During the Lang.NET Symposium, a couple of things "clicked" all simultaneously, giving me one of those "Oh, I get it now" moments that just doesn't want to leave you alone.

During the Intentional Software presentation, as the demo wound onwards I (and the rest of the small group gathered there) found myself looking at the same source code, but presented in a variety of new ways, some of which appealed to me as the programmer, others of which appealed to the mathematicians in the room, others of which appealed to the non-programmers in the room. (I heard one of the Microsoft hosts, a non-technical program manager, I think, say, "Wow, even I could understand that spreadsheet view, and that was writing code?")

During the spreadsheet-written-in-IronPython presentation (ResolverOne), we were essentially looking at new ways of writing IronPython code, thus leveraging all the syntactic power of a programming language with a nicer front end.

During the aspect-oriented talk (the one by Stefan Wenig and Fabian Schmeid), we found ourselves looking at a tool that essentially takes compiled assemblies and weaves in additional code based on descriptors from outside that codebase; in essence, just another aspect-oriented tool.

But combine this with my own investigations into Soot, LLVM, Parrot, and Phoenix, alongside the usual discussions around the DLR, CLR, JVM and DaVinci machine, couple that with the presentation Harry gave about parser expression grammars and the research in the functional community into parser combinators, throw in the aspect-oriented and metaprogramming facilities that the Rubyists and other dynamic linguists go on for days about, and what do you end up with?

Folks, the future is in modular toolchains.

This is an oversimplification, and a radical oversimplification at that, but imagine for a moment:

1. A parser takes your source code (let's assume it is Java, just for grins) and builds an AST out of it. Not an AST that's inherently deeply coupled to the Java language, mind you, but a general-purpose one that stands as a union of Java, C#, C++, Perl, Python, Smalltalk, and other languages. (Note that some of the linguistic concepts in some of those languages may not end up in this AST, but instead operate on the AST itself, a la C++'s template facilities.) Said parser is now finished, and can either output a binary (or potentially XML, though it'd probably be hideously verbose) version of this AST to disk for later consumption, or would more than likely be passed directly along to the next beast in the chain.
2. In the simplest scenario, the next beast would be a code generator, which takes the AST and seeks to export some kind of back-end code out of it. Here, since we're working with a general-purpose AST, we can assume that this back-end is flexible and open, a la the Phoenix toolkit (where either native or MSIL can be generated).
3. In a slightly more complicated scenario, verification of the correctness of the AST (against whatever libraries are specified) is checked, usually prior to code-gen, thus making this particular toolchain a statically-checked chain; were verification left out, it would need to happen at runtime, in which case we'd be talking about a dynamically-checked chain.

   Note that I stay away from the term "statically-typed" or "dynamically-typed" for the moment. That would be a measurement of the parser, not the verifier. Verification still occurs in a lot of these dynamically-typed languages, just as it does in statically-typed languages.

   Assuming the verification process succeeds, the AST can be again, written out or passed to the next step in the chain.
4. Another potential step in the process, usually post-parser and pre-verification, would be an "aspect" step, in which a tool takes the AST, consults some external descriptors, and modifies the AST based on what it finds there. (This is how most of your non-AspectJ-like AOP tools work today, except that they have to rebuild the AST from compiled .class files or assemblies first.)
5. Naturally, another step in the process would be an optimize step, but this has to be considered carefully, since some "high-level" optimizations can be done without regard to code-gen backend, and some will need to be done with regard to code-gen backend; for example, register spill is (from what I've heard, can't say I know too much about this) generally only useful if you know how many registers you're targeting. Plus, it's not hard to imagine certain optimizations that are only generally useful on the x86 architecture, versus those that are useful on other CPU platforms. Even operating systems I would imagine would have an impact here. (It turns out that many compiler toolchains go through a dozen or so optimization steps today, so it's not hard to imagine a "code-gen backend" being a series of a half-dozen or so targeted optimization steps before actually generating code.)
6. Bear in mind, too, that these ASTs should have enough information to be directly executable, thus giving us an interpreter back-end instead of a code-generation back-end, a la the DLR instead of the CLR.
7. Also, given the standard AST format, it would be relatively trivial to create a whole series of different "parser"s to get to the AST, along the lines of what the Intentional Software guys have created, thus blowing open the whole concept of "DSL" into areas that heretofore have only been imagined. You still get the complete support of the rest of the toolchain, which is what makes the whole DSL concept viable in the first place, including aspects and verification and your choice of either interpretation or compilation.
8. While we're at it, bear in mind that this AST could/should also be reachable from within the code itself, thus giving languages that want to operate on their own AST at runtime the ability to do so, because the AST is in a standard format and the interpreter could be bundled as part of the generated executable, thus providing a compile-when-you-can-interpret-when-you-must flavor that is currently the reigning meme in language/platform environments like JRuby. (It would also have the happy side effect of making Paul Graham shut up about Lisp, at least for a while. Yes, Paul, code-as-data, it's brilliant, it's wonderful, we get it.)
9. Nothing says this toolchain needs be one-way, by the way: many of the toolkits I mentioned before (LLVM, Phoenix, Soot) can start from compiled binary and work back to AST, thus offering us the opportunity to do surgery of either the exploratory kind (static analysis) or the manipulative kind (aspect-weaving, etc) on compiled code in a relatively clean way. Reflector demonstrates the power of being able to go "back and forth" in this way (even in the relatively limited way Reflector does so), so imagine how powerful it would be to do this from end-to-end throughout the toolchain.

How likely is this utopian vision? I'm not sure, honestly--certainly tools like LLVM and Phoenix seem to imply that there's ways to represent code across languages in a fairly generic form, but clearly there's much more work to be done, starting with this notion of the "uber-AST" that I've been so casually tossing around without definition. Every AST is more or less tied to the language it is supposed to represent, and there's clearly no way to imagine an AST that could represent every language ever invented. Just imagine trying to create an AST that could incorporate Java, COBOL and Brainf*ck, for example. But if we can get to a relatively stable 80/20, where we manage to represent the most-commonly-used 80% of languages within this AST (such as an AST that can incorporate Java, C#, and C++, for starters), then maybe there's enough of a critical mass there to move forward.

Now all I need to do is find somebody who'll fund this little bit of research... anybody got a pile of cash they don't know what to do with?

Update: By the way, in case you want a graphical depiction of what I'm thinking about, the Phoenix page has one (though obviously it's limited to the Phoenix scope of vision, and you may have to be a Microsoft CONNECT member to see it).