ChatGPT Brings Generative AI to a Mass Audience

On November 30, 2022, the research organization OpenAI launched ChatGPT, an artificial intelligence conversational agent engineered to generate humanlike text from natural prompts. Built upon foundation large language models, the software was refined using supervised fine-tuning, in which human trainers wrote example conversations, and reinforcement learning from human feedback, in which trainers ranked model responses to guide further adjustments. Its release made modern natural language processing directly accessible to the public, attracting one million users in five days, according to OpenAI chief executive Sam Altman, and an estimated one hundred million monthly users within two months, according to the bank UBS, which called it the fastest-growing consumer application in history.

The model acquired its linguistic versatility from extensive pre-training on diverse text corpora, made up largely of hundreds of billions of words scraped from the internet. During reinforcement training, human reviewers established reward benchmarks, while workers in Kenya employed by the outsourcing firm Sama, paid roughly $1.32 to $2 per hour, labeled graphic descriptions of violence and abuse to train a tool that detects toxic content. The resulting interface allowed members of the general public to request creative compositions, ask factual questions, translate languages, and debug software code directly inside an open chat session.