It’s so strange to me why Google is still sharing these amazing improvements instead of just giving them in a black box. Is it because engineers would go to another organisation with the ideas?

Not just that, but often researchers care more about building their CV than working on any specific company. A job is temporary, but the papers you write are forever. As a result, part of the reason why Google could even convince these people to work for them, is because Google cultivated this culture of allowing researchers to publish their research in most cases.

I think if Google didn't do that, it's possible these employee simply wouldn't have even decided to join the company.

If you train with a batch size of 1, is there anything wrong with batch norm? I think batch norm of 1 is better termed instance normalization...