gpt (generative pre-trained transformer)

GPT (Generative Pre-trained Transformer) is OpenAI's family of autoregressive language models that generate text by predicting the next token given all preceding tokens, establishing the foundation for modern large language models and conversational AI systems. The GPT series has progressed through several generations of increasing scale and capability: GPT-1 (2018, 117M parameters — demonstrated that unsupervised pre-training followed by supervised fine-tuning could achieve strong results across diverse NLP tasks), GPT-2 (2019, 1.5B parameters — showed emergent zero-shot task performance, generating coherent long-form text that raised concerns about misuse), GPT-3 (2020, 175B parameters — demonstrated remarkable few-shot learning capabilities through in-context learning, performing tasks from just a few examples without fine-tuning), GPT-3.5/ChatGPT (2022 — fine-tuned with RLHF for instruction following and conversational ability, launching the AI chatbot revolution), GPT-4 (2023 — multimodal model accepting text and image inputs, significantly improved reasoning, reduced hallucination, and broader knowledge), and GPT-4o (2024 — natively multimodal across text, vision, and audio with faster inference). GPT architecture uses the decoder portion of the transformer with causal (left-to-right) self-attention masking, ensuring each token can only attend to preceding tokens. Training objective is next-token prediction: maximize P(t_n | t_1, ..., t_{n-1}). This simple objective, scaled with massive data and compute, produces models with emergent capabilities — chain-of-thought reasoning, code generation, translation, and creative writing — that were not explicitly trained for. Key innovations across the series include: scaling laws (establishing predictable relationships between compute, data, model size, and performance), in-context learning (performing new tasks from demonstrations in the prompt), RLHF alignment (training models to be helpful, harmless, and honest), and tool use (integrating external tools and APIs into generation).