The article addresses the mixed perceptions surrounding DeepSeek’s contributions to the AI field, particularly regarding their model, DeepSeekCoder-V2, released in June 2024. The author critiques how the community either undervalues or overvalues their achievements. They highlight that DeepSeek was among the few teams, alongside Gemini and OpenAI, to successfully implement reinforcement learning (RL) on verifiable rewards early on, marking them as significant players in a competitive space. The author expresses skepticism about the idea that DeepSeek would merely replicate existing models from OpenAI’s research, particularly the o1 model. They emphasize that enough information about o1 was already available publicly, making it unlikely that DeepSeek relied heavily on it for their development. Instead, they suggest that DeepSeek’s RL team possesses the expertise to generate their own approaches, indicating a level of innovation and capability that merits recognition. This perspective challenges the narrative that diminishes their role in advancing AI technology.