Commented on GPT-Migrate · Jul 21, 2023
Hey! Josh here, maker of GPT-Migrate, now one of the top LLM projects on GitHub. Thank you for taking the time to check out the project! Code migration is a painful and arduous task. A couple weeks ago, I had to do a migration for another project I’ve been hacking on. It took a few long days, but as I migrated, tested, and debugged function after function, I started to realize that this process could be largely, if not fully, handled with today’s most powerful LLMs. Last week at a hackathon, GPT-Migrate was born. It works by creating a standardized environment (Docker), translating the code of the source project to the target language/framework, creating unit tests against the source project, testing the newly generated codebase against these, and iteratively debugging based on the error messages and logs. The most common question I hear is, “does it work with language x to language y?” My answer is: “try it!” Usually the tool will create *something* useful, even if it isn’t totally correct the first time (and with human help, it can cross the finish line). I’d encourage you to give it a shot with a codebase that you have, even if just for fun! I’d recommend using GPT-4 32k if you do. Smaller models can work but sometimes the input size can be too large. We’re currently working on an improvement here. What’s next? There are a few “boring” improvements we’ll be making, such as creating more robust unit tests, better and more benchmarks, support for more LLMs, etc. There are also a few exciting things on the horizon that you’ll need to stay tuned to hear about! (Did someone say no more devops? 👀) Cheers and happy migrating!
