Completely agree. Got the same experience. I have the impression that it has gotten worse in recent months

@@omidjung115yep it sucks

I have never used it. I don't like to be too much influenced. Something that spit out code while you are writing is an influencer. I use Chat GPT style AI and review all the code before pasting it.

yes, if the codebase can fit into the model's context window

@@TheDanielLivingstoncontext window myass

@@tomjones8293please, tell us how you really feel...

Im not using copilot yet, but Im using ChatGPT 4.0 while working on large corporate codebase developed by around 20 people. And my experience is that it is useful mostly on those tasks: 1) Question answering - our project has documentation online and public, so ChatGPT knows it and whenever I want to know how the part of the system works from user perspective, I can easily ask. For documentations that ChatGPT does not know, I can use RAG (chatgpt 4.0 with file attachement). 2) Refactoring - it can refactor parts of your code based on your instructions 3) Suggestions - good only for using some simple functions you dont remember how they are named and called. 4) Refactoring according to new frameworks - I was using it to change our components from react class to react functional and 4.0 did 70% of work on its own! 5) Code generation of repetetive code from templates - imagine you have some basic info like data from the server, you have several examples how the form can look like and LLM can generate rest of the code. Of course this can be achieved by metaprogramming like reflection and dynamic forms. But many times, forms have additional nuances that are more easily handled when its written staticaly by having real components here rather then dynamic ones from metadata. LLM can help you generate. 6) Code generation of tests - like previous step, this can be used as tool for generating e2e tests, because they are usualy sequential code. Basicaly any task that has the form of language translation can be done faster with LLMs - which testing is - take the form and translate it to tests - semantics is very similar, syntaxe is different. And the generation can be driven by instructions and examples. Copilot is basicaly ChatGPT 4.0 (althrough Codex is different model, they both perform well in programming), but Copilot has better integration with IDE. So the above points are valid also for copilot.

yup too surface-level

Explain.

@@DannyCookTheProgrammer Easy, these LLMs are usually trained on a couple years old data and not updated.

Its not going to pass. This "hype" is revolutionary. What will pass is coders of the past who do it all by hand. For example how the CNC took over hand machining. Every machinist said the same thing you just said. Now, good luck finding a machinist the same way you would when CNCs first came out. It may not be up to speed right now, it may make mistakes. But that will get better, and the job of a coder will turn into a job of prompting. Prompting and being able to CLEARLY explain what your looking to achieve is gonna be of high regards. Good luck 🤞🏾

Are there any cobol repos on GH?

plenty and sourceforge has GnuCOBOL, free, open source, and runs on Linux/Windows/Mac @@CripplingDuality