LLMs still seem to perform more independent computations (incl. fact lookups) than they have neurons; I don't think there's a way to implement this with only monosemantic neurons.

This feels like a non sequitur to me. Computers do just fine with just 0 and 1. All you need is for "neurons" to be sufficiently compositional... or does that mean they're no longer monosemantic? Seems like we need to nail down precisely what we mean by such terms or we may disagree on language even if we predict the same things.