Many people recently have heard of the amazing capabilities of chatGPT from OpenAI (the recipients of multimillion investments from Microsoft). ChatGPT has incredible AI capabilities including having conversations, answering questions, and even creating stories or writing essays and poems. For example, here I gave ChatGPT a prompt to “create a poem about an ostracized ostrich”:

This blew me away! I was shocked by the dramaticism and beauty within this poem. While I feel I can tell this wasn’t written by a real person, perhaps that is only because I already know it wasn’t?

What in the devil is going on? Well, ChatGPT is a refined model based on OpenAI’s GPT-3. GPT-3 is an extremely capable AI/ML model that performs exceptionally as a natural language processor (NLP) which can understand and generate natural language. It’s been trained on a dataset which included almost everything available at the time on the entire web!

To start incorporating GPT-3 into your applications is actually very straightforward. You can either interact through HTTP requests, through their python bindings, their Node.js library, or a community maintained library. The documentation can be found at https://platform.openai.com/docs/. You can start for free with an $18 credit for your first 3 months.

Within GPT-3 there are 4 main models that have different levels of power and therefore allow you to balance cost and performance. Their most powerful model has training data as recent as Jun 2021. From highest to least performance and what OpenAI suggests each model is good at:

• Davinci – complex intent, cause and effect, summarization for audience
• Curie – language translation, complex classification, text sentiment, summarization
• Babbage – moderate classification, semantic search classification
• Ada – parsing text, simple classification, address correction, keywords

For my capstone project, I am most interested in moderate to complex text classification so I will most likely be utilizing the Curie model (which costs about 1/10 as much as Davinci). I am excited to try using OpenAI API and hope to have some updates in the next blog post!