How I'm actually going to do this
In #001 I said I'd do this by building, not by going back for a degree. Fine words. This is the actual plan behind them: vague enough to survive contact with reality, specific enough to count as a plan.
The internet has roughly ten thousand "learn AI" tutorials, and working through them in order is a wonderful way to spend a year learning things nobody is hiring for. So I'm going backwards. I pull real job descriptions for the roles I actually want (AI engineer, forward-deployed engineer, applied-AI work) and read them for what they keep asking for. The requirements that show up again and again become my syllabus. Learn what the job needs, not what a course wants to sell.
I've built with an API before, to make an earlier tool work, but I did it by instinct: copying patterns until things ran, without really understanding why they ran. So the first job is to go back and build on that from proper understanding of how it actually works, not gut feeling. After that, the rough progression I have in mind is agents and evals, the pieces that turn "calls an API" into something closer to real engineering. I say rough on purpose: I already doubt the order will survive contact with reality. Real problems don't wait their turn, and I've a strong suspicion the later stuff will come barging in early whether I'm ready for it or not. The sequence isn't really the point. The point is that each skill gets learned by building something that genuinely needs it.
Not one big impressive project, but a series of small ones, each a notch harder than the last, the way an actual developer levels up. Ship it, break it, fix it, write down what I learned, then start the next one. Small enough to finish; hard enough to teach me something. The finished tools are the proof. The mistakes on the way to them are the education.
The lazy way to use an AI tutor is to let it hand me the answers. That teaches me nothing and produces a "developer" who falls apart the moment the model is wrong. So I've asked Claude to do the opposite: challenge me, quiz me, push back when I reach for the easy path, and make me justify a decision before it'll agree with it. Used like that it's less autocomplete and more demanding colleague. Knowing how to work with these tools without leaning on them is, conveniently, a large part of the job I'm aiming at.
Building teaches you how; it doesn't always teach you why. So I'm stacking proper structured courses next to the practice, starting with Anthropic Academy's, since they're free, official, and pointed straight at the stack I'm learning. The building shows me what works. The courses tell me why it works that way, and catch the gaps I wouldn't know to look for.
That's it. I fully expect to deviate from every line of this, because plans meet reality and reality usually wins, and documenting exactly that is most of what this journal is for. But it's written down now, which was the whole idea.