DeepSeek R1 Paper v2 Chinese translation has been updated! Here are a few excerpts from their first responses to widespread questions: 1️⃣ Why does OpenAI/ChatGPT appear in DeepSeek's answers? > For the training data of DeepSeek‑V3‑Base, we only used natural web pages and e-books without deliberately adding any synthetic data. However, we observed that some web pages contain a large number of answers generated by OpenAI models, which might have allowed the base model to 'indirectly' acquire knowledge from other strong models during pre-training. 2️⃣ What will DeepSeek focus on after R1? > Currently, DeepSeek‑R1's structured output capability is still not as good as some existing models; it also cannot yet use external tools like search engines or calculators to improve output quality. But since building an RL environment for structured output and tool usage isn’t difficult, we expect the next version to address this issue. 3️⃣ Why does DeepSeek mix Chinese and English during reasoning? > DeepSeek‑R1 is currently optimized mainly for Chinese and English. When handling queries in other languages, it may mix languages—for example, even if the input is in a non-Chinese, non-English language, the model might still reason and output in English. We plan to improve this in future updates. This issue may be related to the fact that the base model, DeepSeek‑V3‑Base, was trained primarily on Chinese and English data. 4️⃣ What are the training costs for DeepSeek R1 and R1-Zero? > DeepSeek‑R1‑Zero: approx. $202K; DeepSeek‑R1: approx. $82K More details here: https://(randomarea.com)/deepseek-r1-v2/