Everyone says “learn AI.” No one says how.

Everyone says, “Learn AI.” No one says how. So students collect the nearest proxy: certificates. I did it too , I have the LinkedIn badges to prove it. They taught me vocabulary. They did not teach me engineering.

The uncomfortable truth is that the best AI education right now isn't coming from universities or expensive bootcamps. It's sitting in the open: model documentation, engineering blogs, papers with public code, and the error messages you get at 2 a.m. when your own project breaks.

The loop that actually works

1. 01Pick a problem that annoys you personally. Not a dataset , a problem. Mine were bad prompts and unreadable medical reports.
2. 02Build the smallest real version. Real means someone else can use it at a URL, not that the notebook runs.
3. 03Hit the wall. The wall is the curriculum: CORS, hallucinations, messy data, cold starts. Every wall is a chapter no course sells.
4. 04Write down what the wall taught you, publicly, in plain language. If you can't explain it, you haven't finished learning it.
5. 05Repeat, slightly bigger.

A certificate says you watched. An artifact says you built. Hiring managers can tell the difference in about four seconds.

Why almost nobody does this

Because the loop has no syllabus, no deadline, and no one to blame. A course tells you you're 60% done. A half-built project just sits there, quietly asking whether you'll come back tomorrow. That discomfort filters out most people , which is precisely why the ones who tolerate it stand out.

I'm not writing this from the summit. I'm a few loops in, and my projects still have rough edges I can list from memory. But every loop, the walls get more interesting , and that, as far as I can tell, is what progress in this field actually feels like.