When DeepSeek-R1 launched recently, it immediately captured the attention of the global artificial intelligence community, prompting major players such as OpenAI, Microsoft, and Meta to investigate its seemingly novel approach to model distillation. Yet, beneath the excitement around distillation lies a more nuanced and impactful innovation: DeepSeek's strategic reliance on reinforcement learning (RL).
Traditionally, large language models (LLMs) have been refined through supervised fine-tuning (SFT), an expensive and resource-intensive method. DeepSeek, however, shifted towards reinforcement learning, optimizing its model through iterative feedback loops. This method dramatically reduced costs, up to 90% compared to traditional methods such as those used by ChatGPT, while delivering comparable or even superior performance in various benchmarks.
