Although the new system is more accurate than its predecessor, it still makes real-world errors. As one example provided by OpenAI, it referred to Elvis Presley as the son of an actor. GPT-4 also lacks knowledge of events after September 2021 and cannot extrapolate. However, OpenAI has improved many safety features of GPT-4, including reducing the model’s tendency to generate forbidden content elements by 82% compared to GPT-3.5, while it only responds to about 29% of sensitive requests such as medical advice and self-harm.
Through ChatGPT, users can experience GPT-4’s ability to generate text from images. This feature allows users to input an image of text, such as a screenshot or a scanned document, and receive a written version of the text in response. This functionality has enormous potential for improving accessibility and efficiency, as it can quickly convert text from an image into a format that is easier to read and manipulate.