4 ms·

Agree 100%. The fact is (well, in my experience anyway) that the people who want programs written don't really know what they want (they only think they do). Th

by defined 10y ago

Agree 100%. The fact is (well, in my experience anyway) that the people who want programs written don't really know what they want (they only think they do). The art of requirements analysis is very difficult to do well, partly because of this, and like you say, when (if) any machine becomes intelligent enough to equal a human in eliciting good requirements, we will all be in a lot of trouble. Right now, is there any natural language system that can appropriately correct the punctuation on these sentences (not meant rhetorically)? SPLINTERS WEAR SHOES PANDA EATS, SHOOTS AND LEAVES

mtdewcmu 10y ago
It's not only cases where the requirements are especially hard to elicit. Computers don't have common sense, so even "obvious" things would be hard.
- TheOtherHobbes 10y ago
  One reason computers don't have common sense because [insert large but hard to calculate percentage] of human interaction is context, learned and scripted social setting/interaction, and tradition, not semantics. And certainly not statistics. 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.
  - mtdewcmu 10y ago
    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.
fragmede 10y ago
It's impossible, as a human, to know what you actually wanted to say with those sentences in isolation. Are you setting up for a joke about a gun-toting panda at diner, who left the scene of a crime? Are you paraphrasing an informational sign at a zoo about a black and white animal? Assuming you're not trying to go for a bit of clever wordplay, however, plug the panda sentence in Google Translate and it'll correctly translate "leaves" to be the plant, and not the verb tense of "to leave", in the chosen destination language. TechCrunch discussed Google's internal representation of language[1] which implies that while the computer doesn't have a concept of the animal and its food sources, it recognizes that "panda" has something to do with "leaves" as being related to plants rather than "to leave". Google Translate's far from perfect, but as poor as it is, it was once considered impossible to advance to where it is now. [1]https://techcrunch.com/2016/11/22/googles-ai-translation-tool-seems-to-have-invented-its-own-secret-internal-language/ https://techcrunch.com/2016/11/22/googles-ai-translation-too...

by defined 10y ago

mtdewcmu 10y ago

It's not only cases where the requirements are especially hard to elicit. Computers don't have common sense, so even "obvious" things would be hard.

TheOtherHobbes 10y ago
One reason computers don't have common sense because [insert large but hard to calculate percentage] of human interaction is context, learned and scripted social setting/interaction, and tradition, not semantics. And certainly not statistics. 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.
- mtdewcmu 10y ago
  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.

fragmede 10y ago

It's impossible, as a human, to know what you actually wanted to say with those sentences in isolation. Are you setting up for a joke about a gun-toting panda at diner, who left the scene of a crime? Are you paraphrasing an informational sign at a zoo about a black and white animal? Assuming you're not trying to go for a bit of clever wordplay, however, plug the panda sentence in Google Translate and it'll correctly translate "leaves" to be the plant, and not the verb tense of "to leave", in the chosen destination language. TechCrunch discussed Google's internal representation of language[1] which implies that while the computer doesn't have a concept of the animal and its food sources, it recognizes that "panda" has something to do with "leaves" as being related to plants rather than "to leave". Google Translate's far from perfect, but as poor as it is, it was once considered impossible to advance to where it is now. [1]https://techcrunch.com/2016/11/22/googles-ai-translation-tool-seems-to-have-invented-its-own-secret-internal-language/ https://techcrunch.com/2016/11/22/googles-ai-translation-too...