For a while I did what I suspect a lot of engineers did: skimmed the headlines, played with a chatbot, and told myself I was "keeping up." I wasn't. You don't keep up with a field this fast by osmosis. So I made it deliberate — I'm upskilling in AI on purpose, the same way I'd learn any other system: by taking it apart.
I come at it as a backend engineer
I'm not trying to become a researcher. I'm trying to understand AI as infrastructure — something I'll have to serve, scale, monitor, and plug into pipelines. That framing helps. A model is, from where I sit, a slightly strange service: it has latency, cost, failure modes, and an API. Those I know how to reason about.
Fundamentals before frameworks
It's tempting to start with whatever library is trending this week. I'm trying to resist that and spend time on the ideas that won't churn — how models are trained and served, what embeddings and vector search actually do, where retrieval fits, why context and evaluation matter. Frameworks come and go; the concepts underneath them age much more slowly.
Build something, then read about it
The fastest learning loop I have is to build a small thing, hit a wall, and then go read about exactly that wall. Abstract material slides off me; a problem I'm currently stuck on sticks. Most of what I've learned has come from a half-broken side project demanding I understand the next concept.
Writing it down
This is partly why this section exists. Explaining something forces the gaps into the open — if I can't write a clear paragraph about how retrieval works, I don't understand retrieval yet. Expect more notes here as I go.
I don't know where AI takes my work. But I'd rather meet it as someone who took the time to actually learn it than as someone who let it happen to him.