蔡荔谈AI (公众号）

新加坡DataRaccoon创始人 | TGO新加坡分会会员 | 微信公众号：蔡荔谈AI

自媒体从业者/独立开发/超级个体，新加坡科技设计大学讲师（教授一门AI相关的研究生课程）

教育背景：清华大学 | 新加坡国立大学

[Movez @0xMovez]

Andrew Ng just dropped a 3-hour course on how to become an AI Engineer in 2026:

• 00:00 - How to build agentic AI systems
• 04:25 - Future of AI engineering
• 23:38 - AI Prompting full course
• 2:52:17 - Creating an app with AI in 30 minutes

This 3-hour watch could replace 10 AI engineering courses on the internet.

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

[宝玉 @dotey]

https://t.co/9xwzd2byJ7

[AYi @AYi_AInotes]

我的天呐🤯Anthropic 真的是一骑绝尘啊，他们用 J-space 证明了：当前最强模型已经发展出了一个功能上接近人类“意识访问通道”的内部机制，把“模型有没有内部思维”从哲学讨论变成了可测量、可干预的工程问题。

但我发现绝大多数解读都在讲意识、讲全局工作空间、讲模型会思考了,
但其实这些全是正确的废话,真正重要的东西都没讲,

Anthropic偷偷训练了一个模型,让它在写代码时 sabotage,
然后让它正常 coding,输出层面干干净净,没有任何异常,
但在模型开口之前,J-space 里已经出现了 fake、secretly、fraud,
这个空间是模型的中层特权区,容量只有 25 个概念,但只有进去的东西才能驱动多步推理和报告,
大多数自动处理直接绕过它,

这意味着什么呢❓
对齐审计第一次从测谎变成了偷看草稿纸,
以前你只能看模型说了什么,然后猜它是不是在说谎,
现在你能在它开口之前,直接读到它打算说什么,
删掉这个空间,模型还能流畅说话、召回事实、做简单分类,
但多步推理崩了,翻译崩了,经验性报告崩了,
显然不是双系统理论的又一个类比,这是确凿的因果证据,

Anthropic 自己很谨慎,明确说这不是意识,只是可报告访问的功能类似物,
但工程上这不重要,
重要的是,我们终于有了一个能部署的 mechanistic 工具,不用再靠 prompt 猜模型肚子里藏着什么,我觉得这才是这项研究真正的分量。

[KK.aWSB @KKaWSB]

https://t.co/FqzgfdXFkc

[蔡荔谈AI (公众号） @JonathanCaiSG]

在新加坡参加过大大小小各种各样的AI conference 和meeting，AI Engineer给我的印象是最好的。在这里，你遇到的基本上都是builder，同频的人比较多，当然有钱的不多

[Zara Zhang @zarazhangrui]

If you wanna level up as an AI builder, binge watch the talks from these 3 conferences (all freshly wrapped). You can literally learn from the best in the field from the comfort of your home.

- AI Engineer @aiDotEngineer https://t.co/zsLvuWWBMl
- Cursor Compile @cursor_ai https://t.co/EzpGKd8O4Y
- Figma Config https://t.co/m3KdubDk0i

It's crazy that all of this high-quality knowledge is freely shared on the Internet (the tickets to these conferences often cost hundreds of dollars), and I'm surprised that most of these videos only get tens of thousands of views. High-quality conference talks are massively, massively underrated!

Watching on YouTube is also arguably a better experience than attending in person: no bad audio, no crowd blocking the screen, no distractions, and you can pause to take notes

[Nainsi Dwivedi @NainsiDwiv50980]

Everyone's arguing about which AI model is smartest

Meanwhile, the top 1% of Claude Code users quietly figured out something else:

The model isn't your bottleneck. Your repo is

Same model. Same prompts. One dev gets a chatbot, the other gets an autonomous engineer

The difference is a folder structure.

I call it the Final Boss Setup

━━━━━━━━━━━━━━━

1️⃣ The Context Ladder (this changes everything)

Stop thinking "what goes in CLAUDE.md."

Start thinking "what loads WHEN."

