Subscribe
Sign in
My Experience With an AI-Powered Coding Editor 🤖
Pros:
- AI-generated code can free coders from routine tasks, such as writing boilerplate code.
- It's great for formatting, refactoring, and suggesting improvements.
- It's excellent at writing comments based on code.
- It's efficient at converting from one language to another, e.g. objc -> swift or Swift -> SwiftUI.
- It's helpful when writing code that you can't remember in detail but know how to do at a high level. Saves you a trip to StackOverflow 😅
Cons:
- It's not great at architecture or structure yet. You still have to do something yourself 😏
- It can only reason about a limited amount of code at a time due to memory limitations.
- Sometimes, you have to be specific about what you want to get the desired results.
- You have to store the best prompts somewhere. I store mine here: https://gist.github.com/eonist/2....
Conclusion:
- Using an AI code editor is worth it. It's a powerful tool to have in your arsenal.
- It will only get better from here, understanding more, making better suggestions, and becoming smarter at understanding what you want.
I'm curious about how you use AI code editors. Which AI code editor are you using or still on the fence? Share your thoughts down bellow 👇
12
Replies
Best
Nice dude thanks for sharing. Im paying for ChatGPT so im using it as my coding sidekick. I havent tried using an AI coding editor yet. Any idea how they compare?
@iskandarchacra Personally I use VSCode with GitHub Copilot chat. It's like having ChatGPT in your code editor (it uses a custom model called codex). If you use cursor.so then it uses GPT4 and you can use your own key. I think the workflow is a bit faster than using ChatGPT and the model will give you answers based on the context of the project you're working on.
@sentry_co wow didn't know i could connect cursor.so to my gpt 4 key. ill try this out, thanks for the tip!
@iskandarchacra Yepp. If you have the "on-demand-usage GPT4 key". Then you can use cursor.so for free. Just go into settings and add the key. Happy cursing 😏
@iskandarchacra @sentry_co That's what I was talking about! One of the best products on the market right now.
My second-hand experience has been very similar. It's awesome for boilerplate code but without seeing the big picture it fails at bigger tasks. I think the real potential of AI copilots will be unlocked once they can get more context from the code.
I wonder if there are open-source models ready for self-managed instances where you feed your complete codebase into the thing and can assist you with a full grasp of what the software is about. (My personal opinion is that AI development is driven in the exact opposite direction of this by incumbents but we'll see if the community can or will do things about it....)
The other thing is: who should review AI-written code? :D Because copilot can do great things when an experienced developer uses it, but a junior might not be able to tell whether the generated code is useful or not.
@geri_mate 1. The limitation right now is hardware. Only so much memory and compute available. Once that progresses, the model can read more of the codebase. Incumbents vs OSS: I think AI has little moat. So we will definitely have OSS models performing as well as OpenAI etc 3. Github is working on AI pull request reviews. More about that here: https://githubnext.com/
@sentry_co I'm just wondering if there are ways to self-host a model with your own specs (basically you can throw as much hardware as you'd like to) and use it as your copilot? Is there a solution for that now?
I think you're right about AI having little moat, that's why incumbents are trying so hard to lobby for sci-fi-esque regulation to prevent Terminators from exterminating our species.
And thanks for the link, I'll give it a better look to see where we're heading in the copilot area
@geri_mate for sure. Local llama2 + RTX 4090 and some elbow grease and you can have your local LLM copilot. Im sure someone will do it. When it happens I'll use that instead. For now we have to use CloudAI (TM)
Thanks for share. at least for now, the biggest advantage of that editors is to perform repetitive and already defined sections. "Architecture" is belong to "human" :)
Sol-A.