[mikekchar]

You can get a FORTH kernel in 2K words. It's also incredibly efficient for your own code since you basically have a dictionary and memory addresses. Thinking about it, in the old days dictionaries stored 8 character identifiers, which will handily fit in a 64 bit word. That means that the dictionary only needs 2 words per entry.

As you imply, the interpreter will be dwarfed by the code needed to talk to the rest of the OS. As an aside, this is why I was initially very excited about the JVM when Java first came around. Compiling down to a FORTH style language should give you pretty impressive benefits.

Virtual machines were very popular for a long time, but I'm not entirely convinced that we've really pushed the concept as far as it can go.