Thoughts on Cloud, DevOps, SRE, MLOps, and HPC — published on LinkedIn.
A measured Python-to-Rust port, one step at a time. Free-threaded CPython lands and a single Rust core still wins by 2.3x — and the first port I wrote was only 1.51x, because the boundary was in the wrong place.
Python has two performance ceilings, not one. Why PyO3 beats a rewrite, what MLIR gives Mojo that plain LLVM can't, and why unified-memory silicon like M5 and DGX Spark makes CUDA lock-in a liability.
After 18 years of DevOps, I thought MLOps was just CI/CD for models. Here's what actually changes — and where your existing skills hit a wall.