Kumo’s Relational AI Model Predicts What LLMs Miss
▼ Summary
– VB Transform is a long-standing event for enterprise leaders to discuss AI strategy, including the latest advancements like generative AI.
– Traditional machine learning still dominates high-value predictive tasks like fraud detection, despite the rise of generative AI for text and data processing.
– Kumo AI’s relational foundation model (RFM) applies transformer technology to structured databases, enabling zero-shot predictive capabilities without manual feature engineering.
– RFM automates predictive analytics by converting relational databases into graphs and using attention mechanisms to learn patterns, reducing reliance on data scientists.
– Kumo’s RFM can power AI agents by enabling real-time, data-driven decisions, bridging the gap between unstructured text processing (LLMs) and structured data forecasting.
Enterprise AI is rapidly evolving, but one critical gap remains: the ability to predict future outcomes from structured data. While large language models (LLMs) excel at processing text, they struggle with forecasting tasks like customer churn or fraud detection, areas where businesses still rely on outdated machine learning techniques. Kumo AI, founded by Stanford professor Jure Leskovec, aims to bridge this divide with its relational foundation model (RFM), bringing the adaptability of generative AI to structured databases.
Traditional predictive modeling requires painstaking manual work. Data scientists spend weeks or months engineering features, combining customer purchase histories, website interactions, and other datasets into a single table for analysis. “If you want to do machine learning today, you’re stuck in the past,” Leskovec explains. The process is slow, expensive, and limits how quickly organizations can act on insights.
Kumo’s solution reimagines how AI interacts with databases. Instead of forcing analysts to preprocess data, its relational deep learning approach automatically transforms any database into a dynamic graph. Customer records, transactions, and other tables become interconnected nodes, preserving their real-world relationships without manual intervention.
The real innovation lies in adapting transformer technology, the architecture behind LLMs, to work with these graphs. Transformers use attention mechanisms to weigh relationships between words; Kumo’s RFM applies the same principle to database connections. Just as convolutional neural networks revolutionized image recognition by learning features directly from pixels, the RFM learns predictive patterns straight from raw tables, eliminating feature engineering.
During a live demo, Leskovec showcased how the model delivers instant, zero-shot predictions. A simple query about a customer’s likelihood to purchase within 30 days returned a probability score and an explanation, all without prior training on that specific dataset. “Point the model to your data, and it gives you an accurate prediction in 200 milliseconds,” he said. The system matches the accuracy of weeks of manual data science work, making predictive analytics accessible to non-experts.
This breakthrough has major implications for AI agents. For autonomous systems to operate effectively in business environments, they need more than language skills, they must make data-driven decisions. An RFM-powered agent could, for instance, predict a customer’s churn risk and adjust its interactions in real time, blending predictive intelligence with natural language capabilities.
Kumo envisions a future where LLMs and RFMs work in tandem: one handling unstructured text, the other forecasting structured data outcomes. By removing the feature engineering bottleneck, the RFM could democratize advanced analytics, letting companies act faster on their data.
A public demo is already available, with plans to launch a version for private datasets soon. For organizations needing even higher precision, Kumo will offer fine-tuning services. The goal? To turn every database into a predictive engine, no data science PhD required.
(Source: VentureBeat)
