A Preliminary Study on the Capability Boundary of LLM and a New Implementation Approach for AGI

摘要: With a preliminary exploration of the capability boundaries of LLM(Language Large Model),we believe that the current mainstream artificial intelligence generally adopts the technical of "attention mechanism + deep learning" + "reinforcement learning", which cannot be applied to those fields that are difficult to a lot of "trial and error". So, to achieve AGI (Artificial General Intelligence) that works in any field, it’s better to change the way we do it. Therefore, we propose a set of machine learning solution different from "deep learning + reinforcement learning". It adopts small samples and cumulative learning, and also realizes the attention mechanism similar to transformer, and also creates a fully connected knowledge network. In addition, it can realize interactive decision making with the environment without using lots of "trial and error" style learning. In addition, humans can preset different innate needs to it to achieve multi-objective balance, thus achieving far higher security than the current artificial intelligence. In this paper, we propose a set of new machine learning techniques which maybe guide humans realizes AGI