微調生成式 AI：新品牌 vs. 成熟品牌

Abstract

生成式人工智慧（生成式 AI）迅速成為品牌溝通的重要工具，在大量訓練數據的支持下，為成熟的品牌提供顯著的優勢。然而，由於新的品牌缺乏既有的數據，他們會在使用生成式 AI 時面臨挑戰，這會影響輸出的品質。本文將探討生成式 AI 在成熟的品牌與新的品牌之間的生成品質差異，並著重於以微調的方法拉近這一段差距。

透過對於 FL Studio（一個成熟品牌）和 Suno AI（一個新品牌）的比較分析，我們研究了無範例提示、人物誌規律和跨部門合作模擬的策略。使用最新的 GPT-4o 模型，我們以對 AI 進行微調生成與品牌身份和目標客群一致的價值主張。

我們的研究發現無範例提示提供基準線，而微調顯著提升輸出品質，尤其是對於新品牌有明顯的改善。行銷長角色扮演與跨部門合作模擬的微調方式可以有效提高新品牌價值主張的相關性和清晰度。

本研究提供品牌溝通上模型微調的寶貴見解，突顯出新品牌可成功利用生成式 AI 的方法。未來工作可以探討更多的調整策略及其在多樣生成環境中的影響，包括多語言和多文化應用。

Generative AI (GenAI) has rapidly become integral in brand communications, providing significant advantages for mature brands due to the extensive training data available. However, new brands face challenges with GenAI due to the lack of pre-existing data, which affects the quality of generated outputs. This thesis investigates the disparities in GenAI output quality between mature (well-established) and new brands, focusing specifically on fine-tuning methodologies to bridge this gap.

Employing a comparative analysis of FL Studio (a mature brand) and Suno AI (a new brand), our research explores strategies such as standard zero-shot prompting, persona pattern, and cross-functional collaboration simulation. Using the latest GPT-4o model, we fine-tune the AI to generate value propositions that align with the brands' identities and target customers.

Our findings reveal that while standard zero-shot prompting offers a baseline, fine-tuning significantly enhances output quality, particularly for new brands. Role-playing as a Chief Marketing Officer and simulating cross-functional collaboration proved effective in improving the relevance and clarity of the value propositions for the new brand.

This research contributes valuable insights into effective AI fine-tuning for brand communications, highlighting actionable methods for new brands to leverage Generative AI successfully. Future work may explore additional tuning strategies and their impact in diverse generative contexts, considering multilingual and multicultural applications.