The article explores the capabilities and limitations of Large Language Models (LLMs) like Llama, Qwen, and Phi, focusing on their token prediction abilities. It argues that while LLMs are not capable of traditional reasoning or thought processes, they display a form of intelligence in their token prediction tasks. The complexity of LLMs arises from their design and architecture, limiting their ability to solve certain types of problems. Experimentation with mathematical expressions highlights the challenges LLMs face in processing and understanding sequential information. The study reveals interesting insights into the performance variations of different LLM models when tasked with specific problems.
https://www.arnaldur.be/writing/about/large-language-model-reasoning