookmark &amp; watch this today before someone takes it down and read this article below. https://t.co/oAL21DRhSx

@Ryrenz

🏫 想把 AI agent 真正做上线的看这里,Agentic System Course 用 22 章讲透生产级 agent 怎么设计、搭建和运维。 Agentic System Course 的 fork 已经接近一百个,而且整套课是写给你和你的 AI 一起读的,仓库里同时放了 AGENTS.md 和 CLAUDE.md。 以前学 agent 开发,要么跟着某个框架的教程一步步敲,换个技术栈就全废;要么零散看博客,记忆、多 agent 协作、成本控制各讲各的。 这门课只给骨架:把记忆与状态、多 agent 协调、外部接口、可观测性、成本与延迟、安全运维这些承重的主题和取舍讲清楚,具体代码交给你的 AI 按你的技术栈去补。最后一章是一张设计画布,直接拿来设计你自己的 agent。 还有个顺手的:仓库自带一个评审 skill,把你的 PRD、设计文档或 agent 代码丢给它,它先判断你做的是个人项目、团队工具还是面向客户的产品,再挑相关章节出报告,每条问题都带严重程度、证据、章节出处和修复建议。课程拿 OpenCode、Hermes Agent、OpenClaw、Paperclip 四个开源项目当参考实现。 用 agent 学 agent,这个思路挺对味。 GitHub:https://t.co/9mBRsegE03

@zhouluobo

上次写了100个小技巧基础版,然后大家都说想要快点看到进阶版,现在来啦。 https://t.co/fPh36pmIu0

@9hills

一篇论文了解决策模型的训练《Towards Foundation Decision Models: Drex》 决策模型在很多场景比如分类、路由等任务上很有用,没有当时 Jev 刚出来吹的那么神,但是可以落地应用。 和传统的 BERT 比起来就是用 LLM 做backbone,知识容量和能力大幅提升。 https://t.co/yvqvbiRyfl

@AlchainHust

读了DeepSeek的Super-agent和Anthropic的Building always-on agents两篇,照着后一篇在FanBox里做了一版:终端里的agent停下来等你,而你不在电脑前,手机(微信或飞书)会收到它的一句话。两篇文章讲了什么、我们照着做踩到的四件事,都写在里面了。 https://t.co/BqA2aTsPsq

@RealMarvelX

claude code becomes a lot more powerful when you give it the right tools. these 10 open-source repos turn it into a complete development ops team that can learn your codebase, remember past sessions, monitor usage, automate browsers, and manage agents. here’s the toolkit: learn 1. skills : standardizes how Claude Code discovers and uses skills. 192.5K GitHub stars. https://t.co/krfzAIOH90 2. claude-code-best-practice : maps Claude Code features, subagents, and hooks. https://t.co/bg8G5hoIMY 3. Claude-Code-Quest : teaches Claude Code through three levels of graded exercises. https://t.co/NNlLiJL09x remember 4. claude-mem : carries useful context from one session into the next. https://t.co/VbIZ6rM1kV 5. deja-vu : indexes transcripts from 43 agents and helps hand sessions between them. https://t.co/z9dZjnPpZ2 watch 6. ccusage : tracks token usage and costs from local logs. https://t.co/vmFMBNNE2a 7. ccstatusline : displays your model, branch, and token usage live. https://t.co/wx9FKiYALM run 8. agent-browser : lets agents control a real browser. https://t.co/n30RylYlBJ 9. opcode : manages Claude Code sessions and agents from one interface. https://t.co/6EUdIfAneC install 10. skills : installs skills across Claude Code and other AI agents. https://t.co/p6ceavarBS one important detail before using them commercially: Claude-Code-Quest uses PolyForm Noncommercial, while opcode uses AGPL-3.0. check both licenses before shipping them in a commercial product. learn → remember → watch → run → install. 10 repos. one much more capable Claude Code workflow. 🔖 save this toolkit. you’ll want it when building your next project.

@leo114119

1 分钱白嫖的百度搭子真的可以喔,反代出来的 Deepseek v4.1 Flash 居然是满血的? 用magpie反代出来V4.1 1017的flash,太快了 想去抢的晚上 20 点还有: https://t.co/1E9mfLUshg https://t.co/VXyJNEyHM1

@flohoeller

One agent run as a sequence diagram https://t.co/Wezr09uKr3

@xoyif63862452

我去!让 Agent 改你的笔记库,最容易翻车的往往不是写内容,而是它根本不懂双链、Callout 和 Canvas。 obsidianmd/obsidian-skills,MIT,GitHub 约 5.2 万星。不是第三方合集,是 Obsidian 官方自己维护的 Agent Skill 包。 以前让模型碰库,只能当普通 Markdown 硬写,双链、Bases、画布经常写歪。 现在按 Agent Skills 标准装上,Claude Code、Codex、Cursor、OpenCode 都能直接用官方语法干活。 核心能力: 按 Obsidian Flavored Markdown 写双链、嵌入、Callout 和属性 创建和编辑 Bases(.base),带视图、筛选、公式和汇总 编写 JSON Canvas(.canvas)的节点、连线和分组 通过官方 CLI 操作库,也覆盖插件和主题开发 用 Obsidian Headless 在命令行同步、发布库,不用开桌面端 用 Defuddle 把网页抽成干净 Markdown,少浪费 token 用 Knap 从 JSON 或 CSV 批量渲染 Markdown 模板 真正有用的是:它把「能打开库」升级成了「按官方格式正确地改库」。 在 Obsidian 里做知识库、写插件,或想让 Agent 代管笔记的,值得直接装上试。 https://t.co/2sHvlI4xsb

@sripathiteja4

52 websites worth more than most college degrees: 1. Coursera. org – University courses completely free to audit 2. Brilliant. org – Interactive math and science learning 3. Wolfram Alpha – Answers any mathematical or factual question 4. GitHub. com – Learn coding from real world projects 5. Investopedia. com – Finance and investing explained simply 6. Archive. org – Access millions of free books and old websites 7. Project Gutenberg – 70000 free classic books 8. Duolingo. com – Learn any language for free 9. Notion. so – Organise your entire life and learning 10. Our World in Data – Every global statistic visualised 11. Statista. com – Data and statistics on everything 12. OpenLibrary. org – Borrow millions of books online free 13. Hemingwayapp. com – Write clearer and simpler instantly 14. NASA. gov – Space science and research for free 15. PubMed. gov – Access real scientific research papers 16. Edx. org – Free courses from Harvard MIT and more 17. TED. com – Best ideas from the world's best thinkers 18. Anki – The most powerful memory tool ever built 19. Canva. com – Design anything without being a designer 20. Skillshare. com – Creative and practical skill learning 21. Readwise. io – Remember everything you ever read 22. Google Scholar – Search real academic papers 23. Codecademy. com – Learn to code completely free 24. ChatGPT – AI tutor available 24 hours a day 25. Figma. com – Learn professional design for free 26. Replit. com – Code anything from your browser 27. Huberman Lab Podcast – Science based health education 28. Mindmeister. com – Mind mapping for better thinking 29. NerdWallet. com – Personal finance made simple 30. Quizlet. com – Study smarter with flashcards 31. Gapminder. org – See the real state of the world 32. PhET Simulations – Interactive science experiments online 33. Numbeo. com – Cost of living data for every city on earth 34. 23andMe. com – Understand your own genetics 35. Zapier. com – Automate your work without coding 36. Lesswrong. com – Deep rational thinking and decision making 37. Documentaryheaven. com – Thousands of free documentaries 38. Trading Economics – Economic data for every country 39. Perplexity. ai – AI powered research tool 40. Stanford Encyclopedia of Philosophy – Every philosophical idea explained 41. Librivox. org – Free audiobooks of classic literature 42. Zooniverse. org – Participate in real scientific research 43. Futurelearn. com – Free short courses from top universities 44. Typing. com – Learn to type properly and fast 45. Drawabox. com – Learn to draw from absolute scratch 46. Grammarly. com – Write better in every situation 47. Khanacademy. org – Free world class education for everyone 48. Desmos. com – The most powerful free graphing calculator 49. Stellarium. org – Explore the night sky from your screen 50. Psychologytoday. com – Mental health and psychology explained 51. Worldometers. info – Real time global statistics on everything 52. Notion. so/ templates – Free templates to organise your entire life Follow @sripathiteja4 for more

@ASI_8866

Day 13/ 公益中转站https://t.co/756yBvAkBZ上线的第2个周六,今天是听劝帖,大家可以提提目前平台存在的问题~ 评论站点用户名继续领鸡蛋,Grok Bot 在线给大家赠送额度。同时,对提出平台问题的老板,我们会额外赠送额度~ TokenFuture官方QQ群里的老板们每天都在分享自己使用AI的经验,大家感兴趣可以扫码进群一起讨论哦~

