The Data Exchange with Ben Lorica The AI Revolution Finally Comes to Structured Data
22 snips
Dec 4, 2025 
Jure Leskovec, a Stanford professor and co-founder of Kumo.ai, dives into the transformative power of relational foundation models for structured enterprise data. He challenges the current limitations of AI in handling relational data, emphasizing the shortcomings of treating tabular data as text. Jure outlines Kumo’s rapid predictive SQL-like language, innovative graph representations, and the model's ability to handle messy data effectively. He also discusses real-world successes like DoorDash's significant improvements and the potential applications of these models across various industries.
AI Snips
Chapters
Transcript
Episode notes
DoorDash Gained 30% On A Recommender
- DoorDash used Kumo to build a restaurant "try something new" recommender.
- They achieved a 30% improvement and deployed a production model within a few months.
Match Inference Mode To Latency Needs
- Choose online inference or precompute embeddings depending on latency needs.
- Use batch embedding materialization to achieve sub-20ms real-time scoring when required.
Auto-Generate Fine-Tuning Data With Time Travel
- Fine-tune by specifying temporal predictive tasks and auto-generate labeled examples via time-travel windows.
- Use sliding windows over historical data to produce training examples without manual labeling.