The plan calls for a mesh of Go probe workers across five cloud providers, designed to spot submarine cables in ordinary traceroutes. Two rounds of adversarial review reshaped the design before I wrote any code.
A public benchmark of ANE, CoreML GPU, MLX, and llama.cpp Metal for text embeddings on M4 Pro and M5 Max: ANE turns out to be the slowest GPU-class path, and an earlier 6,950 sentences/sec MLX number was wrong.