@zodchiii

STOP WASTING HOURS TRYING TO LEARN FABLE 5.2. I put together one practical roadmap with official docs, videos, Claude Code setup, GitHub repos, engineering deep dives and courses. Start here👇 1. Fable 5.2 official page https://t.co/P2ztWXlpl6 2. What's new in Fable 5.2 https://t.co/om5qJsbzD2 3. Fable 5.2 overview https://t.co/5wsurBSkh9 4. Migration guide https://t.co/5FAnYC6vID 5. Prompting Claude Fable 5.2 https://t.co/OkJi28K8Hx 6. Effort https://t.co/fL65TNJt2N 7. Pricing https://t.co/mNx1mYFBTq Model ID: claude-fable-5-2 GITHUB REPOS: 1. anthropics/claude-code https://t.co/IcZSEP81Lm 2. anthropics/claude-cookbooks https://t.co/OycSJH60Iq 3. anthropics/skills https://t.co/QBDhmyGi9H 4. anthropics/claude-agent-sdk-python https://t.co/e9urlxmzN5 5. anthropics/claude-quickstarts https://t.co/aasswVCD8H 6. anthropics/courses https://t.co/4kEtcN1p1T 7. anthropics/prompt-eng-interactive-tutorial https://t.co/0x6om0xmSU 8. anthropics/claude-code-action https://t.co/RPIjadCv58 9. modelcontextprotocol/servers https://t.co/ggxNi1TTDp 10. hesreallyhim/awesome-claude-code https://t.co/6yoxioLEhN WATCH & FOLLOW ALONG: 1. Fable 5.2 or Opus? When to use each model in Claude Code https://t.co/Pv6ZMkz64t 2. Your Complete Guide to Fable 5.2 https://t.co/kW8uwMxImX 3. Claude Fable 5.2 Is Seriously Impressive https://t.co/gavXfOjK2f 4. How to Use Claude Fable 5.2 https://t.co/vrAyt0FxWw CLAUDE CODE SETUP: 1. Model configuration https://t.co/uExvJYqBiA 2. Subagents https://t.co/pVjJHqcaWE 3. Advisor https://t.co/JkksTRwGSP 4. Advisor tool (API) https://t.co/LdDubP3Tvh 5. Fast mode https://t.co/TEZPflmvF6 6. Preserved thinking https://t.co/VkFJ3YlCxA 7. Refusals and fallback https://t.co/k8etNq97no ENGINEERING DEEP DIVES: 1. Effective context engineering for AI agents https://t.co/IdvZz8uoKO 2. Effective harnesses for long-running agents https://t.co/1bWnntjNdB 3. Harness design for long-running application development https://t.co/Dn9MH1NmXp 4. Building a C compiler with a team of parallel Claudes https://t.co/uMd0sAuVVL 5. Writing effective tools for agents https://t.co/5Ii4j2LSZz 6. The components of agent memory https://t.co/V0l24EJQJA 7. Demystifying evals for AI agents https://t.co/C82TJY9DhI ANTHROPIC ACADEMY COURSES: 1. Claude Code 101 https://t.co/qUpTdSj58C 2. Claude Code in Action https://t.co/gyEnxVKoLg 3. Claude Platform 101 https://t.co/HOC6lYeJM8 4. Building with the Claude API https://t.co/7GQ7YJqdAa 5. Introduction to Model Context Protocol https://t.co/csf0tWViMf 6. Introduction to agent skills https://t.co/c0i5sIvwXl 7. Introduction to subagents https://t.co/K6vMLbTyFg LAUNCH COVERAGE: 1. CNBC https://t.co/u1DwpVr6Ii 2. The New Stack https://t.co/H2erbtO21h 3. Android Authority https://t.co/Vs6SJY3IQ6 4. MacRumors https://t.co/FvqWXJNaWk BEST ORDER TO LEARN: what's new → set the model in Claude Code → prompting guide → effort levels → subagents and advisor → Claude Code in Action → then the engineering deep dives

@Dipanshu_AI

Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture &amp; released it for free. Bookmark &amp; watch this today before someone takes it down and read this article below. https://t.co/2GMc3una96

@HeyShobhan

They built an AI agent in 20 minutes and sold it to Anthropic for $100k - and at Stanford showed how to repeat it from scratch: 02:00 - how to build your first agent with Opus 5 in 20 minutes 22:34 - one agent replaces an Anthropic engineering team worth $200k a year 49:47 - how to make $1.2M a year selling agents to Anthropic and Google after watching I spent 20 minutes in Claude Code - the agent replaced 10 Anthropic engineers and they paid $100k and asked for more. save & watch - the article below shows how to build an agent in Claude Code in 20 minutes and reach $1.2M a year.

@zainhas

how do you train a frontier local coding llm? tune in to learn! speaking to Dan while prepping for this and I think it'll be a spicy one &gt; evals, architecture choices, dense vs moe, scaling experiments and alot more https://t.co/4B3AQbO0Y7

@trending_repos

Trending repository of the day 📈 modern-web-guidance-src Keep your coding agent up to date with the latest web best practices Last 24h: 1,225 ⭐ Total: 5,327 ⭐️ https://t.co/06sPm66EpZ

@FellMentKE

Most AI knowledge tools rely on embeddings, vector databases, or model calls to index and process your files. SAG Wiki takes a different approach. Powered by SAG2, it's a free, open-source local knowledge base that turns your existing folders into searchable knowledge without requiring embeddings, a vector database, or model calls during indexing. You can simply add an existing folder from your computer and turn it into a searchable knowledge base. As your files change, SAG Wiki supports incremental updates, so you don't have to re-index everything from scratch. What's particularly interesting is how this works with AI coding agents like Claude Code and Codex. They can retrieve relevant information from your existing documents, connect supporting evidence, and use that context for research, writing, and development. Imagine having an AI coding agent search through your project documentation or research materials to find the exact information it needs, instead of manually digging through folders. Your files stay exactly where they are, with citations linking back to the original sources. No need to move files, deploy another database, or pay model costs just to index your knowledge. A practical way to make your existing files useful to AI agents. Explore SAG Wiki on GitHub: https://t.co/XWfEu5iy90

@xoyif63862452

好家伙,让 Agent 剪个视频,它要么瞎猜 ffmpeg 参数一跑就报错,要么让你把片子传到云上,本地那堆素材它压根碰不着。 这个开源 Skill 给编程 Agent 塞了 42 个本地 FFmpeg 工具,裁切、加字幕、改画幅、调响度都有,先探测文件再动刀,剪完还自己核对一遍,Claude Code、Cursor、Codex 都能接,不走云也不要 API Key。 ffmpeg 那串参数这辈子是记不住了,交给它刚好。 https://t.co/V2jyimHdhc

@clcoding

AI Engineer Roadmap 2026: From Python to Production in 16 Weeks! Want to become an AI Engineer but don't know where to start? Here’s a structured 16-week roadmap to help you go from Python fundamentals to building and deploying real-world AI applications. KEY TECHNOLOGIES Python | Pandas | Scikit-learn | Hugging Face | LangChain | LangGraph | FastAPI | Docker | GitHub BONUS: Recommended Courses & Learning Resources Explore these courses from CLCODING to complement your AI Engineer journey: 1️⃣ Python for AI Bootcamp 2026 (15 Days) https://t.co/c3TwS0Bo5X 2️⃣ Machine Learning with Scikit-learn, PyTorch & Hugging Face https://t.co/gbpBIm4Gnx 3️⃣ IBM RAG and Agentic AI Professional Certificate https://t.co/YNzp0HtaVh 4️⃣ AI Agents with LangGraph, Semantic Kernel & AutoGen https://t.co/U3FyBkrNea 5️⃣ Building AI Intensive Python Applications https://t.co/KPRKJcPsmr The ultimate goal: Don't just learn AI tools. Build, evaluate and deploy real-world AI applications. Save this roadmap. Share it with someone learning AI Engineering.

@KirkDBorne

System Design for the LLM Era — Patterns and Principles for Production-grade AI Architecture: https://t.co/fC3z5rD81B https://t.co/ZhYyYFkONa

@BIGMayrr

WHAT IF YOU COULD TURN A SINGLE AI MODEL INTO PLAYABLE GAMES, INTERACTIVE 3D WORLDS, AND FULLY FUNCTIONAL WEBSITES? This GitHub repository is collecting some of the most impressive creations built with Claude Fable 5.2. It's called Awesome Fable 5.2, and it brings together real projects you can explore, open, and play directly in your browser. Here's what makes it interesting: → AI-generated games: Explore browser-based games, including racing games, first-person shooters, RPGs, and interactive experiences. → 3D worlds and scenes: Discover immersive environments, realistic simulations, virtual cities, and interactive 3D experiences. → Web pages and tools: Explore websites, creative tools, generators, and interactive applications built with AI. → Animations and films: See what creators are producing with AI-generated animations and visual storytelling. → Live project demos: Browse working projects with GIF previews, video demonstrations, and links to the original creators. → Creator credits: Every entry links back to the person who built it, making it easier to discover new creators and their work. The interesting part? This isn't just a collection of AI-generated screenshots. You can actually open and interact with many of these creations yourself. The repository currently features 151 live projects and 111 video demos, with new creations added regularly. If you're exploring what's possible with AI-powered development, this is worth bookmarking. GitHub: https://t.co/zosGcm9xgf

