**The TWIML AI Podcast**
5/21/2026
Relational Foundation Models for Enterprise Data with Jure Leskovec - #768
Short summary
Jure Leskovec discusses Kumo's Relational Foundation Model (RFM2), which trains neural networks directly on multi-table enterprise databases without manual preprocessing. The approach uses in-context learning to generalize to new databases with zero additional training, deployed at Reddit, DoorDash, and Coinbase. The episode also covers AI Virtual Cell, learning data-driven representations from proteins to patients using protein language models and AlphaFold.
- •RFM2 trains directly on raw multi-table databases, eliminating manual feature engineering
- •Uses in-context learning over subgraphs to generalize to new databases without retraining
- •Real-world deployments at scale plus complementary AI for science research
Generated with AI, which can make mistakes.
Is this a good recommendation for you?
