OpenAI unveils state-of-the-art AI system, GPT-4

OpenAI has launched its latest AI model called GPT-4, which excels in understanding images and text. The company claims that this is a significant step in scaling up deep learning. GPT-4 is currently accessible to paying users via ChatGPT Plus, with a usage cap, and developers can sign up on a waitlist to access the API.

The pricing for GPT-4 is $0.03 per 1,000 “prompt” tokens, which are parts of words fed into the AI model, and $0.06 per 1,000 “completion” tokens, which are the generated content. Tokens represent raw text, meaning a word like “fantastic” would be split into “fan,” “tas,” and “tic.”

Interestingly, GPT-4 has been running under the radar, as Microsoft recently confirmed that Bing Chat, its chatbot technology co-developed with OpenAI, is using GPT-4.

Several prominent companies have already adopted GPT-4 for their various applications. Stripe is using it to scan business websites and provide customer support staff with a summary. Duolingo has integrated GPT-4 into a new language learning subscription tier, while Morgan Stanley is developing a GPT-4-powered system that retrieves information from company documents and serves it to financial analysts. Meanwhile, Khan Academy is using GPT-4 to build an automated tutor.

Unlike its predecessor GPT-3.5, GPT-4 can accept both image and text inputs and perform at a human level on various academic and professional benchmarks. For instance, GPT-4 achieves a simulated bar exam score in the top 10% of test takers, while GPT-3.5’s score was at the bottom 10%.

OpenAI has spent six months iteratively aligning GPT-4 using lessons from an internal adversarial testing program and ChatGPT. The result is the “best-ever results” on factuality, steerability, and refusing to go outside of guardrails, according to the company. GPT-4 was trained on publicly available data, including public webpages, as well as data licensed by OpenAI.

OpenAI collaborated with Microsoft to develop a supercomputer from the ground up in the Azure cloud, which was utilized to train GPT-4.

OpenAI, the organization behind the development of language model GPT-4, highlighted the subtle yet significant differences between it and its predecessor GPT-3.5 in a recent blog post. Although the discrepancy might not be obvious during casual conversation, GPT-4 is more dependable, creative, and proficient in handling more nuanced tasks than its predecessor.

GPT-4’s ability to comprehend images as well as text is one of its most intriguing features. For instance, it can caption and interpret relatively complex images, such as identifying a Lightning Cable adapter from a photo of a plugged-in iPhone. Currently, not all OpenAI customers have access to this feature. OpenAI is testing it with a single partner, Be My Eyes, to begin with.

Powered by GPT-4, Be My Eyes’ new Virtual Volunteer feature can respond to queries about images sent to it. The tool can correctly recognize what’s in an image, such as the inside of a refrigerator, and even extrapolate and analyze the ingredients to suggest recipes and provide a step-by-step guide on how to make them.

OpenAI’s GPT-4 is expected to introduce an exciting steerability tooling feature that allows developers to specify AI’s style and task through system messages. With this new API capability, developers can provide specific directions to the AI, setting the tone and boundaries for its interactions.

System messages are essentially instructions that guide the AI’s next interactions. For instance, a message could instruct an AI tutor to always respond in a Socratic style and ask the right questions to help students learn to think for themselves. The AI could also be instructed to tune the questions to the student’s interests and knowledge, breaking down complex problems into simpler parts until the student can understand them.

Despite the introduction of system messages and other upgrades, OpenAI admits that GPT-4 still has flaws. It sometimes makes reasoning errors and hallucinates facts, even with great confidence. OpenAI cites an example where GPT-4 mistakenly described Elvis Presley as the “son of an actor.”

According to OpenAI, GPT-4’s knowledge is mostly limited to events that took place before September 2021, which is when the majority of its data stops. Moreover, the AI system doesn’t learn from its experience and can sometimes make simple reasoning errors. It can also be easily tricked into accepting false statements from users and fail to solve difficult problems, like humans do, including introducing security vulnerabilities in the code it produces.

However, OpenAI claims that GPT-4 has improved in certain areas compared to its predecessor, GPT-3.5. For instance, it is now less likely to fulfill requests related to dangerous chemicals and disallowed content, with an 82% reduction in such responses. Furthermore, the AI system responds more frequently in accordance with OpenAI’s policies to sensitive requests such as medical advice or self-harm related content, with a 29% increase in its compliance rate.

OpenAI is moving full speed ahead with GPT-4, despite the many improvements it requires. The company appears to be confident in the upgrades it has made, although there is much to unpack with the latest iteration of the language model.

OpenAI has stated that it is eager for GPT-4 to become an essential tool in enhancing people’s lives by powering a wide range of applications. The company recognizes that there is still much work to be done to improve the model, and it looks forward to collaborating with the community to build upon, explore, and contribute to the model.

What do you think about the recent announcement from OpenAI regarding their new AI system, GPT-4? Provide us with your opinion in the comment section.