I have to reread about MCMC. Always feels like a bit of magic. How can you sample from the right distribution given so little information?