How difficult, conceptually, would it be to introduce existential types? How much work and complexity does it add to the inference engine? Is there a syntax that you would recommend?