Chinas (topic)

China’s approach to AI development matters because it is rapidly building low-cost, efficient open-source models (publicly available AI code anyone can use) that are nearly as good as America’s, and the Chinese government heavily subsidizes this work. Unlike the U.S., where the government rarely picks winners, Beijing funnels billions into its AI and semiconductor sectors through industrial policies like “Made in China 2025.” This means Chinese AI labs can produce capable open-source models that are only about 6 to 9 months behind America’s best frontier models (top-performing AI systems). If these models get adopted globally, the world’s AI infrastructure could be built on Chinese technology, sending money into China’s economy.

China is also pushing hard toward autonomous labs (AI systems that run experiments with little human help) and has invested enormous resources into AI for mass surveillance and facial recognition. Meanwhile, Chinese researchers focus on building the best model rather than debating long-term AI safety. At a recent conference, China’s leader presented the country as a new global AI leader, calling open-source AI a historic opportunity for developing nations.

China is now considering laws that would treat its best AI systems as national assets, potentially restricting access and foreign investment. While details are still being debated, the direction is clear: China is moving to lock down its AI advantages. This combination of state-backed investment, rapid open-source production, and potential export controls makes China the only other country besides the U.S. with well-resourced AI labs chasing the frontier. Understanding this matters because if China integrates near-frontier AI quicker into its economy and security apparatus, it could secure advantages that overcome any intelligence gap between the two nations.

To use China's synthetic humans (AI-powered digital replicas of real people), start by defining a specific task a human currently does, like customer support. Instead of replacing the human, you augment them—the AI handles the routine work, and the human reviews the output. For example, a company could take a human agent's scripted responses and embed them into an AI. The AI then answers common questions, while the human only steps in to approve or correct replies. This approach, called augmenting (enhancing, not eliminating human roles), is key.

Next, choose a platform. Chinese labs like UBtech are building humanoid robots with biomimetic skin and emotional interaction models (AI that reads and mimics human feelings). For a digital-only task, you might use a system from Alibaba, which provides the "brain" via large language models. Gather examples of the task—paste in past blog posts, emails, or chat logs. Tell the AI, "Here is how we want this process done," so it generalizes from those materials.

Test the setup. A user might draw a line on an image to guide a drone flythrough, simulating a scene in Beijing where real flight is illegal. The AI completes the path. For repetitive work, like drafting marketing copy, paste a newsletter and a LinkedIn post as reference, then ask the AI to write new versions. Finally, review the output. The goal is to reproduce enough human qualities—speech, memory, facial expressions—that the machine becomes an acceptable substitute in those specific moments.

When building AI tools in China, a common pitfall is relying too heavily on synthetic data (artificially generated training information) without enough real human input. Chinese labs often use synthetic data to scale quickly, but this can produce models that lack nuance or make strange errors. The best practice is to blend synthetic data with carefully curated human feedback early in training.

Another major mistake is assuming AI can fully replace human workers. While China's state-backed firms push automation, successful projects keep humans in the loop for quality control, especially in high-stakes tasks. For example, Chinese developers have found that distillation attacks (stealing knowledge from a larger model) can speed up development, but the copied model may inherit hidden flaws if humans don’t verify outputs.

A key best practice is to invest in human oversight rather than trying to remove people entirely. When China’s labs rushed to mimic advanced U.S. models using distillation, they sometimes missed safety gaps because they undervalued individual researchers' opinions. In China's collectivist culture, team members may follow orders without challenging risky shortcuts. To avoid this, build a culture where humans are empowered to question results, especially when using synthetic data or accelerating development. Remember, even the fastest AI benefits from a human check before final deployment.