[laingc]

It comes from the dimensionality of the hidden state of popular NLP models. The number 768 in particular comes from (if I recall correctly) the largest of the original BERT models.

[unbanned]

So it's arbitrary then