The conventional wisdom has been that artificial intelligence is all about algorithms, and therefore restricted by the Turing model of computation. In fact embodied machines (those that interact with the outside world via sensors and actuators) cannot reasonably be modelled as being 'digital' in the conventional sense, or even as Turing machines. Far from being a disadvantage, by understanding and exploiting analog nature of these machines we have the opportunity to increase power efficiency, improve learning and discrimination, and provide better adaptability to changing circumstances.