GPT-4
OpenAI says six months of safety work made GPT-4 less likely to comply with disallowed requests and more likely to answer factually than GPT-3.5.
The company says GPT-4 is 82% less likely to respond to disallowed content requests and 40% more likely to produce factual responses in its internal evaluations. OpenAI says it used more human feedback, input from over 50 experts, and lessons from real-world use of earlier models. It also says GPT-4 helped accelerate safety work by generating training data and improving classifiers. The model is available through ChatGPT Plus and an API, while OpenAI notes remaining limits including bias, hallucinations, and adversarial prompts. OpenAI's note
The company says GPT-4 is 82% less likely to respond to disallowed content requests and 40% more likely to produce factual responses in its internal evaluations. OpenAI says it used more human feedback, input from over 50 experts, and lessons from real-world use of earlier models. It also says GPT-4 helped accelerate safety work by generating training data and improving classifiers. The model is available through ChatGPT Plus and an API, while OpenAI notes remaining limits including bias, hallucinations, and adversarial prompts. OpenAI's note
score 10