OpenAI Releases ChatGPT
A free website where anyone can type a question and get a fluent, human-sounding answer. It's labeled a “research preview” — a careful way of saying nobody, including OpenAI, is quite sure what it's for yet.
OpenAI, a San Francisco research company, put a new website online today called ChatGPT. It's free, and anyone can use it after signing up with an email address. There's nothing to install and nothing to learn: you type a question into a box, the way you'd type a text message, and it writes back in full sentences. Ask it to explain how a mortgage works, draft a birthday toast, or help with ninth-grade algebra, and it will — and it keeps the thread going, so you can say “shorter” or “now make it funnier” and it remembers what you were talking about.
If OpenAI's name rings a bell, it's probably from DALL·E, the program released earlier this year that draws pictures from a written description. ChatGPT is the same idea applied to writing. Underneath it is an LLM — a program that has read an enormous amount of text and learned to predict what words should come next. OpenAI has had a version of this technology for a couple of years, but until now you needed to be a software developer to use it. Today it's a chat window.
The reason it feels different from earlier chatbots — the ones that answer “I'm sorry, I didn't understand that” — is a training step OpenAI describes in its announcement. After the program learned from all that text, people were hired to have conversations with it and rate its answers, and the program was adjusted toward the answers people preferred. In plain terms, it was taught manners: to answer the question that was asked, to admit when it's wrong, to push back on a false premise, and to decline requests it shouldn't help with.
OpenAI is unusually candid about what it doesn't do well. Its own announcement says ChatGPT “sometimes writes plausible-sounding but incorrect or nonsensical answers” — and it says them in the same confident tone it uses when it's right. It can be talked into things by rephrasing a question. It tends to run long. And it has no way to look anything up; everything it knows, it learned before its training ended, which OpenAI says was sometime in 2021. Ask it about last week and it will either say it can't, or make something up.
OpenAI is calling this a “research preview,” which is a careful way of saying nobody, including OpenAI, is quite sure what it's for yet. The stated reason for releasing it free is to see what people do with it and where it breaks. Judging by the screenshots already filling Twitter this evening — poems about cheese, fake Seinfeld scripts, working computer code — people are doing a lot with it.
Whether it's useful or just entertaining is a fair question. For now it feels a bit like Google, except the answer reads as if a person typed it — and you can't be sure the person is right.