TrueNorth chips would probably be a competitive disadvantage. I found Yann LeCun's critique[0] compelling. In short, simulating the brain via spiking neurons is less efficient than just doing the floating point math like regular neural nets do.
While that makes sense on paper, in practice do those chips provide better support for networked, autonomous activities? I haven't heard much about their usages just some PR things about them so I'm genuinely curious.
The machine learning people I know have been very unimpressed with it. From the little I understand, it cannot run the style of neural network that have been breaking records.
Hmm that sounds unfortunate. I'll have to try to look into it further because I didn't even know about them before and they sound really interesting.