GPT

GPT is the name of the AI models that OpenAI makes and that power ChatGPT. You type a question or an instruction, and the model writes an answer by predicting, piece by piece, which words come next. A business can use it to draft an email, summarise a report or answer a customer's question. The letters stand for generative pre-trained transformer.

What do the letters mean?

• Generative: the model produces new text.
• Pre-trained: before it is put to work, it learns from a very large amount of text by predicting the next word, over and over. GPT-2 learned from 40 GB of text from 8 million web pages (OpenAI, February 2019).
• Transformer: the type of neural network it is built on, introduced by Google researchers in 2017 (Attention Is All You Need). See transformer.

How does GPT write text?

GPT splits your prompt into tokens, pieces of words. It calculates how likely each possible next token is, picks one, adds it to the text and repeats until the answer is complete. Each step only looks at the text before it, because GPT is a decoder-only transformer with masked self-attention, as the first GPT paper already describes.

A pre-trained model only continues text. To make it follow instructions, OpenAI trains it further with human feedback (RLHF). In the InstructGPT paper (2022), people preferred the answers of a 1.3-billion-parameter model trained this way to those of the 175-billion-parameter GPT-3. ChatGPT, launched on 30 November 2022, was trained the same way from a model in the GPT-3.5 series.

How did GPT develop?

• GPT-1 (2018): Radford and colleagues trained a 12-layer transformer on more than 7,000 unpublished books, then fine-tuned it for each task (Improving Language Understanding by Generative Pre-Training).
• GPT-2 (February 2019): 1.5 billion parameters. Citing concerns about misuse, OpenAI released it in stages and published the full model only in November 2019.
• GPT-3 (May 2020): 175 billion parameters. It could do new tasks from a few examples in the prompt (few-shot), without further training (Brown et al.).
• GPT-4 (March 2023) accepts images as well as text. OpenAI published no details on its size or architecture (GPT-4 Technical Report).
• GPT-4o (2024) is an "omni" model that takes text, audio, images and video and can answer in text, audio and images (system card).
• GPT-5 (7 August 2025) decides by itself when to answer quickly and when to think longer. In September 2026 OpenAI released the first GPT-6 models.

What is the difference between GPT, ChatGPT and GPTs?

GPT is the model; ChatGPT is the app you talk to, which runs on GPT models. In November 2023 OpenAI also started calling custom versions of ChatGPT "GPTs": ChatGPT with your own instructions and extra knowledge, built without code, for yourself, your company or everyone (Introducing GPTs). In April 2026 OpenAI introduced workspace agents for teams as "an evolution of GPTs" (OpenAI).

Where does it go wrong?

OpenAI's own GPT-4 report says the model is not fully reliable: it can hallucinate, has a limited context window and does not learn from experience. It also lacked knowledge of most events after its training data ended, in September 2021. OpenAI wrote in 2022 that ChatGPT "sometimes writes plausible-sounding but incorrect or nonsensical answers". Check facts, figures and quotes before you use them.

Example

Suppose a webshop asks ChatGPT for a product description of a blue raincoat. The model turns the request into tokens, predicts the first token of the answer, adds it and predicts the next, until the description is finished. Because it writes what sounds likely, it may call the coat breathable when it is not. Check every product claim before it goes online.