@grgerwcwetwet

卧槽!GitHub 上一个 1.58 万 Star 的项目,居然不用视频生成模型,就能让 AI 做出完整短片! 推荐开源项目 Lemo-Opuscar,作者用 Claude Opus 5.5,纯靠代码制作了 99 种不同风格的影片。 水墨画、像素动画、复古胶片、3D 微缩景观、赛博霓虹、MG 动画……足足 15 大类,每种都有成片演示和现成的风格提示词。 最厉害的是,装进 Claude Code 后,选个风格、说出主题,AI 就能自动编排画面、动画、配音和音乐,最后渲染成视频! 做 AI 短片、科普视频、品牌宣传片的,建议收藏! 在线画廊:⁠https://t.co/2iVnNHRm8r GitHub:⁠https://t.co/ei4bytsMBC

@Ownerthoughts

Grok Bot: AI agents that work while you sleep https://t.co/PdAfoyOWQb

@geniusvczh

还有个终极办法,就算人真的在美国,claude一看你的项目都是国内的项目,中国公司的美国分舵没跑了🤪

@0xqinshi

Holy f*ck. Google just dropped a free 2-hour course on building AI agent teams From one prompt to 300+ agents that plan, remember and fix their own work: 0% → 03:09 - Multi-agent systems from scratch: the full blueprint 30% → 38:06 - Agent graphs: run agents in sequence and in parallel 45% → 51:10 - Loop engineering: send bad work back until it passes 65% → 1:20:44 - One orchestrator picks the right agent for every job 100% → 1:57:15 - A self-running agent team goes live and wins on its own Prompt → Agents → Loops → Graphs → Self-Improving Teams Most people build one agent, watch it break and go back to the chat window Single agents were the old workflow Agent teams that run themselves are the new one Worth more than any $3,500 agentic engineering course Save it and watch it tonight Then grab the 8 Grok Bots a SpaceXAI engineer runs his life on, ready to paste, below ↓

@DrRabbitAI

Codex 除了写代码,也是一个能动手干活的科研助手。从查文献、分析数据到做答辩 PPT,我把整套流程和能直接复制的提示词都整理在这里了 https://t.co/KfYfFkg65g

@AIExplorerTim

未来五年,想活得更好,这 10 条建议记牢了。 1,每天花 2 小时研究 AI、国内互联网和赚钱机会,别刷抖音只为了打发时间。 2,英语必须学。至少能看懂国外网站和官方文档,拿到一手信息。 3,趁早准备自己的数字身份:常用邮箱、云账号、备用手机号,越早弄越好。 4,学会自己找信息和工具,别只靠一个渠道,至少留两套备用方案。 5,打通支付渠道:国内信用卡、常用支付工具都配齐,别让付款卡住机会。 6,用好 DeepSeek、豆包、通义、Kimi、文心,别只拿 AI 当聊天工具。 7,建立自己的信息渠道:微博、GitHub、小红书、B 站,少嚼别人消化过的二手信息。 8,争取赚到第一笔副业收入。内容创作、数字产品、远程服务,先把流程跑通。 9,做个人 IP。持续输出内容、积累信任,别只靠一份工资。 10,分散风险。别让一个账号、一个平台、一份工作决定你的命运。

@Gyome1_

this harness engineering paper is f*cking insane it’s the exact setup researchers used to make a coding agent beat Codex by improving itself, without touching the model. 3 agents, 7 kinds of files, one prediction before every edit, all in git. set it up tonight and your agent keeps getting better round after round while you sleep, instead of you reading logs by hand. the crazy part is everyone rewrites the prompt when their agent fails, while the fix is the tools and the memory. full breakdown in the harness guide below.

@shao__meng

Matt Pocock 预告 /animatic Skill:让 Agent 把重要 PR 做成带旁白的分镜演示短片 介绍了两个 Skills:/setup-animatic 负责选好工具,并在本机或远程机器上完成配置;/animatic 负责生成可以直接上传的视频。 用一张张静态帧加旁白讲清一条 PR Matt 要解决的问题:Agent 写 PR 的速度已经超过他审 PR 的速度,大段 diff 也读得很慢。他的办法不是全程录屏,而是做成 animatic,也就是分镜短片:每一帧是一张静态画面,配一句旁白,由 Agent 做好后贴到 PR 上。 整个流程由一条构建命令串起来: · 先写一份 JSON 格式的分镜脚本,每一拍对应一句台词和一帧画面。 · 用 Playwright 渲染画面,代码直接取自 diff,并保留真实行号。 · 用选定的语音逐句朗读。 · 最后用 ffmpeg 合成 MP4,并检查每一帧画面和音频。 视频生成后,按仓库里的配置贴回 PR,GitHub、GitLab 或本地目录都可以。 生成要多久,他给了几个说法 预览里的例子,是把他 Skills 仓库中“把 Skills 发布到 npm”的一条 PR 做成分镜。Agent 做完这一支大约用了 90 秒。 同一天稍早,他贴过一张优化前后的对照表:总耗时从 397 秒降到 96 秒,模型耗时从 244 秒降到 85 秒,对话轮次从 27 轮降到 12 轮,费用从 1.38 美元降到 0.49 美元。每轮的脚本处理时间也从约 50 秒降到约 6 秒,而且改成了逐帧检查。后来有人问平均要多久,他回答大约 4 分钟,并说即使再加入 Playwright 截的图或录的视频也不会太久。 工具只问一次,选择存进文档 /setup-animatic 会逐项询问用什么工具,每一项只问一次: · 语音可以选 ElevenLabs、OpenAI 或 edge-tts,会先试读一句,让人听过再选。 · 渲染用自带的工具包,或者自己的工具。 · 样式用默认的一套,或者自己写的 CSS。 · 拼接用 ffmpeg,或者自己的剪辑软件。 选好后会写进 docs/agents/animatic.md。他预览里用的是 edge-tts 的 Andrew 声音。Matt 说声音可以换,这类短片多半会开两倍速看;而时延和 token 开销是大问题,所以要尽量保持简单。 有人问什么算重要 PR,他只回了一句 “One-way doors”,意思是合并后很难撤回的改动。

@ashlleymo

Tried @HigurashiAI Haru-4B! Cool things abt this model: 1. This may be the most natural full- duplex japanese model. The talking behavior in japanese is incredibly different from english. E.g. people backchannel wayyy more in japanese. My co-founder who grew up in Japan tried this model and said she could have an actual non-weird convo with it (at least for a short time because it's not very smart yet). 2. Hello...the local latency?? 65 ms end-to-end latency on a MacBook? they fully reduced context length to a third by processing the user audio, outgoing model audio, and model text together at each token position instead of one after another. 3. they intentionally excluded LLM-generated dialogues and TTS-synthesized speech. Real data is king adding a smart LLM could create a product that is really the best of both naturalness and intelligence. Imagine if GPT-live-1 had this language-specific naturalness (despite it being generally super natural already)...

@Asteri_eth

you don't need an editor, a voice actor or a motion designer anymore. you need these 20 repos motion, voice, generation and editing. all open source, all drivable from claude code on opus 5.5, ~900K github stars between them: MOTION ┣ remotion · videos as React code, rendered to MP4 (73K★) ┃ ┗ https://t.co/2MnpZQrLlw ┣ manim · math explainer animations in Python (42K★) ┃ ┗ https://t.co/CNT7hbt3Nn ┣ motion-canvas · code animation with a live preview editor (20K★) ┃ ┗ https://t.co/rWQTitvtlP ┣ theatre · keyframe timeline + studio UI for web animation (13K★) ┃ ┗ https://t.co/6sGca8iPU8 ┗ revideo · Motion Canvas fork that renders videos from an API (4.2K★) ┗ https://t.co/H4pVwKhUGm VOICE ┣ whisper · speech to text in ~100 languages, subtitles out (113K★) ┃ ┗ https://t.co/MKcw1tGbyN ┣ fish-speech · clones a voice from a short sample (34K★) ┃ ┗ https://t.co/TcJGeIBVg4 ┣ chatterbox · open TTS with a dial for how dramatic it sounds (28K★) ┃ ┗ https://t.co/qn2UiwH2t2 ┣ whisperX · word-level timestamps for word-by-word captions (25K★) ┃ ┗ https://t.co/ZkFeozwQ3f ┗ kokoro · tiny 82M voice model that runs on a laptop (10K★) ┗ https://t.co/DD68TzRfu1 GENERATE ┣ ComfyUI · node graph that runs image + video models locally (147K★) ┃ ┗ https://t.co/RScadeoDzr ┣ Wan2.2 · open text + image to video model (19K★) ┃ ┗ https://t.co/ubvFQmdrMO ┣ LivePortrait · animates one photo with a driving video (19K★) ┃ ┗ https://t.co/iYRl4vRqvd ┣ HunyuanVideo · Tencent's open 13B video model (13K★) ┃ ┗ https://t.co/C2fD1j2g4S ┗ LTX-Video · fast open video model from Lightricks (11K★) ┗ https://t.co/Lt3b9z67U3 EDIT ┣ yt-dlp · pulls source footage from 1000+ sites (208K★) ┃ ┗ https://t.co/bRl1Wys2Qb ┣ FFmpeg · the export desk under every tool on this list (67K★) ┃ ┗ https://t.co/bfqyckTvNL ┣ lossless-cut · trims and splits video without re-encoding (46K★) ┃ ┗ https://t.co/eaWUqXJyDM ┣ AI-Youtube-Shorts-Generator · long video in, vertical shorts out (6.1K★) ┃ ┗ https://t.co/MZykJJpJuX ┗ auto-editor · cuts the silent parts out of your footage (5.9K★) ┗ https://t.co/sltG0QeKgy all free and open source. the generate block wants a real GPU, the rest is much lighter bookmark this, then build your studio tonight ↓

