Re^5: How will Artificial Intelligence change the way we code?

[BrowserUk]

which could be already sufficient to significantly influence how we code in the future.

I use (and have been using) brute force and GA algorithms to find patterns and correlations in huge datasets for 10 years or more -- there's nothing new or innovative about that.

Eg: Without doing an exhaustive search, most recent(2016) and oldest(2004).

I do think there is scope for using the types of bulk data munging inferencing algorithms that are typified by the kinds of things the press (and others) are currently labeling as "AI"; as part of the development process; but as a part of, rather than as a replacement for.

But to find those applications, you need to look for sources of large datasets that have enough things in common -- that can be relatively easily recognised and categorised -- to provide a training dataset.

Eg. if all the test scripts, of all modules on CPAN that use the Test::* modules/conventions were examined, and those were correlated with their maintenance histories; then it might be possible to first devise a fitness test that could characterise which kinds of tests actually test something worth testing.

I suspect it would confirm my prejudices against certain kinds of tests -- eg. testing whether use My::Module; actually loaded the module -- ever actually caught a problem in the field; and thus you might whittle down the types of tests that are worth adding to the test scripts.

Having determined the types of tests that are worth running, it might then be possible to use that information, and some textual analysis of the code patterns and structure of those tests, so that you could come up with an algorithm to examine new modules and programs and generate those test automatically.

The problem with using deep learning AI for programming is that you need large volumes of essentially similar, but distinctly different, existing data (lots of programs that do roughly similar things but in different ways) in order to train the inferencing engine to the point it can make a decent fist of recognising the patterns. And for the most part, such data does not exist.

Take (say) language compilers. There are maybe 10 or 12 distinct, well-tested & used C compilers on the planet. Even if you could get the requisite permissions from the proprietary owners of all of them to agree to giving you access to their source code, that probably isn't a big enough sample to train the IE to recognise the required code patterns. And you'd be completely out of luck for most other languages where there are only 1 or 2 or 3 distinct compilers. (Perhaps you could do something with the various javascript/ECMAScript engines?)

And most other types of substantial programs there are even less good examples. Maybe there are enough examples of things like editors, 3rd person shooters, spreadsheets and CAD programs; but then you've still the proprietary problem to deal with.