A groundbreaking study published in Nature details the R1 model by DeepSeek, a Chinese AI company, showcasing how its high-level reasoning capabilities drastically reduce training costs by 300 times compared to GPT-4. With a training expenditure of just $294,000—excluding GPU and labor costs—R1 employs a pure reinforcement learning (RL) strategy, diverging from traditional models that rely on human feedback. This innovative approach allows R1 to autonomously derive reasoning strategies from correct answers, achieving accuracy that surpasses human averages in complex tasks like mathematical Olympiads. The model’s training, primarily using Nvidia H800 chips, emphasizes self-developed verification and reflection processes through the “Group Relative Policy Optimization (GRPO)” technique. Unlike OpenAI’s ChatGPT, which focuses on generating human-favored responses, R1 represents a novel paradigm optimized for inference. This peer-reviewed paper marks a significant milestone in large-scale language model research, setting a precedent for transparency and safety in AI development.
Share
Read more