I've been thinking a bit about some common ideas between IKFast, Yu-Chi's research, in how they derive a fast, specialized method for solving an optimization problem that is in general very difficult. Maybe that's just the entirety of optimization, is pushing towards a sort of "pareto front" of specialization and speed. I think for IK it makes a lot of sense to generate robot and even IK-type specific solvers, surely they can be made faster than one that works in the more general case. It kind of feels like what modern C++ compilers do, they take a general purpose algorithm and generate a more specialized version of that algorithm for specific hardware, and so it's faster. I think the interesting question might be "is there one (learning) algorithm that can convert a slow, general purpose solver, into a fast, specialized solver"? Other things to think about might how useful is the "space" in which we specialize