Software Engineering Cannot Be Solved by Code Generation
CS, i.e. SE, is to translate blurry, natural languages
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
CS, i.e. SE, is to translate blurry, natural languages into formal languages.During the process of SE, one of the two scenarios will apply, exhausively:You state your demand in a formal way. Given that it is empirically observed that LLMs seem to behave really well at solving well-defined, formal problems (such as optimization, reverse engineering / cybersecurity, or just general, a-priori math problems), most of the time it will work. But at this point you are essentially writing code yourself, and most SE can be seen as proof by construction.You don't state your demand in a formal way. In this case, LLM starts to guess and fill out the gaps in your natural language. Because of the nature, the formal construction can get pretty arbitrary. Sometimes they happen to be correct as guided by LLM company RL imposing
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。