Finding the shortest program isn't as important as having a simple
metric like compression ratio that can act as a measure of
intelligence.  Engineering is largely about progressive optimization of
some figure of merit.  If you don't have a figure of merit you are
going to be awash in subjective valuations.  I did quite a bit of
practical early work in neural network image segmentation.  While
pruning was crucial for generalization -- and led me to my belief there
was a lot more to Ockham's Razor than a nice rule of thumb -- I didn't
see much evidence of anyone using lossless compression ratio as a test.

Moreover the assumptions you point out are as reasonable as any people
routinely adopt for massive investments in technologically risky
enterprises.  I think Hutter's proof is a really important if for no
other reason than that it provides a strong basis for using compression
ratio as the figure of merit that has been missing from AI.