Two Chinese companies have been generating significant curiosity in the AI world in recent weeks: MoonShot with its Kimi models and Z AI with its GLM models. What is less known in the West, however, are the reasons behind their extraordinary success.
-
What is exceptional about these companies' latest models is that they are breathing down the necks of the most advanced Western models at a fraction of the cost, despite the ban on exporting advanced Nvidia processors to China.
The West accuses these Chinese companies of relying on Western models as a foundation while stealing intellectual property. These accusations ignore several significant technological achievements by these companies, as well as the fact that all Western AI giants are currently facing massive copyright infringement lawsuits themselves.
-
The companies were founded separately by a professor at Tsinghua University in Beijing and his student.
This university is considered one of the best in China, having produced a series of prominent leaders, some in the field of AI. The university's first AI lab was established as early as 1978, and it serves as an incubator that brings together scientific knowledge, research labs, investors, and entrepreneurs under one roof.
Professor Tang Jie founded Zhipu AI—better known by its abbreviation, Z AI—the company behind the GLM models.
Unlike OpenAI's GPT models, which during training read text from the beginning and predict the next word, the early GLM models read text from both directions and learned to fill in the gaps in the middle, making them more efficient and intelligent.
His student, Yang Zhilin, is considered a technological prodigy. He was sent to the United States for his doctoral studies, where he got a close look at the working methods of American tech giants like Google and Meta. Upon returning to China, he founded Moonshot, which developed the Kimi models.
Moonshot's Chinese name translates to 'The Dark Side of the Moon,' a tribute to the famous Pink Floyd album, a band particularly beloved by Yang, who is a drummer and rock enthusiast.
What sets the Kimi models apart is their combination of massive models—with the latest approaching 3 trillion parameters—while maintaining a context window of a million tokens, highly efficient RAM utilization through advanced mathematical methods, and high speed.
Considering that all these achievements originally relied on outdated Western hardware or Chinese hardware, this feat is even more impressive. Regarding their latest model, the company may have managed to get its hands on Nvidia processors for training, a matter that is still under investigation.
-
The Western attitude toward Chinese technology is often suspicious and dismissive, but the foundations upon which these two companies are built are ones that any Western company would be proud to claim.
Academic knowledge accumulated over decades, an incubator that encourages entrepreneurship and facilitates investment, advanced mathematics, and technological excellence—these are the secrets that have made these companies what they are, more than anything else.
Pictured: The library complex at Tsinghua University.
--
👋 Hi, I'm Shlomo Strauss, and my posts are not written by artificial intelligence.
Feel free to follow me for more interesting content on science and technology, or search for 'Shlomo Strauss's Blog' on Google.