China behind in LLM race but it can still win in AI, ex-Tencent AI lead says
Former Tencent AI lead Liu Junfeng argues China’s AI industry lacks foundational breakthroughs like *fanshi*, relying instead on copying US models such as DeepSeek or OpenAI’s innovations, while benchmark scores fail to capture real-world performance gaps. He cited DeepSeek’s V4 model and Anthropic’s Mythos as examples of US leadership in pushing technical frontiers, despite China’s late-mover status in large language models.
Former Tencent AI chief Liu Junfeng has criticized China’s AI industry for failing to deliver original breakthroughs, instead copying US advancements in large language models (LLMs). Speaking to the South China Morning Post, Liu emphasized the term *fanshi*—a technical paradigm shift like OpenAI’s ChatGPT or Anthropic’s Claude Code—as missing from Chinese AI development. He noted that while companies like DeepSeek have made progress, they lack the core innovations driving global AI leadership. Liu’s departure from Tencent in late 2024 raised questions about China’s AI strategy, especially after Hunyuan, Tencent’s generative AI model, debuted just a year earlier. He argued that benchmark scores understate the real-world gap between Chinese and US models, as US firms continue advancing with projects like Anthropic’s Mythos model launched in April. DeepSeek’s latest V4 update also fell short of previous achievements, reinforcing concerns about China’s late-mover disadvantage. The lack of *fanshi* innovations, Liu said, is China’s biggest weakness, with domestic firms replicating US or DeepSeek technologies rather than leading. He warned that without original advancements, China risks falling further behind in the AI race. The comments highlight tensions between China’s ambition to compete globally and its reliance on imitation rather than innovation in foundational AI models.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.