A Computer That Codes Itself

Gary Marcus and Ernest Davis detail attempts to invent one:

What [the Defense Advanced Research Projects Agency (DARPA)] and [programmer Charles] Simonyi are hoping for is a complete paradigm shift. A programmer—and this could be anyone—would simply tell the computer what he needed in plain English, and the computer would figure out the rest. Anyone would be able to program, not just highly trained specialists, and, at least in principle, computers might ultimately produce much more reliable code than their human counterparts.

One big problem with this dream is “that computers still have too little understanding of how the external world works, and therefore too little understanding of how the programs they create will actually work”:

Consider, for example, this seemingly simple, hypothetical programming task: “Add a feature to Google Maps that allows a user to place a simulated boat on a river and have it float downstream.” To do this, you need to know what a river is, what a boat is, and what it means for a boat to float downstream. Any human programmer knows that, but no computer system has the real-world understanding of an average human being. As Tom Dean, a researcher at Google, told us, “Programming is [challenging for artificial intelligence] not because it requires concentration and attention to detail but because the path from the conception of what you want to accomplish to the realization of code that actually accomplishes it requires artistry, insight, and creativity as well as incredible mental dexterity.”

One day computers may have that kind of dexterity and intuition; the DARPA program is a good first step in that direction. But the path to the automated, thinking computer will also require a shift in research priorities, from the currently popular focus on the question “What you can do with Big Data?” back to A.I.’s original, driving one: “How do you build machines that are broadly intelligent?”