China behind in LLM race but it can still win in AI, ex-Tencent AI lead says

For years, Liu Wei was synonymous with Chinese tech giant Tencent Holdings’ artificial intelligence efforts. A distinguished scientist at the Shenzhen-based company, Liu was also head of its Hunyuan team, the firm’s foundational model development unit for the generative AI era. But in late 2024, Liu’s departure from Tencent after more than eight years sparked immediate speculation as to why he left. Hunyuan was introduced only a year earlier – so why did Liu suddenly quit one of China’s most deep-pocketed tech companies so early into the AI boom? Speaking to the South China Morning Post, the term Liu repeatedly used was fanshi, or “paradigm”. Commonly used among AI researchers, the term refers to a technical breakthrough that defines a new era of AI innovations, notable examples being OpenAI’s ChatGPT and Anthropic’s Claude Code. For Liu, the lack of fanshi innovations is the biggest Achilles’ heel of China’s AI industry. “Chinese companies are either copying DeepSeek or US companies at the core technical level,” he said, referring specifically to the development of large language models (LLM), the centrepiece of the global AI race. Since DeepSeek’s breakout moment early last year, there has been recurring speculation about whether Chinese LLMs have caught up with their US counterparts. However, narrowing public benchmark scores do not accurately reflect a gap in real-world usefulness, said Liu. While US industry leaders have continued to push the technical frontier, notably with the launch of Anthropic’s Mythos model in April, domestic Chinese leader DeepSeek failed to reach the same heights it previously did with its latest V4 model. “There is an obvious gap of at least three months, which will probably expand to six months this year because OpenAI’s GPT 5.6 is coming out very soon,” said Liu.

In the AI era, the problem with being a technological follower is the constant risk of having the rug pulled from under you, he explained, invoking the concept of a “dimensionality reduction strike”. Drawn from the acclaimed Chinese sci-fi novel The Three-Body Problem, in which an alien species vanquishes its enemies by reducing the spatial dimensions of their world, the concept has become an industry metaphor for a technical breakthrough that renders competitors obsolete. “For a company to survive in the global AI race, the one thing it must do is to continuously carry out technological paradigm innovations,” said Liu. “If you forgo this ability, then others will do so and take you down.” This reasoning informed Liu’s decision to bet on an alternative AI path instead: video generation. For the last two years, he and three co-founders have been building their start-up, Video Rebirth, which is headquartered in Singapore though most of its core research and development team is based in Hong Kong. Earlier this month, Video Rebirth launched Bach, an AI video engine targeting enterprise and “prosumer” users. The company has raised US$80 million so far and is closing out another funding round, with another round expected by the end of the year. With Video Rebirth, Liu bet that he was better off trying to lead the next wave of AI innovation than just following the US-led LLM fanshi. Chinese companies currently dominate the global AI video generation space, with leading models including ByteDance’s Seedance and Kuaishou’s Kling.

All but three of the top 10 video models on third party benchmark firm Artificial Analysis’ AI video leaderboard are from China, with Bach debuting in sixth place upon its release. In Liu’s view, AI video generation is a technological paradigm being led by China, the core reason being that it is not as capital-intensive as LLM development. US export controls on advanced semiconductor chips have hampered Chinese efforts to develop cutting-edge LLMs by constraining the amount of compute available to them. However, video generation models, which typically have fewer than 50 billion parameters, are nowhere near as computationally costly to train as cutting-edge LLMs sized at trillions of parameters, making them a natural fit for compute-constrained developers. To reserve scarce resources for this compute-intensive LLM race, US firms have turned away from video generation in recent months, including the closure of pioneering video generation app Sora by OpenAI in late March. “We have to take advantage of this precious window in time to establish a clear lead of something like six months,” said Liu. “This would make it very hard for US companies to catch up again even if they wanted to.” Still, video generation is not immune to the global compute supply crunch. Video Rebirth’s focus on the most lucrative segment of professional users – a strategy to become the “Anthropic of AI video generation” – means that it must offer top-tier features such as video clips that run for up to 30 seconds and 1080p resolution, both of which rack up high compute costs. But the company knows it must continue innovating to survive in a fiercely competitive landscape, with plans to extend maximum video generation length to 1 minute and efforts under way to improve picture quality to native 4k. A long-term focus on technical breakthroughs was what drove the company to set up in Singapore “from day one”, according to co-founder and chief operating officer Dan Kong. “We decided that Singapore was the best place for conducting R&D, securing financing and access to overseas computing power,” he said. This strategy has come to the fore in recent weeks following Beijing’s decision to block Manus’ acquisition by Facebook owner Meta Platforms on national security grounds. While Manus is officially registered in Singapore, it had developed its products in mainland China. However, unlike Manus, Video Rebirth has never conducted research and development on the mainland, with Singapore being its home base since the beginning. The company’s key backers include AMD Ventures, the venture capital arm of US chip giant Advanced Micro Devices. “I think when a Chinese company is able to make very good technology and products overseas using global resources and tap global markets, this should be a point of pride,” he said. For Liu, Video Rebirth is fundamentally an R&D-driven start-up. The long-term plan is to move from advanced video models to eventually world models – real-time interactive models that can accurately simulate the physical world – with a prototype called Olympus already under development. In the third quarter of this year, the company will release Bach 2.0, which will have a completely different technical architecture from the AI video industry standard today of diffusion transformers. This, said Liu, is indicative of how he and his company can ultimately prevail through fanshi defining innovations. “My advice to Chinese entrepreneurs is that they need to be serious about technology,” he said. “In AI, being serious means doing real innovation, which means having hardcore technology.”