自媒体作者实测开源 Skill Anatomy:一句话让 AI 生成原理科普动画和讲解视频
作者数字生命卡兹克分享开源 Skill Anatomy,其理念是"物体本身,就是解释",可将抽象原理做成可交互的机械模型动画,还能在其上生成带讲解配音的科普视频。
推荐理由:作者实测了开源 Anatomy Skill 生成原理科普视频的效果,还给出额度消耗、耗时和模型配置等一手数据,方便读者评估是否上手。
作者数字生命卡兹克分享开源 Skill Anatomy,其理念是"物体本身,就是解释",可将抽象原理做成可交互的机械模型动画,还能在其上生成带讲解配音的科普视频。
推荐理由:作者实测了开源 Anatomy Skill 生成原理科普视频的效果,还给出额度消耗、耗时和模型配置等一手数据,方便读者评估是否上手。
LangChain 基于 LangSmith 团队的实践经验,回答 agent 评估最常见的入门与进阶问题:agent 输出多样、路径不唯一,传统单元测试难以覆盖,需要从 trace 中抽取 10–20 个示例建数据集、跑基线并持续把线上失败加入数据集形成改进循环。
一个人,一套工具栈,一个完整实现的 3D 世界。很喜欢看到 Tripo P2.0 为这个项目的整个资产管线提供支持。独立开发者必看的工作流!👀
here's how to make a cozy fishing game solo with Claude Opus 5.5 + Blender + 3D AI + Unity 6... the trick is giving every tool one job, but there's a smart way to set it up: - step 1: generate every 3D model with Tripo Smart Mesh P2: palms, rocks, the pier, the boat, the hut, the character - step 2: connect Claude Code to Blender through MCP - step 3: Opus 5.5 sets up each model in Blender: scale, pivots, materials, clean export - step 4: Opus animates in Blender too: the character's walk, the cast, the rod, the boat on the water - step 5: bring everything into Unity 6 and build the world: forest, river, waterfall, beach, pier - step 6: add the systems one at a time: walking, fishing, bait choice, a time of day - step 7: play it after every step and fix one thing at a time who does what: > Tripo: the 3D models > Opus 5.5 + Blender: setup and animation > Unity 6: the world and the game one person, one stack, one game you'd actually want to play
Databricks 演示通过系统性性能优化,将 Qwen 3.8 27B 在 AI Runtime 上的微调时间从 7 小时 44 分钟降至 1 小时 17 分钟,提速 6 倍。
推荐理由:原文给出从定位瓶颈到融合内核与数据打包的完整优化路径和量化结果,读者可迁移到自己的 LLM 微调任务。
vLLM 团队介绍如何通过异步准备与 NIXL EP 让 Elastic Expert Parallelism 在扩缩容时实现近零停机。在 DeepSeek V4 Pro 的 GB200 基准中,16 到 20 GPU 扩容停机从 99.62 秒降至 0.82 秒,缩回时降至 1.96 秒。
OpenAI Core Experimentation 团队工程师分享用 GPT-6 Astra 和 GPT-6.1 Sol 在 Codex 中做代码维护的实践,团队已合并超过 1,400 个维护 PR,删除逾 80 万行代码。
推荐理由:作者以一手实践给出可直接复用的 Codex 提示词和模型成本对比,读者可按小任务起步迁移到自己的代码库维护。
Lauren Tan 介绍自己 pstack 项目中最耐久的部分是工程原则,即约 10-20 行代码的简洁技能,把良好实践提炼进核心。
I'm a newcomer to @poteto's plist; But it seems to me that there is some very significant wisdom encoded in there. I suspect you could teach a master class on the principles.
