August 29th.Nguyen Yi FengIn the most recent issue of the Science and Technology Hobbies Weekly, the process of answering questions by large models was broken down into three mechanisms: parameters, reasoning and networking. In his view, training began with the use of parameters to establish a mathematical relationship between the words, on the basis of which the model answers produced more probabilistic elements; a conclusion that the parameters did not contain directly could be obtained through logical reasoning。
He summarized the parameter mechanism as "compressed-generated": the training stage encoded knowledge relationships into weights, and the generation phase relied on these weights to find the most possible next word. The more parameters, the more they are usually described, the more training and reasoning costs rise, so that the model cannot store all knowledge directly in the parameters。
The reasoning mechanism assumes the part of drawing new conclusions from existing facts. When, for example, birth and mortality rates are known in one city, the net growth rate can be calculated by logic and need not be an independent memory. The Networking Mechanism is responsible for the completion of real-time information that neither parameters nor reasoning can give, such as the stock-take index of the day, obtained by Agent or the application frame for external search。
This article is "The One."AI THE THIRD EDITION OF THE SMALL KNOWLEDGE SERIES, THE FIRST TWO OF WHICH DISCUSS THE MEMORY NEEDS OF LARGE MODELS AND AI CACHES, IS WRITTEN FOR ORDINARY READERS WHO WISH TO UNDERSTAND HOW LARGE MODELS WORK WITH LOWER TECHNOLOGY THRESHOLDS。