As Knuth himself said, literate programming facilitates the task of thinking about a problem without having to use a formal language to construct such understanding bottom-up. Therefore literate programming can, for example, produce much less stack depth in a program while making it easy to understand and maintain. This is quite an interesting aspect with regards to performance of the resulting code, that is often overlooked. He also claims that if TeX had been written without it it would not have a depth of 4-5 subroutines on the call stack, but around 50-100 [1]. He also states that he would not have been able to produce MMIX [2] without literate programming because he would not have been able to reason about it's design bottom up [1].
So, a lot of complexity is added to a program just so that it can be understood and maintained in a formal language; but he also states that many modern comment styles go a long way towards literate programming.
Having worked on large codebases I concur that using literate programming to construct these seems far-fetched and too academic, yet I am enticed by the idea of optimizing code by literate programming since I have seen far too many abstractions for the purpose of maintaining easy to read code.