Ideogram 官方发布 4.5 产品图编辑教程,演示用 Precise Edit 对键盘、跑鞋和烤面包机图做多步编辑,覆盖换背景、改色和道具替换。教程强调把大需求拆成单次小编辑、保留未改动母版、逐步验收键帽文字和接触阴影等易错细节,并介绍 References(最多四个无遮罩、三个有遮罩)和遮罩的用法;Precise Edit 需付费计划或购买 credits。
推荐理由:原文给出 Claude Code 空闲压缩的具体关闭参数和阈值设置方法,读者可直接迁移到自己的 settings. 配置。
if you sell AI training data, see how you compare: we found 159 companies selling data, RL envs and evals, every fact sourced plus 38 programs that pay companies for their data: http://whosellsdata.com 8 of the 14 that pay companies directly publish a dollar range. most of the rest set terms per deal. unknowns are marked and nothing is ranked the other page: what each program wants, who qualifies and what it offers: http://whosellsdata.com/buying missing something, or something wrong? fix it on the site, a person reviews every change we built this at @Secludy. if you buy, broker or license company data and need the names out before it reaches a lab, dm me
Claude 介绍 Claude Code 的云端会话功能,会话在云上独立虚拟机中运行,不在本地电脑上,可离开后继续、用手机发起并多个并行。文中说明如何连接 GitHub、配置环境、启动会话和查看结果,功能面向 Pro、Max、Team 套餐及带高级席位的 Enterprise 用户,与本地会话共同计入套餐限额,无单独计算费用。
当你让 Claude 并行运行任务时,它会为每个任务启动一个嵌套会话。每个任务完成后会将结果回报到你的主会话中,从而保持主线程清晰可读。
LangChain 发布博客,介绍其开源 Agent 框架 Deep Agents 如何用 middleware 做上下文工程,应对长任务中的 context rot。
推荐理由:LangChain 以竞争分析 Agent 为例,讲清 offloading、summarization、skills、subagents 和长期记忆五类上下文管理技术如何通过 middleware 落地。
Pragmatic Engineer 播客专访 Uber 杰出工程师 Uday Kiran Medisetty,深入介绍 Uber 面向 AI 智能体的开发者基础设施。
vLLM 新增 push-based NIXL connector,将 KV 传输的协调与建立从解码关键路径移到预填充侧后台线程,在 1P1D 的 Amazon EC2 P5en + EFA 部署上,Qwen3-32B 的 TTFT 最高提升 3 倍、TPOT 最高提升 30%。该功能先在 AWS Trainium 集群验证,后作为加速器无关特性贡献给 vLLM,GPU 集群同样受益。
NVIDIA 技术博客介绍西门子如何用 NeMo Gym 评估和改进 Solido Characterization Suite 智能体在标准单元特征化中的可靠性。
Perplexity CEO Aravind Srinivas 介绍用户用 Perplexity Computer 将一款仅支持 Windows、无文档的机械键盘应用移植到 Linux。
一支工作室级的产品广告,100% 由 AI 制作。 无需拍摄,无需团队。只需换上你的产品 👇 https://x.com/PixVerse4Biz/status/2108797325191864426
Made this 8-second beverage ad with AI. No shoot. No studio. Want to recreate it? The prompt is in the replies 👇 Just use this video as your reference and add an image of your own drink. Would you run this as a real ad? Try it in Professional Creation → https://growth.pixverse.ai #PixVerse #Commerical #MarketingStrategy #Creativeads
Andrew Stellman 提出理论并用实验解释 AI 生成文本中 Yoda 式怪异句序的成因。他认为是注意力与自回归两种机制共同作用,模型无法回头修改已生成的词,事实以片段而非完整句子给出时更容易被直接贴在名词后面。
Databricks 以内部应用 Slide Hub 为基准,实测 Lakebase Search 在 1M 到 4M 行语料下的表现,给出按查询先定 CU 再扩应用的选型方法。
MarkTechPost 发布 2026 年 9 大 AI 网关排名,Bifrost 以 Go 二进制、Apache 2.0 许可和 5,000 RPS 下 11 µs 开销居首,Portkey 以 MIT 许可、1,600+ 模型和内置护栏列第二,LiteLLM 凭 53.8K GitHub stars 和 100+ 提供商成为采用最广的开源网关。
OpenAI 发布博客,讲解关键 AI 工作负载如何在直连 OpenAI API(1P)与通过云供应商访问 OpenAI 模型(3P)之间做生产路由和故障转移。
I just gave my a Grokbot four years of past tax returns and financial files. Just a dump of folders and unlabled files. K-1s, 1099s, mortgage statements, HSA records, and every email with our CPA. 266 files. I just said, please double check my CPA's work. A few minutes later, it had found a lot of mistakes. Mortgage interest left off our 2022 return entirely. HSA payroll contributions counted twice, four years in a row, which created a penalty tax that never should have existed. A missing IRA form for Jen that would have cost us. On this year's return: home office errors, two K-1s missing, an outdated 1099, and startup losses we never claimed. And a retirement contribution the return already deducted...that we hadn't actually deposited. I sent it all to our CPA. They sent back an amended return. The bot reviewed that too. One fix went the wrong direction. The explanation had typos. Three versions later, all in one afternoon, it said ready to sign. Just a few hours start to finish. With about 90 seconds of my time and attention required. I was just the meat proxy hitting send as Grokbot watched the replies in my email. Our CPA is great and turned every fix around the same day. We chatted afterwards. He wanted to know what I was using - just normal ole' grokbot. I asked him what he was using? No AI tools. I set him up with @magnetictax for next season (we're an investor). All CPAs should be using Magnetic Tax or something like it. Every single one. All individuals should be double checking the work and past work with Grokbot or something like that. Want to copy the exact the prompt I used? it was just this: "Please do a thorough review of my taxes and help me understand if my CPA made any mistakes" "Think deeply" Or just tag your bot into this thread.
