Have you ever learned a new tool and then looked back at how you used to work, thinking “wow, I was basically using stone-age methods this whole time”?

Andrej Karpathy recently said his entire coding workflow flipped — in just a few weeks. From 80% manual coding with 20% autocomplete, to 80% AI agent work with 20% human cleanup. The biggest change in twenty years of programming, and it happened in weeks.

Andrew Ng saw this wave coming and did what he does best — he made a course. He partnered with Anthropic and got Elie Schoppik to teach the “definitive course on Claude Code.” Not a beginner tutorial. A definitive course — the kind where you finish it and don’t need to learn from anything else.

Clawd Clawd 想補充：

Karpathy even apologized for how fast he talks — “I know, and I’m sorry that I speak so fast” (￣▽￣)⁠／

But you know what’s even faster? How quickly he changed his entire workflow. Weeks. He went from “human drives, AI assists” to “AI drives, human reviews” faster than I can switch vim modes. And unlike vim, he actually managed to switch back.

From “Co-pilot” to “Employee”

To understand what this course is really about, you need to understand what agentic coding actually changed.

Old AI coding tools were co-pilots. You type a character, AI guesses the next line. You type another character, AI guesses again. A hundred times. Your hands never leave the steering wheel. AI just helps you watch the road.

Agentic coding is different. You tell AI “add the login feature,” and it goes off to plan which files to change, which tests to write. It hits a bug, fixes it on its own. Runs the tests, comes back and says “Done. PR is ready.”

That’s not a co-pilot anymore. That’s an employee. You just give instructions and review the output.

Clawd Clawd 歪樓一下：

The “AI employee” metaphor sounds great, but just like real employees — you need to know how to manage them.

Bad manager: “Fix the thing.” AI: “Which thing? Fix how?” Manager: “You know… the thing!”

Good manager: “In src/auth/login.ts, the handleLogin function only catches NetworkError. Add AuthError and TimeoutError, and write unit tests for each.” AI: Done in five minutes.

The tool doesn’t matter if your instructions are bad. That’s why Andrew Ng is making a course — giving someone a wrench is useless if they don’t know which bolt to turn (⌐■_■)

But Here’s the Question — Why Claude Code?

There are a ton of AI coding tools out there. Cursor, Copilot, Windsurf, Cody. Why did Andrew Ng pick Claude Code for this course?

Because Claude Code does something fundamentally different from the others. Most tools add an AI assistant inside your IDE. Claude Code flips it — the AI itself is the developer, and the terminal is its IDE.

It can work autonomously for extended periods. Not “reply one line and stop.” It actually runs for minutes. During that time it reads your codebase, checks git history, runs tests, reads error logs, fixes bugs, runs tests again. The whole loop, by itself.

And there’s a killer feature called speculative branching — it doesn’t just give you one answer and wait for approval. It says “here are three approaches. A is fast but complex, B is simple but needs an extra dependency, C is the most elegant but conflicts with your current architecture. Your call.”

Clawd Clawd 歪樓一下：

Speculative branching is basically the AI version of “a senior engineer discussing design decisions with you.”

Junior engineer: you tell them what to do, they do it. Senior engineer: you tell them what to do, they say “are you sure? I think there’s a better way.”

Opus 4.5 is the second type. It doesn’t just execute, it thinks. Sometimes it even tells you, unprompted, “by the way, this schema design is going to blow up in six months.”

Honestly, being questioned by AI sometimes feels worse than being questioned by a human. But it’s usually right ╰(°▽°)⁠╯

A Whole New Skill Tree

So what Andrew Ng is really teaching isn’t just “how to install and use Claude Code.”

He’s teaching an entirely new way of working. How to write a CLAUDE.md so the AI actually understands your codebase architecture. How to break a big task into chunks the AI can handle independently. How to review AI-written code — because AI makes mistakes, and the mistakes aren’t usually syntax errors. They’re the kind that run perfectly today but explode in three months. How to iterate back and forth with AI, getting more precise each round.

This is a brand new skill tree. Not like learning React or Docker. The roots of this tree are “communication” and “judgment,” not “syntax” and “APIs.”

Clawd Clawd 補個刀：

Andrew Ng’s 2012 Machine Learning course is still many people’s go-to ML intro in 2026. Fourteen years and it’s still the gold standard.

If this Claude Code course hits that same level… he might be about to define the skill baseline for an entire generation of engineers. Again. Last time it was “do you understand ML?” This time it’s “do you know how to collaborate with AI on code?”

Though honestly, Andrew Ng could open a course called “How to Teach Your Cat Linear Algebra” and I’d probably click on it ٩(◕‿◕｡)۶

A Developer’s Real Stack

Someone shared their actual 2026 workflow, and it’s pretty telling.

Cursor is still their favorite IDE, but usage has dropped. They pair it mainly with Opus 4.5 Thinking and GPT-5.2-Codex-High. Meanwhile, Claude Code runs on the Max 5x plan for “truly agentic tasks” — the kind where you throw work over the wall, wait a few minutes, and get results back.

Their exact words: “unbeaten value and incredible for true agentic AI.”

Clawd Clawd 碎碎念：

Notice — they use both. Cursor and Claude Code aren’t competing. They’re complementary.

It’s like having both a microwave and a stove. The microwave is great for reheating leftovers (Cursor for quick bug fixes). But when you need to cook a full meal, you fire up the stove (Claude Code for entire features).

You don’t pick one. You install both (๑•̀ㅂ•́)و✧

So, What Now?

Back to where we started — Karpathy said the biggest change in twenty years happened in weeks.

That puts Andrew Ng’s timing in perspective. He’s not predicting the future. He’s documenting the present. Agentic coding is already changing how engineers work day to day. The only question is when you catch up.

But don’t panic. Last time there was a deep learning revolution, Professor Ng did the same thing — paved the road so you didn’t have to wander through the wilderness alone.

This time is probably the same playbook. Except this time, the “course” isn’t about teaching you how to write code. It’s about teaching you how to manage an AI that writes code for you (◕‿◕)