Liu Xiaochen

#ArtificialIntelligence #MachineLearning
#ComputerVision #3Dreconstruction
#3Dmeasurement #ImageProcessing
#Robotics #Programming

[宝玉 @dotey]

Anthropic 7 月 10 日发布了一场关于 Agent 基础设施的对谈。Claude 平台工程负责人 Katelyn Lesse、产品负责人 Angela Jiang 和产品经理 Jess Yann，分享了几个来自一线的观察。

【Agent 的“脚手架”正在变薄】

几个月前，搭建 Agent 往往需要写大量流程控制代码：先执行 A，满足条件再进入 B，遇到不同情况还要切换不同分支。流程越复杂，系统越容易出错。

随着模型的推理和工具调用能力增强，这些编排层（harness）正在变薄。开发者不用再规定每一步，只需给出目标和基本边界，让模型自己决定怎么完成。

与此同时，一种更高层的编排方式开始出现：让多个 Agent 同时解决一个问题，从中选出最佳方案；让一个 Agent 提方案，另一个负责挑错；或者在 Agent 卡住时，请另一个能力更强的 Agent 提供建议。

重点正在从“控制每一步”，转向“设计 Agent 之间如何协作”。

【衡量 Agent “投入产出比”（ROI，Return on Investment），先看一个人快了多少】

Angela 建议，企业不要一开始就规划上百个自动化流程，而应该先看一个具体的人：用了 Agent 之后，他的工作速度和产出提高了多少？

验证有效后，再从个人推广到团队，最后才处理跨部门流程。前期重点看速度和生产力，等应用逐步成熟，再衡量收入、成本和用户指标。

很多企业做 AI 转型时，喜欢先画一张宏大的自动化蓝图。问题是，流程涉及的部门越多、规则越复杂，落地阻力就越大。从个人开始，更容易看到效果，也更容易持续推进。

【工程团队没消失，但每个人的角色都变了】

Katelyn 观察到，Anthropic 的工程团队和半年前相比，人员构成没有太大变化，但协作方式已经不同。

过去通常由技术负责人决定架构，其他工程师领取任务、编写代码。现在，更多工程师会参与产品和架构决策，再分别指挥 Claude 完成具体工作。

Agent 的作用也不再只是“帮忙写代码”。她提到 Shopify 的 River 系统，已经把需求文档、开发环境、代码实现和 QA 测试串成了一套端到端的 Agent 工作流。

【个体变强，不等于团队自然变好】

Agent 降低了开发和试错成本，也可能带来新的问题。

过去，一个团队会先讨论十个方案中哪个最值得做。现在，每个人都可以快速做出十个原型，甚至全部上线，让市场决定谁胜出。

这样做速度很快，但如果缺少统一方向，产品很容易无序扩张。Agent 能显著放大个人能力，却不会自动解决团队的协调、取舍和决策问题。

来源：https://t.co/fEuKIzIlfk

[codila @0xCodila]

Ex-Google engineer just dropped 1-hour course: loops, self-improving AI, memory systems - from scratch:

00:00 - the self-building agent
03:01 - soul.md runs everything
30:16 - RAG memory: pull 20 messages, not 2,000
31:48 - the loop that knows when to stop
35:14 - find the bug, fix the prompt
50:22 - how Claude compresses your memory

1 hour of his guide beats any paid agent course

watch & bookmark - then read Karpathy's loop method below

[AYi @AYi_AInotes]

强烈推荐大家都去看这个跟workbudy产品负责人Jason的深度对谈播客，
另外关于腾讯workbudy，我说个暴论，你看100篇那些AI写的教程文章，都不如抽1个小时好好看看腾讯官方出品的这期播客，太他么深度了，干货拉满！

Jason分享了一个关键判断让我印象特别深：2021年他曾计划培养孩子学习编程，到2026年发现这个方向已经失效了，在AI的加持下编程本身不再是稀缺技能，洞察力、审美品味、逻辑清晰的语言表达能力，才是AI时代真正的核心竞争力。

还有这个文科生的故事，英语文学专业的前媒体人，一行代码都不会写，就因为自己攒了7000多首杜比全景声的歌，Excel卡到打不开，找外包又说不清楚需求，被逼得没办法开始摸AI编程工具。
你以为是AI两下给他把App生成出来了？根本不是，
他说前前后后开了40多个对话窗口，文件拆得乱七八糟，环境部署从头啃，踩了无数坑，磨了好几个月，才真的把那个杜比音乐App做出来，现在都更新到3.2版本，正经上架App Store了。

那他为什么没放弃？
他说AI会在他做对一点点的时候，蹦出来一句“你这个思路很清晰，像个成熟的产品经理”。
卧槽，就这一下兄弟们应该懂吧，
我们这种在否定式教育里长大的东亚小孩，太吃这一套了。

