Who's using AI to help them code? And how?
I came across this article about how to use Github's Copilot, and it got me thinking...
Who's using AI to help them code? Share your AI coding stack and how you use it so we can all learn from your experience! 🤓
Trending on Indie Hackers
A Combination of Github Copilot and GPT4 works amazing. I literally spend 10 times less time on tasks. also GPT4 is brilliant for writing bash or python scripts for migration or some other purposes
This is the right combo folks.
Could you please elaborate on how you combine them? I though ChatGpt can do what Copilot does.
Well, there are some tasks that you really need explanation, in this case I use GPT4, because obviously Copilot cannot output the reasoning behind writing a code. Copilot is just good in code repetitions. Like if you are using the same pattern of code in all over your code, then copilot does a good job in recreating the same pattern. But With GPT 4 you can get to write scripts and the reasoning behind them which most of the times its useful to customize it furthermore. Also I have used GPT 4 recently for some software architecture decisions. It also helps, but its not the smartest. so a room for improvement there for GPT4
thanks!
+1 on the bash scripts! It's amazing how much automation on my dev environment I've been able to get up and running in a couple hours.
Yes you're right
I'm a senior software engineer with a degree and about a decade of experience. I've founded 2 code based startups including one I launched today (Poket Dev on product hunt, an unlimited software development subscription service).
I've played around with LLM's extensively with coding. I find they are pretty good at small tasks, like if you ask it to write a simple function. They fail very quickly given larger scope or having to write more than 50 lines. I regularly see them make trivial mistakes that they are also unable to fix. I don't think we're quite there yet, but I do think we are months to years away from truly impressive results.
Having said that, I find them AMAZING at high level software architecture. All the time when I come up with an idea of something to build I will run the idea through chat gpt and ask for suggestions on tech stack, design, architecture, etc. Not all of its suggestions are relevant, but it will often give amazing ideas I hadn't thought about. These ideas usually save me quite a bit of time or enlighten me to tools/libraries I didn't previously know existed that solve some of the problems for me.
I've tried this, but it seems like ChatGPT's designs are very generic, even when given context. Can you share some example prompts that you've gotten decent results with?
I use GPT4 daily for development and system administration. I ask for code snippets for various things, sometimes I'm just curious to see the code for a specific task on Windows and Linux. Or how the code looks in Bash, Batch, PowerShell, Python.
The research is much faster than with Google. You'll also get a concise response if you use a well-defined prompt.
Although it has its flaws and the dialogue could be smoother, like a dialogue with a colleague you are working on a task with, I haven't regretted the GPT4 subscription.
I use Tabnine.
But also there is good Github Copilot
I am a Software Engineer with about 8 years of experience. I use ChatGPT to generate code for me... WHY? If I have a project spec in front of me and I know exactly what to code and I would not waste any time on thinking/solving etc. just type - the chatGPT can do it faster than that... This is the main reason I use it - if you are a programmer that learned how to talk to AI (yes, you have to learn to talk with it to get proper responses) then you can give it correct instructions on what needs to be done - then you just have to check it over, correct it, test it and off you go.
I just made a Go API service with Postgresdb in under 30h (after work sort of thing).
I am quite glad that AI is finally here, it really freed me from the D2D bul****t tasks.
Truth be told.
I already understand the logic of what I want to do, why spend hours writing code when I can just tell ChatGPT exactly what to do and get actual code for it?
Saves time and I can focus on other aspects of the project
Exactly, I value my free time above all. AI is here to help us out and get rid of these garbage tasks.
I've been using the Chat Copilot Beta for a few weeks now and it's increased my productivity immensely. It's the power of Copilot and its understanding of code, coupled with being able to iteratively instruct it through chat - all without content switching and just staying in VSCode.
Honestly can't imagine ever not having access to this!
Strangely I had the opposite experience -- you've convinced me to go back to it later today.
I have Chat Copilot Beta as well but in my few days of usage it was giving junk responses. For example, if I copied a chunk of code and asked for 'clean up' suggestions or to add a function, it would return code that stripped fundamental styling and functionality right out of my code.
I do love vanilla copilot though, and can't imagine life without it anymore.
Note on the above; I'm a pretty hacky coder so these results could be due to my reference code being junk in general.
I'm using AWS CodeWhisperer (yes, because it's free...) and after an initial "meh" phase, I really like it.
I rarely have it write full algorithms, and when I do I spend a lot of time looking at it to see if it's actually correct.
But, it's just a much smarter auto-complete. E.g.
I think though, that the "copilot" moniker is wrong. In classical pair-programming the current setting is more like ME being the copilot - at least when I'm using it
I have had the same good experience with AWS CodeWhisperer. Specifically around writing a comment on what I want to do and allowing the suggestions to fill in the blanks. This has easily increased the speed at which I write things from scratch by 10x. I manage multiple projects in multiple languages and it helps me remember subtle syntax differences between languages that I often forget.
I was spoiled by CodeWhisperer by the time I tried Github Copilot, which left me disappointed. Suggestions directly in my IDE help me maintain flow state and avoid distractions or context switching required to do prompt engineering with GPT4 or Copilot.
I use Copilot, but only a bit. It'll fill in a line for me here and there.
I also use ChatGPT to help explain APIs to me that I'm unfamiliar with. It'll occasionally hallucinate, but in most cases, it's still faster than reading through the docs.
For reading through docs I would suggest something like collectiv AI. Their service provides a QA system on top of any existing github repository or documentation. I have heard our developer team using it quite frequently.
