The numerical ecosystem could really use a modern, optionally-multithreaded BLAS under a BSD-like license with a priority on

• Correctness
• Out-of-the-box single-binary functionality (e.g., runtime kernel selection, runtime thread control)
• Speed
• Portability

...in roughly that order.

OpenBLAS is currently the library that's closest to providing these things, but there are a number of improvements possible. Fixing these might make some good concrete targets for people to go after:

• The path leading to getting a generally-useful build is lined with tricky booby-traps (e.g., automagic capping of the maximum number of threads and the famous NO_AFFINITY).
• There are concerns about lack of tests. That link lists a number of specific bugs that made it past the existing test suite and still are not tested for; in general it would be very useful to build up a set of comprehensive BLAS/Lapack tests that includes tests for realistic problem sizes.
• It's not possible (?) to override CPU detection at runtime, which makes it hard to run comprehensive tests.
• The use of AT&T-syntax inline asm (?) prevents the use of MSVC; using intrinsics instead might be more maintainable and certainly more portable. MSVC now supported
• ...any more?