GPT-2 large language model

GPT-2 is a large-scale, general purpose language model that generates text. Developed by OpenAI, GPT-2 was released as a limited beta in February 2019. The complete version, comprising 1.5 billion parameters, was published in November 2019.

Created to perform a variety of purposes, such as providing prompts, summarising text, and answering questions, GPT-2 was trained on BookCorpus and a dataset of 8 million web pages scraped from the internet.

Reaction

On its release, researchers, practitioners and commentators generally praised GPT-2's ability to generate realistic, plausible writing and translate text. Its flexibility was also welcomed.

Some were less enthusiastic, pointing out the model's tendency to generate gibberish, incoherent lengthy text, and to answer questions inaccurately. Potential misuses of GPT-2, including to generate convincing misinformation and disinformation were also raised.

Others took issue with the fact that training and developing GPT-2 and similar models is highly resource-intensive and consumes  huge amounts of energy.

Transparency

OpenAI said that it would keep GPT-2's dataset, code, and model weights private 'Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale'.

The decision proved contentious, with some people supporting the decision by arguing GPT-2 was too dangerous to publish in the wild.

Others argued that the threats posed by GPT-2 were over-stated, the approach was opaque, and would not allow the software to be properly tested.

OpenAI later published the code on Github.