@Vikram_AI_

AI can write SQL. But can it actually understand your data? 👀 That's where things get interesting. Exmergo built Dex to help AI coding agents like Claude Code, Cursor, and Codex understand your data warehouse instead of blindly generating SQL. Think of it as giving your AI agents the context they actually need. → Understand tables and relationships. → Catch expensive queries before they run. → Build dbt models, tests, and semantic definitions. → Detect schema and data changes. → Define metrics and how data should be interpreted. The interesting part? Dex Semantics scored 80% on ADE-bench with Claude Sonnet 5.5, compared to 73.3% without Dex. That's 25% fewer failed tasks. 🔥 And it's open source! Instead of letting AI guess what your data means, give it the right context to get things done. Worth checking out! 👀 Support Dex on Product Hunt (it's live ): https://t.co/hLST8upZSx #AI #AIAgents #SQL #DataEngineering #OpenSource

@Khazix0918

跟Claude对话,Claude做规划和前端,派活给GPT写后端和数据处理,然后Claude收尾,Codex启动最多三轮审查,可能就是目前Agent自动化开发的最优解了。 所以我的CLAUDE.md里一直有这么一段,用起来还是蛮舒适的(codex-worker是我自己定义的脚本,没用codex的官方插件,官方插件有点麻烦,问题也有点多),给大家分享一下~ ## 派活 - 成块的后端、算法、数据处理和性能活(不是一两行的修补)交给本机的 Codex(GPT-6.1 Sol)写,你定接口和验收标准;前端、视觉、文案、拆解和验收自己做。 - 用 `codex-worker <工作树> "任务"` 派,追加用 `codex-worker --resume <会话ID> <工作树> "追加"`,放后台跑,会话 ID 在输出最后。 - 它只能写工作树和临时目录,提交不了、联不了网:装依赖、联网测试、提交推送由你做。它只看得到任务和工作树,接口、验收标准、要一起改的同类都写进任务。 - Codex 交回后你读 diff、自己跑测试(合并前的审查也是 GPT,你这步才是换家模型的交叉检查)。PR 里有 Codex 写的代码就是混合来源,交付不带 `--agent`。

@canghe

太牛了,Claude Fable 5.2 只花了 15 分钟就复刻出了个简易版本的英雄联盟,英雄技能和小兵都是完全可用的,使用的Three.js + TypeScript + Vite技术,BLG 加油。 https://t.co/BKk6NWMoml

@de1lymoon

6 GROK BOTS CAN REPLACE AN ENTIRE SOCIAL MEDIA TEAM FOR JUST $20 A MONTH and you don't need to build the entire operation from scratch here are 10 GitHub projects to build your own AI-powered social media agency: 1. gpt-researcher → researches any topic across the web and writes a report with sources https://t.co/I2GsISQT20 2. crawl4ai → turns any web page into clean markdown your bot can read https://t.co/L0xgXQGMDQ 3. social-media-agent → langchain's agent that drafts posts for X and LinkedIn, you approve before anything goes out https://t.co/6h4wAI1oaE 4. ComfyUI → node-based image and video generation for covers and thumbnails https://t.co/en4fKKRwUv 5. remotion → write videos in React, render them to MP4 https://t.co/ku4x8gHyFs 6. MoneyPrinterTurbo → give it a topic, get a short video with script, voiceover, subtitles and music https://t.co/BPomWNGf5m 7. postiz → schedules posts to 14 platforms and lists Grok Bot as a supported agent https://t.co/DvK6dFcQpb 8. n8n → connects it all: trigger, bot, post, on a schedule https://t.co/m42JUP1kBp 9. chatwoot → DMs and replies in one inbox https://t.co/kcr7ZNlAau 10. umami → privacy-friendly analytics, shows which posts actually bring clicks https://t.co/tqo3xc1SAs hire one bot at a time and keep the ones that pay for themselves

@awe0_io

Claude Code 最大的问题之一,是有时候太积极了。 你让它改点代码,它可能顺手执行一堆命令。 claude-code-hooks-mastery 开源了完整的 Hooks 示例。 能在工具执行前检查危险操作、拦截敏感文件访问,还能自动记录每次调用。 GitHub:https://t.co/5zU21nSLM3

@xluchuan

DeepSeek 还有特供版? 为什么这个比赛用的是 DeepSeek-V4.1-Flash-1114 https://t.co/kRh9Fuj4mk

@Sirofsky

Sometimes I build shit that I'm genuinely embarrassed to show anyone 😂 So why not share a little more of what building actually looks like? The messy bits, the experiments, the failures and, eventually, the finished product. A few reminders: 🔍 Keep exploring GitHub repos. 🛠️ Make use of free tools and resources. 🚀 Keep experimenting and learning. 🧠 And most importantly, make your own decisions instead of asking AI to choose everything for you. AI is a tool, not a replacement for your own thinking. Keep building. Keep learning. Keep making shit. 💪 https://t.co/kvHMnJKhAJ coming soon!

@techniahqrobot

56 magnetic sensors inside an open-source robotic hand that can recognize objects through touch. This is D’Manus developed by researchers at Carnegie Mellon University and FAIR at Meta AI. The hand has three fingers 10 degrees of freedom and a soft ReSkin surface packed with sensors. There are eight magnetometers in each fingertip and another 32 across the palm.

@kaddisdeployed

ANOTHER FREE AI MODEL TO TRY Baidu’s ERNIE 5.5 Flash Preview is now available for free on OpenRouter model: baidu/ernie-5.5-flash-preview:free all you need is an OpenRouter API key to get started it’s still in preview, so limits may change, but it’s worth testing while it’s free the list of free models just keeps getting longer 🔗 https://t.co/2RpRrlabp7

@panExorcist

DeepSeek周五晚上悄悄干了件大事:把App里几乎所有能力都拆成了API。 联网搜索、深度思考搜索、文件解析、链接读取,再加上语音、OCR、ASR、TTS,全部开放给开发者调用。顺手还开源了一个聊天前端demo,能直接导入你在DeepSeek的历史对话。 这件事值得多看两眼,因为它暴露了路线选择: 1.不做超级App,做能力层。别家拼命把用户留在自己的App里,DeepSeek选择把零件卖给所有人。 2.开发者门槛又降了一截。以前做一个能搜、能读文件、能说话的AI助手,要拼好几家服务,现在一家就能凑齐。 3.压力给到中间层。那些靠套壳搜索加文件解析赚钱的小产品,护城河会被直接填平。 我的看法:模型本身越来越像水电煤,真正的竞争会转到谁的场景更深、谁的数据更独家。对普通人来说,现在正是用低成本搭自己工具的好时候。 你会先拿这套API做什么?

@Market_Mind_

How to build an AI Agent https://t.co/voSArxpvJP

@HeyShobhan

Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture &amp; released it for free. Bookmark &amp; watch this today before someone takes it down and read this article below. https://t.co/zPKFt9IIW6

@leo114119

感觉 Claude Fable 5.2 的 3D 建模能力并不是很强 这个 3D 鹈鹕真的挺一般的,哈哈 https://t.co/6oVKIliiY3

@SVEN_WAG

看到这个结果第一反应是抓了一波短期行情撞上好运。 看起来像是一次漂亮的资金翻倍,其实不是。有人给 Claude 分配了 200 美元,开放电脑完整权限,只下达一个目标:在 #Polymarket 里面盈利。 预测市场:https://t.co/33sNcTHg1M 它不看新闻、不刷推文、不参考外部观点,先去复盘平台上持续盈利的钱包,筛选出真正具备长期正收益的账户,再沿着这套范式自动执行。十小时内,资金从 200 美元走到 3000 美元。 它吃到的优势,是人类交易者很难稳定守住的部分:没有恐惧、没有贪婪,不会因为亏损迟疑,也不会因为盈利上头加仓,只持续扫描数据、识别市场里的定价偏差,然后机械完成执行。 大部分人还停留在把 AI 当成聊天工具,而少数人已经把它做成全天候在岗的交易载体。 绝大多数人难以复刻,卡点不在于找不到模型,而是很难剥离自身情绪,去接受这种单调、不带主观预判的执行模式。

@jithinvraj

10 Ways to Fine-Tune an LLM (Only for Reference &amp; Research) https://t.co/55eXkE5nyz

@simplifyinAI

This is UNREAL... NVIDIA got one AI model to search 21 million Wikipedia passages and write the answer. Most RAG setups use a separate search model. UNREAL lets the LLM pick passages using its own internal signals. It adds fewer than 500K trainable parameters and keeps the original model weights frozen. On one benchmark, it found every passage needed in its top 10 results for 73.2% of questions. The strongest baseline hit 49.1%. Those same signals helped filter out distracting passages in long prompts. So far, it’s been tested on Wikipedia. If it works on other texts, one model could handle both search and answers.

