Masoptimizer

The paper is sort of talking about these risks from, like, these learned optimisers and optimizes. And i'm wondering, like it is mess optimiser the right category? If you think about the sutabular masoptmiser, how likely would you be to actually fine the sort of tabular masoptimizer? I think the answer is very unlikey. It has to incode all this sort of really complex, like, results of athorization explicitly in a way. This takes up so much space and explicit memorization. But not just memorization, like, what your deployment disfigured, what your sort of, like, out of distribution behaviour would be,