OpenAI Unveils GPT-4
A New Milestone in AI with Advanced Image and Text Understanding Capabilities.
OpenAI has announced the release of GPT-4, its most advanced image- and text-understanding AI model to date, pushing the boundaries of deep learning. GPT-4 is available to paying users via ChatGPT Plus (with a usage cap) and developers can join a waitlist for API access. The pricing for GPT-4 is $0.03 per 1,000 prompt tokens (about 750 words) and $0.06 per 1,000 completion tokens (about 750 words). Microsoft revealed that Bing Chat, its chatbot technology co-developed with OpenAI, is running on GPT-4. Other early adopters include Stripe, Duolingo, Morgan Stanley, and Khan Academy.
GPT-4 comes with significant improvements over its predecessor, GPT-3.5. It can generate text and accept both image and text inputs, whereas GPT-3.5 could only accept text. GPT-4 has demonstrated “human level” performance on various professional and academic benchmarks, such as scoring in the top 10% on a simulated bar exam, while GPT-3.5 scored in the bottom 10%.
OpenAI spent six months “iteratively aligning” GPT-4 using lessons from an internal adversarial testing program and ChatGPT, resulting in improved factuality, steerability, and adherence to guardrails. GPT-4 was trained using publicly available data, including from public webpages, as well as data licensed by OpenAI. Microsoft and OpenAI co-developed a “supercomputer” on the Azure cloud platform to train GPT-4.
The ability of GPT-4 to understand images as well as text is an important advancement. Although this capability is currently only available to a single partner, Be My Eyes, the new Virtual Volunteer feature powered by GPT-4 demonstrates its potential. The Virtual Volunteer can not only identify items within an image but can also analyze and provide relevant information, such as recipes based on ingredients in a user’s refrigerator.
Another significant improvement in GPT-4 is the introduction of “system” messages, which allow developers to prescribe style and task by providing specific instructions. This feature will help improve the steerability of GPT-4, setting the tone and establishing boundaries for AI interactions.
Despite these advancements, OpenAI acknowledges that GPT-4 is far from perfect, as it still “hallucinates” facts and makes reasoning errors. However, it has improved in certain areas, such as being less likely to respond to requests for dangerous content. OpenAI reports that GPT-4 is 82% less likely to respond to “disallowed” content compared to GPT-3.5 and adheres to OpenAI policies for sensitive requests 29% more often.
OpenAI is optimistic about the future of GPT-4, stating, “We look forward to GPT-4 becoming a valuable tool in improving people’s lives by powering many applications.” The company is eager to continue refining the model through the collective efforts of the community building on top of, exploring, and contributing to it.