Large language models (GPT lineage and rivals)
Transformers pretrained on vast text that generate language, code and reasoning; the general-purpose engine of the 2020s AI boom.
- 2018

First version
OpenAI (Alec Radford et al.); later Anthropic, Google DeepMind, Meta, DeepSeek and others
Image: OpenAI · Public domain · Wikimedia Commons - 2019

GPT-2
Tenfold scale-up (1.5B parameters) writes coherent long text and performs tasks without task-specific training.
Image: OpenAI · Public domain · Wikimedia Commons - 2020

GPT-3
175B parameters; few-shot learning from prompts without fine-tuning.
Image: OpenAI · Public domain · Wikimedia Commons - 2022

ChatGPT (instruction-tuned with RLHF)
Human-feedback training turns the raw model into an assistant that follows instructions in conversation.
Image: OpenAI · Public domain · Wikimedia Commons - 2023

GPT-4
Multimodal: accepts images as well as text and scores near top human levels on professional exams.
Image: OpenAI · Public domain · Wikimedia Commons - 2024

Reasoning models (OpenAI o1)
RL-trained chain-of-thought spends more compute 'thinking' before answering.
Image: OpenAI · © OpenAI (official launch artwork; no r… · OpenAI o1 launch page, 12 Sep 2024 (Wayback cap…
Contested: ULMFiT and ELMo (2018) also pioneered pretrained language models; what counts as 'large' has no agreed threshold.