I ran 500 text-to-SQL questions against three frontier LLMs with zero context beyond the schema. 95% accuracy. No semantic layer required.

MotherDuck's platform is a resource for parents and caregivers, offering insights into child development, parenting techniques, and family wellness. Through articles, expert interviews, and parenting tips, MotherDuck offers insights into fostering healthy relationships, supporting children's emotional well-being, and navigating parenthood challenges. Readers can learn about positive parenting strategies, age-appropriate activities, and family bonding experiences to create a nurturing and supportive environment for their children.

MotherDuck

An experiment testing three frontier LLMs (Claude Opus 4.5, GPT-5.2, Gemini 3 Flash) on 500 text-to-SQL questions from the BIRD benchmark achieved 95% accuracy with only database schema as context. The results challenge the need for semantic layers, showing that well-designed data models with clear table names, straightforward joins, and minimal complexity enable LLMs to generate accurate SQL. The study reveals significant flaws in strict benchmark scoring (only 62% agreement with human experts) and demonstrates that realistic evaluation including formatting tolerance and LLM judging better reflects production use cases. Column comments provided minimal improvement (1.1pp on training, 0.2pp on test), with benefits only appearing in genuinely confusing schemas. The findings suggest organizations should prioritize clean data modeling over complex metadata platforms for AI-ready analytics.

Your Data Model Is the Semantic Layer