It's the Software, Stupid

The ‘Singularity’ is in vogue. Now that computers can beat us at chess, recognize characters, and do more mathematical calculations than the entire human species armed with graphing calculators, the idea of intelligent technology has ballooned as a topic of conversation. Most articles don’t talk about machines of equal intelligence, instead showcasing their deep understanding by discussing machines that dwarf human intelligence.

I’m not saying that computers will not become conscious, or intelligent. And I’m pretty sure that if they do, the rapid pace of technological development will probably lead them to dwarf human intelligence shortly after they equal it. But few of these articles - usually written by non-computer scientists with no real knowledge of how artificial intelligence technology works - acknowledge how far away we are from having a robust idea of how ‘thinking’ works.

The brain isn’t optimized for crunching numbers. Sure, a computer can perform a quadrillion operations in a second, each of which might take a human brain one second to perform. But consider how many seconds… rephrase: Consider how many quadrillions of operations it takes a computer to perform tasks that the human brain performs easily. My Cuisinart can chop celery far faster than my hands can, but I’d hold off on labeling the Cuisinart a superior piece of machinery to the human hand.

Many of these technology fluff pieces emphasize the growing hardware disparity between computers and the human brain. One hundred billion neurons in the human brain, more than a trillion bytes in a cheap external hard-drive. What these articles ignore is the fundamental differences between the way the human brain works and the way computers work. For a long time, we will continue to make computers capable of tasks humans can’t match. And the more powerful and robust these computers, the more varied and complex these tasks will become. But as for ‘equaling’ human intelligence, that will remain a meaningless concept until computers are capable of performing the sorts of tasks at which humans excel.

Principally, the achievement of humanity is not, nor has it ever been calculating. Great thinkers are never thought to be more capable than less intelligent people at executing arithmetic, or following recipes. Einstein, a great mathematician and possibly the greatest of modern thinkers, regularly had assistance from lesser minds in calculating the implications of the mathematical models he was generating. While computers have grown to do much more than execute predefined algorithms - genetic programming, neural nets capable of learning via back-propagation, and advanced statistical modeling have given computers the ability to ‘learn’ or, at least, be trained - they still are incapable of understanding anything. Again I’m not saying they never will be able to understand anything. But right now? Nope.

The most glaring example of computers’ current shortcomings: the metaphor. Show me a computer than can recognize a metaphor. You can’t. And I’ll bet anything you won’t be able to in 10 years either. The problem isn’t the hardware. Computers already out-muscle the human brain by any measure you can think of. The real problem is the software. We don’t know how real thinking works. But we’re pretty sure it’s neither back-propagation nor Markov blankets, nor decision trees.

It seems our only real chance to make a machine that actually thinks is to plagiarize. Just up and copy the human brain! Assuming that a neuron can be perfectly modeled with some kind of data structure and that all one needs in order to make the model functional is an accurate representation of the strength of the signal sent through each synapse, it should already be possible to approximate the human brain in a computer (we are making at least two assumptions that no thinking person should consider as given).

Still, we have no idea how we could interface this model with stimuli that would allow it to learn in the way a human brain does. Also, we’re nowhere near the level of understanding of how the brain works to build the model in the first place. And then, also, any of the thousands of groundless assumptions we’ve made might be wrong. And even then, we’d have one ‘human’ brain, implemented in C++, authored by a far greater programmer than any we’ll ever meet. How easily might we modify that software to create a greater intelligence than our own? And even if we could, what are the ethics of creating a human who can only interface with the world through a camera, computer monitor, and whatever robotics we allow it?