If you submit an improvement to atlas-comm@cs.utk.edu, we hope that other ATLAS users will try it out, and give comments. Ultimately, someone in the ATLAS group (currently, that would be me) will decide whether to include it in the standard tarfile or not. Such a decision will be based on how hard it is to get to work with our distribution (this pain should be minimal if the instructions in this note are followed), how much of an improvement it is, etc. Obviously, there are speed improvements that some users may utilize that the ATLAS group can't indulge in (examples include speedups that possibly lose accuracy, such as, for instance, using Strassen's algorithm for matrix multiply). If there appears to be a great demand for such lossy optimizations, we may consider ways of allowing users to select if they want to risk them.
Note that even if we are ungrateful morons who do not release your submission, you, and anyone who wants to use your kernel routines, can use the ATLAS infrastructure to get a complete, fast BLAS implementation.