Also skip the first loss values being generated since they're biased towards "not lost" due to the initialization.
