Breaking news
Pakistan Khabar

China Uses US AI Models to Train Its Own Systems, Unusual Method Revealed

China Uses US AI Models to Train Its Own Systems, Unusual Method Revealed

Shenzhen (Web Desk): China has begun training its domestic artificial intelligence systems using outputs generated by leading US AI models in an effort to reduce the impact of American restrictions on exports of advanced AI chips.

According to a Reuters report, Chinese organisations are obtaining answers, computer code and solutions to complex problems from advanced US AI models and using them to train their own domestic systems.

The report reviewed more than 300 Chinese research papers, intellectual property documents, company announcements, job advertisements and government records. Its findings indicate that this technique, known as distillation, is no longer limited to research experiments and is increasingly being adopted at an industrial scale.

Under this approach, powerful US artificial intelligence models are used as teachers, providing answers to various questions and solutions to complex problems. The information and examples generated are then used to train comparatively smaller domestic models. Chinese companies are also developing software to automate the process, while specially trained workers are engaged in preparing training material.

According to the report, the Chinese government is also encouraging this approach to reduce the cost of developing and deploying artificial intelligence and make the technology more effective across the country's vast industrial sector.

How AI Distillation Works

In artificial intelligence, distillation involves transferring selected capabilities of a large, powerful model to a relatively smaller one. The process does not reproduce all the features of the original system, nor does it completely eliminate the need for advanced computing resources. However, it can help develop models capable of performing specific tasks at a lower cost.

The technique could also help China reduce its dependence on US technology, particularly at a time when Washington has restricted Chinese access to advanced artificial intelligence chips.

The issue is emerging as a new point of contention amid the existing economic and technological tensions between the United States and China. The development comes ahead of the Asia-Pacific Economic Cooperation summit scheduled to be held in Shenzhen on November 18 and 19, where US President Donald Trump and Chinese President Xi Jinping are expected to meet. Meanwhile, Washington has announced further restrictions on China's access to advanced chips manufactured by companies including Nvidia.

In September, US officials accused Chinese companies of systematically carrying out large-scale distillation to obtain the capabilities of American artificial intelligence models. Earlier, during meetings held in Chengdu in July, US Commerce Department officials had also raised the issue with Chinese organisations, warning that such activities could, in certain cases, amount to intellectual property theft.

Chinese officials, meanwhile, have defended distillation as a standard practice in the development of artificial intelligence. However, they have said they are prepared to take action wherever the process amounts to a violation of intellectual property rights.

Chinese Systems Trained Using Responses from US Models

According to a former Alibaba Group engineer, the technique is particularly beneficial for Chinese organisations facing shortages of computing resources, as an expensive response obtained once can be reused repeatedly. A complex question can be submitted to a US model, its answer saved, and the resulting example used to train a domestic model.

In a research paper published in November 2024, researchers at Shanghai Jiao Tong University said they had selected approximately 5,000 responses generated by OpenAI's o1 artificial intelligence model. They subsequently used GPT-4o mini to translate the responses into another language, producing 10,750 bilingual training examples. These examples were then used for further training of Alibaba's Qwen 2.5-72B model.