Re:Programming won’t be automated, or it already has been
by mtdewcmu on Thursday March 23, 2017 @01:48PM
(#13942161)
There are many tasks related to programming that would be maximally hard for computers. It would be kind of like a gordian knot of all the intractable problems in AI all woven together.
You can't parse conversations effectively without a social context, because a lot of the meaning is implied by the context. The words themselves are more like pointers into the context, and are never a full description of the meaning. (See also: "Make me a sandwich".)
So when someone on HN says "Time flies like an arrow, fruit flies like a banana" the context tells a human reader that the real meaning is "This is a meta-sentence illustrating a point about grammar, context, ambiguity, and NLP. It does not mean I am saying anything specific about flies, time, arrows, fruit, or bananas."
The contextually correct response would be to continue with a different comment about AI. Or maybe to make a context-bending joke. (But not on HN, where the context doesn't allow that.)
Every so often someone in AI rediscovers this idea and tries to model it, but for some reason it keeps being forgotten. It's been around since at least the 1970s, but so far as I can tell modern statistical ML largely fails to be interested in it.