> If all the formal semantic models for a language are unwieldy then you've probably got a non-simple language.
Now, "simplicity" is a mental construct, a language UX construct. To handle this, I think of "unwieldy" as a bit of a technical term. What does it mean to be unwieldy? It means that there is significant non-ignorable complexity.
Significant here must be defined almost probabilistically, too. If there is significant complexity which is ignorable across 99/100 real-world uses of a language then it really should win some significant points.
Ignorable complexity is also an important concept. It asks you to take empirical complexity measures (you mention Kolmogorov complexity; sure why not?) and temper them against the risk of using a significantly simpler "stand-in" semantic model. I accept that the stand-in model will fail to capture what we care about sometimes, but if it does so with an acceptable risk profile then I, pretty much definitionally, don't care.
Now that I've weakened your idea so much, it's clear how to slip in justifications for really terrible languages. Imagine one with a heinous semantics but a "tolerable" companion model which works "most of the time".
From this the obvious counterpoint is that "most of the time" isn't good enough for (a) large projects (b) tricky problems and (c) long support timelines. Small probabilities grow intolerable with increased exposure.
---
But after all this, we're at an interesting place because we can now talk about real languages as being things with potentially many formally or informally compatible formal or informal semantic models. We can talk about how complexity arises when too few of these models are sufficiently simple. We can also talk about whether or not any of these models are human-inelligible and measure their complexity against that metric instead of something more alien like raw Kolgomorov complexity.
So here's what I'd like to say:
> Languages which hide intolerable complexity in their semantics behind surface simplicity are probably bad long-term investments.
and
> Languages which have many "workably compatible" semantic models, each of which being human-intelligible, are vastly easier to use since you can pick and choose your mode of analysis with confidence.
and
> Value-centric semantic models (those ones with that nasty idea of "purity" or whatever) are really great for reasoning and scale very well.
In particular, I'm personally quite happy to reject the assertion made elsewhere that value-centric semantics are not very human intelligible. On the other hand
> Simple operational semantic models are also pretty easy to understand
I just fear that they scale less well.