Can we trust open weight models?

Sounds like we ought to treat models as primitives of a kind of software-defined CPU of some larger machine than a convenient, complete package.