Making LLMs Better at Creative Writing using Entropy

This article explores how to enhance the creative writing quality of large language models (LLMs) by addressing the common issue of outputs feeling generic or 'average.' The author, an AI engineer, proposes modifying the sampling process to incorporate information about future entropy—the diversity of future token choices—when selecting the next token. The article explains the role of logits and samplers in LLM text generation, compares greedy sampling with other methods, and aims to reduce the 'LLM wrote this' vibe. It targets developers and AI enthusiasts interested in fine-tuning LLM behavior for better creative output.