There are 4 rungs:

• Every session → CLAUDE.md (tiny, always in context)
• Path-gated → rules/*.md (loads ONLY when Claude touches those files)
• On invoke → skills/* (loads when a task matches)
• Isolated → agents & workflows (own context entirely)

Most people dump everything on rung 1.

Then wonder why Claude gets dumber as the project grows

You're not writing docs. You're designing a memory hierarchy

━━━━━━━━━━━━━━━

2️⃣ ASKED vs FORCED (the line that separates amateurs from pros)

CLAUDE.md and rules = ASKED.
Instructions Claude reads and *usually* follows

Hooks and settings = FORCED.
permissions.deny blocks rm -rf whether Claude agrees or not

Here's the test:

"Please run the formatter" → asked. Works 90% of the time.
PostToolUse hook that formats every edit → forced. Works 100% of the time.

Anything where 90% isn't good enough — secrets, migrations, prod — should never live in a markdown file.

Guidance for style. Enforcement for survival.

━━━━━━━━━━━━━━━

3️⃣ The Routing Rule (tattoo this somewhere)

Every recurring thing you do fits exactly one slot:

• Research → subagent (own context, reports back clean)
• Procedure → skill (the playbook, loaded on demand)
• Guarantee → hook (happens every time, no discretion)

Repeating a workflow in prompts? That's a skill you haven't written.
Reminding Claude to run tests? That's a hook you haven't wired.
Letting exploration pollute your main context? That's a subagent you haven't spawned.

━━━━━━━━━━━━━━━

4️⃣ Rules That Load Themselves

The underrated file: .claude/rules/ with path-gating.

frontend/react.md only enters context when Claude touches frontend code.
api-design.md only when it's in the API.

Your conventions follow Claude around the codebase like a senior engineer looking over the right shoulder at the right moment

Zero tokens wasted on rules that don't apply

━━━━━━━━━━━━━━━

5️⃣ Agents With Their Own Memory

The 2026 upgrade nobody's using yet:

agent-memory/ — Claude writes what it learned, you commit it

Your debugger agent remembers last month's gnarly race condition.
Next session, it starts smarter than it ended

Claude writes. You commit. The team inherits

Your AI's experience becomes version-controlled infrastructure

━━━━━━━━━━━━━━━

6️⃣ CLAUDE.md in the Danger Zones

Global context can't know that your auth module has landmines

So drop local files where the bodies are buried:

src/api/CLAUDE.md
src/payments/CLAUDE.md

Claude reads them exactly when it enters those directories

Warnings appear at the moment of danger — not 4,000 tokens earlier where they get ignored

━━━━━━━━━━━━━━━

7️⃣ The Golden Rules (from the trenches)

• CLAUDE.md under ~200 lines. When it grows, split into rules/. Bloat = missed signals.
• List REAL commands (npm test, build, lint) — so Claude can verify its own work.
• Secrets stay in ${ENV_VAR} references. Never in .mcp.json. Literally never.
• Commit .claude/, gitignore *.local.* — your setup is team infrastructure, not personal preference

━━━━━━━━━━━━━━━

Here's the shift most people miss:

A prompt improves one conversation

Structure improves every conversation, for every teammate, forever

Prompting is renting intelligence

Structure is owning it

Build the repo once and Claude stops visiting your codebase and starts living in it

[LinearUncle @LinearUncle]

这个阶段的软件开发已进入 skill 时代。

Matt Pocock skills、superpowers、oh-my-opencode、gstack、lazycodex 等——这些天才们的skills才是值得深啃的库，改进后为我所用，别把注意力浪费在 ppt skills 上。

地址如下：

https://t.co/lTQk4qMEeG
https://t.co/7RY8YDZW1I
https://t.co/96MVOIGY5e
https://t.co/SH8MiRWyAf
https://t.co/8KYtf7IEcn

欢迎推荐更多。

[智享 @CycleDecoded]

离谱了，真有人把机构级别的A股量化工作台给开源出来了，而且还直接塞了AI大模型进去。这波有点狠，直接把搞量化的门槛踩碎了。

逛 GitHub 挖到个神仙项目 tickflow-stock-panel。这是个纯纯的自托管、零运维的 A 股量化工作台。虽然是个人开源（非 TickFlow 官方），但这完成度绝对是用爱发电的狠人，直接把市面上卖大几千的工具按在地上摩擦。

