The computer on your desk is very nearly useless. You can use it the way I, as a hilariously over-engineered typewriter-with-integrated-postal-service, or as a television set, or to store all the photographs you took so carefully and never remember to look at again.
Your computer and its more powerful siblings are confined to desktops and server closets because, despite the exquisite refinement of all the sensors we connect to them, they understand almost nothing of the world around them. In their ability to comprehend and make use of what comes into their sensors, they are hopelessly outclassed by the idlest toddler.
Once, this was the fault of what passes for their minds — early processors were so very limited that they were confined to working inside desktop calculators or inside the guidance systems of ballistic missiles. After fifty years of Moore’s Law, the limiting factor is no longer the processor, but the human engineer sitting at the keyboard: an everyday task like ironing a shirt or making an omelette cannot be automated by a programmer, not even one who makes very good omelettes. The difficulty is that although we can make the omelette, we have barely more insight into how we crack and pour those eggs than a tulip does into how it photosynthesizes. Thus, we can automate simple things which we can explain clearly, like arithmetic or ballistic trajectories, but not the mysterious process by which fresh shirts appear in one’s wardrobe.
Since computers are so limited, the entire industry, despite the headlines it creates, accounts for only a few trillion — a few per cent — of the global economy. However, we have opened — just a crack — a doorway into a world where the key limitation (us engineers) has been removed, or at least greatly relaxed. Instead of traditional software, (excruciatingly detailed instructions, hand-coded one line at a time by skilled & pampered artisans), a new breed of engineer, conceding their intellectual limitations, has adopted a new strategy: give the machine a goal, great flexibility in the means of solving it, and coach it until its performance is sufficiently good for a viable product (or at least another funding round).
Through the window of machine learning, we glimpse a world where machines step out of the server closet and start being useful. Almost none of the work we want to be done can be accomplished inside a server closet by a machine that is blind, deaf, and incapable of learning. As these limitations are overcome, we move closer to a world where computers are not merely incidentally pervasive (doing trivial things inside a watch or a microwave oven), but in which they truly transform the value of the machine they control (gliding an iron over my shirt, loading my dishwasher). If machines confined to a closet are worth a few trillion dollars a year, imagine what their share of GDP might look like when they become more broadly useful. Instead of a headline-grabbing industry with annual sales of a few trillion dollars, we’ll have a headline-grabbing industry an order of magnitude or two larger than it is today — a world where a trillion-dollar market cap makes a firm a minor player, and corporate revenues match the tax-takes of middle-tier nations.
It sounds like the setup for a novel, but as demos turn into prototypes, and prototypes into products, that server closet door is starting to open.