> "The theory behind these models so aggressively lags the engineering"

The problem is that 99% of theories are hard to scale.

I am not an expert, as I work adjacent to this field, but I see the inverse - dumbing down theory to increase parallelism/scalability.