interesting stuff. how does verifiable AI computation actually work in practice. curious about the technical implementation