More than the Alibaba detail, I think it gives us a focus on how China has gotten its act together in AI after two years of lagging behind the United States. What models do they have, what State plan do they have (because they surely have it), how are they solving the drama of data centers, what results are they having?
Models that “reason”. Among the latest developments in the AI segment in China are AI models that “reason” as does OpenAI’s o1 model. In the Asian country we already saw a first competitor with DeepSeek, and these days we have met the second. As indicated on TechCrunch, a promising new model called QwQ-32B-Preview has now been launched by Alibaba to compete with o1. According to those responsibleperformance is notable “in mathematics and programming” but they recognize that there is still room for improvement in language understanding.
Socialist AIs. Those responsible for the Chinese government have made things clear to their technology companies. Those who develop AI models must guarantee that their systems “embody fundamental socialist values.” The Cyberspace Administration of China (CAC), a government supervisory body, expressed this to ByteDance, Alibaba, Moonshot and o1.AI, the large Chinese companies that work on this type of technology.
Machines that think how the government wants. These companies are subject to a process of review and testing of their models by government officials. The evaluation of the various developments includes a series of questions related to Chinese political sensitivities and its president, Xi Jinping. The country, which already controlled open access to the internet with the “Great Firewall of China” – no websites and platforms considered toxic – now intends to do the same with Chinese generative AI models, which must create appropriate content. If you ask them, for example, what happened in Tiananmen Square in 1989, they will excuse themselves by saying that they cannot answer a sensitive topic.
Source: Alibaba
Centralization versus decentralization in the US. The Chinese strategy is characterized by its ambition to control all AI development in the country. It is the government that directs funds and resources to specific AI projects that align with national objectives. In the US, the development of AI is completely decentralized and falls on private companies, although of course there are developments by government agencies that also collaborate with the country’s large technology companies, for example in the military field.
They don’t stop investigating. And ITIF study reveals how in recent months China has become a global leader in publishing research articles on AI, although the impact of these publications is less than that of those published by US institutions.
And their models are no longer (so) behind. LLMs developed in China are also catching up of the models of large Western technology companies quickly. Lo indicate various benchmarks that compare the performance of these models in standard tests, although there is no particularly reliable way to know if one model is better than another. Qwen-72B, Alibaba’s LLM, is among the best according to Clement Delanguethe creator of Hugging Face, and open models from China they are advancing in a remarkable way.
And that despite the difficulties. This turn of events is surprising, especially considering the restrictions to which China is subject due to the trade war with the US. Access to key resources like NVIDIA chips is limited, but Chinese developers and researchers seem to be managing to overcome all those obstacles.
Image | idnaklss with Midjourney
In Xataka | China already has a 6nm, 45 TOPS ARM SoC ready. It’s what you need to build your own AI PCs