I mean, it’s still a fairly common pattern to use the l-value yielded by the assignment operator in for-loops in C.

I’d tend to agree. There are enough barriers to training large models without artificially increase them just because the largest players can afford it.