Stanford Teaches You to Build an LLM From Scratch. Just Don't Let the AI Do It For You.

These days, finding a developer who doesn’t ask an AI for help writing code is harder than finding one who does. So here’s a twist: there’s a university course that asks students to build the very thing powering those AIs — a large language model — entirely from scratch. It’s Stanford’s CS336. And the most interesting part is that this course hands students a separate set of rules for how to use AI coding tools. The dilemma of teaching AI in the age of AI is compressed right here.

What Makes CS336 Different

The official name is “Language Modeling from Scratch.” It means exactly what it says. Students write the tokenizer, stand up the transformer architecture by hand, run the training loop, and clean the data — the entire pipeline, built personally.

That’s the opposite of most AI courses, which focus on how to use models someone else already built. CS336 is about opening the black box. The goal is to make you understand what’s happening inside a model, one line of code at a time.

The appetite for this isn’t confined to a classroom. On YouTube, a video titled “Build a Small Language Model From Scratch” has racked up more than 130,000 views. People clearly want to go one step deeper than calling a finished API.

So Why Put Guardrails on AI Coding Tools

Here’s where it looks contradictory. A course that teaches you how to build AI puts conditions on using AI to do it. The modern convention is to drop a guide file like CLAUDE.md into a project — a written contract that spells out what an AI tool is and isn’t allowed to do. CS336 sets up its rules the same way.

The reasoning is simple. Ask an AI to “implement a transformer” and you get working code in five minutes. The problem is that the student learns almost nothing during those five minutes. The entire point of CS336 — the experience of writing it yourself, hitting a wall, and debugging your way out — evaporates.

So the guideline isn’t “don’t use AI.” It’s closer to “do the core learning by hand, and lean on AI only for the supporting work.” Separate the part where you build understanding from the part that’s just mechanical repetition.

The Real Dilemma of Learning AI With AI

This isn’t really about one course’s policy. It’s a question facing education broadly. When an AI can solve the assignment for you, what do you still need to learn by hand?

The answer keeps converging on two things: muscle memory and the ability to verify. To judge whether an AI’s transformer code is right or wrong, you first have to understand transformers yourself. Copy down the AI’s output without that understanding, and the code might run — but you’ll have no way to check it.

It’s telling that a recent lecture from Stanford’s CS230 on “AI career advice” has been viewed more than 420,000 times. In an era when AI seems to do everything, the question “then what do I actually need to bring to the table” feels urgent. CS336’s guideline is one practical answer: fundamentals by hand, applications with tools.

What This Means for the Rest of Us

CS336’s approach travels well beyond campus. Every developer using AI coding tools carries the same tension. The smarter the tool gets, the more it matters whether the user has a baseline to judge what comes out of it.

What’s notable is how this balance is increasingly being written down. Whether it’s a CLAUDE.md in a company codebase or a syllabus guideline in a course, there’s a clear move toward documenting which territory you hand to the AI and which you keep in your own grip. Everyone seems to feel that an unspoken understanding isn’t enough.

One honest caveat: this isn’t a hot topic that blew up in the community over the past month. It’s closer to a slow, persistent worry running through education. Better read as a trend than a single event.

In the end, it narrows to one question. What do you hand off to the AI, and what do you keep your hands on until the very end? If you’ve never drawn that line for yourself, this is a good moment to look hard at how you actually work.