Coding on steroids: StackOverflow > ChatGPT > Cursor > ...

Coding on steroids: StackOverflow > ChatGPT > Cursor > ...

There is a lot of excitement about using LLM for coding, including from prominent LLM experts and great coders like

.

In this short post, I want to compare 3 ways I have used to help me when coding: StackOverflow, ChatGPT and and IDE (I have used Cursor, a new IDE based of VSCode with built-in LLM capabilities).

StackOverflow vs ChatGPT

With StackOveflow, you are joining somebody else's conversation about their problem. If you are lucky, their problem is similar enough to yours and you can lean from the proposed solution. But it is very rare to be able to copy-paste the code and have it work for you.

With ChatGPT, this is your conversation about your problem. In most cases, the piece of code spit out by ChatGPT will work for you.

ChatGPT vs IDE

With ChatGPT, you are having an out-of-codebase experience. ChatGPT only knows about whatever you put in its context. And adding a subset of your codebase –?say a few files –?is not totally straightforward. Whatever code ChatGPT spits out must be stitched back to your codebase, unless you are starting from scratch and ChatGPT creates the first version of your file(s)

With an IDE, you have an in-codebase experience. The IDE often infers the files to bring to the context or makes it super easy for you to reference them (Cursor lets you bring files by using the "@" sign, the same way you tag a person on social media). A nice feature of Cursor is that it creates a RAG-index of your entire codebase and can therefore brings into the context the relevant code snippets from your code base.

Whatever code the IDE spits out is nicely integrated in your codebase: this is like reviewing a pull-request from the LLM, with a nice diff UI, the ability to accept individual changes and undo/redo them as you see fit.

To summarize

StackOverflow-like = somebody else's conversation about their problem

ChatGPT-like = your conversation about your problem, out-of-codebase experience with limited context and generated code to be stitched back to your codebase

Cursor-like = your conversation about your problem, in-codebase experience with seamless context and output code as easy and rich as a pull request


要查看或添加评论,请登录

社区洞察

其他会员也浏览了