28-10-2026, 03:30 PM
(This post was last modified: 28-10-2026, 05:09 PM by npcompl33t.)
I'm not arguing either way, personally I think multiple hands and the gradient are not mutually exclusive. However, it shows that the Hand information is already known to correlate with the gradient, regardless if it was one or multiple hands.
a statement like 'If the hands explain almost all of the gradient, the gradient cannot tell hands apart from time.' is problematic, because the Hands are stratified across the time gradient. If your model is reproducing text at 94%, it is learning the gradient.
Saying that 'hands explain all the gradient', is not accurate, and is stretching the model to make claims beyond what it was designed for, and beyond what it is capable of asserting. Many of these features are tightly intertwined. The model was not designed to test this type of distinction, and had no controls to keep these factors independent.
It is definitely neat that the model can reproduce voynich text, but it isn't capable of supporting these types of claims. the type of model that was used here frequently will learn such features unintentionally as a result of the way they are trained. It was trained on the data without any controls to keep these factors independent from one another.
a statement like 'If the hands explain almost all of the gradient, the gradient cannot tell hands apart from time.' is problematic, because the Hands are stratified across the time gradient. If your model is reproducing text at 94%, it is learning the gradient.
Saying that 'hands explain all the gradient', is not accurate, and is stretching the model to make claims beyond what it was designed for, and beyond what it is capable of asserting. Many of these features are tightly intertwined. The model was not designed to test this type of distinction, and had no controls to keep these factors independent.
It is definitely neat that the model can reproduce voynich text, but it isn't capable of supporting these types of claims. the type of model that was used here frequently will learn such features unintentionally as a result of the way they are trained. It was trained on the data without any controls to keep these factors independent from one another.