(no title)
ultrablue | 1 year ago
The results have been dismal. Fine-tuning prompts is a good exercise for requirements gathering, complete with the tedium that we all love. Sadly none of the AIs that I used were capable of remembering the current requirements with any degree of confidence. ChatGPT4 does a pretty good job until it forgets everything randomly. Both it and CoPilot failed to remember simple instructions like "Please don't send the updated requirements until I ask for them." I assume so that I can waste tokens?
For code creation, both AIs consistently failed to include features that were in the requirements and in pseudo code that they provided and that I approved.
I did enjoy the change of pace of prompt engineering. It's fun while it's happening and while the AI behaves. But it gets very old very quickly saying things like:
Me: We just added <some requirement>, but I don't see it in the most recent version of the requirements.
It: Sorry Dave, you're right. Here's the latest requirements:
<Requirements still without the one in question.>
Hallucinations abound, not just adding things at random, but forgetting things, sometimes with a great deal of resistance to incorporating the forgotten thing.
Also, once you do get to code, don't expect to fine tune it. If you see a bug or oversight in AI-provided code and point it out, it won't correct the code. Instead it appears to re-generate it from scratch. Will there be new, unasked for features? Could be. What about the correction -- is it fixed? Maybe. And maybe in a completely new way.
I have found some success with CoPilot and basic research. In essence it's a smarter search engine. So prompts like this can result in useful leads:
I've got some data that has <a description of some feature in the data>.
Questions:
What is the correct term for <some feature>?
What python library would you suggest to help me analyze <the feature>?
It's quite possible that my prompt-fu is simply weak and ignorant. And using COTS free-ish AIs I'm getting what I pay for. But I would say that as far as using an AI as an all-purpose junior goes, we're not there yet.
No comments yet.