The article distinguishes between two types of post-training for large language models: capability elicitation, which reweights existing behaviors, and capability creation, which expands the model's behavioral range. This nuanced understanding could influence future research and development strategies in AI training methodologies.