(Author’s note: This was written in mid-2023 but never published.)
As an analogy, imagine a parrot named Sam.
For the last fifty years Sam has lived inside a bar near universities, churches, and a park fieldhouse, all of which host varied discussion and interest groups. There’s also a TV behind the bar which is never shut off, but does have the channel changed occasionally to suit the audience.
All sorts of folks come into the bar, from professors to construction workers, conspiracy theorists to knitting groups. They have varied conversations that Sam hears and remembers, along with everything he hears on TV.
Can Sam the parrot create commercial jingles in iambic pentameter Yup. What about reciting the Declaration of Independence in the voice of a pirate Sure, parrots are super smart.
Should you trust Sam to write a legal brief with case citations Nope. What about asking him to provide a report on recent stock market trends No way. If you do either of these, will the result sound knowledgeable (but likely incorrect) Absolutely.
This is the most important thing to remember about generative AI (such as ChatGPT): systems like these are built to SOUND great and deliver text that is amazingly fluent and well-organized, but NOT to deliver reliable answers or specific information (like a knowledgeable sassy parrot). Also, should you trust Sam to keep your secrets safe You should not.
Definitions
Just like a good contract, let’s start with some definitions for clarity. We’ll first break down this sentence: “ChatGPT is an application that leverages LLMs like GPT-3.5 and GPT-4 to interact with users using plain language prompts.”
ChatGPT– The name of an application launched by OpenAI on November 30, 2022.
Chat– The application interface that connects ChatGPT users to the powerful GPT models that drive results. Many powerful models existed previous to these, but the easy-to-use chat functionality of ChatGPT is what drove its explosive growth and adoption.
GPT– Generative Pretrained Model
Generative– Describes something that makes or creates some sort of output. The output can be text, images, sounds, computer code, etc.
Pretrained– Denotes a model has already been programmed with (trained on) useful data. For ChatGPT to work, engineers had to (1) Build the model, (2) Train the model, (3) Build the chat application interface, (4) Connect the model to the chat interface, and (5) Iterate the rules of the interface along with the model settings to optimize performance.
Model– An adjustable program with “if this, then that” logic that can intake, process, and output data. For GPT, the model is designed to take in text prompts and output generated text (Dall-E takes in text and generates images).
More to come, but for now just trust ChatGPT as much as you would Sam, if you asked a question in a bar and a parrot started squawking away.
