A key premise here is that overparameterized models can learn a theory seems off base though in the sense that nothing in these models is actually statistically identifiable, meaning there are many different sets of parameters that would yield the same outputs
