“The historical approach of creating AI benchmarks to target specific, isolated human skills has been largely unsuccessful because models tend to overfit to them.”