We're currently doing an overhaul to our Clojure numerical libraries, particularly those wrapping other libs for performing non-linear equation solving or optimization. Most of these external libs seem to be pretty poor in terms of API, and built on algorithms written for the computing constraints of the 70s-80s. If you look at the code, a lot of it is just screwing around with trying to minimize memory usage etc. and/or duplicating FORTRAN constructs etc. Seems like the state of the art should be better now that we have GPUs, cheap memory, etc. but I'm not having much luck finding things. Any suggestions? One thing I've considered is that given the rich computing resources we have, maybe the thing to do is just roll the basic algorithms rather than drag in large external libraries.