Artificial intelligence (AI) has made tremendous progress in recent years, and one of the most remarkable developments is the emergence of ChatGPT. This AI-powered language model is generating significant interest due to its ability to hold smooth and natural conversations with users.

In this article, we will explore, in an accessible manner, how ChatGPT works, including its training and operation.

 

Summary:

  1. What is the difference between GPT and ChatGPT?
    • Definition and functioning of GPT and ChatGPT.
  2. How does ChatGPT work?
    • Training with internet sources, human supervision, AI supervision, and human language
definition gpt

1. What's the difference between GPT and ChatGPT?

What is GPT?

The GPT model was initially introduced by OpenAI in June 2018 with the release of the first model, GPT-1, trained on over 500 billion words. Since then, several iterations and improvements have been made. GPT-2 was unveiled in February 2019, and GPT-3 in June 2020. These models have undergone significant evolution in terms of size, capabilities, and performance.

“GPT-3 is not only capable of writing correctly in multiple languages, but it is also an encyclopedic model that integrates numerous references to the real world (people, events, scientific knowledge), which it reproduces with varying degrees of accuracy. (...) GPT-3 doesn’t have much inhibition and can generate anything as long as the text is superficially coherent.” — “ChatGPT: How Does It Work?” by Pierre-Carl Langlais, Hypotheses, 07/02/2023.

 

Since March 2023, GPT-4 provides more precise and reliable answers compared to its predecessor.

GPT ("Generative Pre-trained Transformer") is a large language model (LLM). This technology is actually familiar to everyone. In our messaging apps or search engines, we use intuitive text (also known as predictive writing), which relies on a pre-existing dictionary and probabilities.

GPT uses predictive writing to generate responses that are likely coherent by choosing the most probable sequence of words.

 

What is ChatGPT?

 

ChatGPT is the conversational agent based on the GPT language model, specifically designed to provide answers and engage in conversations with users.

definition chatgpt

 

2. How does ChatGPT work?

When you give ChatGPT an instruction or a priming phrase (called a “prompt”), the model analyzes the context and generates a response using the knowledge it has acquired during training. It tries to produce a sequence of words that seems appropriate in the given context.

chatGPT training visualization
Cheating in the age of ChatGPT: findings and solutions
chatGPT data analysis

Training with Internet sources

GPT is trained on vast quantities of textual data from the Internet, articles, newspapers, books, social networks, forums... It learns to predict the probability of a word or phrase appearing based on a given text. ChatGPT uses a neural network architecture called Transformer, which is designed to capture the relationships between words in a text. 

GPT is trained to give the response expected by a human.

chatGPT Analysis Developer

Training with human supervision

GPT is trained to give several answers to a question. Humans classify and evaluate the relevance of these answers by giving a “reward”. The aim is to reinforce the consistency of the answers, and avoid drift and errors. 

Reward AI is developing. It understands how to classify and evaluate responses. This process is called Reinforcement Learning.

Queries analyzed by AI

Training with AI supervision

In this final phase of training, GPT provides several answers to a question under the control of a reward AI. This AI evaluates the responses to refine them.

Humans verify that the given answers are not sensitive and align with current thought patterns. This step is called "fine tuning."

Algorithm chatgpt

Training with human language

After all this training, the technology is asked to answer questions using human language. ChatGPT is the conversational version of GPT. It's what we call a Chatbot.

ChatGPT works interactively by exchanging turns with a user. It uses previous information, including the user's messages and the model's previous responses, to generate answers.

Discover AI detection

In conclusion, ChatGPT is a conversational tool that answers your questions based on training using billions of sources.

The content generated by ChatGPT remains detectable thanks to an AI Checker.

This ChatBot is widely used by students and teachers. Can we say it is a promising advancement or a step backward for education? What is certain is that AI continues to evolve.

Here are some of Gartner.com's key predictions (information for professional decision-makers) regarding generative AI:

  • By 2025, generative AI will produce 10% of all data (today, it is less than 1%) and 20% of all test data for consumer use cases.
  • By 2025, generative AI will be used in 50% of drug discovery and development initiatives.
  • By 2027, 30% of manufacturers will use generative AI to enhance the efficiency of their product development.
    — Explanation of Generative AI Models Altexsoft, 10/13/2022.

 


To learn more:


Information: This informative article, which does not require personal reflection, was partially written with the help of ChatGPT. We revised the automatically generated content (correcting repetitions, refining phrasing, adding details, adding citations, and verifying the accuracy of the information).