ChatGPT turns a language model into a continuing conversation
OpenAI releases ChatGPT as a free research preview, presenting a dialogue system that answers follow-ups, revises mistakes and refuses some unsafe requests.
OpenAI released ChatGPT today as a free research preview. The web interface lets a user carry on a written conversation with a model that remembers earlier turns, responds to follow-up questions and can revise an answer when challenged.
The system is related to InstructGPT and was trained with human feedback. OpenAI says trainers wrote conversations and ranked model responses, while additional safeguards aim to reject harmful requests. The company also plainly lists limitations, including plausible-sounding false answers and sensitivity to wording.
A conversational wrapper makes language-model capabilities accessible without programming or specialized prompts. Students and workers will quickly test it on explanations, drafts and code, where fluent prose can disguise factual errors. This preview is therefore as much an experiment in human judgment as in generation: every useful answer still needs checking.
Source: OpenAI ↗