No transcript available for this episode.
This episode examines a modality specific shift in predictive modeling, where relational deep learning from research associated with Jure Leskovec at Stanford and the Kumo team treats databases as graphs for graph transformer attention. It covers Sakana AI’s Fugu model, a reinforcement learning trained orchestrator that selects different models at each reasoning step to optimize cost performance, and the Namazu post training effort that decensors, debiases, and re aligns open models while managing catastrophic forgetting. The briefing then turns to governed composable customer data platforms signaled in Snowflake’s Modern Marketing Data Stack report, two contrasting agentic deployment patterns from Gap Inc. and NTT DATA, and a Thomson Reuters survey showing high weekly AI usage but widespread failure to extract value. Operationally, it connects model iteration cadence, benchmark fragmentation, and infrastructure governance to the deployment and accountability decisions that determine real world value.