MLOps.community Relational Foundation Models: Unlocking the Next Frontier of Enterprise AI // Jure Leskovec // #348
55 snips
Nov 25, 2025 
Jure Leskovec, a leading AI researcher and Chief Scientist at Kumo.AI, discusses relational foundation models that revolutionize how enterprises harness structured data. He explains the importance of relational data over document-centric AI and proposes raw-data learning to replace feature engineering. Jure highlights using graph neural networks for efficient database representation, the advantages of relational models in recommendations, and successful implementations like DoorDash's 30% accuracy boost. He also emphasizes the cost-effectiveness and efficiency of these models, transforming the landscape of enterprise AI.
AI Snips
Chapters
Transcript
Episode notes
Learn From Raw Relational Data
- Relational Foundation Models learn directly from raw relational databases instead of hand-engineered features.
- They represent multiple tables as graphs so neural nets can combine signals automatically for better fidelity.
Graphs Capture What LLMs Miss
- LLMs capture common-sense but miss organization-specific relational signals learned from enterprise data.
- Graph-based models capture ‘guilt by association’ and relational structure that sequence models often miss.
Use Pretrained Then Fine-Tune For Performance
- Point a pre-trained relational foundation model at your schema to get instant predictive answers without months of feature work.
- Fine-tune the model on your specific task and database when you need top-tier, task-specific performance.