February 19, 2026
On reading papers vs. understanding them
The gap between finishing a paper and being able to implement it is enormous. Notes on why.
Reading a paper and understanding a paper are different activities. I keep confusing them.
Reading: you follow the sentences. The math looks plausible. You nod at the conclusion. You close the tab feeling like you've done something.
Understanding: you can re-derive the key equations without looking at them. You know what would happen if you changed assumption X. You can implement it from memory and debug it when it breaks.
The gap is enormous. I spent 2.5 hours reading the Adam paper before I started implementing, and still missed the bias correction entirely on the first pass. It only showed up when my loss curve diverged from the PyTorch baseline.
The implementation is the test. If you can't implement it, you haven't understood it. If your implementation doesn't reproduce the results, you haven't understood the details that matter.
This is going to take longer than I thought.