Or are the model author’s, i.e the blog author with a vested interest, getting better at optimizing for the test while real world performance aren’t increasing

Or are the model author’s, i.e the blog author with a vested interest, getting better at optimizing for the test while real world performance aren’t increasing as fast?