GPT-4 accepts images and improves difficult language tasks
OpenAI introduces GPT-4, a multimodal model that accepts text and images while showing stronger performance on professional and academic benchmarks.
OpenAI introduced GPT-4 today, making text access available through ChatGPT Plus and a waitlisted programming interface. The model can accept both text and images as input, although image capability is initially being tested with selected partners.
The company reports large gains over GPT-3.5 on several exams, including a simulated bar examination, while describing ordinary conversation as a subtler improvement. GPT-4 was trained to follow instructions and incorporates months of safety work and feedback from outside experts.
Benchmark scores offer evidence of broader capability, but they do not remove familiar weaknesses. OpenAI says GPT-4 can hallucinate facts, make reasoning errors and lacks knowledge of events after its training cutoff. Its wider context and visual input expand possible uses, making careful evaluation and disclosure more consequential wherever an answer directly affects people.
Source: OpenAI ↗