The world is being quietly rearranged by people who write very long documents.


The title they went with Streaming Model Cascades for Semantic SQL Noisy translates that to

Database queries now use cheap AI to handle most data, escalating only hard cases to expensive models


Researchers built two algorithms that route database queries more efficiently: a fast, cheap language model handles most rows, and only uncertain cases get sent to a slower, more expensive model for verification. This cuts the cost of running AI-powered database searches at scale without losing accuracy.
Data warehouses are increasingly using language models to search and understand data, but running AI on every row is prohibitively expensive. This solution uses a two-tier system — fast approximation first, expensive verification only when needed — which is how real systems will have to work if AI-augmented databases become standard infrastructure. The practical effect is that semantic search on massive datasets becomes economically viable, which means companies can now build applications that would have been too costly to run before.
Watch whether major database vendors (Postgres, Snowflake, BigQuery) adopt these cascade algorithms in production. If they do within 18 months, semantic SQL moves from research toy to infrastructure. If adoption stalls, the cost problem remains unsolved and these databases stay specialized tools.

If you insist
Read the original →