"implement the lambda calculus on a handful of combinators, then implement lisp on the lambda calculus, then write stuff in that lisp"

Having done this years ago as a student (not Lisp, but a simple language that macro expanded into lambda calculus) - it is both mind-blowing that you can do anything (including recursion) with just S & K but also impressive as to just how phenomenally inefficient this is... though as you say things do get a lot better as your "instruction set" gets larger.