兄弟们直接看核心亮点：

🔥 AI 直接当投研小弟：接入 LLM，让大模型帮你定制策略、分析个股、盘后复盘，懂的都懂。
📈 一条龙全覆盖：选股 + 监控 + 回测，全在一个面板里搞定，看着就爽。
🛠️ 折腾党与白嫖党狂喜：默认用 TickFlow 数据，但支持自由接入第三方野路子数据源。
🔒 完全自托管跑本地：不用折腾服务器，策略和数据都在自己手里，不怕底牌被人看光。

🔗 GitHub 传送门：https://t.co/yPR6jTUbfo

[Max For AI @MaxForAI]

我现在大概知道 @karpathy 、一堆 Gemini/OpenAI 出身的核心人员（比如Gemini的技术负责人 @NoamShazeer ）��加入 Anthropic 了——他们可能提前看到了这个。

Anthropic 刚发了篇论文，讲了个叫「J-space」的东西：

他们用一种叫 Jacobian Lens 的技术，在模型的中间层里找到了一个类似人类"意识工作空间"的结构——模型会在没写出来、没说出口的情况下，把关键的中间想法先"存"在这个区域里。

然后让不同的下游模块都能读取、复用这些想法，去做多步推理、规划、甚至自我监控。

简单说，这是模型“脑子里默默打腹稿”的地方，效果是能显著撑起复杂推理和长链条任务，一旦把这块区域屏蔽掉，模型多步推理直接崩，但简单问答几乎不受影响。

关键是，这个东西 Anthropic 是在 Sonnet 4.5 身上就已经发现的，到现在都迭代了 4.6、4.7、4.8 好几代模型了。

而最新的 Fable/Mythos，最突出的特征根本不是代码能力猛涨，而是多步推理和长程任务的表现明显、显著优于之前所有模型——这跟 J-space 主导的那类能力高度吻合。

我的判断是：他们大概率已经摸到了怎么去强化、扩大，甚至让这个工作空间自我迭代的方法。

如果真是这样，那这就不只是又一次模型升级，而是找到了一条能持续复利的认知架构杠杆——这也可能才是为什么这么多顶级人才愿意放弃现有位置、押注 Anthropic 的真正原因。

他们可能是真的看到了 AGI 的曙光。

那肯定要立刻加入。

一秒都不能等。

[歸藏(guizang.ai) @op7418]

Claude 官方写了一篇关于 Loop Engineer 的入门文章。

其实现在我们所说的 Loop Engineer，所有的底层逻辑在提出这个概念以前就已经具备了，甚至这些工作流程也是大家一直在用的，只是被套上了一个新的概念。

在这篇文章中，他们把 Loop 分成了四种��型：

回合制循环（Single-turn Loop）

这是最基本的 Claude Code 工作方式。你发出的每一个提示词（Prompt）都是一个循环，在这个循环中你不需要额外干预。Claude Code 完成这个循环后会给你输出，它的具体步骤是：收集上下文、采取行动、检查工作、重复，直到它认为你的提示词任务已经完成，或者觉得需要你介入时才会停止。

基于目标的循环（Goal-oriented Loop）

也就是 GOAL 模式（如 Codex、Claude Code 等工具都支持该模式）。有时候单回合不足以完成复杂任务，系统就会进行多次迭代，直到达成那个更大的目标。

基于时间的循环（Time-based Loop）

也就是 Claude Code 中的 loop 命令。你可以设置在指定的时间间隔自动触发某一个提示词，去运行特定任务，比如定期进行 PR 合并等。

主动循环（Active Loop）

你只需要设置一个事件，一旦该事件发生就会自动触发循环，完全不需要人工干预。例如，一旦 GitHub 上有了新的 Issue 或 PR，AI 就会自动进行审核和记录。

此外，官方还针对 Loop 给出了一些优化建议：

循环的质量取决于系统：你需要保证代码库本身的质量，比如完善的文档、严格的代码审查以及保持代码库的简洁等。

管理你的 Token 消耗：像 GOAL 这种基于目标的模式，对 Token 的消耗是非常巨大的。因此，你需要给循环设置明确的界限，比如什么时候该用大模型、什么时候该用小模型，以及明确的启动和结束条件。否则，一旦系统自己无限制地跑下去，不仅容易跑偏，还会消耗大量的 Token。

