The 2026 World Artificial Intelligence Conference and the High-Level Meeting on Global AI Governance took place in Shanghai from July 17th to 20th. During these days, two key terms were frequently mentioned: tokens and computing power. What do these terms mean?
A "token" can be considered the smallest unit for AI to understand human language. It's akin to teaching an infant to speak; you wouldn't start with an entire article but break it down into sentences, and further into individual words or characters. An interesting fact: approximately three English letters equal one token, while one Chinese character roughly equals one to two tokens.
Processing each token consumes computing power. Tokens are like bricks, and computing power is the effort to move them. The more tokens used, the greater the computing power required. Currently, China's daily token invocation volume reaches hundreds of trillions.
How are these trillions of daily token invocations consumed, and what is their relationship with computing power? Tokens represent the workload for AI, while computing power is the driving force.
Professor Liu Zhiyuan from Tsinghua University explains that tokens are the smallest units of information processed by artificial intelligence. Just as humans read, think, and converse using words, for AI, understanding, thinking, and generating answers consume vast numbers of tokens. One can understand tokens as the workload AI must complete, computing power as the driving force, and electricity as the essential energy supporting the operation of that computing power.
China is accelerating the construction of a nationwide integrated computing power network. Furthermore, China's computing power deployment is not limited to Earth. In June of this year, Beijing established its first Space Computing Power Industry Innovation Center. Many might ask: isn't building a computing center in space even more costly?
Professor Liu Zhiyuan explains that in the past, satellite data followed a "space data, ground computing" model, meaning satellite data had to be transmitted back to Earth for processing using ground-based computing power. As the number of satellites and the volume of space data increase, the bandwidth for transmitting data back to Earth is clearly insufficient. Deploying space-based computing power, utilizing solar energy to achieve "space data, space computing"—processing data where it is generated in space—addresses this. While the initial investment for "space data, space computing" is high, it may pave the way for future "ground data, space computing" capabilities.
At this World AI Conference, multiple new intelligent agents were showcased. Many feel that large language models are already quite handy, and professional intelligent agents answer questions that large models can also handle. So, what is the necessity of transitioning from large models to specialized intelligent agents?
Professor Liu Zhiyuan notes that over the past few years, the primary application form of large models has been conversational assistants. As these models become more specialized, they are beginning to enter various professional fields to assist humans directly in their work. A very important development direction is the advancement of programming intelligent agents. Software engineers globally now largely rely on AI to assist in software development work, which is an application of professional intelligent agents. Looking ahead, as AI masters knowledge in more specialized domains, it can play a role as a new quality productive force across more disciplines within human society. This is certainly a crucial direction for future development.
Comments