Nevertheless, she persisted

I dipped my toes into Github Copilot (henceforth, "the llm") these past couple days and it’s been interesting.

I’ve first used it at work because I’ve heard llm’s are good was the minutiae of syntax. I was working in a templating language. llmwas great at generating the correct syntax, and switching back and forth between the template and the code was a breeze. It was also aware of what functions were available in the templating language, which was nice, using them to suggest snippets of code based on variable names. It correctly guessed my intent several times. There were a couple places where it forget closing quotes, but that was easy to notice and fix.

Today I used it to write a github actions workflow, so automate publishing a project of mine. The project contains a few different packages (in go, ruby, and rust), and I wanted to publish them all at once. It successfully generated the workflow file (installing toolchains, wiring up secrets and environment variables). However, it hallucinated a publishing step (go publish) that didn’t exist, presumably copying the ones for ruby and rust.

Overall, very impressed with the llm. Spicy autocomplete feels like the future.

#Tweet #Dev