很多时候你不敢开始，其实不是懒也不是笨，是脑子里那个“你不行”的声音太大了。

咱们从小到大做什么都先被挑错，刚迈出一步就有人说你这不对那不好，久了你自己都信了，你就是做不成事。

但是AI不judge你，不骂你，不会翻个白眼说这都不会，你做对一点它就真心实意夸你，直接把那个压了你几十年的声音盖过去了。

播客里WorkBuddy的产品负责人Jason说，2021年他还在规划以后要送孩子学编程，2026年回头看，这个想法已经不成立了。

我以前以为，只是因为写代码的门槛被AI踏平了，现在才懂，被踏平的根本不止是代码门槛。

以前我们想做个东西，不仅要会写代码，还要有颗大心脏——被bug卡一周，被前辈骂写的什么垃圾，被环境配置搞到崩溃，90%的人在第一步就被劝退了。

现在AI不仅帮你写代码、查bug，还给你当啦啦队，把你那点被打击了几十年、碎得渣都不剩的信心，一点点给你拼回来。

你看现在满屏都是MCP、Agent、Vibe Coding这些新名词，好像你三天不背就要被时代抛弃，大家课买了一堆，一个工具都不敢打开，但实际上哪有那么玄乎啊🤣

当然了，如果你从小在夸奖里长大，内心强大到根本不需要别人给你托底，可能get不到这个点。

但如果你和我一样，总怕自己做不好，怕出错被笑话，迟迟不敢迈出第一步，真的不用上来就想做个上架的App，

不如今天就花半小时，让AI帮你写个批量改文件名的小脚本，或者整理下你乱了半年的文件夹。

你会发现，你根本不是不行，只是之前从来没人，在你做对一点点的时候，认认真真跟你说一句“你做得真棒”🥹🥹

焦虑这东西，站在岸边看永远觉得水很深，脚伸下去才知道，原来不仅水不深，还有人在旁边扶着你，跟大家共勉！

评论区是我对51分钟的文字版总结👇

[Amto @XAMTO_AI]

剑桥这波直接放大招了！！

AI & ML 经典教材全集免费公开，PDF随便下。

想入坑机器学习又不想被割韭菜买天价课的，这十本啃完，地基基本就稳了。

顺序排好了，从入门到进阶：

1️⃣ 机器学习理解，讲透理论和算法，入门首选
🔗 https://t.co/YesyCaPPeu
2️⃣ 机器学习数学基础，数学欠火候的先补这本
🔗 https://t.co/OYdtp3jdrs

3️⃣ 机器学习算法的数学分析
🔗 https://t.co/tjRW5NvCIS
4️⃣ 深度学习理论原理
🔗 https://t.co/rSvcdo0kKu

5️⃣ 神经网络与机器学习
🔗 https://t.co/8zcDBX1Kja
6️⃣ 图深度学习
🔗 https://t.co/zn1066GJkg

7️⃣ 机器学习的算法视角
🔗 https://t.co/YQ86hPbAMc
8️⃣ 概率论：理论与实例
🔗 https://t.co/mx26GfzTFX

9️⃣ 应用概率基础
🔗 https://t.co/xyf6PxsY9X
🔟 高级数据分析
🔗 https://t.co/F2TsKWreDk

说实话，这些书难度都不低，别指望躺着刷完。

但只要你真下定决心啃两三本，比在群里听人吹AI牛逼强一百倍。

[CyrilXBT @cyrilXBT]

Anthropic just dropped 5 workshops on building self-improving agentic systems from scratch:

00:00 - Ship your first Claude agent
36:44 - Build memory for Claude agents
1:05:06 - Make your agent autonomous
1:26:46 - Set up a proactive agent
2:03:35 - self-improving agents (tools,skills)

These 3-hours of free Claude workshops will replace 10 paid agentic courses.

Watch today, then read article below on how to build a self-improving agentic system with Fable 5.

[Tushar 🚀 @tushar_elric]

Watched this and it’s incredible. Crazy amount of info compressed into a dense 20 minutes.

This is exactly how you enjoy Twitter Premium without actually paying for it.

[Roan @RohOnChain]

My friend applied to 200 tech jobs in two years. No PhD. No Stanford.

Last month Anthropic offered him $750,000.

I asked him how he broke in from zero.

He sent me a course that was never supposed to get out. A 3-hour video to build a full LLM from scratch.

A developer teaches you exactly how LLMs like ChatGPT and Claude are actually built.

I watched it last night.

Halfway through, I realized it's embarrassingly simple to break into an AI lab.

Bookmark this and read the article below.

