Where I would start with AI today
Uni colleagues keep asking me where to begin with AI. Here is the honest answer, in five steps, none of which is a course.
People in my study group keep asking me where to start with AI. The question usually means “which course should I buy”, and my honest answer disappoints them, because it is shorter and cheaper than a course. Here it is, written down so I can send a link next time.
Step one: pick one model and use it every day. Any of the big ones, it genuinely does not matter which. Use it for real things: your homework, your code, your emails, explaining your own spreadsheets back to you. Do this for a month. You will learn more about what these systems can and cannot do than any course would teach you, because you will catch them being brilliant and being confidently wrong on things you actually know about. Both lessons matter, and the second one only lands when it is your own topic.
Step two: keep a list of what you cannot explain. Every time you nod along to a term, write it down. Attention, embeddings, context window, hallucination. The list will grow fast and that is fine. It is not a list of your failures, it is your personal curriculum, and it is better than any generic roadmap because every entry on it is something you actually met in the wild.
Step three: read explanations before papers. I wasted evenings trying to read famous papers cold. The trick is embarrassingly simple: read a good blog explanation first, then the paper. For how these models work, start with The Illustrated Transformer, and only then try Attention Is All You Need. The blog gives you the picture, the paper gives you the precision. In that order it works. In the other order it is decoding notation with no picture in your head. (By the way, this picture-first idea is exactly what I am trying to build with Sushi & Syntax. You are looking at step three.)
Step four: build one small, real thing. Not a tutorial project. Something over your own material: your notes, your files, a dataset from your studies. Make the model answer questions about it. It will work in the demo and then fail on the third real question, and figuring out why is where the actual learning lives. My first attempt confidently quoted the wrong half of a table at me. I learned more from that one failure than from a month of reading.
Step five: ignore the news cycle. A new “best model ever” ships roughly every three weeks. If you chase them you will spend your energy on comparison threads instead of on learning. Check a weekly summary, keep using your one model, and switch only when something clearly beats it at your own daily tasks. The fundamentals you are learning do not change between releases. That is what makes them fundamentals.
There is no step where you buy a course, and no step where you read about AI without touching it. That is the whole trick. The people I know who got good at this all did some version of the same loop: use it daily, collect your gaps, close them one at a time, build something that can embarrass you.
And when you understand something properly, explain it to someone else. Nothing exposes a half-understood idea faster than a friend asking “but why?” twice in a row. This site is my version of that step.