I think you missed the main thrust of his complaint: back when everyone was hacking in FORTRAN and COBOL, you really didn't need to know much more than the language, its standard library, and stuff specific to the problem you were solving. At worst maybe you'd have to learn BLAS and LAPACK if you were doing numerical stuff. The APIs/overhead/junk memorization situation has been steadily worsening since then.
But at least now we don't have to take a stack of punched cards to the Priest with an Offering (bribe) and wait a day just to get a printout saying that the damn thing dumped core.
You still have the option of using FORTRAN and COBOL with BLAS and LAPACK, if you're satisfied with the result.
The problem is that today we demand much more than what those earlier tools can give us. Such demands result in an inevitable (?) increase in complexity of the aggregate toolset. The tools are capable of more, and give you more options. So it seems that learning and using those tools effectively would be more challenging.
But it's not a one-way street. Along with this increase in complexity have been efforts to simplify, streamline, and structure these tools and ease their use. But both the demand for increased ease of learning/use and the demand for power are moving targets. Hopefully with time both could be maximized. We're just not there yet. Remember, computer science is still in its infancy.
But at least now we don't have to take a stack of punched cards to the Priest with an Offering (bribe) and wait a day just to get a printout saying that the damn thing dumped core.