Stonebraker says enterprise AI needs trusted databases and realistic testing

Michael Stonebraker, who introduced Postgres in 1986, argues that enterprise AI needs dependable systems built around structured data, and databases governed by open communities. He says language models still struggle with complex joins and that AI demonstrations on public benchmarks can give businesses an unrealistic picture of how well text-to-SQL works on their own data.

Using real queries against MIT’s warehouse of 1,400 tables, his team reported zero per cent accuracy at first; adding context and AI techniques raised it into the teens, and describing the schema brought it into the 30s. Stonebraker says his Beaver benchmark has likewise seen no results beyond the 30-per-cent range. His Rubicon system uses models to interpret text, then passes structured results to a tabular engine; he says it is ten times cheaper and more accurate than model-centred agent frameworks. He also argues Postgres’s independence from any single company is valuable for enterprise AI.