Scott commented on the Turing Machine vs lambda calculus approaches in:
http://www.scottaaronson.com/blog/?p=2725#comment-1085117
showing a clear preference for the "physical" operation of Turing Machines.
In the end, complexities are not that far apart, considering that
https://gist.github.com/anonymous/a64213f391339236c2fe31f874...
describes a 27 state Goldbach checking TM, which takes no more than 378 bits to encode, while
https://github.com/tromp/AIT/blob/master/goldbach.lam
describes an equivalent 267 bit Lambda Calculus term, which can also be represented as this
https://github.com/tromp/AIT/blob/master/goldbach.gif
lambda diagram, as explained in
The proof was actually written in Laconic, and then compiled down to a TM. We already know certain things about TMs (like busy beaver and the halting problem) so you don't have to re-prove those things about your new machine if you can just prove that it's equivalent to a TM. It's mainly used as a convenient way to keep score. If some other proof can be proven equivalent to a TM with fewer states, then you have a better upper bound.
To me, binary combinatory logic and binary lambda calculus seem like much more sensible languages for performing these kinds of algorithmic information theory experiments.