ChatGPT is OpenAI's AI chat app, launched in November 2022. You type or talk to it and it answers with text, images or voice. Inside, it uses OpenAI models from the GPT family: to answer with text, the model generates tokens from what it learned and from the conversation.
Hi! How can I help?
7 tokensA question from the app. Pick an answer.
Are a chat app and the model it uses the same thing?
The app is the product you interact with. The model is one of its components.
One app can switch models or use several for different tasks.
The app can use a remote or local model, and switch it. Installing the app does not imply downloading the model.
GPT stands for Generative Pre-trained Transformer. "Generative" because it builds new text from patterns it learned, though it sometimes repeats passages it read: it is generative AI. "Pre-trained" because it first learned from a huge amount of text and only later was tuned to hold a conversation. And Transformer is the kind of neural network it uses, introduced by Google researchers in 2017.
Your message, along with the earlier part of the chat and some OpenAI instructions you do not see, is split into tokens. The model works out probabilities for the tokens that could come next, picks one, adds it and repeats until it is done. That is why the answer shows up bit by bit. (The split in the picture above is an example: it depends on the model.) To make it good at chatting, after pre-training it was tuned with sample dialogues and with people who rated its answers.
What it remembers from the chat lives in its context window: in a very long conversation, the early parts can lose detail. Memory, if you have it turned on, saves a few facts about you and puts them back into the context. The model itself does not change while you talk.
ChatGPT is the product: the app, the website and what it adds around the model, like searching the web, reading a file or making an image. The model inside keeps changing: OpenAI releases new versions and the app switches to them, sometimes picking one or another depending on the task or your plan. Claude and Gemini work the same way: a chatbot on the outside, an LLM on the inside.
The model answers with what it learned and what is in the chat. To talk about something recent it needs current information: a news story you paste in or a web search. When ChatGPT searches, it shows you the sources: check them. If it shows no sources, do not assume it searched.
"If it sounds sure, it is true." It can hallucinate: make up a fact, a quote or a link that sound right. Check what matters. "It learns from what I tell it." The model does not change while you chat; what the company later does with your chats depends on your settings and its policy. "ChatGPT is AI." It is one app among many; artificial intelligence is a much larger field.