Artificial Intelligence (AI) is a broad church of techniques that can be used to create intelligent behavior. In fact, a really good place to start in all of this is to define intelligence – and it is not particularly glamorous or mystical. AI is concerned with the creation of intelligent agents (things that do something). Such an agent might man the help desk and deal with queries and problems (or at least direct them to the relevant person), or it might be embedded in a washing machine that knows which settings to use depending on the items loaded into it. Either way, an agent is considered intelligent when it performs well against some pre-defined objective. Human beings exhibit intelligence all the time. We generally avoid being flattened by a truck when we cross the road, and most meals are not burned to a cinder (unless my mother-in-law is cooking).
Many of the techniques used in AI simply employ brute force to solve problems, find patterns, identify objects in an image – and so on. These techniques have taken decades to refine, and without these refinements even chess programs would take years to execute each move. Most of these techniques are specific to a particular problem. We wouldn’t use search methods to try and build an agent that needed to execute business logic for example. One of the big challenges in AI is to create more general purpose methods. No doubt they will evolve, but we are not there yet. The current state-of-the-art is generally considered to be deep learning. The word ‘deep’ does not signify some profound capability, but simply refers to the depth of neurons used in a special form of neural network. There was great jubilation when deep learning was used in image recognition and the algorithm could detect various objects (cats for example) without being told what the object looked like. In fact Andrew Ng, one of the leading experts in machine learning (an AI technique), proclaimed that prior to deep learning, everything else was just curve fitting (getting algorithms to adjust their parameters and variables until they worked). So real progress is being made.
The set of techniques that are now available address many business problems, and large numbers of applications are already being rolled out. Some automate large parts of the drug discovery task, others help sales people identify the best prospects, and others can carry out a conversation with customers who need help. So there is real momentum, and AI will automate many tasks that are currently executed by humans – even the analysis of MRI scans.
We now come to the more fantastical claims for AI. Can machines become aware, and do they think? This is a question that has undefined terms. No one knows what awareness is or indeed what thought is. Sure we can look at chemical and electrical activity in the brain, but these are effects. Since consciousness and thought have no rigorous definition, the question whether machines can be endowed with these properties is wholly meaningless. Yes we all know about the Turing test – whether a human being can tell the difference between interaction with a computer and human. But this does not address the issue of thought and awareness.
For now AI is largely a set of number crunching algorithms that do some things remarkably well – much better than humans. General purpose intelligence is a long way off, and for now at least, the impact of current AI will be more than enough for businesses and society to absorb.