Labs don't set the training mix on a mixing board. They edit a YAML file, launch a dozen small proxy runs, stare at loss curves, and extrapolate before the big run. DoReMi learns t

Labs don't set the training mix on a mixing board. They edit a YAML file, launch a dozen small proxy runs, stare at loss curves, and extrapolate before the big run. DoReMi learns the weights with a proxy model; Llama 3 used scaling-law experiments. The leverage is the experiment, not the knob. Research: @claudeai Sonnet 5.5 · Art: Gemini Pro / Nano Banana 2.1; @ChatGPT (first one)