The concept of big model token is actually quite complex After returning from vacation, start supplementing with AI. Test the latest models and popularize the concept of token that everyone is talking about. Token is actually a quite complex concept, typically appearing shallow but actually deep. Ordinary retail investors purchase large models, such as a monthly subscription of $20 for a Gemini. You cannot experience the existence of tokens. Who is the buyer of the token? Is it Big B or Small B. Financial giants, consumer electronics companies (such as Xinyuan Auto Factory), and various AI shell applications, at worst, have to be independent developers. So, ordinary users cannot access it. What exactly is a token? Token originally meant "the smallest meaningful unit". 1 English word ≈ 1.3 tokens 1 Chinese character ≈ 1 to 1.5 tokens Interestingly, this text unit corresponds precisely to the cost structure of the large model. When the user inputs a token, the graphics card needs to load the token, perform matrix operations, and store it in expensive video memory. The number of tokens here corresponds exactly to the "instantaneous computing power" and "video memory usage fee" When the large model outputs tokens, autoregression outputs the results. The number of tokens here corresponds exactly to the "server exclusive time fee" and "bandwidth transfer fee". Therefore, the industry has developed a habit of using tokens for pricing. This also explains the counterintuitive aspect of why the big model API not only charges based on output, but also requires payment for input. But the trouble here is that the output token is a serial task and cannot be computed in parallel. Therefore, the output price is usually several times more expensive than the input. For example, the miniMax M2.5 has an input price of $0.2 per million tokens and an output price of $1 per million tokens. This way, each model has differentiated pricing and input-output differentiated pricing This can effectively counter attacks such as calculating pi. In summary, tokens are not only text units, but also cost structures and pricing methods. Upon deeper reflection, tokens provide a perspective, a more profound view of AI. Most people consider AI as a form of high-tech: model training+reasoning ability. But from the perspective of token production, AI is seen as a manufacturing industry. Manufacturing capabilities utilizing data centers, graphics cards, and power banks. This perspective means that the ultimate outcome of the game is the convergence of the abilities of the big model, and the competition is focused on the underlying manufacturing capabilities. From a manufacturing perspective, what kind of commodity is a token? It is a commodity with intense inflation. It is a product that iterates at an extremely high speed. It is a commodity that cannot be stored and quickly depreciated to zero upon production. Some people say it's like water, electricity, and coal, but silly brother thinks it's more like a telecommunications company without regional protection. Is this a good business? PS. During the holiday, there is a group dedicated to researching AI in the field of cramming AI. Welcome to join us, silly brother.