Can we get an AI to write better? (www.strangeloopcanon.com)

A new project, dubbed "Horace," is exploring how to enhance the writing quality of AI models by identifying quantifiable “signatures” of good prose. The initiative stems from the observation that while current large language models (LLMs) generate competent text, their output often lacks stylistic depth. The researcher hypothesizes that effective writing is characterized by deliberate deviations from the expected, akin to musical rhythms, allowing for moments of surprise and engagement. By analyzing classic literature, the project seeks to measure aspects such as token-level distributions and cadence statistics, ultimately aiming to teach models how to mimic literary styles more effectively. This endeavor holds significant implications for the AI/ML community as it challenges the prevailing notion that more data and larger models are the only paths to improvement. By dissecting the components of writing, such as "surprisal" and "cohesion," the research offers a framework for generating text that resonates more closely with distinct authorial voices. The results indicate that even smaller models can learn to identify different authors' styles, suggesting a future where writers might guide AI to create text with a targeted stylistic signature. This approach could lead to advancements in narrative construction, allowing for a more sophisticated application of AI in creative domains.