Stonebraker: text-to-SQL hits 0% accuracy on MIT's real warehouse

Postgres creator Michael Stonebraker and PostgreSQL core team member Bruce Momjian said text-to-SQL scored zero-percent accuracy on MIT's 1,400-table warehouse, versus over 80% on Spider and BIRD benchmarks. They argue LLMs handle joins of structured data poorly and that extensible, vendor-neutral systems will win.

• Text-to-SQL tops 80% on Spider and BIRD but scored 0% on MIT's data
• Agentic add-ons lifted accuracy only into the 30s percent range
• Their Rubicon system uses models for text and a tabular engine for joins
• PostgreSQL 18's August 2026 patch release fixed 28 CVEs