• 00:00 - intro to LLMs
• 05:43 - LLM transformer architecture
• 40:24 - training the LLM
• 1:30:27 - modernizing the LLM
• 2:33:53 - scaling the LLM

[Avi Chawla @_avichawla]

Karpathy's Agentic Engineering finally has proper tooling!

(built by Google)

Karpathy defined agentic engineering as the discipline that separates production agent work from vibe coding. The core skills he listed were spec design, eval loops, and security oversight.

The problem is that production agents don't run on the agent code alone. They need a model provider, a retrieval layer, an eval system, a deployment target, and observability.

And normally, these are scattered across different dashboards, config files, and tools, each with its own setup and learning curve.

The solution to production-grade Agentic Engineering is now actually implemented in Google’s Agents CLI.

It covers the entire lifecycle in a single interface, covering scaffolding, evaluating, and deploying ADK agents.

The setup command injects 7 ADK-specific skills into a coding agent's context, which lets it handle scaffolding, evals, deployment, and enterprise registration through natural language.

I mapped the whole thing end-to-end in the diagram below. Here is what each layer is doing:

> Agent Orchestration defines the agent using the ADK, and the A2A protocol handles coordination when one agent needs to invoke another. ADK is model-agnostic, so the models cover Gemini, Gemma, and third-party models like Claude, via Model Garden.

> The data layer handles retrieval. Vector Store stores the embeddings for RAG, and the ingestion pipeline loads and chunks docs into it. Monitoring and storage happen alongside it.

> The evaluation layer runs the agent against test scenarios with LLM-as judge scoring. It runs before deployment.

> For deployment, you can choose between Agent Runtime, Cloud Run, etc., depending on how much control you need. Agents CLI configures across all targets.

> Finally, teams that need staging and production environments, the IaC and CI/CD layer handles infrastructure as code and deployment pipelines.

Every one of these layers is a service that a dev would otherwise configure manually.

But Google's Agents CLI ties them together to make the full agentic engineering lifecycle practical.

The diagram below makes this clear, and I worked with the Google Cloud team to put this together.

Agents CLI GitHub repo → https://t.co/ho8fGyd8GF

(don't forget to star it ⭐ )

Akshay wrote a hands-on article covering all six steps from install to enterprise registration.

It includes the eval scorecard, the instruction loophole the eval caught before deployment, and what the deployment process actually looks like end-to-end.

Read it below.

[Xudong Han @Xudong07452910]

现在很多人聊 Research Agent，默认期待是：

读论文、找 gap、想 idea、做实验、写 paper。

但耶鲁大学的这篇论文问了一个更深入的问题：

LLM 生成的研究想法，和人类研究者真正做出来的 paper idea，差距到底在哪里？

论文的核心概念叫 research taste，可以理解成「研究品味」：你通常会发现什么样的问题，用什么方式把它变成一个 contribution。

它的实验设计挺巧。

作者从真实论文里抽取 human idea，再反推 4 到 8 篇最可能启发这个 idea 的 prior works。然后把这些 prior works 的标题和摘要给 LLM，让模型在同样的文献上下文里生成一个新的 research idea。

这样比较的重点，就从「单个 idea 看起来新不新」推进到「一批 idea 的分布像不像人类研究者」。

论文一共用了 11,683 个 human ideas，覆盖 ML 会议论文和 Nature Communications 里的自然科学论文；模型包括 Claude、Gemini、GPT、Qwen、DeepSeek 等 9 个设置。

结果很有意思：人类 idea 的分布明显更宽。

在人类论文里，只有 12.1% 的 idea 属于 bridge opportunity，也就是把不同文献、方法或证据流连接起来。

但在 LLM 生成的 idea 里，这个比例变成了 47.1% 到 64.2%。

方法层面也类似。

人类论文里，synthesis / unification 只占 5.1%；LLM 这里是 22.5% 到 38.7%。

简单说，模型很喜欢把研究问题理解成：

这里有两个东西，可以把它们整合一下。

论文还发现，给模型更多上下文也没有明显解决这个问题。full-paper context 版本没有让分布更接近人类；开启 thinking mode 甚至可能让模型更强化这种 bridge-and-synthesis 倾向。

这篇最有意思的地方在于，它没有简单说 LLM idea 好或不好，而是指出了一个更隐蔽的问题：

LLM 可以生成很多看起来合理的研究想法，但这些想法的类型可能高度集中。

这对 AI for Science 和 Research Agent 很关键。

对于真正强的 Research Agent，难点在于能不能形成更宽的 problem-finding 能力：发现 failure、识别机制、提出测量工具、构造系统、修改局部假设，而不总是回到「整合两个已有东西」这条安全路径。

自动科研如果缺少这种多样性，最后可能会变成一种很流畅的同质化机器。

📎 arxiv: https://t.co/GrrulIuJrF

[Rohan Paul @rohanpaul_ai]

A 178 page survey study for refreshing math and generative AI foundations from University of Huddersfield.

