Summary: It has actually been a few months since my site came into existence. But being a newcomer to Lemmy I thought I’d post my experience once again here.
I had close to zero experience in web site development. I had never written a line of code in PHP nor used a PostgreSQL database, let alone creating and managing one.
However, I thought this lack of experience made me a good candidate to test just how powerful ChatGPT is. After two weeks of on-and-off construction, I finally completed a completely functional website that serves as an “online guest book” and is open for everyone to try out. A feat that I probably could never have achieved without any help.
Here are some of the amazing highlights of how ChatGPT helped:
-
Debugging - I took the approach of using a website design software and incorporating snippets provided by ChatGPT. Very often, that would lead to unknown errors, and I just found myself copying and pasting the entire file and giving one single word of instruction to ChatGPT - debug. Time and again, it managed to pinpoint the errors after a few back and forths.
-
Geolocation and other features - I just told ChatGPT what I wanted to do, and it pointed me in the right direction very quickly. In the case of geolocation, it led me to the right library to use that I had no idea about (geoip geolite2), walked me through the procedure to install it on my NAS, and got it up and running within something like one hour. I am absolutely certain it would have taken me days if not weeks to get it going given my programming background or lack thereof.
-
Backend admin site (that only I get to use so no fancy formatting required) - I did not even have to write a single line of code for it. I just told ChatGPT what I wanted the backend admin site to do, and it churned out 4 files for me just like that (with the usual problem of stopping midway through then having to encourage it to continue). I told ChatGPT what errors I encountered with the files, and it kept revising the code until it started running smoothly after a few tries. Two hours later, the backend admin site was done.
Anyway, give this site a try and see what you think: https://www.stringtone.com. The concept is simple, and all of the intelligence and many of the security measures came directly from ChatGPT.
It has been a fun project, but yes, I still have no clue how I can construct something similar without getting ChatGPT’s help.
It’s pretty good at making functions and basic program layouts like make a go program to do x with a docker compose file to run it but it hardly ever runs correctly the first time. You’re right it can get confused or just totally hallucinate or refactor longer code.
Disclosure I did have it write me a html/css professional webpage and all the text and a prompt for an ai image it needed for the layout in a single paragraph prompt. It all turned out pretty good sounding and looking. I was surprised I replaced my old one with it.
HTML is markup though and not a coding language. They’re much simpler to follow imo
Yeah I know but it’s pretty much the same for chatgpt html/c/ASM/python all the same to a language model
Mmmm
I have to disagree. Namely because it’s infeasible to actually know what’s going on in the inner layer of the neural network so saying that two different items are processed the same way using the same routes.
If this were true, it’s scores would be the same across the board - something are inherently “harder” (takes longer to calculate/traverse the neural net for, or are actually impossible)
¯\_(ツ)_/¯ it’s a language model it’s all the same for it. Granted it depends on it’s training set for how accurate it is. And the context window is pretty small with chatgpt, codex does a better job with longer code samples since you can increase the context window size and the training set is more code based.
Codex is not available anymore though that’s my qualm.
I don’t think copilot is codex either the more I think about it I’m pretty sure that’s GitHub’s or Gotland own thing
It might even be ms/azures own model
I wasn’t really disagreeing outside that to the model everything is just language. I just want to keep saying not magic just a large language model with a huge manicured training set.
Have you looked into running llama.cpp at home with a model trained on code? Some of them work pretty well and you can play with some stuff chatgpt abstracts away. You can even try to train/fine-tune model on your code if you have enough and a relatively new GPU.
It’s slow but all you need is lots of ram and a CPU no GPU nessassary.
But is it actually good code or is it more basic or small chunk only code?
I feel like sometimes it just needs to refeed its own code but idk. Gpt4 feels jank as fuck lately anyway.
I just wanna make cool stuff 😭
I’ve never trained one on code or used a code model. I’ve just come across them in research. Like this https://huggingface.co/Salesforce/codegen2-16B
release gpt was a better coder than gpt-turbo and 10 questions every 60 min was never enough for long code with gpt4 since it also forgets.
Edit I would ask here !localllama@sh.itjust.works