Moore’s Law states that the level of technology/computing power available doubles every 18-24 months. This exponential growth has held true the last thirty years and shows no signs of slowing down. There is a physical limitation which will be reach but I have no doubt that by that time we will switch from silicon-based transistors to some other means. E.g. Just the other day scientists announced the world’s first molecular transistor consisting of a single molecule of benzene [LINK].

So what’s next? Well if computers get stronger/smaller/faster you might think we’ll wake up one day to find “Skynet has become self-aware”. Unlikely. Let’s set self-awareness aside for a minute and look at computing power. The brain has about 100 billion neurons (10^11 estimated) all inter-connected. It works by passing chemical and electrical signals back and forth in pathways. The strength of these pathways has a relation to the brain’s capabilities to process information. E.g. some pathways control motor function, ability to play music, abstract thinking and so forth. The ability to form new pathways and build new neurons is known as brain plasticity.

The world population today is roughly 6.5 billion and of those roughly 1.5billion have internet access.  Let’s assume some of those people share computers so conservatively there are 1 billion computers in the world.

Now, let’s take the average computer, an Intel Pentium D has 230 million transistors but let’s go lower than that, let’s assume 100 million transistors per computer, which is the average from the 90′s. Each transistor might equate to a neuron. I’m going to leave storage out since RAM and disc space is tough to estimate.

That’s 100,000,000 X 1,000,000,000 or 10^17 artificial neurons in the world. With all these machines interconnected, we far exceed the raw computing power of the human brain even in today’s technology.

Setting computing speed aside we could say the hardware is there but the software is not. These machines are all focused on different tasks.

Let’s imagine you could teach such a system to learn. We have already done this to some degree, e.g. machines can learn to play chess, recognize voices, recognize faces and so on. But these are all trivial and on the grand scheme of things not very useful.

If I want to play “Crossroads” on the guitar, I can pickup a guitar and a book, look at the notes and practice long enough until I can play that one song. In doing so I will burn in the necessary pathways to play that particular song. It’s the tactical approach.

But there’s a better way, a “meta” way which is if I want to play “guitar”, what I need to do is augment myself in some way to be able to play the guitar regardless of the song. I need to understand musical theory, to listen to existing music and flood my mind with reference material. I need to enhance my fingers, strengthen them, exercise them and so forth.  Take some lessons. At the end of all this I’ll have gained a more general knowledge of the guitar and hopefully be able to play. I will have a “meta” knowledge which will not teach me any songs but once presented with a song I will have a system by which I can figure out how to play it. E.g. listen for the root-note of the chord, look for 3-chord progressions, 12-bar blues etc.

This all assumes that I have the raw material already within my mind. For example some folks have no sense of timing, or no ear for pitch. This is hardwired in their brains and no matter what you do you cannot change it. At least I have not seen it. So we as humans are limited in some capacity.

A machine on the other hand has more freedom. If it decides the best way to play guitar is with 6 fingers instead of 5 it could simply add an appendage. It is not limited as we humans are. Similarly if a machine decides the best way to learn the guitar is to rip out it’s “timing” module and replace it with a finer grained implementation of same it can do that.

To learn a song is one thing. Computers can do this today with a very narrow focus and even this they struggle with. Now, imagine the circumstance where the machine is asked to learn a song and realises the best way to achieve this is to first learn how to play guitar; “meta” side of a given task.  This would be a quantum leap from what we have today.

There is more though. Learning the “meta” side of anything is limited by what? Answer: Our ability to learn.

1. The computer decides to learn to play a song on the guitar, the best way is to learn how to play the guitar.
2. In order to learn how to play the guitar the best way is to learn how to learn.

Once a machine can  improve the way it learns another threshold has been breached. It learns faster. Not only that, it learns how to learn. And does THAT faster. This is a key threshold which we as humans will never reach.

This leads to exponential growth in the rate of learning. Learning on the part of the machine becomes faster and faster. Not only that but new ways to learn are found. The machine self-augments itself recursively leading to ever increasing improvement.This would be a watershed event and is known as The Singularity. Ray Kurzweil has predicted it will happen in 2045.

It’s important to note that the Singularity does not require self-awareness.

It’s commonly held that once the Singularity occurs, the machine(s) will reach a point where they eliminate mankind, either by accident or because we will have inadvertently set goals for the machine which require our demise.

Asimov tried to address this by laying out the three laws of Robotics:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey orders given to it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with either the First or Second Law.

I don’t think this is the right approach though. For example all three laws would fail to apply if a super intelligence evolved to the point where it no longer fit the definition of a Robot. Our minds cannot conceive the super intelligence and we must assume that we cannot constrain it in any way.

It’s difficult to discuss such topics without assigning human characteristics to the machine whatever it might be. This of course makes no sense since we have no possibility of knowing what form it will take or what its value system or world view will be (if any).

The natural tendency is to assign human traits to the intelligence, e.g. a desire to dominate, to conquer, a quest for longevity and so forth.These traits could be seen as a product of our evolution, things which we value since they further our likelihood of survival.

This begs the question, as human beings, just what is our value system? At a macro level it might be viewed that we seek to rule the world or some other materialistic goal.  Or simply that we want to provide for our families.  It is true at a macro level to say we  desire to sustain our own species. More accurately though you might say that our genes favour characteristics which allow themselves to be sustained. Richard Dawkins talks about it at length in the Selfish Gene. Another way to state it might be that our behaviours on the macro scale are simply the tail wagging, the dog in this case is the mechanism by which genetic material is replicated.

Getting back to the Singularity, if we want to avoid being turned into batteries by the super intelligence, or ground down into fertilizer, the question becomes how do we encode a value system into the very makeup of the machine so that we can never arrive at a circumstance whereby the computer decides the best way to learn to play Crossroads is by chopping a human’s head off and playing guitar with their body?

Perhaps a topic for another day…