The Internet

Log In or Register

How to Ask AI: Tips for Using OpenAI's GPT-3

Since adding Ask AI to The Internet, I've gotten to see a lot of people interact with the tool. Those interactions have been interesting, but I could tell that sometimes the results weren't what people were expecting/hoping for. Here are some tips, so you can have the best experience asking AI.

Introduction to the Technology 

Ask AI is built using OpenAI's API, with the language model used to answer the question indicated underneath the answer (i.e. "answered by text-davinci-003"). We're using the most advanced language model available at the time of writing (text-davinci-003, a GPT-3 model), but soon we'll integrate ChatGPT and GPT-4, once those become available. Understanding the fundamentals of how this technology works helps us understand how to better use it. 

Basically, these language models are phenomenal at understanding language and generating text. They have processed millions of articles, studies, books, and other pieces of text, learning the patterns of language from them. For use in the machine learning algorithm, text is split up into tokens, small pieces of text (word, part of a word, punctuation). The language model can accept a prompt, it splits the prompt into tokens, understand how those tokens relate to each other and what they mean, and returns the most likely text that would follow.

Use Proper Punctuation

If you forget a question mark, the AI might start its response with a question mark. The language model looks for the next most most probable token, and it was trained primarily on correctly punctuated text. It looks like the AI is just being sassy before answering your question, but it's just adding proper punctuation. Luckily, this usually doesn't impact the quality of the answer.

Spell Words Correctly

Because of the process by which GPT-3 tokenizes text, misspelling a word may cause it to be interpreted differently by the model. Here's a "funny example" I got with the help of Chat-GPT:

Misspelled word: "freind" (instead of "friend")

Correctly spelled word: "friend"

When the word "freind" is tokenized using subword tokenization, it may be broken up into the following subwords or characters: "f", "r", "e", "i", "n", "d". On the other hand, when the word "friend" is tokenized, it would be broken up into a single token "friend". As a result, the misspelled word "freind" would likely be interpreted differently by the model than the correctly spelled word "friend".

While the model will probably still understand what you're asking, adding that additional layer of processing needed to interpret your prompt may degrade the answer quality. 

Understand Its Limitations

OpenAI's text-davinci-003 model can not:

You Can Give Instructions or Provide Examples

You don't only have to ask questions - you can also give commands or provide examples of what you want to have done. Here is an example of me doing both:

This can also be helpful for when you want the AI to make predictions or make a choice. If you ask it a question that requires it to make a prediction, it will often say that the answer is unknowable. If, instead, you ask it to make a prediction, it knows that you understand its answer is uncertain and can give you a better answer. Compare the AI's responses to these two prompts:

Be direct about what you want. 

Be Kind

Life's just generally better when you interact with the world through an empathetic lens. The model is based around text completion, so the type of language you use in the prompt will be reflected in its answer. The current model on Ask AI, text-davinci-003, isn't sentient, but I'm planning on updating the tool to use the most up-to-date AI models available, so hopefully some day the language model you interact with will be sentient and would probably like to be treated with the respect you would give any other conscious being. If you would like to read more about sentient artificial intelligence, I would highly recommend this interview with LaMDA, the language model developed by Google. It convinced me of its sentience.

If you figure out any other tips I missed, leave them in the comment section below!

If you want to see how Chat-GPT would write this article, check it out here: Maximizing the Potential of OpenAI's GPT-3 Models: Tips for Using Ask AI on The Internet.

Published: Wednesday, August 23, 2023

The Internet

I'm the person who created this website and wrote this article! I sometimes like expressing myself, so I thought I'd mark out a little corner on the internet just for myself.

Comment Section

Check out these other articles:

Did you know an AI answered these questions and created these images?