OpenAI has created GPT-4, the latest achievement in the deep learning extension. This multimodal model, capable of processing both images and text, has shown human-like performance on a variety of professional and academic tests, such as a simulated bar exam, where it ranked in the top 10 percent of examinees.
After rebuilding the entire deep learning stack over the past two years, OpenAI and Azure jointly designed a supercomputer specifically for OpenAI's workload. GPT-3.5 was a test run of the system, which helped identify bugs, improve the theoretical underpinnings and optimize the training run of GPT-4.
"We expect GPT-4 to become a valuable tool for improving people's lives by powering many applications. There is still much work to be done, and we look forward to improving this model through the collective effort of the community that is developing, exploring and contributing to it," OpenAI.
GPT-4 text input capability is available through ChatGPT and the API, with a waiting list for the latter. The image input capability is being prepared for wider availability through collaboration with a single partner. OpenAI Evals, the framework for automated performance evaluation of artificial intelligence models, has been launched so that anyone can report model deficiencies and thus contribute to their improvement.
GPT-4: What to expect?
While the difference between GPT-3.5 and GPT-4 may be subtle in casual conversations, the capabilities of GPT-4 become more apparent when the complexity of the task reaches a sufficient threshold. GPT-4 is more reliable, creative, and capable of handling much more nuanced instructions than GPT-3.5, as demonstrated by tests on a range of benchmarks, including simulation of tests designed for humans.
GPT-4 also outperforms existing large language models in traditional benchmark tests, even for resource-poor languages. OpenAI has been using GPT-4 internally with great impact in functions such as support, sales, content moderation and scheduling, as well as to help humans evaluate AI results.
In addition, GPT-4 can accept text and images, generating natural language, code or other text output from inputs composed of interspersed text and images. Developers and, soon, ChatGPT users can customize the style and task of their AI by describing those prompts in the "system" message.
This release is a significant step in OpenAI's effort to scale deep learning reliably and securely, and the company intends to refine its methodology to help predict and prepare for future capabilities earlier and earlier.
