ChatGPT Was the Weaker Model OpenAI Released First on Purpose
OpenAI released the weaker model first and held the stronger one back for a later launch, after a last-minute rename created ChatGPT.
Invest Like the Best with Patrick O'Shaughnessy · Sam Altman on AGI, Compute, and Human AgencyDevelopers were hunched over a bare testing screen, typing out sample back-and-forths just to trick an early artificial intelligence model into chatting. The tool was called the playground, and it was not designed for conversation. Users had to show the model a few examples of what chatting even meant, and then the model would chat back.
That awkward workaround caught the eye of Sam Altman, the entrepreneur leading OpenAI, the company behind the models. At the time, OpenAI was trying to sell access to its GPT-3 model, and the business was thin. The model now looks strikingly dumb, and Altman said the only commercial use that really worked was copywriting, where a customer might pay a marketing firm 20 dollars for a landing page or pay OpenAI 20 cents to have the computer draft it.
This is the story of how the most famous chatbot started as a sidestep. OpenAI had a weak moneymaker that could write ad copy, and a clumsy test box that users kept using to chat. Altman said he had learned a lesson from YC, which is Y Combinator: "if you notice your users doing something," go down that path. So the team decided to build a good chatbot, because that was what people were already trying to do.
Then the plan changed. The team finished a much stronger model, GPT-4, and began using it internally. Altman recalled the reaction inside the lab as, "This is a big deal." A chat interface plus GPT-4 at once felt like a lot, and there were hard questions about fake news, offensive answers, and whether the company would get in trouble. So OpenAI staged the rollout, putting out the chat interface with GPT-3.5 first.
It was not meant to be a product, and it was not expected to be a huge hit. The working title was Chat with GPT-3.5. A few hours before launch, the team renamed it ChatGPT and put it out as a research preview, hoping it would help the public catch up to what was coming. The idea was to release the preview, then follow a few months later with a real product built on GPT-4.
Altman said outsiders felt the preview crossed a threshold. Altman recalled people saying, "Okay, this is awesome." There was not yet that much practical utility, but it was the first time many people could feel progress for themselves, with something they simply enjoyed using. He called himself a massive texter his whole life and said he knew what it was like to just "start chatting in a text box."
Questions people ask
What was the playground developers were using?
The playground was a developer tool for testing GPT-3. You typed a prompt, adjusted settings, and saw what the model generated. Because the model had not been tuned for dialogue, users had to paste in sample conversations to steer it, which is why chatting felt like a trick.
Why was copywriting the one business that worked for GPT-3?
Short marketing copy is formulaic, low risk, and easy to edit, so even a weak model could be useful. A flawed draft still saved time versus paying a firm full price. Open-ended questions, coding, and factual work required more reliability than GPT-3 had.
Why hold back GPT-4 if it was already working inside OpenAI?
OpenAI leaders thought GPT-4 was a large jump and worried about fake news, offensive content, and other misuse. Releasing chat with the older, weaker GPT-3.5 let them test public reaction and safety issues first. A full GPT-4 product followed months later.
What does it mean to launch something as a research preview?
A research preview signals an experimental release for learning, not a polished, supported product. Companies use the label to gather feedback, study safety, and set lower expectations. In this case, the preview became so popular that it effectively became the product.