We know algorithm to convert each PDA in the underlying grammar that generates the language the PDA recognize. But we have also the algorithm for creating a two state pda from a grammar. This algorithm use the first state for all the computation, with rules based on the stack and the input symbol, and when the stack is empty goes to a final state. This means that, for all the generic PDAs we can create a minimal PDA with only two states (and a lot of rules for the first state). Am I right?