China Develops AI Model with 1 Trillion Parameters Using Domestic Chips
AI Taking over the world
MOSCOW (Sputnik) – Chinese telecommunications giant China Telecom has developed two large AI language models using chips of its own production, one of them with 1 trillion parameters, Hong Kong-based newspaper South China Morning Post reported on Monday. The open-source TeleChat2-115B and a second unnamed model have been tested on tens of thousands of chips of its own production, which has become an important milestone against the backdrop of tightening US restrictions on China’s access to advanced semiconductors, including the latest artificial intelligence chips from Nvidia, the newspaper said. The development was carried out by the AI research institute of China Telecom. Telechart-115B contains over 100 billion parameters, while the unnamed model has 1 trillion parameters. The number of parameters involved in the training process determines the complexity and efficiency of AI models. 
WorldUS Orders NVIDIA, AMD to Restrict Sales of Advanced AI Chips to China1 September 2022, 10:20 GMTIn October 2022, the United States announced a series of rules restricting the export of equipment and components for the production of advanced chips to companies located in China. A year later, the US Department of Commerce issued a series of new restrictions on the export of such semiconductors. These rules changed the definition of AI chips and introduced additional licensing requirements for their shipments to more than 40 countries to avoid their resale in China. In China, there are more than 100 LLMs (large language models) with over 1 billion parameters. They can be used in electronic information transmission, healthcare, transportation, and in any other field. In two years, the size of large language models has increased thousands of times, and the cost of computing power continues to fall. LLM technology is advancing rapidly and can handle text, language, and vision, but to achieve AGI (artificial intelligence that can think and act like a human), next-generation models must have larger, more complex, and multi-level logical thinking.
