we're already doing that, it's called distillation and how models like deepseek are trained.