Last week I finally understood decision trees.
Not the textbook version. The real one. A tree is just the checklist you already run in your head, turned into code. Ask the question that splits the group best, then ask again inside each smaller group. Ask, split, repeat. That's it.
I felt smart for about an hour.
**Then I read one more line: **change a few rows of your data and the whole tree can rebuild itself into something different. The thing I just learned was shaky. The fix is ensembles, which is a whole other layer I haven't touched yet.
That's the part nobody warned me about. Learning AI isn't one mountain. It's layers, and every layer makes the one below it look incomplete.
The floor keeps dropping
Here's the pattern I keep hitting.
I learn a thing. I feel like I get it. Then I learn the thing that sits on top of it, and suddenly the first thing was just the warm-up.
Linear regression felt like the whole game until logistic regression showed up. Logistic felt solid until trees. Trees felt clever until I found out they're unstable and ensembles exist. I haven't even reached the layer after ensembles and I already know it's there.
If you're starting out, here's what would've saved me some panic: feeling like you understand something is not the same as being done. It usually means you've cleared one layer and the next one is about to load.
That's not discouraging. It's just the shape of the thing. Once you expect it, you stop feeling stupid every time the floor drops.
The rabbit hole is real and it looks like progress
The dangerous part is what happens next.
You learn trees, you hear "ensembles," so you open a tab. That tab mentions XGBoost. XGBoost mentions gradient boosting. Gradient boosting mentions a paper. Three hours later you have eleven tabs open and you've built nothing.
It feels like progress because you're reading hard things. It isn't. It's just falling.
**What I do now: **when I hit the next layer mid-learning, I write it down and close the tab. "Ensembles, look at next." Then I go finish the thing I was actually on. The note means I won't forget it. Closing the tab means I won't drown in it today.
Depth is good. Depth at the wrong time is just a nicer way to procrastinate.
"Which tool should I use" has a boring answer
The other thing that almost broke me is the tool zoo.
I spent a week on APIs. Just for two systems to talk to each other, there's REST, GraphQL, SOAP, gRPC, and WebSockets. Each one has fans who'll tell you it's the right one. As a beginner you sit there thinking you have to learn all five before you're allowed to build anything.
You don't.
REST covers the large majority of cases. gRPC is for fast backend to backend stuff. WebSockets are for live two-way things like chat or live prices. The "best" tool doesn't exist. There's only the right tool for the job in front of you, and most of the time the job is boring and REST is fine.
Most teams overthink this. Reaching for gRPC or GraphQL early adds tooling pain for almost no gain. Once I saw that, the panic of "I don't know all the tools" turned into a calmer question: what does this specific problem actually need?
There's still a scarier version of this I haven't solved. What if there's a tool that's perfect for my problem and I just don't know it exists? No clean answer yet. Right now I solve it by building the dumb version first, then asking people who've done it longer what they'd have reached for. The dumb version is rarely wasted. It teaches you what the problem actually is.
Sometimes the answer isn't AI at all
This one surprised me.
I kept assuming automation meant an LLM somewhere in the pipe. It doesn't. A lot of "automation" is just basic conditions. If this, then that. No model, no fine-tuning, no GPU.
Half the maturity in this field seems to be not reaching for the fancy thing when a plain one works. An LLM is the answer to fewer problems than the hype suggests. Learning to tell the two apart is its own skill, and it isn't a technical one.
Be skeptical of clean numbers
Last layer, for now.
I read a research paper comparing two big data tools. The numbers didn't add up. It claimed more memory and processing power than its own machines had, and never explained why. Another paper claimed a 56 to 69 percent improvement, but I only had the summary, and summaries always show the best result against whatever weak thing made them look good.
Reading papers taught me less about the tools and more about not swallowing claims whole. When something is presented clean and impressive, that's usually the cue to ask what got left out. Not because people lie, but because nobody advertises their own gaps.
What I'd actually tell a beginner
I'm three weeks deep, so take this as field notes, not a map.
Pick one real problem. Find the simplest thing that solves it, even if it's embarrassingly basic. When you hit a rabbit hole, write it down and close the tab. Match the tool to the job, not to the hype. And when a number looks too clean, poke it.
I still don't know what I don't know. Next month's version of me will probably think this post missed the point. ngl, that's kind of the whole thing about layers.
The floor's going to drop again. I'm starting to think that's the job.
Top comments (1)
The "poke clean numbers" bit — yeah that's the same trap I keep running into 🤣 Solid first post. If three weeks looks like this, curious what three months brings