OpenAI has developed GPT-4, the most recent milestone in its efforts to scale deep learning. GPT-4 is a big multimodal model (accepting picture and text inputs and producing text outputs) that demonstrates human-level performance on various professional and academic standards despite being less proficient than humans in many real-world settings.

For example, it passes a simulated bar exam with a score in the top 10% of test-takers, whereas GPT-3.5 scored in the bottom 10%. In addition, OpenAI spent six months iteratively aligning GPT-4 utilising lessons from their adversarial testing programme and ChatGPT, resulting in their best-ever results (but not flawless) regarding factuality, steerability, and avoiding going outside of guardrails.

Over the previous two years, OpenAI redesigned its whole deep learning stack and, in collaboration with Microsoft Azure, constructed a custom supercomputer for our workload. They trained GPT-3.5 as a "test run" for the system a year ago. OpenAI discovered and resolved several flaws and enhanced its theoretical foundations. So reliable was their GPT-4 training run that it became their first significant model. OpenAI intends to refine its methods to enable them to foresee and prepare for future capabilities with ever-greater foresight, which they view as essential for safety as they focus on scalable reliability. 

OpenAI will release ChatGPT and the API for GPT-4's text input feature (with a waitlist). In addition, they are working closely with a single partner to begin preparing the picture input capacity for broader distribution. OpenAI is also open-sourcing OpenAI Evals, its framework for automated evaluation of AI model performance so that anyone can report flaws in their models to aid in their future enhancement.

OpenAI anticipates that GPT-4 will become a vital tool for enhancing people's lives by powering several applications. OpenAI looks forward to improving this model through the collaborative efforts of the community as it builds on top of, explores, and contributes to it.