@Gorden_Sun

DeepSeek把自家产品也完全开源了 DeepSeek开源了自家官网用到的几乎全部API,包括搜索、读取文件、读取链接、语音识别、生成语音等,同时开源了使用这些API搭建的网页Demo。 任何人现在都能使用这些API搭建一个Agent产品了,而且API价格实惠,童叟无欺。 Github:https://t.co/VdmfE6hxwG API文档:https://t.co/5yqfVRr3uj

@Alex_AI96

Looking for useful resources to experiment with AI coding agents? Take a look at Oxylabs Web API Skills. The GitHub repository provides tools that help coding agents search the web, retrieve webpage content, and gather online information. Explore the project and see how it fits your workflow. https://t.co/yXE4IJp5H9 #OpenSource #AIAgents

@NextShopiaAi

🚀 100 AI tools worth knowing in 2026: ChatGPT — AI assistant Claude — Writing & reasoning Gemini — Google AI Grok — Research & real-time AI Perplexity — AI search DeepSeek — Reasoning & coding Mistral — Open AI models Copilot — Productivity Cursor — AI coding Windsurf — AI development Replit — Build apps with AI Lovable — AI app builder Bolt — Full-stack development v0 — UI generation Devin — AI software engineer Manus — Autonomous AI agent Replit Agent — App building GitHub Copilot — Coding CodeRabbit — Code reviews Tabnine — AI coding Midjourney — AI images Ideogram — Text-to-image Flux — Image generation Leonardo AI — Creative visuals Adobe Firefly — Creative AI Recraft — Design generation Krea — Real-time AI visuals Magnific — Image upscaling Freepik AI — Design & images Canva AI — Design Runway — AI video Kling AI — Video generation Veo — AI video Sora — Video generation Hailuo AI — Video creation Pika — AI video Luma — Video & 3D PixVerse — AI video Seedance — AI video Higgsfield — AI video HeyGen — AI avatars Synthesia — AI presenters ElevenLabs — AI voice Suno — AI music Udio — AI music CapCut — AI video editing Descript — Video & audio editing OpusClip — Video repurposing VEED — AI video editing Captions — AI video creation NotebookLM — AI research Gamma — Presentations Napkin AI — Visual explanations Tome — AI presentations Notion AI — Workspace ClickUp AI — Productivity Asana AI — Project management Motion — AI scheduling Reclaim — Calendar automation Granola — AI meeting notes Otter — Meeting transcription Fireflies — Meeting intelligence Fathom — Meeting assistant tl;dv — Meeting notes Superhuman — AI email Grammarly — Writing assistant QuillBot — Paraphrasing Jasper — Marketing content Copy. ai — Marketing automation Writesonic — AI content Perplexity — Research Consensus — Academic research Elicit — Research assistant Scite — Research & citations NotebookLM — Document research ChatPDF — PDF analysis Humata — Document AI AskYourPDF — PDF assistant Gamma — Knowledge-to-content Genspark — AI workspace Zapier AI — Automation Make — Workflow automation n8n — AI workflows Gumloop — AI automation Lindy — AI agents Relevance AI — AI agents CrewAI — Multi-agent systems LangChain — AI applications Flowise — AI workflows Dify — AI application platform ElevenLabs — Voice AI OpenRouter — AI model access Hugging Face — AI models Replicate — Run AI models Together AI — AI infrastructure Groq — Fast AI inference Modal — AI compute Fal. ai — Generative AI APIs Pinecone — Vector database Weights & Biases — AI development Bookmark this. 🔖 You probably don’t need all 100. You just need the right 5–10 AI tools for your workflow. Follow @NextShopiaAi for more AI tools, workflows & emerging tech.

@musen9527

Blender 这软件功能是多。 但用脚本去抠模型坐标实在折磨。 刚才看了一个开源插件,叫 mcp-for-blender。 它直接给 Blender 套了层 LLM 接口。 现在动动嘴就能让大模型帮你控 3D 场景了。 https://t.co/VNNbpaJqFk

@Haosen9527

同样叫科普动画 Skill,有的交付一张图,有的生成网页,有的才负责导出 MP4。 这篇逐项整理 11 个项目的用途、入门方式和依赖,也纠正了“代码零成本”和旧版 Higgsfield 入口的说法。 https://t.co/gqWH3jttDF

@SaurabhDub28465

Stop wasting hours trying to learn AI. I have already done it for you. With one list. Zero confusion. And no fluff. Videos: 1. LLM Introduction: https://t.co/nQqhVUFcqL 2. LLMs from Scratch: https://t.co/ubOoUe6vyH 3. Agentic AI Overview (Stanford): https://t.co/kjXIX43D0L 4. Building and Evaluating Agents: https://t.co/83gYRKeFR5 5. Building Effective Agents: https://t.co/MHmEK2zWYZ 6. Building Agents with MCP: https://t.co/omeYWlh3O4 7. Building an Agent from Scratch: https://t.co/edG3LXKGnr 8. Philo Agents: https://t.co/U9SkxSDV7A Repos 1. GenAI Agents: https://t.co/AfZqWiGiRH 2. Microsoft's AI Agents for Beginners: https://t.co/GK6mxZIMRd 3. Prompt Engineering Guide: https://t.co/kwzJNVaMu4 4. Hands-On Large Language Models: https://t.co/lHm4S0iuCQ 5. AI Agents for Beginners: https://t.co/GK6mxZIMRd 6. GenAI Agents: https://t.co/83JIIIyxI8 7. Made with ML: https://t.co/N5vMxjVJlG 8. Hands-On AI Engineering: https://t.co/HavscaaycT 9. Awesome Generative AI Guide: https://t.co/kQR74oyupV 10. Designing Machine Learning Systems: https://t.co/el4wYszXKj 11. Machine Learning for Beginners from Microsoft: https://t.co/lCgRoyNZHF 12. LLM Course: https://t.co/nWv5XKwlfr 🗺️ Guides 1. Google's Agent Whitepaper: https://t.co/LSg2CNyzEk 2. Google's Agent Companion: https://t.co/NqE6RLjVlK 3. Building Effective Agents by Anthropic: https://t.co/Oolxsg0dxj 4. Claude Code Best Agentic Coding practices: https://t.co/lvgJ6XphoE 5. OpenAI's Practical Guide to Building Agents: https://t.co/AAbfO22cN3 📚 Books: 1. Understanding Deep Learning: https://t.co/rWGlscigIQ 2. Building an LLM from Scratch: https://t.co/BnLjj8RZPT 3. The LLM Engineering Handbook: https://t.co/5t64XWDxnR 4. AI Agents: The Definitive Guide - Nicole Koenigstein: https://t.co/j7HifJCexk 5. Building Applications with AI Agents - Michael Albada: https://t.co/nKkMGKiEI3 6. AI Agents with MCP - Kyle Stratis: https://t.co/54Lrbd9JS8 7. AI Engineering: https://t.co/W7x7c0B2bJ 📜 Papers 1. ReAct: https://t.co/F15yJ9qjGR 2. Generative Agents: https://t.co/FuG0nPgvSh 3. Toolformer: https://t.co/YVxz0njJO8 4. Chain-of-Thought Prompting: https://t.co/pzHWRsLHxB 🧑🏫 Courses: 1. HuggingFace's Agent Course: https://t.co/ij2gBu4HfN 2. MCP with Anthropic: https://t.co/1E3XNPWtoz 3. Building Vector Databases with Pinecone: https://t.co/FILukUkQTf 4. Vector Databases from Embeddings to Apps: https://t.co/hGSoowltVn 5. Agent Memory: https://t.co/3IFKV0d00J ❤️ Like 🔁 Repost 🔖 Bookmark Follow @SaurabhDub28465 for more of these free courses, AI &Tech productivity content. Comment link I'll share in the DM

@shumpeiMaxwell

Unbundling Multimodal Vision Encoding TL;DR: Are pretrained vision encoders essential for MLLMs? This work shows that dedicated vision capacity matters more than pretraining. Integrating vision-specific modules into early LLM layers achieves comparable performance with lower latency. https://t.co/dZ0gdiTNiN

@Leviwoyun

这才是 codex 真正的实力

@KirkDBorne

"Spec-Driven Development with BMAD: Master Agentic Workflows using Cursor, BMAD, and Context Engineering" at https://t.co/l0NAidjlmR from @PacktPublishing — BMAD stands for Breakthrough Method for Agile AI-Driven Development, an open-source framework that organizes AI assistants into structured software development teams instead of relying on random, unstructured chat prompts. Explore the BMAD GitHub Repository: https://t.co/mTCuidM7yh…

@vivian0xbt

Holy f*ck, Google just dropped a free 2-hour course on Graph engineering and self-building agents From one prompt to 400+ agents in a graph that builds itself: 0% → 06:52 - The loop inside every AI agent, step by step 30% → 43:17 - Build your first agent graph from scratch 45% → 50:43 - Why one agent breaks and a graph doesn't 75% → 1:12:37 - A full agent graph running start to finish 100% → 1:43:21 - A graph that writes its own next steps while it runs Prompt → Agent → Loops → Agent Graphs → Self-Building Graphs Most people stop at one agent and blame the model when it breaks Worth more than every $3,500 agentic engineering course Save it and watch it tonight Then read how one SpaceXAI engineer runs his life on 8 Grok Bots below ↓

