ChatGPT Launch: GPT-3.5 Dialogue, Free Access, Early Growth

The underlying model was not new

ChatGPT was fine-tuned from a model in the GPT-3.5 series that had already finished training in early 2022, not a new base model built for the occasion. [1]

Two days before ChatGPT shipped, OpenAI had already put another GPT-3.5 model, text-davinci-003, into its API and Playground, aligned with the same reinforcement-learning-from-human-feedback approach as InstructGPT. OpenAI researcher Jan Leike called it higher-scoring on human preference ratings "without being fundamentally more capable" than the model it replaced. [3]

Atlas interpretation: So the capability jump readers later credited to ChatGPT had, in large part, already shipped days earlier through an API few outside developer circles used. What changed on November 30 was mostly who could reach the model. [1][3]

What ChatGPT's training actually added

OpenAI called ChatGPT a sibling model to InstructGPT: same reinforcement-learning method, different data. Human trainers wrote conversations by playing both the user and the assistant, and that dialogue set was mixed with the existing InstructGPT data, converted into the same conversational format. [1]

That training let the model hold a back-and-forth: answering follow-up questions, admitting mistakes, challenging an incorrect premise in a prior turn, and refusing some requests, rather than only completing a single prompt the way the davinci API models did. [1]

Atlas interpretation: MIT Technology Review's framing captures the size of the change: OpenAI trained GPT-3 to master the game of conversation and invited everyone to come and play. That is a real improvement in behavior, but a different thing from a leap in underlying capability, and the two got conflated in most of the coverage that followed. [2]

A product change, not just a training change

OpenAI released ChatGPT as a free research preview at a public web address, with no API key, no billing account and no per-token pricing to work out first, the barriers that had kept the GPT-3.5 API models mostly in developers' hands. [1]

What the user-growth numbers actually count

The "one million users" figure traces to a founder's tweet on December 5, 2022, five days after launch. It was an unaudited count from OpenAI itself, with no published definition of what counted as a user, a signup, a visit, or something else. [4]

The other widely repeated figure, roughly 100 million users within two months, is not an OpenAI number at all. It is a UBS analyst note from February 2023, built on third-party Similarweb data showing about 13 million daily unique visitors to the ChatGPT website in January, extrapolated to a monthly-active-user estimate and compared to TikTok's nine months and Instagram's two and a half years to the same mark. [5]

Atlas interpretation: Both numbers describe traffic to a free web product, not paying customers, retained users, or any audited metric, and neither was measured the same way. What they demonstrate reliably is the speed of the on-ramp: a browser tab anyone could open, not a change in what the model could do. [4][5]