📄 中文摘要
在运行 AI 产品的 18 个月后，发现 GPT-4 和 Claude Sonnet 并不总是适合所有任务。作者将两款产品从前沿模型切换到小型语言模型，结果显示延迟更低、成本更低，并且在某些特定任务上准确率更高。具体的实施步骤和原因被详细阐述。第一个产品 AgriIntel 利用 AI 对传感器数据进行分类，并将其路由到相应的推荐工作流。该分类任务涉及根据土壤湿度、温度、营养水平和天气预报等传感器读数，判断所需的农业决策类型。
为何我将两款产品从 GPT-4 切换到小型语言模型
📄 English Summary
Why I Switched From GPT-4 to Small Language Models for Two of My Products
After 18 months of operating AI products, it was found that GPT-4 and Claude Sonnet are not always suitable for every task. The author transitioned two products from frontier models to small language models, resulting in better latency, lower costs, and in one case, higher accuracy for a specific task. The article details the exact steps taken and the reasons behind this decision. The first product, AgriIntel, utilizes AI to classify incoming sensor data events and route them to the appropriate recommendation workflow. The classification task involves determining the type of agronomic decision needed based on sensor readings such as soil moisture, temperature, nutrient levels, and weather forecasts.