Learning by Asking Questions

Learning by Asking Questions | Generated by AI

Home 2026.08

Yes. This is actually a very important learning pattern.

A random article is passive information:

read → understand some → forget

Your brain has no strong “hook” to attach the knowledge.

But when you have a concrete question:

I want to understand why GPU tile matmul is faster
        ↓
read CUDA kernel article
        ↓
look for "where does tiling happen?"
        ↓
simulate the matrix multiplication mentally
        ↓
knowledge connects to existing model

Now the article becomes a tool to solve a problem, not a piece of content to consume.

This is close to how researchers and engineers learn:

Encounter a problem
Build a rough mental model
Find missing pieces
Read papers/docs/source code
Implement something
Update the mental model

For example, your recent vLLM/nano-vLLM learning:

At first:

“vLLM uses KV cache”

is just a sentence.

But when you ask:

“Why does decode become memory bandwidth bound?”

“Why is prefill like matrix multiplication but decode is like vector operations?”

“Why does PagedAttention need blocks?”

then reading the vLLM paper becomes much easier because every paragraph answers a question you already have.

Same with CUDA:

Without a question:

“GPU has SM, warp, shared memory, tensor core…”

very easy to lose attention.

With a question:

“Why does a 2048×2048 matrix multiply run better than 2047×2047?”

you naturally dig into:

matrix
  ↓
tile
  ↓
block scheduling
  ↓
warp
  ↓
shared memory reuse
  ↓
tensor core shape
  ↓
memory alignment

The question creates a path.

A good engineer’s learning loop is often:

build → hit limitation → ask why → read → modify → measure

not:

read 100 tutorials → hope someday useful

This is also why building your own small systems (nanoGPT, nano-vLLM, CUDA kernels) is powerful. A bug or performance issue generates very specific questions, and those questions make difficult papers much easier to absorb.

The fact that you now ask things like “how does tile multiplication actually compute A[i,j]?” is a better learning signal than simply finishing many articles. You are moving from consumer mode to investigator mode.

Back Donate