Nathan Lambert, head of the post-training team at the Allen Institute for AI (AI2) and a leading authority on RLHF (reinforcement learning from human feedback), has published an observational report from visits to several top AI labs in China. His findings reveal a fractured competitive mentality in China's domestic large language model (LLM) ecosystem: the entire industry fears and resents the financial muscle of ByteDance and Alibaba, yet universally acknowledges DeepSeek as the true technical leader.
Lambert notes that ByteDance, as China's only closed-source frontier lab, terrifies competitors with its resource advantage — it's seen as a monopolistic giant that will eventually buy its way to dominating the market. DeepSeek, by contrast, sets the technical direction with impeccable research taste and earns pure respect from peers, even though many believe it's "not built to win commercially." Meanwhile, Chinese developers heavily rely on Claude for coding assistance and are desperately hungry for Nvidia's compute power.
When explaining why Chinese models are catching up to the US so quickly, Lambert argues that the moat is cultural, not technical. Today's LLM training is an extraordinarily tedious systems engineering effort. The US research culture worships "star-making," where individual researcher interests often conflict with overall model optimization — Silicon Valley-style scientific arrogance hobbles collaboration (the Llama team experienced such turmoil).
In contrast, while OpenAI and Anthropic barely let interns touch core operations, Chinese labs rely heavily on a massive workforce of current students. These young people never experienced the early AI hype, nor do they carry the philosophical baggage of debating "humanity's destiny." They are ruthlessly pragmatic and willing to take on the dirtiest tuning work.
Lambert's report shatters the stereotype that Chinese models only copy open-source projects, and reveals a harsh reality: the LLM race is no longer a research contest of "genius ideas" but an engineering war of attrition where discipline wins. China, powered by its unburdened student army, is rapidly closing the gap.