OpenAI, a leading artificial intelligence research organization, has announced the launch of its latest deep learning model, GPT-4. This new model is a large multimodal model that accepts both image and text inputs and produces text outputs, exhibiting human-level performance on various professional and academic benchmarks.
GPT-4 has already achieved significant milestones, such as passing a simulated bar exam with a score around the top 10% of test takers, an improvement over its predecessor GPT-3.5, which scored around the bottom 10%. OpenAI spent six months aligning GPT-4 using lessons from their adversarial testing program and ChatGPT, resulting in their best-ever results on factuality, steerability, and refusing to go outside of guardrails.
OpenAI has rebuilt their entire deep learning stack over the past two years and co-designed a supercomputer with Azure specifically for their workload. They trained GPT-3.5 as a “test run” of the system, which allowed them to find and fix bugs and improve their theoretical foundations. As a result, GPT-4’s training run was unprecedentedly stable, becoming their first large model whose training performance they were able to accurately predict ahead of time.
While OpenAI is releasing GPT-4’s text input capability via ChatGPT and their API, there is currently a waitlist for access. The image input capability will be prepared for wider availability by collaborating with a single partner. It is worth noting that GPT-4 is only available in the paid version.
OpenAI is open-sourcing OpenAI Evals, their framework for automated evaluation of AI model performance, to allow anyone to report shortcomings in their models to help guide further improvements.
GPT-4 is a significant milestone in scaling up deep learning, and OpenAI has already achieved noteworthy progress in its performance. With its release, OpenAI continues to focus on reliable scaling and aims to predict and prepare for future capabilities, something they view as critical for safety.