Chen Boyuan Speaks on Zhihu: Behind the Scenes of OpenAI's Chinese Language Optimization

The Man Who 'Fixed Chinese' at OpenAI Finally Speaks Up

Recently, Chen Boyuan — a somewhat legendary figure in China's AI circles — posted an update on Zhihu that quickly drew community attention and discussion. As one of the core members responsible for optimizing Chinese language capabilities within OpenAI, every public statement from Chen is regarded as a precious window into OpenAI's Chinese language strategy.

For the tens of millions of ChatGPT users in China, "the man who fixes Chinese at OpenAI" is already a familiar moniker. From GPT-3.5 to GPT-4, and now GPT-4o, every leap in Chinese language capability has been inseparable from the efforts of Chinese engineers and researchers like Chen Boyuan.

From 'Functional' to 'Fluent': The Arduous Path of Chinese Optimization

Early GPT models were widely criticized for their poor Chinese performance — inefficient tokenization, excessive token consumption, semantic misunderstandings, and stiff, unnatural expressions. These issues significantly undermined the product experience for Chinese users. As a highly complex language, Chinese differs enormously from English in grammatical structure, cultural context, and expressive conventions, posing extremely demanding requirements for large language model training and optimization.

Chen Boyuan's work tackled precisely these core challenges. From improving the Chinese tokenizer, to curating and cleaning high-quality Chinese corpora, to fine-tuning Chinese output quality during the RLHF (Reinforcement Learning from Human Feedback) stage — every step required talent with both technical depth and linguistic intuition.

Indeed, GPT-4o's significant improvements in Chinese capability are plain for all to see. Whether in literary creation, professional translation, or complex logical reasoning tasks, the model's Chinese performance has made a qualitative leap over its predecessors. The dramatic improvement in token efficiency has directly reduced costs for Chinese users — a direct result of tokenizer optimization.

The Role of Chinese Talent in the Global AI Landscape

Chen Boyuan's experience reflects a broader phenomenon: Chinese researchers are playing an increasingly important role in the world's top AI laboratories. From OpenAI to Google DeepMind, from Anthropic to Meta FAIR, Chinese scientists and engineers are everywhere. They have not only made outstanding contributions at the general technology level but are also playing an irreplaceable role in advancing AI's multilingual capabilities and cross-cultural understanding.

Notably, this work of "fixing Chinese" goes far beyond simple translation or localization. It involves design considerations at the model architecture level, strategic configuration of training data, multi-dimensional construction of evaluation systems, and a deep understanding of the needs of Chinese-speaking users. It is a systematic endeavor that requires finding a balance between technology and the humanities.

Reflections and Lessons for Domestic Large Language Models

Chen Boyuan's public remarks on Zhihu have also prompted many industry insiders to re-examine the approaches domestic LLM developers have taken in building Chinese language capabilities. Compared to OpenAI's strategy of gradually expanding multilingual capabilities from an "English-first" foundation, domestic models such as Tongyi Qianwen, ERNIE Bot, and Kimi are naturally built around Chinese, giving them an inherent advantage in the nuance of Chinese understanding and generation.

However, OpenAI's Chinese optimization experience also offers important lessons:

• Data quality over quantity: A robust mechanism for curating high-quality Chinese corpora is essential
• Evaluation systems need localization: English benchmarks cannot simply be applied to measure Chinese capabilities
• User feedback loops: Real user feedback is a key driver of model iteration
• Cultural context understanding: The ultimate test of language capability lies in grasping the deep logic of culture

Looking Ahead: The Next Chapter for Chinese Large Language Models

As global LLM competition enters deeper waters, Chinese language capability has evolved from an "add-on feature" into a strategically competitive dimension. OpenAI's continued investment in Chinese optimization itself speaks to the enormous value and strategic significance of the Chinese market.

Chinese technical talent like Chen Boyuan is becoming a critical bridge connecting Eastern and Western AI ecosystems. Their technical insight and cultural understanding not only help overseas models better serve Chinese-speaking users but also provide valuable reference perspectives for the development of domestic large language models.

We look forward to Chen Boyuan sharing more observations and reflections from the OpenAI frontlines on Zhihu. In this era of rapidly evolving AI, every authentic voice from the inner circle is invaluable.