A fellow bootcamp member wrote a good summary on this — sharing the link: https://velog.io/@hanlyang0522/weight-init%EC%9D%84-%ED%95%98%EB%8A%94-%EC%9D%B4%EC%9C%A0
The gist: as long as you don’t initialize weights to 0, there’s no problem. Most frameworks handle it automatically, and when using PyTorch’s built-in initialization functions (as the instructor did in the lab), it matches the default initialization anyway.
The caveat: initializing to 0 can prevent learning entirely.