He said that we already have the material capability to build an artificial brain with superior capacity to execute more FLOPS (Floating Operations Per Second) than the human brain (@ 20 billion) but that FLOPS are not the key determining factor in superior cognitive ability. That's because electro-mechanical computers use linear networks, and organic brains use associative networks. 

IOW, the human brain functions like a network that actually has 20 billion individual computers all connected and adding information to the computation in real time.  

Eventually a large enough associative network will be built which can defeat a human brain in any computation, so long as the right "question" can be asked. 

That is likely when the Technological Singularity will begin and human beings will discover whether or not artificial intelligence is friendly to Humanity. 

With all due respect to SF author Isaac Azimov, the notion that AI can be constructed to be unable to harm humans has already been shown to be laughable. Asimov was a great writer but a mediocre scientific mind at best.