Foo bar machines
If AI is to one day superannuate programmers by taking over the task of writing code, then I submit there’s a chasm in the road ahead for Machine Learning researchers: aesthetics.
When some sixty percent of resources are spent maintaining/supporting codebases (rather than writing up fresh code), it’s obvious how economic incentives will trip up AI in replacing human programmers.
It takes a pillage
In a world where AI writes source code, we’d be foolish to think that the code yielded is 100% correct. But to be kind, for the sake of this post, let’s assume that such code is close to correct, but still, bugs linger within.
What happens when one of these bugs rears its ugly head? Well, when systems go down in the world today, humans are on call to introspect & intervene.
And herein lies the crux of the problem during a transition period when humans and machines are pair-programming: if there isn’t some transcript of the source code that reads incredibly simple, no human will be able to parse it cost-efficiently.
Dumb is hard
In a previous post, I pleaded that rock-star, fancy-schmancy code ought to be avoided, and dumbed-down code embraced at all costs. If that sounds unsophisticated, take consolation in this: writing lucid source code is quite hard. (Needless to say, the costs of opaque source code are innumerable.)
Writing clear code is hard because nearly every decision made requires a judgement-call about the beauty of what results: the gradient one moves along in crafting such elegant code isn’t along some cleverness axis, but rather moves along one rooted in aesthetics.
If humans are going to be involved in any part of the code-support process in the near term, then Machines will necessarily have to craft lucid, no-frills source code.
Beauty is ‘whoa’ hard
For AI to understand aesthetics seems so far down the road, that I would challenge any expert to take a wild guess on when we can expect it (while keeping a straight face). Machines that ‘get’ jokes and robots that know to pull a string rather than push is one level of sophistication. But for AI to cobble together correct code that also reads simply & elegantly seems like a brand of hard fit for the bravest of researches to tackle.