2 events
ChatGPT opens, and a million people try it in five days
On 30 November, OpenAI released ChatGPT: free, open to anyone who signed up, and framed as a conversation. A million people used it within five days, far beyond what the company had expected.
Technically it was not a new model. It was the existing GPT-3.5 line reshaped into a chat and tuned on human feedback. The same capability, handed over differently, reached a different set of people.
Two days earlier, on the 28th, OpenAI had released text-davinci-003, a stronger version of the same GPT-3.5 line.
This block is written by the editors. It is kept separate from the sourced record below.
OpenAI released a new model, text-davinci-003, on its API. As the latest in the GPT-3 line, it improved instruction following, longer-form generation, and writing quality. Jan Leike, who led the alignment team, explained that it shared the same underlying model as text-davinci-002 but was more aligned and scored higher on human preference ratings. It belonged to the same GPT-3.5 generation as ChatGPT, which was released two days later.
OpenAI released ChatGPT, a conversational AI, as a free research preview. Built on a GPT-3.5-series model fine-tuned for dialogue with reinforcement learning from human feedback (RLHF), it reached one million users within five days — an unprecedented pace at the time for something billed as a research preview.