This is the first in a series of articles about ChatGPT. In this article we will explain ChatGPT (and the technologies behind it) in simple non-technical terms
You can read the next articles in this series afterwards:
What is ChatGPT?
Since ChatGPT was released a few months ago (November 30, 2022), it has caused a lot of excitement on the internet.
However much of the content available about ChatGPT tends to fall into two extremes:
1. “ChatGPT will change everything and many jobs will become obsolete”. e.g., 300 million jobs affected by AI.
2. “Look at this example of how dumb ChatGPT is” e.g., Best/Worst/Funniest ChatGPT Responses.
The reality is, of course, somewhere in the middle. We all remember how the Internet was supposed to reduce employment around the world.
ChatGPT has shown that many tasks that our society used to (mistakenly) believe were out of the realm of computers are actually possible. For example, a computer having a conversation with a person where the person cannot tell that the other side is human or computer, a computer being able to create original content, a computer being able to understand emotions from users and using emotion in its conversation, and a computer being able to have a back-and-forth conversation with a person.
In terms of the “funny responses” from ChatGPT, just like a foreigner speaking English for the first time, ChatGPT will make some mistakes. Humans are not very clear when we talk and it takes a lot of iteration and context for us to understand each other. A number of the limitations of ChatGPT are due to our questions not being clear or specific since human communication is frequently neither.
ChatGPT is a tool (technically a set of tools as described below). Much of the future of ChatGPT will be determined by people like us in each industry figuring out how to use it and how to address the concerns around its use.
Note that ChatGPT is the name of the product that OpenAI has built. There are other companies like Google that are building products on the same underlying technologies (Google Bard). In the article we use ChatGPT to refer to the set of capabilities that the ChatGPT product has but similar capabilities will be available from other vendors so you are not tied to OpenAI products.
Disclaimer
The goal of this article to be simple enough for a non-data scientist to understand. As a result some simplifications are used. I’ll be writing more technical articles in the near future for the data scientists so don’t fret if the simplified explanations below are not completely accurate.
Technologies behind ChatGPT
First let’s start with clarifying what ChatGPT is.
ChatGPT is actually just a chat based interface on top of Large Language Models (LLMs), Generative Pre-trained Transformers (GPT), Generative AI (GenAI) and Reinforcement Learning with Human Feedback (RLHF). Don’t worry if these terms are not known to you; I will describe them below.
It is important to not limit the capabilities of the underlying technologies to just chat. Chat was just the use case that made the technologies accessible to people on the internet. The technologies behind ChatGPT can be used in traditional applications that are not chat based. Not every applications needs to become chat based!
NOTE: In the rest of the article I used the term “ChatGPT” to mean the underlying technologies and NOT the chat based interface that internet users use today.
Large Language Models (LLM)
A large language model (LLM) is a language model consisting of a neural network with many parameters (typically billions of weights or more), trained on large quantities of unlabeled text using self-supervised learning or semi-supervised learning.
Ok that is the technical explanation. Simply LLM is a machine learning model that can understand language used by humans.
Language models have existed for a long time. One of the earliest was ELIZA in 1960s. However, recent advances in processing power in the cloud, in more efficient training methods for neural networks and in parallel training techniques have enabled us to create and use larger and larger language models.
When you start learning a new language, you struggle until you learn about a thousand words (technically word families or lemmas) and then are able to have a decent conversation in that language. To be a native speaker, you have to know about 20,000 word families.
Due to the advances in technology mentioned above, large language models have reached sufficient size where they can be used to understand and converse in languages.
Generative Pre-trained Transformer (GPT)
Generative Pre-trained Transformers (GPT) are a family of LLMs (Large Language Models) used for natural language processing (NLP) tasks such as language translation, text summarization, and question answering.
Simply put, GPT is a language model that has been trained on a corpus of knowledge.
OpenAI is the company that created a GPT called ChatGPT.
GPT-3, released in 2020, is a whopping 175B parameter model pre-trained on a corpus of more than 300B tokens. GPT-4 was released on March 14, 2023. Unlike GPT-3.5, which focuses primarily on text, GPT-4 can analyze and comment on images and graphics.
If you’re interested in more technical details, you can read this article: Why is ChatGPT so good?
Generative AI (GenAI)
GPT is a type of Generative AI Model.
Generative AI produces new content, chat responses, designs, synthetic data or deep fakes.
On the other hand, Traditional AI tends to focus on detecting patterns, making decisions, honing analytics, classifying data and detecting fraud.
Reinforcement Learning With Human Feedback (RLHF)
The basic idea is to train an additional reward model that rates how good a model’s response is from the perspective of a human to guide the model’s learning process. Then use this reward model to fine-tune the original language model using reinforcement learning.
Explained simply, RLHF is a technique that allows us to capture feedback from humans and use that to improve machine learning models.
While RLHF has existed for some time, our increase in the ability to capture human responses, process them quickly and be able to generate responses in real time has increased dramatically in recent past which has enabled us to use RLHF more broadly.
RLHF is applicable outside of Generative AI models since it can be used to improve all kinds of machine learning models by using human feedback. Even if you don’t use ChatGPT, you should be looking at RLHF to improve the performance of any user facing machine learning model.
Key Capabilities of ChatGPT
- Prompting (being able to ask questions in natural language)
- Conversational (iterative back-and-forth conversation)
- Summarizing (creating summaries automatically from large text content)
- Inferring (infer additional knowledge that is not present in the provided text content)
- Transforming (converting the format of content from one form to another e.g., English to Spanish or clinical to non-clinical)
- Expanding (converting a “to-the-point” response into a more natural expanded response)
- Personalizing (targeting the persona of the end user e.g., doctor vs nurse vs administrator vs patient etc and the persona that the software should adopt e.g. act like an assistant, as a teacher, etc)
- Understanding and expressing emotion
Continue to How to Use ChatGPT.