Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph

Why watch

Tejas Bhakta explains how Morph used AI agents to write custom GPU kernels combined with bare-metal hardware tweaks to achieve 3x inference speedup on cheaper GPUs. The video covers how humans provide high-level ideas while agents handle tuning (block sizes, parameters), the importance of giving agents hardware and model context, common reward hacking issues like CUDA graphs, and how kernel optimizations compound. Target audience: ML infrastructure engineers, inference engineers, and developers working on model optimization.