GPT (Generative Pre-trained Transformer)

GPT denotes the generative pre-training recipe introduced by Radford et al. (2018): pre-train a decoder-only Transformer on unlabelled text, then fine-tune. Brown et al. (2020) scaled this to GPT-3 (175B parameters) and showed strong few-shot, in-context learning, establishing the template most modern LLMs follow.

OpenAI's ChatGPT is the best-known product built on this lineage, but "GPT" now broadly names the decoder-only, next-token-prediction approach.