The Little Book of Generative AI Foundations.

[鸟哥 | 蓝鸟会🕊️ @NFTCPS]

MIT 那本《计算机科学数学》，真要入坑计算机科学的，这书基本就是天花板了。

别被吓到，虽然它厚到离谱，整整 1048 页，但人家定位就是给初学者看的入门读物。

我知道你在想啥：1048 页也叫入门？我第一眼也懵。可翻进去就懂了，它不是让你一口气啃完，是把最底子的东西掰开揉碎讲明白，慢慢来完全跟得上。

打算真学计算机的，别老收藏各种速成课了，这种硬货啃一本，顶你刷十个短视频。

链接：https://t.co/6Gjui33xpj

[Shubham Saboo @Saboo_Shubham_]

Static agents won't survive 2026.

Self-learning agents are WHAT you need, and the best ones get better every time someone uses them.

Let's understand this.

Your AI Agents actually learn in 3 places:

1. Model (the weights): owned by the AI labs, the hardest layer to improve.

2. Harness (the loop, tools, prompts): medium difficulty. You can improve it, but it takes some real work.

3. Context (memory + skills): plain text you can read and edit, the easiest place to start.

You already use all three in Claude Code and Codex.

But here's what almost everyone skips:

Your agent should learn from your users, not just its own runs.

Every time a person fixes the agent's call, that's a labeled example no benchmark can give you. Worth more than a hundred synthetic tests.

Capture it. Feed it back. That's the loop that compounds.

[Morlex @0xMorlex]

Andrej Karpathy explained the 5 shifts that turned LLMs from chatbots into agentic systems:

00:00 - Memory turns chat into personal AI
6:41- Multimodal AI can read the world
16:58 - Thinking models solve harder tasks
24:51 - Search makes LLMs live
30:58 - Tools turn LLMs into workers

This is not another video about “prompt engineering.”

It is a 40-minute roadmap for the next AI workflow: memory / vision / reasoning / search / tools.

Watch today, then read the article below on how to turn LLMs into self-improving agent loops.

[Anatoli Kopadze @AnatoliKopadze]

Anthropic engineer:

"You're not supposed to prompt Claude. You're supposed to build a system that prompts itself."

In 45 minutes she breaks down how Anthropic builds agents that remember, learn from their mistakes, and get smarter with every run.

Worth more than any paid course you'll find on building agents.

Watch the session, then read the guide on building loops below.

[Codez @0xCodez]

Google just dropped a 1-hour course on agentic engineering from scratch:

00:00 – How to build your first AI agent
08:24 – Build agent memory (short, persistent, long)
28:34 – Agentic loops, long-running AI agents
40:04 – How to build MCP (MCP vs API)
1:00:22 – Multi-agentic systems

This 1-hour watch will replace 10 paid agentic courses on the internet.

Watch it today, then read how to build a self-improving agentic system in the article below.

[Miles.Ma @ma_zhenyuan]

太离谱了，今天拿着这个去企业宣讲，成了40万的单

你把这个删了吧，别让太多人看到了

[Geek @geekbb]

Hugging Face 出的极简智能体，手把手教你从零造一个属于自己的 coding agent。学习写 agent 的入门玩具。

https://t.co/7nBeX7B2DE

[AI Will @FinanceYF5]

Andrew Ng刚放出一条3小时的"如何在2026年成为AI工程师"教程视频。

内容据称覆盖搭建agentic AI系统、AI工程的未来方向、完整的AI提示词课程，还有一段用AI在30分钟内做出应用的实操演示

[codila @0xCodila]

Claude Code team just dropped a free course on loop engineering with Fable 5:

00:00 - how Claude Code works under the hood
05:01 - the agentic loop explained
16:21 - the feature 99% of devs miss: auto mode
19:01 - why voice beats typing
32:34 - auto code review with draft PRs
58:39 - Fable 5 for non-code work

this free course replaces every paid Claude Code tutorial

watch today, then read the article below on loop engineering by Karpathy

[Moysei @0xMoysei]

Andrej Karpathy quietly published 9 rules for building AI agents.

Rule 1: stop writing prompts.

"If you find yourself iterating on a single message at 3 in the morning, you are still in the prompting era."

A friend who runs agent infra at a trading firm read it once and deleted half the harness his team built last quarter.

The whole paper argues most agents die from a weak harness, not a weak model. Everything you added to compensate for the model becomes dead weight the moment the model improves.

The rule near the middle, about letting the loop delete its own work and start over, is the part he screenshotted. It contradicts how almost everyone builds right now.

The closing section on where the bottleneck goes next is the whole paper in one line.

He said he read it twice, second time with his own repo open beside it.

Everyone is still tuning prompts. Karpathy already moved on.