Following the research path from GPT, GPT-2, and GPT-3, our deep learning method leverages more information and more calculation to produce increasingly sophisticated and capable language designs.

We invested 6 months making GPT-4 more secure and more aligned. GPT-4 is 82% less likely to react to demands for disallowed material and 40% most likely to produce accurate responses than GPT-3.5 on our internal assessments.

Safety & positioning

Training with human feedback We included more human feedback, consisting of feedback sent by ChatGPT users, to improve GPT-4's behavior. We also dealt with over 50 experts for early feedback in domains including AI security and security.

Continuous enhancement from real-world usage We've applied lessons from real-world use of our previous models into GPT-4's safety research and tracking system. Like ChatGPT, we'll be updating and improving GPT-4 at a regular cadence as more individuals use it.

GPT-4-assisted security research GPT-4's advanced reasoning and instruction-following capabilities accelerated our safety work. We utilized GPT-4 to help create training data for design fine-tuning and repeat on classifiers throughout training, evaluations, and monitoring.

Built with GPT-4

We have actually teamed up with companies constructing ingenious items with GPT-4.

GPT-4 deepens the conversation on Duolingo.

Be My Eyes uses GPT-4 to transform visual ease of access.

Stripe leverages GPT-4 to improve user experience and battle fraud.

Morgan Stanley wealth management releases GPT-4 to arrange its vast understanding base.

Khan Academy explores the capacity for GPT-4 in a restricted pilot program.

How Iceland is using GPT-4 to protect its language.

More on GPT-4

Research.
GPT-4 is the current milestone in OpenAI's effort in scaling up deep knowing.
View GPT-4 research

Infrastructure.
GPT-4 was trained on Microsoft Azure AI supercomputers. Azure's AI-optimized infrastructure likewise allows us to deliver GPT-4 to users worldwide.

Limitations.
GPT-4 still has numerous known limitations that we are working to attend to, such as social predispositions, hallucinations, and adversarial triggers. We motivate and assist in openness, user education, and wider AI literacy as society adopts these designs. We likewise intend to expand the opportunities of input individuals have in shaping our models.