Can LLMs Build a 10x Faster Database? GenDB & BespokeOLAP [PWL BK]

Two papers posted to arXiv on the same day, from two different labs, with the same idea: use an LLM coding agent to synthesize query execution code instead of running queries through a general-purpose engine. The results: 10x faster than DuckDB on TPC-H, with per-query wins up to 1,466x.

Alex Kouzemtchenko, CTO of Espresso AI, breaks down GenDB (Cornell) and BespokeOLAP (TU Darmstadt) at Papers We Love Brooklyn: how the pipelines work, what the ablations actually show, what it costs, and why LLM-generated query code isn't production-ready yet.

PAPERS
GenDB: https://arxiv.org/abs/2603.02081
GenDB repo: https://github.com/SolidLao/GenDB
BespokeOLAP: https://arxiv.org/abs/2603.02001
BespokeOLAP repo: https://github.com/DataManagementLab/...

Recorded at Papers We Love Brooklyn, June 2026. Papers We Love is a community that reads and discusses computer science papers. https://paperswelove.org/

Come to the next one! https://luma.com/2p3was2k

Espresso AI automatically cuts Snowflake and Databricks costs by 40-70%. Free savings estimate, no engineering effort: https://espresso.ai
#databases #llm #duckdb #sql #ai