OpenAI releases GPT-4

OpenAI/GPT models

OpenAI released GPT-4, a large model that accepts images as well as text. On a simulated bar exam it scored around the top 10% of test-takers, where GPT-3.5 scored around the bottom 10%. OpenAI rebuilt its training stack with Microsoft Azure and could predict the model's performance before training finished; text input shipped first in ChatGPT and the API.

• Accepts images as input alongside text; can describe humor in unusual images and answer diagram questions.
• Scores in top 10% on simulated bar exam, contrasting with GPT-3.5 in bottom 10%.
• Available with 8,192 and 32,768 token context windows, significant improvements over GPT-3.5.
• More reliable, creative, and able to handle more nuanced instructions than GPT-3.5.