OpenAI believes DeepSeek used a process called “distillation,” which helps make smaller AI models perform better by learning from larger ones.