After seeing it on Twitter and mentions by a few co-workers, I burned a couple of evenings and took the "ChatGPT Prompt Engineering for Developers" mini-course taught by (Andrew Ng - Coursera and Isabella Fulford - OpenAI).
As with any deeplearning.ai course: the instructors incorporated a simple, Pythonic approach in Jupyter notebooks with informative lectures to utilizing ChatGPT (currently Turbo 3.5) programmatically. The course ended with a build of a functioning chatbot to handle pizza orders that was surprisingly robust and easy-to-build.
I'd recommend anyone wanting to understand ChatGPT more to take this free course.
Here are 3 things that stood out to me:
1. Hallucinations are a thing and they don't go away easily
A short section delivered a hard lesson: you can't trust ChatGPT. Asking ChatGPT about a product name that was very similar to a real product had it generate a description for a product that didn't exist, but - even after putting in checks into follow-up prompts - the result suggested it did.
Every professional I know whose used ChatGPT for something serious ran into this problem. I wouldn't go as far as "trust but verify", but rather I wouldn't expect ChatGPT to share insights that I didn't feed it.
2. Building context and developing prompts to give the model "time to think"
I've seen a bit about prompt engineering, but very little on how to get it to deliver more detailed answers. The lesson they gave on Iteration showed how not just to "engineer" a prompt like you see on those annoying Twitter threads, but to programmatically incorporate feedback to better build context towards more useful results.
3. Utilizing parameters
My biggest takeaway is that not only you can adjust the "creativity" of the model through temperature, but approaching ChatGPT programmatically allows for a host of options. You can find a complete list of the parameters in the API spec, but it shows we have a long way to go with this model and with generative AI
If you're still not convinced to take a few evenings to go through this free course, @shanshaji is writing a series and diving a little deeper into some examples based on his experience with the course.
Top comments (0)