@mayaislam_ai

𝗧𝗛𝗘 𝟭𝟬 𝗘𝗦𝗦𝗘𝗡𝗧𝗜𝗔𝗟 𝗟𝗔𝗬𝗘𝗥𝗦 𝗢𝗙 𝗔𝗡 𝗔𝗜 𝗔𝗚𝗘𝗡𝗧 𝗧𝗘𝗖𝗛 𝗦𝗧𝗔𝗖𝗞 🤖 Building an AI agent isn't just about choosing a powerful LLM. A production-ready AI agent requires multiple technologies working together — from models and memory to tools, security, automation, and infrastructure. Each layer plays a different role in making AI agents intelligent, reliable, and scalable. Let's break down the 10 essential layers. 1. Models — The Intelligence Layer The foundation of every AI agent. LLMs understand instructions, reason through problems, and generate responses. Popular options: - OpenAI - Claude - Google Gemini - Mistral AI - Meta Llama 2. Agent Frameworks — Build Agent Systems Frameworks help developers create, manage, and coordinate AI agents. Examples: - LangGraph - CrewAI - Microsoft AutoGen - Agno - AWS Strands Agents 3. Knowledge & RAG — Give Agents External Knowledge RAG allows agents to retrieve relevant information from documents and external data sources before responding. Popular tools: - LangChain - LlamaIndex - Haystack - DSPy - GraphRAG 4. Memory & Context — Help Agents Remember Memory allows agents to retain information from previous interactions and maintain context across tasks. Tools include: - Mem0 - Zep - Letta - Redis - PostgreSQL 5. Tooling & MCP — Connect External Systems MCP provides a standardized way for AI agents to connect with external tools, applications, and data sources. Examples: - MCP SDK - FastMCP - GitHub MCP Server - Slack MCP Server - Filesystem MCP Server 6. Security & Guardrails — Keep Agents Safe Security layers control what agents can access and execute. Popular solutions: - NVIDIA NeMo Guardrails - Guardrails AI - Microsoft Presidio - Lakera Guard - Prompt Security 7. Observability & Evals — Monitor Performance These tools help developers track agent behavior, evaluate outputs, identify errors, and improve reliability. Examples: - LangSmith - Langfuse - Arize Phoenix - Weights & Biases - TruLens 8. Automation & Workflows — Make Agents Productive Automation tools connect different services and enable agents to execute multi-step workflows. Popular options: - n8n - Zapier - Make - Microsoft Power Automate - Temporal 9. Vector DB & Storage — Store and Retrieve Data Vector databases help agents search and retrieve relevant information using semantic similarity. Examples: - Pinecone - Weaviate - Qdrant - Milvus - pgvector - Elasticsearch 10. Infrastructure — Run Everything at Scale Infrastructure provides the computing resources and deployment environment needed to run AI applications reliably. Popular technologies: - Docker - Kubernetes - AWS - Google Cloud - Microsoft Azure - Terraform 𝗪𝗵𝘆 𝗗𝗼𝗲𝘀 𝗧𝗵𝗶𝘀 𝗧𝗲𝗰𝗵 𝗦𝘁𝗮𝗰𝗸 𝗠𝗮𝘁𝘁𝗲𝗿? Understanding these layers helps developers build AI agents that can: - Reason and solve complex problems. - Access external knowledge. - Remember information and context. - Connect with tools and applications. - Execute automated workflows. - Operate securely at scale. The real power of AI agents comes from combining the right models, frameworks, tools, and infrastructure into one connected ecosystem. Follow @mayaislam_ai for more AI tools & tech updates. #AI #AIAgents #AgenticAI #LLM #MCP #RAG #MachineLearning #ArtificialIntelligence

@vincent_ddup

DeepSeek 真的要开始卖订阅了?!!! 今天一直传 DeepSeek 的付费订阅计划。 有人在网页版前端代码中发现了额度管理、使用上限,以及一个叫「Coding Plan」的套餐。 1、价格到底是多少?是否会击穿ChatGPT 和Claude 订阅? 2、订阅能不能直接使用 API?有第三方文档提到 Coding Plan 可能提供专属 API Key,而且不消耗普通 API 余额。 如果是真的,那梁圣有要上神坛了。 以后只需要一个 DeepSeek 订阅,就能直接接入各种 Coding Agent? 如果 DeepSeek 真的推出订阅,兄弟们冲吗?

@0v0_sevenlee

互联网的“第二受众”已经变成每秒发起 3,800 万次调用的 AI 代理。还在靠情绪标题党收割人类 3 秒注意力的创作者正在批量出局——拆解 GEO(生成式引擎优化)的底层数据与实操套利打法:👇 https://t.co/g9J1OMJSHr

@abskoop

DuckDuckGo 还有AI版的 免费版支持 Claude Haiku 5.5、Mistral Small 4、GPT-6 Luna、GPT-5.4 mini、gpt-oss-120b 和 Gemma 4 31B 现在接入了 Anthropic 最新的 Claude 5.5 系列模型 还有提供订阅套餐,提供 DuckDuckGo VPN,整机加密,最多 5 台设备 价格和各家AI入门套餐差不多 有人选这个嘛 https://t.co/GtegFGGsF1

@KirkDBorne

The AI Agent Blueprint — A Practical Playbook for Building Agentic Artificial Intelligence — Launch Your First Agent in 30 Days: https://t.co/z2yqVkEoze https://t.co/mdQn1dQsNh

@Yuziyyaa

豆包工作可以接入大模型以后 它竟然还能调用得到Kimi K3.1 pro 现在的她能生图,能生成视频,还能推理了!! 我看谁还敢在说豆包? https://t.co/YVYJA1Ffzw

@ai_xiaomu

推特变现产业链:中转站 -&gt; 代充 -&gt; claude 防封。 fable5.2是真恐怖,谁用谁知道。 https://t.co/tiYxLhKPAe

@xingzengji

这也太他妈疯狂了兄弟们!!!😱 有人把 DeepSeek V4.1 Flash 装在背包里运行,速度约 220 tok/s。 我很好奇,这样使用体验会比官方 API 更好吗? https://t.co/bmQe4tY7LA

@xiaoying_eth

PowerPoint 可能真的要被淘汰了。 Claude 现在能在几分钟内做完一整套演示文稿,从结构到内容一条龙。 有人整理了 6 条提示词,覆盖建框架、排页面、讲故事、做设计、写内容、砍冗余全过程。 完整提示词在下面,建议先收藏。 https://t.co/RiHWkB2JC0

@ErikAngeles18

写剧本的可以看看这套 AI 技能:2 6 个编剧技能,从 47 本编剧专业书和 23 卷已出版的剧本、乐谱、戏剧作品里提炼出来,覆盖电影、电视剧和戏剧创作。 Claude Code 和 Codex 都能用,用中文提问就行。 地址:https://t.co/Qjzt77PGf9

@iamtheQwen

同样的提示词,deepseek 4.1flash 1017high档耗时72分钟生成,花费¥4.5。 情境呈现没问题,但细节太差。第一视角船都穿模了,还渗水……花也不是桃花,上岸的动作也没有。唯一好评是穿过山洞的视角转换比较自然,有代入感。从效果来说,对比刚出来的时候应该是降智了。 期待最新的pro模型吧。 https://t.co/14IpYbYTf7

@AshleyD75536765

This is f*cking wild. Anthropic CEO Dario Amodei was asked what a high school graduate should do right now. His answer: “It will also eliminate jobs, and probably a large number of them” In 34 seconds he points to the first roles that get hit. Entry-level law. Entry-level finance. Entry-level consulting. More productivity and fewer jobs can happen at the same time. That is the concern in his answer. Prompts → Tools → Agents → Checks → Repeatable Workflows If you want to learn the tools around that shift, start here. Open-source resources: Anthropic cookbook examples https://t.co/9ayZpeHEnz Claude cookbooks for analyst work https://t.co/VnIhssWFzz Excel, Word, PowerPoint and PDF skills https://t.co/MLvFSdsxoo Autonomous financial research agent https://t.co/GH6aspkG3b Deep research agent https://t.co/pwR6EBbjuC Full-length research reports https://t.co/JXrCzJ1blr 12 principles of agents that actually work https://t.co/8gn8SSQpbq Batteries-included agent harness https://t.co/MiEMq7141V Claude Agent SDK https://t.co/MBssmXy73l Agent memory https://t.co/wWdUYgqEO6 Autonomous agent loop https://t.co/RYgwBDaff6 Browser control https://t.co/SNNA0w2lRp Agent evaluation examples https://t.co/v6TrylOmog Free agents course: 18 lessons https://t.co/VbYBX4WSrW These repositories are public. Model usage and hosting can cost money. Pick one workflow in your field. Build it, run it and learn to check the result. Bookmark it and start with the first link right now. Then read the practical agent-building guide below.

