GPT-4 arrives, a model that sees pictures and passes a lawyers' exam

On 14 March 2023, OpenAI released GPT-4, a new model for ChatGPT. The company described it as more reliable and able to handle far more nuanced and complex requests than its predecessor, GPT-3.5.

To show the difference, OpenAI gave the model real professional exams. On a simulated bar exam, the test people in the US take to qualify as lawyers, GPT-4 scored around the top 10% of test takers. GPT-3.5 had scored around the bottom 10% on the same exam.

Another first: GPT-4 understands images. You can show it a photo or a diagram and ask what is in it. At first this was tested with a single partner, Be My Eyes, a service that helps blind and low-vision people make sense of their surroundings.

ChatGPT Plus subscribers could use GPT-4 right away, with a limit on the number of messages. Developers could join a waitlist to connect the model to their own apps.

Why it matters

GPT-4 showed that ChatGPT's success was no accident: the next model, built the same way, was again markedly smarter. The industry fully bought into a simple recipe: more data and more compute, better results.

For many professions, it was a first warning. If AI can handle the lawyers' exam, what happens to the work of lawyers, doctors, programmers? The question has stayed on the agenda ever since.

For the curious

OpenAI calls GPT-4 a large multimodal model: it accepts text and images as input and responds with text. At launch, image input remained a research preview and was not available to everyone. Access to the API (application programming interface, the way other programs connect the model to their services) was granted through a waitlist.