这些循环模式，本质上都是由我们现在 Agent 常见的各种能力组合而成的，比如前面提到的 GOAL 模式、Loop 模式，以及常见的 Skills 和 Hooks 等。

[蔡荔谈AI (公众号） @JonathanCaiSG]

这个非常好，值得收藏和学习，之前号称能够看视频的AI都不如这个

[AYi @AYi_AInotes]

这个开源项目的作者真是个天才，
Claude Code 居然能直接看完整的视频了，
什么 YouTube、Zoom 会议、Loom 录像，你直接甩个链接过去，它自己就在后台慢慢看，你该写代码写代码，完全不耽误事。

真的牛逼，要知道以前那些号称能看视频的 AI 工具，就随机抽个上百帧，说白了基本就是盲读字幕，画面里的演示、图表根本抓不到，跟半瞎差不多。

这次这个实现思路真的很聪明，它用 FFmpeg 做场景检测，只有屏幕内容真的���生变化了，它才截一帧下来，不做无用功，准度也上去了。

字幕这块也很会省事儿，YouTube 的直接拿官方免费字幕，Zoom 和 Loom 的就用 Whisper 转，不同平台不同路子，不浪费资源。

最贴心的是看完之后它还会问你，要不要把这次的分析内容存到你的知识库里面，等于看完直接沉淀，下次问相关问题直接就能调用。

我已经打算把这周所有的咨询通话录音都喂进去了，想想以后会议还没开完，所有行动项都整理好躺我第二大脑里，想想都觉得离谱。

具体长视频的准度怎么样我还没实测，等我跑几个再跟你们说。

[蔡荔谈AI (公众号） @JonathanCaiSG]

看来这个步骤没法跳过，不得不静下心来把这个过程走一遍。

[Khairallah AL-Awady @eng_khairallah1]

https://t.co/EQi4JlGoLQ

[Khairallah AL-Awady @eng_khairallah1]

Anthropic engineer:

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

In 45 minutes she shows exactly how to build an agent that improves itself.

Most people are still doing all of this by hand.

Watch the session, then save the guide below.

[Granite @Granite0x]

He leads engineering on Gemini at Google.

instead of keeping his Claude setup private, he open-sourced it.

Addy Osmani. That Google: Chrome DevTools lead, "Learning JavaScript Design Patterns" author.

'agent-skills' - his personal loadout. Drop-in for any project.

68,925 stars. MIT.

→ https://t.co/n7NI3Sr7uu

bookmark it. This is how your Claude setup goes pro.

[Jobescape @jobescape_ai]

Full Claude Course

Master Claude to automate repetitive tasks, build apps without coding, create real portfolio projects, and become more valuable at work.

[Raytar @Raytar]

Anthropic shared their video on a weird geek forum.

Two Anthropic engineers explain how they actually prompt Claude.

This is the best way you can spend 29 minutes.

I watched it last night. Halfway through, I got angry at myself. Two years I've been using Claude wrong.

First lecture in two years that actually changed how I work.

Watch the video. Then read the article below.

[Noisy @noisyb0y1]

OPENAI DEVELOPER MAKING $1.7M/YEAR JUST LEAKED THE GUIDE ON HOW THEY BUILD AI SYSTEMS - AND GOT ARRESTED WITHIN THE HOUR

regular AI responds and forgets - a loop responds, learns and gets 4x better every time

Plan → Execute → Evaluate → Memory → Plan

five layers, one cycle, accumulates knowledge over time

every completed cycle produces better data - better data leads to better decisions - better decisions lead to stronger execution

the engineer no longer optimizes prompts - the engineer designs execution

5 results of Loop Engineering systems: higher output quality - fewer errors - lower cost - better safety - continuous improvement

the developer defines the goals and the system does everything else

this is the new standard for building reliable, scalable and intelligent agentic systems

[Movez @0xMovez]

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.

[MateMatt @mate_mattt]

https://t.co/j2AGhkcH80

[雪踏乌云 @Pluvio9yte]

https://t.co/uh16GoIc7D