GPT-2

GPT-2 is a 1.5-billion-parameter transformer language model released by OpenAI in February 2019. It demonstrated that a single unsupervised model could perform a wide range of language tasks — translation, summarization, question answering — with no task-specific training, a capability the paper framed as multitask learning emerging from scale.

OpenAI initially withheld the full model, citing misuse concerns, and released it in stages (124M, then 355M, 774M, finally the full 1.5B in November 2019). That staged release became one of the most debated decisions in AI history and set the template for how frontier labs communicate capability risk. The model and code were eventually released under a modified MIT license.