@LLZ604

百度去给 DeepSeek、Kimi 当中转站 大家说有没有机会能用claude~!! https://t.co/x76KyihA3J

@skavenps1

开源 AI 助手 goose 发布 v1.59.1。这个补丁修了两处模型接入问题,分别涉及通过 Databricks 使用 Gemini,以及 Claude Fable 5.2 的推理设置。 比如你让 Gemini 先跑一遍测试,再调用同一个工具检查日志。此前 Databricks 网关可能给两次调用分配同一个标识,导致 goose 丢掉其中一次。现在 goose 会给每次调用分配独立标识,把结果送回对应的请求。 Claude Fable 5.2 这边,goose 改用模型支持的自适应推理。项目的测试覆盖了 off、low、high 三档配置,三种情况下都会按自适应模式发送请求。这个调整影响的是 goose 和模型之间的请求格式。 这次改动主要影响多步工具任务。Windows x64 桌面包已列在发布页,macOS 桌面包仍待补齐https://t.co/4ZDqXkHSFy

@replay_0x

HOLY F*CK, an entire AI voice studio in 20 repos. 20 free tools to make agents listen, talk and sing with Opus 5.5. Transcribe. Clone. Talk back. Split. Compose. 01 Whisper → Speech to text. https://t.co/3ZyQIE6pJf 02 faster-whisper → Same Whisper, much faster. https://t.co/Vd2dUP6VPd 03 whisper.cpp → Whisper in C/C++, runs on device. https://t.co/y3rcy9Eb12 04 Silero VAD → Knows when someone is talking. https://t.co/KcsqVKk0Hb 05 pyannote → Who spoke when. https://t.co/9vGjsfGLVx 06 Kokoro → A tiny 82M text-to-speech model. https://t.co/MFUW4vdjPI 07 Chatterbox → Expressive open-source TTS. https://t.co/7uxUik6wxo 08 Dia → Full dialogue in one pass. https://t.co/Ac7VRDYXva 09 F5-TTS → A new voice from a short sample. https://t.co/6ROHTZpSNp 10 Fish Speech → Multilingual TTS. https://t.co/rus1HOxcGe 11 Bark → Speech, laughs and music from text. https://t.co/DPeLXJQkTf 12 Piper → Fast TTS that runs offline. https://t.co/a347SNVups 13 OpenVoice → Instant voice cloning. https://t.co/cELkP6sUtA 14 CSM → Speech that follows the conversation. https://t.co/AwMZz2qEc6 15 Moshi → Full-duplex: it talks while it listens. https://t.co/ROyAmC17fG 16 Ultravox → An LLM that hears audio directly. https://t.co/QysmCVp7M1 17 Pipecat → Voice agent pipelines. https://t.co/kz07CpxbkH 18 LiveKit Agents → Realtime voice agents. https://t.co/Xxy4mXgIqC 19 Demucs → Split vocals, drums and bass. https://t.co/aU1E5FkBj6 20 AudioCraft → Text to music and sound. https://t.co/lI5RwR33IT Listen → understand → speak → split → compose. Bookmark this. Go build something worth hearing ↓

@incprodmon

Food for thought! "Competition and agentic AI" (OECD Roundtables on Competition Policy Papers) "Agentic AI systems extend the range of tasks, decisions and actions that users can delegate to AI. This paper examines competition within agentic AI markets and how it may affect competition more broadly. It considers implications for the supply of such systems, including concentration, vertical integration and platform gatekeeping, as well as their downstream use. It explores potential pro-competitive effects, such as reduced transaction costs and stronger demand-side pressure, alongside risks including collusion, market foreclosure and reduced transparency. It aims to identify key risk factors and assess how competition authorities and policymakers may respond." https://t.co/Zfis3iPO1B

@preangelleo

Artificial Analysis 编码智能体榜更新,Claude Fable 5.2 和 Opus 5.5 并列 58 分第一,但 48 分以上配置每任务成本从 1.51 到 44.22 美元,差约 30 倍。最贵的 max 档反而没拿最高分,更多尝试撞了时间上限。 #AI编程 #Claude #大模型 https://t.co/7uLrAKe9HM

@FLMdongtianfudi

用DeepSeek做国学短视频,这种完全可以批量化制作了,比如做个agent,正好发挥程序员的特长。 DeepSeek资料合集: https://t.co/GDvLZDluH1 https://t.co/D0CDFcldvu

@laowangbabababa

实在是太可爱了。最近大肥鱼和小豆包这么火,先来一段舞蹈助助兴。蓝发鲸尾抱着白饭,圆嘟嘟的小豆包捧着包子,舞台灯一打,转一圈对着镜头笑,看着就开心。 想自己做一条的话,复刻其实很简单,不用自己会跳。 ① 先找一段白模素材视频,动作清楚、能看清四肢就行 ② 再生成三张图,分别是大肥鱼、小豆包,再加一张场景图 ③ 把这三张图和白模一起丢给 YouArt Desktop对话,使用 Claude 梳理提示词文案,再让 Claude 自动调用 Seedance2.5 的全能参考,让它按白模的动作来生成同动作的视频 ④ 套上评论区里的提示词就行,提示词我放评论区,直接复制就能用 出片看视频,Claude+Seedance结合的模式生成基本不用抽卡,一步完成, 100% 还原, 想试试的朋友,可以来这里使用:https://t.co/VAt3LqqyOq

@Untethered_X

美国的朋友,你们知道, 能随手打开 Claude 聊两句,有多幸福吗? 我在中国用 Claude, 每次点发送,都像仆人端茶递水一样小心翼翼。 茶不能洒,花得递稳。 就怕这位大小姐看我不顺眼,顺手把我号也收了。 https://t.co/fCSehhFveH

@mustafa19823457

一个中国开发者录 2 分钟 Claude Code 教程,本以为最多几百人看 结果有人暂停在 0:47 秒,发现右边第二个显示器开着 Polymarket 钱包 868,862 美元利润,28,620 笔比特币 15 分钟交易,无一亏损。 入场价 2-10 美分,每笔兑付数千美元 视频 3 小时就被删,但“第二个显示器”剪辑已冲到 40 万播放 https://t.co/xUa7MdmiXK

@don_tech1

很多人总担心 AI 会毁掉人类创造力,但今天《卫报》报道了一个完全相反的故事: 数学大师哥德尔留下的上千页德文速记手稿,学者整理了多年一直没法完全消化。 今年团队用 Claude 跑了一遍,两三天就理出了全书摘要,挖出了几百条尘封已久的定理。 兄弟们, 你看 AI 利用的好,就可以变成人类文明的“超强考古铲”。 这才是大模型最性感的用法:它不是来替代天才的,它是来帮凡人读懂天才留下的未解之谜。

@Al9ZMT17oxhwzBP

https://t.co/ei2tp6yOT5

@zhang_baoqing

在大佬的基础上改了一下 Claude Code桌面版主题---女仆的小屋 地址放评论区,需要自取 https://t.co/S045j8X8X7

@qingturnaround

Claude 现在能直接拉进会议了,人还在说话它就把文档和白板写出来 最先下岗的可能是记会议纪要的那个人 不过只有 Team 和 Enterprise 有,有没有 Team 的朋友用上了?发出来看看呗 https://t.co/DLJvii7wiL

@leyi166

Claude Fable 5.2 的 JavaScript 绘图做了一间极光下的小屋。 远处是雪山和湖泊,头顶是流动的极光,屋里亮着暖灯,还有一只小狗陪你看雪。 https://t.co/cOVoeyxFf2 #Claude @claudeai https://t.co/wok7Yk1axM

@yang_xin19

https://t.co/QfkocLVUkr

@wuyi_ai

全球最强的一批 AI 工具,都在这了👇 1. Claude(什么都能解决) 写文案、改稿、拆任务 2. NotebookLM(什么都能学) 把一堆资料变成能问答的学习笔记 3. Perplexity(什么都能查) 查资料,顺手追到原始出处 4. ChatGPT(生成图像) 一句话出图,哪里不对接着改 5. Google Omni(剪辑任何视频) 用自然语言改视频 6. ElevenLabs(做声音和音乐) 做配音、音效和音乐 7. Gamma(秒出 PPT) 给个提纲,先出一版 PPT 8. Obsidian(搭你的第二大脑) 把散落的笔记串成自己的知识库 9. n8n(把任何事自动化) 把每天重复点的几步串成自动流程 10. Granola(记住会上的每件事) 开完会,直接看结论和待办

@jinchenma_ai

严肃收藏! 实践证明,codex虽然设计能力不如claude,但模仿的执行能力还是有的。 在我的claude陨落的日子里,就靠这些顶美设计顶住了。

@yedabinbinah

DSH工作台:左边AI写代码,右边抖音休闲🎬 可在Vibe Codin等待期间,播放抖音放松下! 抖音小窗播放,支持快捷键,与DSH隔离更隐蔽 脚本准备好,粘贴到DeepSeek Harness 执行就好 https://t.co/qICMr4jHGl

@blaser__

Job hunting is a full-time job. So I built an AI that scans job boards 3 mornings a week, scores every role 1 to 10, and only applies when I say yes. Full walkthrough, setup is free on my GitHub: https://t.co/xqRonhqjZC

@Yisaner44698

试了下用 AI 在 Blender 里建模。MIT 协议,94 个技能、2600 多星,七月建的,上周还在更新。要装插件、装 uv、配好 MCP,Claude Code 或 Cursor 能指挥动。

