As I understand current movement in Racket, PyPy or JS, them use self interpreter as knowledge machine at first run|trace. Code that executed once newer be optimized. Only parts of code which frequently executed (a.k.a 'hotspots) undergoes optimization steps. Which one should be specialization (selection|substitution) of procedures with well defined argument | return type signatures exactly. Static type inference as thing we have well defined in strong typed langs, for example in OCaml. Inference as process in Lisp may be special form of EVAL.