OpenAI has officially announced the release of its latest text-generating model, GPT-4, making it available through its API. The company is granting access to existing OpenAI API developers with a proven track record of successful payments. The plan is to gradually expand access to new developers by the end of this month and subsequently increase availability limits, depending on the availability of computing resources.
OpenAI revealed that since March, millions of developers have expressed their interest in accessing the GPT-4 API, leading to a growing number of innovative products that leverage this advanced model. The company envisions a future where chat-based models can cater to a wide range of use cases.
One notable improvement of GPT-4 over its predecessor, GPT-3.5, is its ability to generate text, including code, and accept both image and text inputs. OpenAI trained GPT-4 using publicly available data from sources like web pages, as well as licensed data. Although GPT-4 performs at a "human level" on various professional and academic benchmarks, it still exhibits imperfections. For instance, it may produce inaccurate information and reasoning errors with an unwarranted sense of confidence. Additionally, GPT-4 lacks the capability to learn from its own experiences, which renders it incapable of solving complex problems such as introducing security vulnerabilities into the code it generates.
The image-understanding feature of GPT-4 is currently being tested with a single partner, Be My Eyes, and OpenAI has not disclosed when it will be available to a wider customer base.
OpenAI plans to enable developers to fine-tune both GPT-4 and GPT-3.5 Turbo, another recent text-generating model, with their own data. This functionality, which has been available for several of OpenAI's other models, is expected to be introduced later this year.
The competition in the field of generative AI has intensified since the announcement of GPT-4 in March. Recently, Anthropic expanded the context window of its flagship model, Claude, from 9,000 tokens to 100,000 tokens. In comparison, GPT-4 held the previous record with a context window of 32,000 tokens. Models with smaller context windows often struggle to maintain the context of recent conversations, leading to digressions from the topic at hand.
In addition to the GPT-4 announcement, OpenAI has made its DALL-E 2 (image-generating) and Whisper (speech-to-text) APIs generally available. The company plans to deprecate older models accessible through its API to optimize compute capacity, as it has experienced challenges in meeting the high demand for its generative models, partly due to the popularity of ChatGPT.
Effective January 4, 2024, certain older models, including GPT-3 and its derivatives, will no longer be accessible. These models will be replaced with new "base GPT-3" models, presumably designed to improve compute efficiency. Developers currently using the older models must upgrade their integrations manually by January 4. Those wishing to continue using fine-tuned old models beyond that date will need to fine-tune replacements built on top of the new base GPT-3 models.
OpenAI has committed to providing support to users who have previously fine-tuned models, aiming to facilitate a smooth transition. In the coming weeks, OpenAI will contact developers who have recently utilized the older models and provide more information once the new completion models are ready for early testing.