Oct 10, 2026

From 15% to 90% GPU Utilization: Fix the Data Pipeline, Not the Model

Tarun Sunkaraneni from Amazon AGI shows how a multimodal SFT pipeline that spent about 85% of its time waiting on data was brought to roughly 90% GPU utilization. He walks through concurrency with asyncio and Ray actors, prefetching, zero-copy transfer via Ray's object store, and a network-card bottleneck that appears only at scale, and argues that the hardest part of training throughput is often keeping GPUs fed rather than optimizing the model.

Aug 25, 2026

The 12 KB File That Replaces Weeks of Training (with Tristan Handy)

Tristan Handy, founder and CEO of dbt Labs, details the evolution of analytics engineering from a 2016 study into a tool used by over 100,000 data teams. He explains his decision to use SQL over Spark for accessibility, the concept of "progressive complexity," and how dbt projects transform raw data into modeled tables using a Directed Acyclic Graph. Handy elaborates on the critical role of the semantic layer in ensuring consistent metric definitions for both human users and AI agents, especially in large organizations. He introduces the dbt Fusion Engine, aiming to bring type safety and universal SQL understanding, and discusses how 12-kilobyte skill files can revolutionize large-scale dbt migrations, reducing them from years to weeks by enabling AI agents to absorb vast amounts of expert knowledge.