However, not all software engineering groups have the capabilities to produce such highly optimized binaries, and there is always the risk that a user's particular usage patterns will differ enough from the expected patterns that they will lose the benefit of this extensive optimization. However, in an interpreted or byte-code language a lot of the same information needed for optimization is available to the run-time. A run-time designed for optimization may be able to take advantage of that, creating super efficient code paths based on actual usage. This model is more difficult to implement but potentially more robust than statically optimized compilation (and also has the potential to take greater advantage of differences in hardware, a statically compiled native binary doesn't have the ability to morph its optimization based on whether its running on a single core Atom or a 6-way Core i7 cpu, or some 100-core monster of the future, but a run-time potentially can).
In the average case most of this is just theory, but the potential is very real.
Some worthwhile background reading:
Trace Trees: http://www.ics.uci.edu/~franz/Site/pubs-pdf/ICS-TR-06-16.pdf
A blog post / talk from Steve Yegge on dynamic language performance and other topics: http://steve-yegge.blogspot.com/2008/05/dynamic-languages-st...
BTW these technologies are not interpreters, but compilers (but runtime compilers).