I'm more curious about going from text to Prela instead of going from text to SQL and measuring any difference in the performance there. On one hand models have been trained on a lot of SQL on the other hand they are really good in mathematical reasoning too so thinking in Perla might be a natural fit for them.
Yes, maybe not the language itself, but the ideas behind it. Tarski's Algebra of Relations is actually a better model for modern columns stores than the standard relational algebra, because a column is a binary relation from the primary key into its value.