Learning Method

Learning Method

What "I learned it" means

For a concept to count as learned, you need all four:

1. Explain it — you can explain it at a whiteboard, in 5 minutes, without notes
2. Point to it — you can point to the code/line in the project that this concept corresponds to
3. Defend it — you have an answer to "why this way, why not another way"
4. Write it — you wrote it in your own words on the Glossary page

If only 1 holds: you memorized it. If only 2 holds: you copied it.

Why the explanation test?

Because the illusion of understanding is real. While watching a video, everything looks clear; the moment you try to explain it, the parts you don't know suddenly surface. Explaining is the cheapest and most merciless test of knowledge.

The learning loop (per concept)

read/watch (30-60 min)
   ↓
write it in your own words (10 min)   ← don't skip this, the most critical step
   ↓
apply it in the project or run an experiment
   ↓
explain it to your partner
   ↓
record it in the glossary ✅

What we go deep on, what we get familiar with

This distinction is the backbone of the plan. If you mix it up, you end up superficial in both areas after six months.

Go deep (the vertical bar)	Get familiar with (the horizontal bar)
Distributed systems / server side	Deep learning
Relational databases	Neural network math
Docker and container orchestration	Model training
Security	Model architecture design
Observability
Applied LLM & agent engineering

"Getting familiar" is also a standard: you must be able to explain what a transformer is, why it works, and when it's used. You just don't need to derive its math.

Source discipline

Pick at most two sources per topic. A third source feeds the feeling of procrastination, not learning. Source list: Resources.

When consulting AI

Using Claude/GPT is allowed and smart — but with this rule: don't have it write the code, don't have it make the decision.

✅ Good use	❌ Bad use
"What's happening in this EXPLAIN ANALYZE output?"	"Write this endpoint for me"
"What's the weak point of this design?"	"Which architecture should I pick, you decide"
"Quiz me on this concept Socratically"	"Fix this bug" (look yourself first)
"Is there a mistake in this explanation I just gave?"	Copying code you don't understand

The test is simple: can you defend every line that goes into the repo? If the answer is no, that line shouldn't go in.