Currently, even if you were to program such a thing, it would still have to obey certain rules and can't learn outside of the variables you define as input. A machine can't process data that it isn't prepared to process. AKA a machine/program can't learn outside of the variables you tell it to learn. It also has no will of it's own.
So, we are in no danger for now. We would be in danger if we discovered exactly how the human brain works and could replicate the neuron's network. At that point we would he creating life and a very dangerous one. On the other hand off switches can always be implemented or safeguards. It is still dangerous and it would be unethical.
Thinking about it... maybe what the guy is warning about is giving the wrong input variables to the program. That could be dangerous too, because the program has no conscience and in my human replication scenario they would be the same as us but with a much sturdier body.
