Generative Pre-Trained Transformer, abbreviated GPT, refers to a transformer-based language-model approach trained first on broad text or multimodal data and then adapted for useful interaction. Pretraining teaches statistical patterns, concepts, and relationships by predicting tokens; instruction tuning and preference optimization can subsequently make the model more helpful and controllable. GPT systems can draft, summarize, translate, answer questions, write code, and support tool-using applications. The acronym is associated with OpenAI model families, but it also describes the core design idea. A GPT output is generated from learned probabilities and context, so fluency does not guarantee factual accuracy, current knowledge, or sound judgment.
OpenAI’s next version of its Sora video generator will default to including copyrighted content unless owners explicitly opt out – drawing criticism from media creators.