I just submitted a post about my Shopify development + chat gpt
I've been using ChatGPT quite a bit for building boilerplate React code + Storybook components. I've also been using it to optimize SQL queries and summarize complex SQL queries back to me, to ensure I'm written what I expected to. I feel supercharged.
My challenge was to code the entire service with ChatGPT only, here is what I built https://architectai.app
I use Chat GPT-4 primarily for smaller projects, as I am still learning React. I request that it uses comments to explain how the entire codebase works, because it can be challenging to understand code that you didn't write yourself. However, on larger projects, I have observed that it becomes somewhat less reliable, and it becomes more of a fight to correct design and functional issues.
One thing I have noticed is that it is proficient at analyzing code. You can simply copy your entire code and paste it into Chat GPT. I find this particularly useful in React. If you copy multiple components, it will notify you of any errors that may exist between them.
Gotta have to confirm with our developers but from the back of my mind we definitely use Perplexity AI (600 gpt4 powered answers daily costing the same price), Github CoPilot, Github Chat (recently released) and collectiv AI (asking questions over a github repository). Hope that helps you out.
What is the pricing for Perplexity AI? Can't find it on page.
Its $20 but you will get $10 discount for your first month. Its been a game changer really especially for our marketing team as we need to experiment a lot.
I've been using CoPilot for more than a year now, and it's pretty much seamless, I even forgot I was using it. You can write a comment describing what you want the code to do, and it generates the code for you. There can be some flaws at times, so you need to know what the right code would look like.
Also GPT-4 is great, I can give it some code and ask it to clean it up or add comments or things like that, it's better at these things than CoPilot. It's ultimately something that saves time, so you can focus on the things that actually matter.
We are using GPT for generating json from our sources based on our templates.
I use GPT4 from Botmate, which made my life easier
Nekton.ai can write the code to automate small tasks and run them in the cloud.
Raycast is brilliant and far closer to how I think we'll be interacting with GPT in the future compared to using the browser (it's essentially Apple Spotlight on steroids and then some).
I'm pretty stingy with paying for tools let alone annual deals but GitHub Copilot and Raycast are the only 2 tools I happily placed a yearly subscription for the instant they became available.
Take today for example, I was looking for an algorithm that would compare the semantically similarity of a list of strings against itself. Open Raycast > Ask GPT (10 secs) and it gives me a neat function. However I pull the thread a bit more to see how optimised it could be reg. time complexity and gave me a fantastic function that could handle enormous-sized lists (which was part of the initial problem anyway).
I use ChatGPT to debug and help implement certain concepts that would normally take some time on Google to work my way around. There is also Copilot as well
https://github.com/AntonOsika/gpt-engineer
I have not yet tried it, but hey 30K stars on GitHub! I believe this is a very natural continuation of the GPT ideas, I think this is how we are going to develop apps in the near future - probably even this year
I have been trying Bard for a while now.
Just a heads up for everyone here. Don't bother- it sucks.
Stick with ChatGPT
Doesn't it ever.
Yeah, ChatGpt is a good tool. I do some frontend with it, because I am backend/infra person.
Copilot is widely used in the dev communities I am in, especially devops people love it. But i don't use it for now, 100$ a year seems like a stretch because I can use ChatGpt for it, and ChatGpt can do a whole lot more than just coding.
I haven't tried Copilot or Tabnine, but I coded my first micro-SaaS project with the help of ChatGPT. Whenever I had an idea, I explained it to ChatGPT and tried getting a more efficient code than mine.
More often than not it worked. But for complex problems, it's just frustrating and often goes into a loop of mistakes.
A lot of developers i know are using AI / ChatGPT as part of their daily programming tasks. Especially, ChatGPT has been very helpful in fixing the issues in code, whatever the language is.
I have personally used AI to build a few AI related products :)
I'm trying out tabnine. So far it's been a mixed bag. It's code suggestions can be helpful, but I regularly get suggestions that seem to ignore the code context and look like they come from elsewhere. For example, I'll have a structure with particular data fields. Instead of suggesting the field with the right datatype, it'll suggest something that looks like it came from an open-source framework. **
NOTE: I haven't got round to trying out their paid tier, which is more advanced.
** EDIT: Just encountered one of these instances. I have a structure with function pointers. Instead of suggesting code that calls the function pointer, it suggests I set the pointer to true (i.e., treat it as a boolean). That's obviously wrong.
A happy GitHut Copilot user here 🙋♂️. Its unobtrusiveness and help is more than worth the annual fee.
I stopped using GitHub Copilot and now I use Codeium.
it's free and good.
I found Copilot and Kodi to be meh. Maybe they would be more effective as a learning tool. I primarily use AI for ideation, verifying ideas, boilerplate like generating models (especially if they have a lot of relations), and debugging confusing errors.
You should use AI. I think that programmers who don't use AI should change their developing paradigms using AI. If you google, stackoverflow a lot when you program, start ask same thing at chatgpt. I use copilot, chatgpt a lot. They help me a lot when I code from scratch. I explain what I want in language, get results, if result is bad, re-ask and iterate. It saves a l-o-t of time!
I use copilot, but there is a learning curve. it seems to get in the way sometimes, but i'm gonna continue with it.
Not for programming, but I use GPT to write all my sales copy (then I edit it to be safe). Thinking about connecting GPT to Zapier to increase efficiency too.
Using AI for pair programming is awesome. I've never really used copilot, but AIXcoder is great.