Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph
Morph achieved 3x model speedup using Autoresearch to automate GPU kernel parameter optimization
“your job is to have good ideas”
Tejas Bhakta from Morph describes applying Andrej Karpathy's Autoresearch framework — an agent-driven while loop — to automate GPU CUDA kernel optimization, achieving 3x model inference speedup. The framework excels at fine-grained parameter search (block sizes, memory tiling) but requires human engineers to supply the high-level architectural ideas. The practical formula: human identifies the inefficiency, Autoresearch searches the parameter space and validates correctness and speed, outperforming manual tuning at scale.