Every time a new AI model is released, its performance figures are often highlighted. However, the number of parameters or the size of the model is rarely disclosed.
For example, imagine you are trying to decide which model to integrate into your service. Model A is high-performance but slow to respond and expensive. Model B is not as capable but is cheaper and faster. Is Model A superior to Model B because of its intelligent design, or simply because it's massive? It would be easier to choose if this were clear, but the crucial size information is kept secret.
However, a method has been developed to estimate the approximate size of a model simply by posing questions to it. Surprisingly, the key to this method lies in classic literary works that everyone knows. This article explains the mechanism and the differences in strategies among various companies that have emerged from it.
