3 ms·

William Gibson once said, "The Future is already here. It's just not evenly distributed." I tried out Github Co-pilot with Rust and with Elixir. Co-pilot did

by Dowwie 4y ago

William Gibson once said, "The Future is already here. It's just not evenly distributed." I tried out Github Co-pilot with Rust and with Elixir. Co-pilot did some things well and other things were horrible. It often recommended Ruby code snippets as Elixir and suggested implementation based on APIs that didn't even exist. When different libraries have APIs and modules similarly named, Co-pilot would share an API as if it were a universal truth. What an underwhelming experience. Skepticism of AI code generation is the result of trying it out and finding the AI tool failing to meet our expectations. It's not unfounded skepticism but that which is based on anecdotal experience. Herein lies the problem. Programmers are experiencing far better experiences with Co-pilot in other languages. As to what those languages are, I'd like to know, but if you're experiencing the future today, I'd like to know what your workflow involves.

kybernetikos 4y ago
I've had genuinely valuable interactions using both bash and javascript, and also some configuration languages for major systems (various forms of yaml, etc). It probably just has far less information on Elixir and Rust in its corpus. It's actually pretty good at understanding the commonly used tools to do a job and which command switches you should use. E.g. ask it to write a script to create slowed tempo versions of a piece of music at 70%, 80% and 90%, it'll probably come up with some bash that runs sox and does a solid job of the requirement without you needing to work out which tool to use and what its parameters are.
valine 4y ago
For me the most impactful use case is prototyping small apps that I sort of know how to build, but don't know the relevant libraries off the top of my head. My most recent example: I got an Onnx file from a guy with a small machine learning model, that for various reasons he needed to run on an android tablet. I had never used the Onnx inference library on android and only had a vague idea of how to pull it in. Without ChatGPT that prototype probably would have taken me 5 or 6 hours to throw together. I'd need to look up the maven repository, research the API, figure out the syntax for creating tensors in Kotlin, write methods for loading android resource files, ect. With ChatGPTs help it took maybe half an hour. Nothing about the code it produced was complicated, the time save came from not having trial and error my way through learning the library.

by Dowwie 4y ago

kybernetikos 4y ago

I've had genuinely valuable interactions using both bash and javascript, and also some configuration languages for major systems (various forms of yaml, etc). It probably just has far less information on Elixir and Rust in its corpus. It's actually pretty good at understanding the commonly used tools to do a job and which command switches you should use. E.g. ask it to write a script to create slowed tempo versions of a piece of music at 70%, 80% and 90%, it'll probably come up with some bash that runs sox and does a solid job of the requirement without you needing to work out which tool to use and what its parameters are.

valine 4y ago

For me the most impactful use case is prototyping small apps that I sort of know how to build, but don't know the relevant libraries off the top of my head. My most recent example: I got an Onnx file from a guy with a small machine learning model, that for various reasons he needed to run on an android tablet. I had never used the Onnx inference library on android and only had a vague idea of how to pull it in. Without ChatGPT that prototype probably would have taken me 5 or 6 hours to throw together. I'd need to look up the maven repository, research the API, figure out the syntax for creating tensors in Kotlin, write methods for loading android resource files, ect. With ChatGPTs help it took maybe half an hour. Nothing about the code it produced was complicated, the time save came from not having trial and error my way through learning the library.