Now we're thinking about model safety in the right way. Would love to see experiments on how much of a speedup this actually gives vs. a more naive finetuning approach.