Start learning

Start here

What to learn first, and why the order matters

A course can be full of good explanations and still fail, because a screen you weren't ready for teaches nothing — and you blame yourself rather than the order. Order is most of the difficulty.

Why "just build something" fails here

It's good advice for web development, where you can see what you built and the feedback is immediate and honest. It's poor advice for AI, for one reason: a broken AI system still produces output.

A broken web page looks broken. A model that has learned the wrong thing returns confident, plausible, well-formatted answers, and nothing on the screen says otherwise. Build first and you can't tell a working system from a broken one, so the loop that makes project-based learning work never closes.

Two build-and-check loops side by side. With a web page, the breakage is visible, so the loop returns to a fix. With a model, the output looks fine either way, so the return path is broken and nothing is learned. A broken web page You build it It looks broken You fix it the loop closes A broken model You build it Output looks fine You learn nothing the loop never closes
Project-based learning runs on visible failure. AI hands you a confident answer whether or not anything worked.

So understanding comes first here in a way it doesn't elsewhere. You need enough of a model in your head to judge what you built before building teaches you anything.

The two failure modes an order has to avoid

Understanding without building gives you someone who talks about AI fluently and has never met a real dataset.

Building without understanding gives you someone who copies snippets, gets something running, and can't debug it — because debugging needs a picture of what should be happening.

So interleave them, at a fine grain: understand a mechanism, immediately do something that depends on it, and let the doing expose the parts of your understanding that were fake.

The order, and the reason for it

Five stacked layers, widest at the bottom: understand AI, then computing and math, then machine learning, then building and shipping, then robotics at the top. 5 · Robotics 4 · Building and shipping 3 · Machine learning 2 · Computing and math 1 · Understand AI
Every stage is graded on the vocabulary of the one below it. Start halfway up and each new topic arrives as a noun to memorize.

1. Understand AI

What learning from data is, how it differs from writing rules, why a model can be confidently wrong. Why first: every later stage assumes this vocabulary.

2. Computing and math foundations

Code, and the math AI leans on — not all of math, the parts that keep turning up. Why second: meeting a matrix operation after you've seen why models need one is a different experience from meeting it cold. Same content, different retention.

3. Machine learning and modern AI

Classic algorithms first, because they're simple enough to understand completely and the intuitions transfer upward; then language, vision and generative models. Why here: this is where you find out the algorithm was never the hard part. The data and the evaluation were.

4. Building and shipping

Turning a model into something real, safe and maintainable — everything that's true of production systems and false of tutorials. Why here: the gap between "it works in a notebook" and "it works for users" is where most of the profession lives, and you can't see it until you've stood on both sides.

5. Robotics and physical AI

Intelligence with a body, in a world that hasn't read your documentation. Why last: robotics compounds everything above and adds physics, timing and noise. Reality doesn't negotiate with a plausible answer.

Two orders that are wrong, and one that only looks it

  • Math first, everything else later. Sensible-sounding, and it loses more beginners than any other plan. Math with no question attached is just symbols.
  • Newest thing first. The latest architecture is built out of ideas from every earlier stage, so starting there means memorizing a diagram.
  • A hard idea early, on purpose. This one only looks wrong. A hard idea you've been given the tools to attack is worth a lot — as long as the difficulty is the idea and not missing vocabulary.

"But I only care about one thing"

Fair. Plenty of people want just enough to use AI well in the job they already have, with no intention of training a model. That's a shorter route, covered in Learning AI while holding down a full-time job — and it still needs stage one. "How do I use this tool" without "how does this tool fail" produces someone who can't tell when it has handed them something wrong, which is the risk of using AI at work.

Next