OpenAI launches GPT-4

OpenAI released GPT-4 on March 14, 2023 through ChatGPT Plus and a limited API rollout. It described a model better able to follow complex instructions and tackle difficult tasks than GPT-3.5. Public access initially covered text; image input remained a research preview with a selected partner. OpenAI ↗

The accompanying technical report described strong results across professional and academic benchmarks. GPT-4 scored around the top tenth of test takers on a simulated bar examination. The report also stressed that the model remained less capable than people in many real-world situations: success on an examination was a bounded capability result, not a demonstration of dependable performance across an entire profession. OpenAI ↗

The model could interpret images alongside text and produce written answers, including explanations of diagrams and documents. Developers could also use system instructions to shape its behavior. Together, these changes extended the range of tasks one general-purpose model could handle, although the announced capabilities and what users could access on launch day were different. OpenAI ↗

OpenAI reported six months of safety work and testing by more than 50 external experts. It nevertheless warned that GPT-4 could invent facts, make reasoning mistakes, generate insecure code and be induced to bypass safeguards. The release therefore paired a substantial capability advance with continuing limits on reliability and control. OpenAI ↗