There was a fascinating article from Cambridge University this week which describes how a robot has designed better ‘babies’ for itself. A few commentators dismissed this as ‘AI hubris’, but, for me, this is a very significant development in the advancement (good or bad) of artificial intelligence.
Having just finished reading Nick Bostrom’s book, ‘SuperIntelligence‘, I am convinced that the human race needs to plan carefully for the time that AI exceeds human intelligence – this is the point that the acceleration in intelligence will proceed at an unimaginable rate, and it may be a very short period of time before the AI has appropriated all of the Earth’s (and beyond) resources to fulfil it’s objective (which could be something as simple as making as many paperclips as possible). If this sounds a little crazy, just read the book.
So, the Cambridge experiment looks like the first, early, step to an AI creating, and (crucially) improving on, its progeny- effectively this is AI evolution. Take that concept and look (not very far) into the future and one can almost imagine a Bostrom-esque AI explosion. One that includes not only software, but the ability to interact physically with the world. Now do you get why this is pretty significant?
In my mind, the article gets a few things wrong in the comparison of AI reproduction with natural selection. Natural selection is completely random – every gene has the potential to mutate, not just the ‘weaker’ ones. The model that the scientists have created is a hybrid of ‘natural selection’ and an ‘ultimate designer’ – only the weaker progeny mutate whilst the stronger ones are left alone. But, the overall idea of a robot (and a physical robot at that) making a two-fold performance improvement in its progeny over just 10 generations is significant enough in itself.
Robots building better robots, who build even better robots. Each ‘generation’ is 10 minutes long. After 10 generations performance is twice as good. Do the maths…