Today’s foundation models excel at text and images—but they miss the relationships that define how the world works. In every enterprise, value emerges from connections: customers to products, suppliers to shipments, molecules to targets. This talk introduces Relational Foundation Models (RFMs)—a new class of models that reason over interactions, not just data points. Drawing on advances in graph neural networks and large-scale ML systems, I’ll show how RFMs capture structure, enable richer reasoning, and deliver measurable business impact. Audience will learn where relational modeling drives the biggest wins, how to build the data backbone for it, and how to operationalize these models responsibly and at scale.

An end-to-end DenseNet-121 pipeline on the MedNIST dataset was rebuilt using NVIDIA’s GPU-native tools, replacing traditional CPU-based stages like Pillow, OpenCV, and PyTorch DataLoader. The GPU workflow delivered 3.3× higher throughput, ~3× lower latency, better memory efficiency, and higher hardware utilization on a Tesla T4. The post also outlines future gains through TensorRT, INT8 quantization, RAPIDS cuDF, and GPUdirect Storage to push medical imaging pipelines closer to real-time performance.

Jeff Huber drops some hard truths about “context rot” — the slow decay of AI memory that’s quietly breaking your favorite models. From retrieval chaos to the hidden limits of context windows, he and Demetrios Brinkmann unpack why most AI systems forget what matters and how Chroma is rethinking the entire retrieval stack. It’s a bold look at whether smarter AI means cleaner context — or just better ways to hide the mess.