Sparked by a blog post by Github on Copilot productivity enhancements and developer happiness, there was lively reflection on LLMs as a developer tool. One developer brought up the error scenario where ChatGPT would get stuck in a conversation loop. As a remedy, I offered this advice: sometimes it feels like it’s cornered and can’t get out. In such cases, I feed it a relevant excerpt from the documentation just to get it going again - thanks to in-context learning 💪🏻. There may still be gaps in transferring, but these can be rectified by posing mere suggestive questions…
I confirmed that ChatGPT can be used for Rubberduck Debugging and also for giving a rough description first and iterate on that. There are things that get in the way though:
- ChatGPT Plus message limit for GPT-4
- limited context size: ChatGPT appears “forgetful” at some point
- some programming languages are better suited that others. My impression was that Python works better than Qt/QML
#2 must also be relevant to Copilot - in some way, regarding “[copilot-x] can see the context of the project”. Large inputs are usually “indexed” first and then selectively chosen for presentation to GPT. But that can induce said “forgetfulness” and oversights.
Finally, I emphasized on making sure to use GPT-4 for best results, rather than GPT-3.5 which continues to be the default:
