Inhye Jin, Jonathan A Tangsrivimol, Erfan Darzi, Hafeez Ul Hassan Virk, Zhen Wang, Jan Egger, Sean Hacking, Benjamin S Glicksberg, Markus Strauss, Chayakrit Krittanawong
Front Artif Intell. 2025 Jun 19:8:1576992. doi: 10.3389/frai.2025.1576992. eCollection 2025.
Abstract
DeepSeek has introduced its recent model DeepSeek-R1, showing divergence from OpenAI's ChatGPT, suggesting an open-source alternative to users. This paper analyzes the architecture of DeepSeek-R1, mainly adopting rule-based reinforcement learning (RL) without preliminary supervised fine-tuning (SFT), which has shown better efficiency. By integrating multi-stage training along with cold-start data usage before RL, the model can achieve meaningful performance in reasoning tasks along with reward modeling optimizing training process. DeepSeek shows its strength in technical, reasoning tasks, able to show its decision-making process through open source whereas ChatGPT shows its strength on general tasks and areas requiring creativeness. Despite the groundbreaking developments of both models, there is room for improvement in AI landscape and matters to be handled such as quality of data, black box problems, privacy management, and job displacement. This paper suggests the future of AI, expecting better performance in multi-modal tasks, enhancing its effectiveness in handling larger data sets, enabling users with improved AI landscapes and utility.