OpenAI has officially launched GPT-4, which had been rumored previously. This latest version of its chatbot AI program boasts many new features and performance enhancements, as detailed in the accompanying blog post.
The most significant addition to GPT-4 is its ability to respond to queries that combine both text and images. The blog states:
Over a range of domains—including documents with text and photographs, diagrams, or screenshots—GPT-4 exhibits similar capabilities as it does on text-only inputs. Furthermore, it can be augmented with test-time techniques that were developed for text-only language models, including few-shot and chain-of-thought prompting.
Unfortunately, the option to input images is currently only accessible to those in the “research preview” mode, meaning it’s not yet available to the general public. It’s unclear when this feature will become widely available.
Furthermore, users can modify the tone of GPT-4’s responses in the new version. OpenAI blog states:
Rather than the classic ChatGPT personality with a fixed verbosity, tone, and style, developers (and soon ChatGPT users) can now prescribe their AI’s style and task by describing those directions in the “system” message. System messages allow API users to customize their users’ experience within bounds significantly.
In addition, the blog post included several performance charts that demonstrated the enhancements to GPT-4. One of the examples involved OpenAI testing the chatbot’s proficiency through simulated exams. The blog states:
We proceeded by using the most recent publicly-available tests (in the case of the Olympiads and AP free response questions) or by purchasing 2022–2023 editions of practice exams. We did no specific training for these exams.
According to the above-shown chart, GPT-4 achieved notably superior exam scores than GPT-3.5 in most cases. However, as with earlier iterations, GPT-4 has the constraint of being unable to retrieve current subject information, limited to only as recent as September 2021. OpenAI has also claimed that this latest version “substantially decreases hallucinations compared to prior models,” although it “cannot guarantee complete reliability.”
Access to the GPT-4 API is currently reserved for developers who sign up for a waitlist. For the general public, limited access to GPT-4 is available through ChatGPT Plus, which costs $20 monthly. The accompanying blog post suggests that OpenAI may introduce an additional subscription tier allowing greater volume levels of access.