Open Weights Isn't Open Training
Summary
Open Weights isn't Open Training argues that post-training trillion-parameter models on open-source ML stacks reveal deep inefficiencies that simple patching cannot fix. The author documents attempts to post-train Kimi-K2-Thinking, discusses the limits of existing open-source tooling, and chronicles a sequence of memory and architecture challenges, along with iterative fixes. The piece emphasizes the debt in open-source ML infra and the need for robust, well-integrated training stacks over patchwork solutions.