OpenAI releases GPT-4, its first model to accept images
OpenAI released GPT-4 on 14 March 2023 to ChatGPT Plus subscribers and to developers through a waitlist. The model accepted images as well as text. OpenAI reported it placing near the top 10 percent on a simulated bar exam, against roughly the bottom 10 percent for GPT-3.5.
Why it mattered GPT-4 became the fixed point every competing laboratory measured itself against for the rest of the year, and the threshold the pause letter named eight days later.
OpenAI had put ChatGPT in front of the public four months earlier. GPT-4 was the model that followed it. The release on 14 March 2023 reached subscribers to ChatGPT Plus first, and developers through a waitlist for the API.
The change a reader met first was a second kind of input. GPT-4 accepted images as well as text, so a photograph, a chart or a page of handwriting could form part of the question put to it. What came back was still text.
OpenAI led its announcement with examinations. It reported the model placing near the top 10 percent on a simulated bar exam, where GPT-3.5 had placed near the bottom 10 percent, and it published a technical report of about a hundred pages on arXiv the same week. The scores were the company’s own, measured on its own systems and reported by it.
Anthropic released Claude the same day, to much the same set of customers. Eight days after that, an open letter asked every laboratory to stop training anything more powerful than GPT-4 for six months. None did. Google’s answer arrived nine months later as Gemini.
For the rest of 2023 GPT-4 was the point of comparison. Competing laboratories described their models by how near they came to it, and the letter calling for a pause defined the systems it wanted stopped as those more powerful than GPT-4.