Anyone want to work on anything? Here's an idea - take text and
develop a universal natural language understanding system that
produces standoff logical form assertions about the text, including
all the usual suspects, plus additionally derived knowledge
delineating possible (defeasible) meanings about text-spans. Also have
assertions about the relationship between the contents of the text and
the containing file or other source. So for instance, if you have a
text-span in which is estimated to exist a self-contained sentence,
you can assert that, as well as provide some evidence for it by
linking to a memoized result of running a grammar parser on it like
the Link Parser. That's the first step. The goal is to have the
computer be a deliberative agent in the execution of tasks. You
provide it with a model of the types of commands
(shell/emacs/perl/APIs/etc) it can use, along with handlers for
invoking these commands. Then, you obtain grammars for these languages
and embed the parse results of sample commands in those languages into
the NLU system, augmented by type information and other
semantics. Finally, you develop a classical AI planning or intelligent
agent domain, which embeds the grammatical constraints into the
planning domain and so provides the ability to generate commands using
AI planning. So for instance, you might parse "mv file.txt file2.txt",
and so derive facts that there is an <expression> containing a
<command> with <arg1> and <arg2> thus, and that (equals <arg1>
"file.txt") etc, (isa "file.txt" FileName) (exists-file "file.txt")
and (isa "file2.txt" FileName), and then link this to an action object
in the planning domain. Then by AI planning and other domain specific
knowledge, such as derived from scanning the filesystem, you might
compose a command to move a file somewhere. The effects of the
operations can be learned automatically by diffing the set of
environmental observations, or explicitly via either conversion from
man files, or from tacit programming of the preconditions and
postconditions of effects. (such as here
http://arxiv.org/pdf/1210.4889.pdf ) The end result should be a system
that can initiate software commands/programming. For instance, the use
of ssh could be added, as well as standard syntax for many CS-related
concepts, and ultimately, you could have an intelligent agent that
could transport itself from computer to computer, and provided with
more knowledge, could defend or attack a network, manage system
routines, and generally interface with a personal planning system for
the automatic achievement of user goals stated in a natural language.