The first starts with an idea. You read something, and a clean hypothesis pops up: what if we replace X with Y to get a better model It's exciting. You sketch it out and start building toward it. Most of us got into research because of this feeling.

The second way is less romantic. You don't start with a specific idea, but you start with the best thing that already exists. You take the strongest baseline you can find, you reproduce it honestly, and then you push it until it breaks.