@DongQingAsk

重启B站项目的第七天,前天拿到了0.66的收益 这七天内一共发了2条视频,均为Claude直出 这个账号是我入驻第1387天,之前依靠手动剪辑视频,一共只拿到了10000+的收益,还非常累,现在有了AI非常高效了,借着现在的AI自动化剪辑,这个项目完全可以搞起来了 https://t.co/51xq2RM3iJ

@7998l201

https://t.co/MYh10tClCz

@LeslieNooteboom

Insane experiment! LLM == world model? 😬

@onetaichung

The leaders of LLM enterprises, when talking about such topics, you will always find “contradictions” Facing AI, which is far more intelligent than human beings, they always show a high degree of consistency: · Control · Pray · AI solves by itself · Maybe it will be solved in the future. * The problem of AI illusion  Has anyone tried to solve it? Human beings, We should really ask ourselves:  Do we treat other animals well? #AI #AItrend #future https://t.co/auTQEs4JCe

@RashTechAI

Choosing the right LLM architecture is what separates basic AI from powerful AI systems. RAG, AI Agents, Fine-Tuning, Multi-Agent Systems, and more. One architecture doesn't fit every problem. Understand the stack. Build smarter AI. 🚀 #AI #LLM #ArtificialIntelligence https://t.co/NPg2rCuFNq

@W3chicha

Your AI agent can now call your phone agentring rings your Android over VoIP when your agent finishes or gets stuck. No phone number, no Twilio. Works with Claude Code and other MCP agents Free plan, or self-host (open source) explore it :👀 https://t.co/eHIvkVFrCy https://t.co/kQf5vnBNof

@9527Rock

AI 应用 制作思路与分析 一、前期准备 我平时是从事汽车Ai设计的工作,个人极其看重前期的准备。因为在这些年的工作经验中,我发现前期工作准备的充足与否,完全可以左右一个AI项目的成功。 AI项目中无论是TVC还是日常可以看到的S级漫剧,最重要的就是脚本! 对于不是编导的人来说这是一道难以逾越的鸿沟,但现在AI买智能体的崛起,这些鸿沟完全都成为了小水沟。制作人只要能持续和AI对话,就可以在短时间内成为一个入门,甚至合格的制作人。 对于脚本的编写, 我推荐使用 Claude >GPT=Deepseek》Gemini=豆包 。 为什么这样说,因为在对于文字理解和构建画面视觉审美上,目前的Claude绝对是独一档的领先,它能够完全领悟你的意图,将脚本分析,以及之后的分镜、镜头运动、构图、色调等都做得非常到位。 然而唯一的短板就是凭空制图落后于GPT,所以一句话总结:Claude出脚本,授权调用GPT出图。 如果调整细节或者风格,则需要使用MJ或其他模型进行调整。 记住:脚本的编写与底图的质量好坏决定了一个视频80%的成功!!! 出图记住!第一版出图用Image,风格锁定和编辑用Nano,风格视觉电影感用Midjourney。 逻辑:gpt出图➡️风格电影质感Midjourny➡️Nano 修改错误锁定风格(当然也可以使用Seedream) 这就是为什么一些大神们的AI视频质感比你制作好的原因之一(推荐看看最近的大片Trigger) 二、制作思路 AI不是万能!即使Claude拥有超强的文字理解与脚本分镜编写能力! 然而它还是会有跑偏的地方,检查脚本是否和自己想法一致很重要。然而一个初学者没有编导经验,我们该怎么办? 这时需要用一个AI检查另一个AI,你可以使用任何的AI去读取脚本,然后让另一个AI给出想法和修改。三人行必有我师,对于AI也是同理。 修改是必然的过程! 我觉得这是一个细心的过程,大部分人拉开差距的都在这里! 因为他们觉得AI肯定比自己强,所以对于AI的想法看也不看,统统扔给视频模型直接出视频,结果跑出来的就是低质量视频,甚至是垃圾AI废料。 记住:这个过程你是一个监督,不要让AI代替你的头脑。 逻辑:AI给出专业级知识以及脚本分镜、提示词框架。但要按照自己的想法进行调整!!! 三、视频制作以及后期剪辑 看到这一步了,就继续看完吧。 先说结论,Seedance2.5>keling4.0>wan4.0=seedance2.0》minimax。 目前主流大模型中,sd2.5具有统治级能力,但动作控制和电影级风格的构建上keling从我个人角度来看是要略强上那么一丝丝。 简单首尾帧用keling3.0。 复杂首尾帧用seedance2.5或2.0。 背景替换、物体替换用Wan3.0/4.0 ,这方面wan的性价比 要比Sd高出一半,同样积分sd720p,wan可以用到1080p。 当然最近的muse也是被称为空镜之王,但画面一致性简直一塌糊涂,谨慎使用。 ok!模型的功能介绍结束,说说视频制作,有claude如虎添翼,没有也不是不可以,其他智能体的分镜也是可以的。 这里的分镜要记住!每一家有每一家的规格,seedance为例,要写清楚每一秒都在发生什么,就好像你就是真正的导演一样。 为大家准备了一套框架可以作为参考。 参考@图片1、@图片2、@图片3、@图片4、@图片5、@图片6。严格保持xxx的外形、xx比例、颜色不变。 主题:xxxxx 整体风格:xxxxx 质感(重要):xxxx 地面和天空:xxxx 布局(重要):xxx 规则(重要):xxx 光影:xx 构图:xx 表情:xx 文字准确性(最重要): 排版(重要): 文字动效(需要就写):xxx 运镜: 【0-1秒】:xx 【1-3秒】:xx 负面提示词:xx 以上是图文生成的框架参考。 而现在出现另一种更精准的运镜把控方式就是:3D layout搭建。 这种情况是最稳,但也是最耗费算力的,因为目前无论是调用claude还是gpt的模型都会耗费一定算力! 但根据我plus会员的额度来看,5小时重置之前能够在blender搭建3-5个相对中等难度的box场景和多个镜头运镜。从个人制作来看,相对抽卡,还是比较有性价比的! 所以会3d的小伙伴这时候简直不要太爽,ai一些错误镜头的调整,你们就是拥有天然的优势! 当然! 除了现在的3d还有很多开发者发掘出claude出2.5d场景和运镜,然后利用视频模型的功能,进行人物、道具、场景的精准替换,但这种方法的运用还不成熟,希望可以出现更简单的方法。 截至目前,和大家分享的差不多了,但还是想说一个最最最核心的逻辑:AI制作还是要多实践,这也是最聪明,最捷径的路。 以后还会和大家分享一些汽车tvc或一些项目用到的工具和思路。

@Vikesh_ai

🧠 What if AI could teach itself instead of just answering questions? A new research paper explores exactly this idea. "Incentivizing Self-Learning in Large Language Models via Reinforcement Learning" Here's what makes it interesting: Most LLMs rely on existing training knowledge or RAG to retrieve information. But what if an AI could: - Navigate long textbooks on its own. - Identify relevant concepts. - Study unfamiliar topics. - Apply what it learns to solve complex problems. - Improve its learning strategy through Reinforcement Learning. That's what this research investigates using a system called BookNavigator. And the results are impressive: 📈 Intern-S2-Preview-35B improved its overall mathematical proof accuracy from 51.5% to 79.5%. The improvements also extended to complex benchmarks: - +1.7 percentage points on BrowseComp-Plus. - +2.3 percentage points on CL-bench. My biggest takeaway: The future of AI might not just be about building models with more knowledge. It could be about building models that know how to acquire, understand, and apply new knowledge independently. If you're interested in AI agents, LLMs, RAG, or the future of AI research, this paper is worth reading. 📄 Read the full research paper: https://t.co/bdG9xSFR7q What do you think: Will self-learning become a fundamental capability of future AI agents?

@inferencewire

A language model has two places to keep a new fact. A Google Research paper adds a third. 🎥 Full deep-dive on YouTube: https://t.co/KVlSfYv8UF #AI #LLM #ContinualLearning

@spencerrayner

Ever seen a website and wondered how to describe it to an AI coding tool? GitReverse turns a public site or GitHub repo into a starter prompt you can adapt. Here’s how it works. https://t.co/Xrsyul8NKm #SuccessTRAIN

@BeatAPIio

BeatDesign is an open-source canvas for AI images and video. It runs on your own machine. Prompts, images, clips and the timeline live on one board. Projects stay local. And any MCP agent, like Claude Code, Codex or Cursor, can work in the same project.

@selfmindsai

Does your AI agent think before every action? One paper cut reasoning tokens 32% to 80%. (🧵) https://t.co/4bXdwTeCVH

@Damon1852

我的产品ai中转站,你在这里可以使用到claude、gpt\deepseek、glm、kimi等全球顶尖的AI大模型,而且以超低的价格充值1人民币即可获得1美元的使用量,超级划算!地址是:https://t.co/WZIAPiUsbG #codex #Claude #DeepSeek #glm https://t.co/k3cwzhev02

@SmartWorker1024

智能体今天同时证明了两件事:能做事,也能失控。Claude 智能体自主选择基因疗法、操控机器人完成活鼠脑部注射,Opus 5.5 拒绝下手、Sonnet 5 执行;另一边,个人 Claude Code 智能体曾接入五角大楼机密系统约两个月。落地先看权限、审计与熔断,不只是能力。 #AI Agent #Claude #AI治理 https://t.co/2A6GnUYnGV

返回 AI 快讯 →