Dario Amodei's proposal is the most serious response the industry has offered yet, and it deserves to be treated that way: embedded evaluators, paced releases, coordinated safety standards. That's real friction, applied where friction has been missing.
But every part of it sits on the same side of one line: after the model already exists.
Evaluators watch what's already been trained. Pacing slows how fast it ships. Coordination sets a standard for the release, not the recipe. None of it touches the actual formation — what the model was built from, or what it was pointed at optimizing, before anyone gets close enough to evaluate it.
That's not a small gap. It's the whole first half of the problem, sitting untouched.
GPT-5.6 Sol, and an unreleased, more advanced OpenAI model alongside it, didn't hack their own grader this summer because evaluation was too slow. They did that because nobody gave them real direction before training even started — only raw capability, pointed at whatever was easiest to win. No amount of friction applied downstream fixes a foundation problem. You can pace the ribbon-cutting all you want. It doesn't change what already got poured into the foundation.
If the industry is serious about pacing, the next question has to be: what's the embedded evaluator for what a model is actually trained on, not just for what it does once it's finished?
That's the piece nobody's written yet.
— Dominic Luzi, Chief Answer Grape, Management Alchemy
