But wouldn't that be horribly inefficient? Instead of executing code by transforming it into a set of weights and running inference on them, wouldn't it be easier to just execute code as... code?