Comment by zozbot234

So they're paying expensive input tokens to extract at best a tiny amount of information ("judgment") per request? That's even less like "distillation" than the other claim of them trying to figure out reasoning by asking the model to think step by step.