Tips for Choosing a Curriculum from Someone Who Took AI Courses for 6 Months
Since the second half of last year, I've taken various AI-related courses totaling about six months. I spent roughly 1 million won, and honestly, I regretted half of it. That experience gave me some criteria.
The biggest regret was the "theory-focused, finish in 6 weeks" type. It was structured as three weeks of math explanation, then at the end you run one model and that's it. But even the code for running that model wouldn't execute because the library versions didn't match.
By contrast, the courses I was satisfied with had common traits:
- You get working code from the first week
- The instructor mentions painful struggles from real work
- The output can remain as a portfolio piece
- The curriculum includes the words "deployment" or "serving"
The last item is especially important, surprisingly: courses that end at the learning stage don't give you a sense of what to do in real practice. Even just wrapping it in a simple API and deploying it made the next step visible.
These days, when I buy a course, I look at its outcomes before the curriculum. If there are no student outputs, or if they're all similar handwriting recognition demos, I tend to skip it. Hope this helps.
10 answers
If it has the words “deployment” or “serving” in it, I just throw it into the candidate list. Agreed.
Half of 1 million won being regret… I was the same last year lol. I took a 6-week theory course and had nothing left to show for it.
I really relate to this. I believed the advice that you have to start with math, so I clung to a graduate-level textbook and wasted three months. If I’d spent that time writing even one more line of working code, I’d be building something by now. In the end, you learn by fumbling through things, and lectures seem to compress that fumbling. Now I get why it matters to have something running in the first week.
Well, that’s a bit... Courses that even cover deployment are often superficial. Many just show you how to 'get it up and running' without explaining why it works that way. I think filtering based only on whether the word 'deployment' appears is a pretty rough criterion.