“Another interesting limitation of speech recognition systems -- and machine learning systems more generally -- is that they are more optimized for specific problems than most people realize. For example, understanding a person talking to a computer is subtly different problem from understanding a person talking to another person. If you train a machine learning system on data from Google of people speaking to a search engine, it will perform relatively worse on Facebook at understanding people talking to real people. In the case of Jarvis, training an AI that you'll talk to at close range is also different from training a system you'll talk to from all the way across the room, like Echo. These systems are more specialized than it appears, and that implies we are further off from having general systems than it might seem.”