I had my first experience last week of a useful conversation with an AI. I was trying to edit some text but was tired and struggling to find the right words. I gave ChatGPT an example of the style of writing I was aiming for, then asked it to summarise the text I was struggling with. It did a good first draft and helped me get past the writer’s block.
Since then I’ve started to take it more seriously as a tool. I’ve got some catching up to do to use it well. Tyler Cowen and Alex Tabarrok have created these tips for using GPTs to teach and learn economics, though the tips apply across many other tasks.
What does ChatGPT do well?
- Read, transform and manipulate text
- Improve writing
- Improve coding
- Summarize ideas
- Answer analytical questions with short causal chains
- Solve simple models
- Write exams
- Generate hypotheses and ideas
Where is its performance mixed?
- Finding data
- Directing you to sources
- Summarizing papers and books
Cowen and Tabarrok have a good, concrete set of tips for ‘prompt engineering’:
- Surround your question with lots of detail and specific keywords
- Make your question sound smart, and ask for intelligence and expertise in the answer
- Ask for answers in the voice of various experts.
- Ask for compare and contrast.
- Have it make lists
- Keep on asking lots of sequential questions on a particular topic.
- Ask it to summarize doctrines
- Ask it to vary the mode of presentation
- Use it to generate new ideas and hypotheses
Finally:
When working with GPT models, many of your conventional assumptions about what computers can and cannot do will be challenged. Interacting with a GPT differs from programming or using previous software. Instead, it resembles collaborating with a bright and knowledgeable research assistant, albeit one from a different culture.
Read the whole thing here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4391863.
Peter Henderson 
Leave a comment