[6ren]

aside: What about context sensitive/markov completion, backed by say github/google code? And if the prediction was certain enough, offer multi-line completions. This would be helpful in languages/situations that need boilerplate for common tasks, that you'd usually google-paste.

Or maybe howdoi-based completion http://news.ycombinator.com/item?id=5027021

[seanmcdirmid]

One of the projects I dream about doing is to use github/google/... code as a corpus to feed into a deep neural network, which could then perhaps do what you are talking about. However, more work needs to be done on how to represent the code and train the models.

There has been lots of work on intelligent code completion in the SE community. It doesn't seem to be very effective yet, however.

[nano_o]

One can also take types into account for completion. There is a Scala Eclipse IDE plugin that does it: http://lara.epfl.ch/w/insynth (it's a research project).