> for any problem there is an ideally sized model that when trained will produce the lowest error rate.

You studied ML before discovery of "double descent"?

https://youtu.be/z64a7USuGX0

I did, however I have also observed ideal sized models since then in algorithms designed with knowledge of it.