AI Models and Kernel Optimization Techniques

The author argues that AI models have become orders of magnitude better than humans at kernel work, citing examples where early checkpoints optimized the majority of kernel code for later development. The author notes that tasks like profiling, rewriting, and benchmarking are well-defined and verifiable, making them ideal for agentic hill-climbing, and outlines specific optimization techniques including kernel fusion, programmatic dependent launch, and host loop minimization. Open article →