Once it grows past a certain size, you need the right kind of IDE, etc. to cope with the conceptual sprawl. People start duplicating functionality because they aren't aware things already exist, minor details that handle historically recurring bugs get overlooked in the noise, it takes much longer for new people to understand the codebase, etc. A 100k codebase is not just ten times harder to maintain than a 10k codebase.
Not that I'm defending LOC. If one has to count something, counting tokens seems less susceptible to noise.
Besides, the article talks about implementing the same algorithm twice in different languages, if you went about that in a completely different way then that would of course be problematic. But when comparing the expressiveness between two languages you'd assume the same major structures would be present.
It also doesn't account for language misfeatures that introduce significant additional cognitive burden for the programmer but don't appear much in source code. For example, manual memory management in C++ results in a bunch of "delete x" calls in code, which would get GZipped down to very little, but impose a very high cognitive cost on the programmer. PHP's inconsistency in argument order doesn't show up at all in code length, but also presents a big cognitive tax that sends you back to the reference manual all the time.
Personally, I think that the best metric of language productivity is "The amount of information you have to keep in your head in order to write code as fast as you can type." Unfortunately, that's nearly impossible to measure.
After all, there is not much to remembering that two functions exist for the same type, it's not like you have to remember that there is something 'completely' new, just a routing that is almost the same as another.
Caching also means that if you execute less code you will have better performance.
The actual instructions can be interpreted serially, in fact when you're 'desk checking' that is exactly what you do, to unwind the way the code is written to the serial stream that the processor is executing.
I don't think you can objectively measure readability. It is tightly coupled with concrete syntax, which is highly subjective.
That's not the best metric to begin with (I used to think so too, but someone here pointed out that gzipped source is already much better), and besides that, there are other ways of looking at it besides 'size', such as readability.