Thinking vs Understanding

Thought was prompted by seeing this karpathy tweet

https://x.com/karpathy/status/2049907410303865030

It seems like understanidng is more fundamental.

You can think of thinking as doing computation on the fly. But understanding is the deep neurons that understand the shit.

Understanding = the model of the world in your head.
It is your compressed intuition: what things mean, how pieces connect, what matters, what does not.

Bad understanding + lots of thinking = spinning wheels.
Good understanding + little thinking = fast intuition.
Good understanding + good thinking = elite engineer.

Before understanding ViT:
“Why is it 128? Why linear projection? Why Conv2D? Why CLS?”

After understanding:
“Oh, every patch becomes a token vector. Embed dim is just the token width. CLS is a learned query-like summary token. Conv2D with kernel=stride=patch size is just a fast patchify + projection.”

Then thinking becomes easier because the concepts are already compressed.

So the hierarchy is: 1. Exposure — you see examples.
2. Understanding — you build the right mental model.
3. Thinking — you manipulate the model.
4. Judgment — you know what matters.
5. Taste — you choose elegant/simple solutions.

For becoming truly great, I would prioritize:

Understanding first, thinking second.