Copilot is great, but a hundred bucks for what is basically a smart autocomplete seems a bit much - mostly, I hate the fact that the code is constantly transmitted to github (my repos are mostly local) - are there any reasonably convenient options for doing this without github looking over my shoulder all the time? I’m using VSCode but not wedded to it.
The alternative is… YOU!
Invest in yourself, get training/practice, and eventually you will become strong enough you realize you don’t need any autocomplete! :)
Thanks! Respectfully, I think I’m OK on that side of the equation. But you’re right - you should invest in your own learning and self-directed growth - this applies to all facets of life, not just writing python modules.
I like using copilot. Now that we aren’t using punch cards to write monolithic BASIC and we have an internet to work with, most of the brain work in programming is component-based integration. AI makes typing out code a LOT faster, so I won’t be ditching it to resume writing out for-loops end-to-end. I just don’t want every line of code available to github and definitely don’t want to fund the walled AI model if I can find a way around it.
Copilot isn’t really about needing autocomplete, it’s about not wasting time writing predictable long lines of code. One example I use is MPI calls, they take tons of parameters and doesn’t really require much brainwork, instead of writing it all, you let copilot do it and just quickly read it to check.
IMO consistent repeated “writing predictable long lines” means the code smells of abstractions that can be improved, i.e. if autocomplete is really saving someone that much time, there are likely even worse problems.
https://github.com/fauxpilot/fauxpilot is an OSS alternative, don’t know how viable it is. https://codeium.com is commercial, but has a free tier that seems to offer most of what CoPilot offers.
Jetbrains Fleet is about to roll out their own AI assistant and it looks very promising.
Fairly skeptical about Jetbrains’ Model, so it’s basically a wait and see.
from the FAQ:
How much will AI Assistant cost in JetBrains IDEs? The AI assistant is currently free to use during the preview phase. We’ll be providing the licensing and pricing model at a later date.
It looks as though it would be similar to github - they host the model and you pay a subscription for real time access while you’re coding.
Still, I’d feel ideologically more at peace paying Jetbrains for IDE-as-a-service than I do slinging github a hundred smackers just to use their model from somebody else’s IDE.
A public / open source solution would be ideal.
I know, just not sure the quality it offers and I am 99% sure that they train it on the piles which is a huge nearly 1 petabyte dataset that includes GitHub code. They are likely doing the same thing GitHub did when training copilot.
deleted by creator
Me too I use Rider and I really like how they put the emphasis on the “integrated” part of “IDE”.
Having that said, Fleet looks pretty good and I’m excitedly waiting for it a stable release.
I use VS Codium, which seems to have a lot of home-phoning disabled although I really don’t trust Microsoft code still somehow doing its thing now and then
Tabnine is pretty good and has a free tier.
thanks, yes - looks like a free entry point for autocomplete - the ‘pro’ tier is probably similar to copilot, by the looks of things - about the same price too.
This project seemed promising from the initial description:
https://github.com/CodedotAl/gpt-code-clippy/
but looking at the issues and history, it doesn’t seem to be all that promising practically, so presumably avoiding corporate AI is not easy (yet).
I tried a few and Codeium is the best
For a research project, I’m currently looking into open-source GitHub Copilot alternatives and monitoring options. As someone else already said, you can use the FauxPilot self-hosted backend. It works pretty well, but you have to host it yourself. You can use this backend with a frontend extension, such as the official GitHub Copilot plugin (with telemetry disabled) or with the FauxPilot VSCode extension. However, I found that the FauxPilot extension is very lackluster in terms of features compared to the official extension (On top of my head, it does not have any caching mechanism, nor does it cancels the requests already made if you continue typing (they just pile up), and does not have any sort of filter for the completions. All of these lead to way too many generations being requested from the model.
TL;DR : If you want to use the GitHub Copilot extension, without GitHub looking over your shoulder, you can use the FauxPilot backend and opt-out of the telemetry from VSCode/Copilot (I don’t know which one needs to be disabled) (or do as I did for research purposes, and change the telemetry url in the source code to a localhost one).
Thanks for taking the time to respond - it sounds like an interesting research project. This indeed looks like the ideal long-term solution, where the model is hosted locally and no code is sent to the internet while it’s under development. I’ll take a closer look at this - thanks again.
An issue you’ll run into here is that all the really fine tuned advanced ones are going to run in the cloud somewhere, meaning you’ll be transmitting code to someone. You could attempt a local or selfhost LLM, but doing that would require you to have or rent a beefy GPU with a ton of memory to run a good enough model.
Why would you use copilot in the first place! Is it that hard to write code in a simple text editor like vim?
At work we use AWS Codewhisperer: https://aws.amazon.com/codewhisperer/
Of course that does not fix the fact that the code is not transmitted and a company cannot look over your shoulder :p
Thanks! Codewhisperer looks pretty good - especially the price - It looks similar to copilot functionally but it’s free for individual use.
for anyone interested, there’s a pretty decent demo here https://www.youtube.com/watch?v=CxRhACSZsJU
It also integrates with VSCode, so it looks like I can just register with AWS and replace copilot immediately. Still transmitting everything in the IDE to Amazon real time, but production systems are in EC2, so it feels slightly less intrusive than github.
You could try HF Code Autocomplete VSCode extension. That would allow you to use open-source LLMs like https://github.com/bigcode-project/starcoder.
thanks - that looks like full independence and would be good to try. I’ve been using codeium, which was suggested here and it seems as good, if not better than github copilot and it’s free for solo developers.
I appreciate you taking the time to make the suggestion - I’ve learned quite a lot about open source software options in a few weeks here on Lemmy and find it’s been an excellent resource for technical info and suggestions.