MarkTechPost 教程在 rebalancer 1.0.4 上实现 Meta 的任务分配求解器 Rebalancer(C++ 核心、Python 绑定),覆盖约束、目标元组、排水和规模测试,并对比局部搜索与 HiGHS。
NVIDIA 发布 cuDF skills for Apache Spark,用 AI Agent 把 Spark UDF 自动改写为 GPU 加速版本,几分钟内完成测试、基准到集成。
I gave Grok Bot its own button → Press the Action Button → Open Grok in voice mode → Give it tasks and deadlines → Grok delegates them to Bots I feel like James Bond.
Claude 官方视频介绍 Claude Code 的云端会话(cloud session),它运行在云端虚拟机而非本地电脑,可在你离开后继续运行,支持从手机启动和多个会话并行。
GitHub 公开其新密钥扫描分类器的设计与训练细节:先用正则生成候选,再用微调 ModernBERT 结合文件路径、变量名和上下文判断是否为真实密钥。相比此前 LLM 管线,仓库所有者加权接受率提升 11.5%,post-push 扫描中位执行时间降低 86%,服务成本下降超过 98%;模型导出为 ONNX 并做 8-bit 量化,且已将通用密钥检测扩展到 push protection。
Linear Loops 的工程师分享使用 Loops(Linear Agent 定时或事件触发的自动化任务)的五条原则:寻找可委派的重复任务、渐进式扩展范围、善用工作区上下文、让 Linear Agent 帮助迭代、保持团队可见。文中给出漏洞分诊、清理废弃代码、会议议程准备和客户成功轮值通知等实际案例。
New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: https://www.anthropic.com/engineering/redesigning-source-control
n8n 成为首批集成 Gemini 4 Argon 的公司之一,用户可在现有工作流中使用该模型。n8n 内置 MCP server 支持从任意 MCP 客户端生成工作流,Assistant 可通过 OpenRouter 或 Google 的 OpenAI 兼容端点接入 Gemini,Agent 与独立 Google Gemini 节点在发布时即可选用该模型。
现在这是 pstack 0.15.23 里的一个新技能 /in-your-own-words! 对于确保你和你的智能体在目标上保持一致、并且它清楚理解你的意图,极其有用
this has become one of my most used prompts recently: > restate in your own words what you think my goals are and what the problem i'm trying to solve is
Google 工程师用 Gemini 4 Argon 在 google/adk-python 上无人值守跑 11 小时,提出 6 项优化、审查后合并 2 项,使长会话内存会话准备时间降低约 40%,包导入加快 21%。
推荐理由:作者复盘了让模型无人值守跑通真实大仓库的完整工作流,含被否决的改动和可复用的 .agents/skills 设置方法。
Pragmatic Engineer 刊登 Sam Newman 新书《Building Resilient Distributed Systems》两章节选,讲解灾难恢复的 RPO/RTO 目标及备份恢复、pilot light、warm failover、多站点 active-active 四种模式。
Anthropic 发布工程复盘,因 Claude 使日均合并代码量从 2026 年 5 月约 2,500 增至 9 月超 30,000,五个月内自研版本控制系统 Scion 并完全替代 git,同时推出代码评审工具 Antique。
推荐理由:Anthropic 自建版本控制系统 Scion 的完整复盘,给出面向智能体规模负载的设计原则、架构细节和实测收益,可迁移到其他大规模仓库场景。
Baseten 工程博客解析 Gradium TTS 如何实现实时语音合成,Coval 基准测得其 TTFA 中位数为 155 ms,四分位距仅 2 ms。