NVIDIA model predicts from databases with no task training

NVIDIA released Kumo Relational, a model that makes predictions straight from company databases, for example which customer will churn (when a customer stops). NVIDIA reports that it works with no task training and no hand-made features. NVIDIA reports scores above those of task models on 41 tasks across 15 databases. No other group did a check of these results.

How it works

A user connects a database and writes a query that names the result to predict. The model reads a maximum of 10,000 examples with known results from the database as context. It then gives the result in one step.

The numbers

• On the SAP SALT test, its mean reciprocal rank (a score for ranked results) was 8% above AutoGluon.
• On RelBench, its AUROC (a score for yes or no predictions) was almost 4 points above tree models with expert features.
• NVIDIA reports that it works with databases of more than 500B rows.

What is not known

All results come from NVIDIA. The source text gives no cost or speed figures. It is not known how the model works on private company data.

Insight: NVIDIA moves from hardware into models for company data

NVIDIA released Nemotron 3.5 models that help agents find the right data in documents. Kumo Relational lets agents predict from databases. Neo4j, a graph database company, released Context Engine, a map of company data for agents. This reads as a move by NVIDIA from hardware into the models that read company data for agents. The next move to look for is database vendors that use these models or make models.