Make this forward pass as fast as possible
Summary
Ian Barber analyzes whether ML compilers remain necessary as agent-driven optimization evolves kernel development. The post covers TVM's call for a minimal, stable IR foundation, Together Compiler's cost-model and verification framework, and examples of how Codex-assisted code mods can affect Triton kernels. It argues that the bottleneck is the search and verification process rather than traditional compilation, and discusses open-source experiments and implications for ML practitioners.