GPT-4: A New Language Revolution for AI

May 17, 2023

In the rapid development of artificial intelligence, language processing technology has always played an important role. The OpenAI ChatGPT series has already changed the way we interact with machines, and the latest GPT-4 takes this revolution to new heights. This AI model utilizes cutting-edge machine learning algorithms to provide unprecedented natural language processing capabilities. It can not only understand and generate text but also engage in deep conversations. This article will take you on a deep dive into GPT-4 and explore this new language revolution in artificial intelligence.

What is GPT-4?

GPT-4 is the latest natural language processing model released by OpenAI on March 14, 2023. Compared to its predecessor, GPT-3, GPT-4 has demonstrated human-level performance on many professional tests and academic benchmarks, surpassing over 90% of candidates in simulated lawyer exams. However, it is worth noting that GPT-4 still has room for improvement in real-world scenarios, with issues of fictional facts and reasoning errors.

One major innovation in GPT-4 is the addition of support for visual inputs, enabling the model to generate text outputs based on both textual and image prompts. However, this functionality is currently in the research stage and not yet available to the public.

With the release of GPT-4, OpenAI has also upgraded its conversation model, ChatGPT, and the new ChatGPT Plus now supports GPT-4, providing more powerful conversational capabilities.

What’s New in GPT-4?

The development of GPT-4 aims to enhance the model’s “alignment” capability, which means being more closely aligned with user intent while ensuring the output’s truthfulness and minimizing potentially offensive or risky content.

1. Performance Improvements

GPT-4 has shown improvements in the accuracy of factual answers compared to GPT-3.5. The model exhibits fewer instances of “hallucinations” where it produces factual or reasoning errors, improving its performance on OpenAI’s internal factuality benchmark by 40% compared to GPT-3.5.

Additionally, GPT-4 enhances “controllability,” the ability to change its behavior based on user instructions. For example, you can instruct it to write in different styles, tones, or voices. Try providing prompts like “You are a talkative data expert” or “You are a concise data expert” and ask it to explain a concept in data science. You can read more about designing excellent prompts for GPT models.

Furthermore, the model has improved in following the guardrails. If you ask it to do something illegal or inappropriate, it performs better at refusing such requests.

2. Visual Input Usage in GPT-4

One significant change in GPT-4 is its ability to handle image inputs (currently limited to research previews and not publicly available) alongside text. Users can specify any visual or language task by alternating between text and image inputs.

Example of GPT-4 visual input:

GPT-4 reply：

The image shows a package for a “Lightning Cable” adapter with three panels.

Panel 1: A smartphone with a VGA connector (a large, blue, 15-pin connector

typically used for computer monitors) plugged into its charging port.

Panel 2: The package for the “Lightning Cable” adapter with a picture of

a VGA connector on it.

Panel 3: A close-up of the VGA connector with a small Lightning connector (used for charging iPhones and other Apple devices) at the end.

The humor in this image comes from the absurdity of plugging a large,

outdated VGA connector into a small, modern smartphone charging port.

How to use GPT-4?

So how can you make the most of GPT-4’s capabilities?

Here are two quick ways to access GPT-4:

Getting on the GPT-4 API waitlist
Subscribing to ChatGPT Plus for access to GPT-4

1.Getting on the GPT-4 API waitlist

To access GPT-4 directly, you can join the API waitlist. Once you gain access, you can use the API to integrate GPT-4 into your applications and services. However, it’s important to note that this option may not always be available as access is provided on a case-by-case basis and may come with a significant cost.

Here are the steps to join the GPT-4 waitlist:

Step 1: Visit the OpenAI website and click on “Products.”

Step 2: Choose “GPT-4,” and you will be redirected to the GPT-4 page.

Step 3: Click on “Join the API waitlist,” and it will take you to the GPT-4 API waitlist page.

Step 4: Fill in the required information and click the “Join the waitlist” button.

Please note that OpenAI is rolling out GPT-4 gradually, and they carefully evaluate each application, so it may take some time.

2.Subscribing to ChatGPT Plus for access to GPT-4

Another way to use GPT-4 is by subscribing to ChatGPT Plus, which provides access to GPT-4 for various applications, including text generation, conversations, and content creation.

While this method makes it easier to use GPT-4, it does require a subscription fee.

Here are the steps to quickly upgrade to ChatGPT Plus subscription:

Step 1: Visit ChatGPT.

Step 2: In the bottom left corner, you can find the option to “Upgrade to Plus.” Click on it.

Step 3: You will see a pop-up comparing the free and ChatGPT plans.

Step 4: Click on “Upgrade plan,” and you will be redirected to enter your credit card details and billing information. Fill in the details and make the payment.

Step 5: Once the payment is completed, you will have access to both GPT-4 and GPT-3.5.

Step 6: Select the GPT-4 model from the drop-down menu on the ChatGPT interface and start generating content.

Additionally, OpenAI will continue to update the model, so stay tuned for the latest changes.

GPT-4 Performance Benchmarks

OpenAI evaluated GPT-4 by simulating exams designed for humans, such as the Uniform Bar Examination for lawyers and the LSAT, as well as SAT exams for university admissions. The results showed that GPT-4 achieved human-level performance on various professional and academic benchmarks.

OpenAI also benchmarked GPT-4 against traditional machine learning models, including 57 multiple-choice questions across subjects, everyday event commonsense reasoning, and elementary science multiple-choice questions. In these tests, GPT-4 surpassed not only existing large language models but also most state-of-the-art models, including those specifically designed or additionally trained for benchmark tests.

OpenAI further tested GPT-4’s ability to handle other languages by using Azure Translate to translate the MMLU benchmark test, which includes 14,000 multiple-choice questions across 57 subjects, into various languages. In 24 out of 26 languages tested, GPT-4’s English handling abilities outperformed GPT-3.5 and other large language models.

Limitations of GPT-4

OpenAI explicitly acknowledges that although GPT-4 has made significant advancements in many aspects, it still faces challenges such as “social biases, producing incorrect information (hallucinations), and reacting to adversarial prompts.” These issues are also present in previous versions of the GPT models.

In other words, GPT-4 is not flawless. There are areas where it still needs improvement and optimization. However, OpenAI has made it clear that they are actively addressing these issues. These efforts include in-depth research on reducing social biases in the model, improving its ability to discern factual information to minimize hallucinations, and enhancing its ability to recognize and respond to adversarial prompts.

OpenAI’s open and honest approach not only demonstrates their commitment to product quality and user experience but also showcases their profound understanding of AI ethics and social responsibility. OpenAI is taking concrete actions to fulfill its promises by continuously refining and improving GPT-4’s capabilities to serve the public and society.

The Price of GPT-4

For more information, please refer to the official description of OpenAI

Models	GPT-4 API Basic Version	GPT-4 API Plus Version	ChatGPT Plus
Price	1.$0.03 / 1K prompt tokens 2.$0.06 / 1K completion tokens	1.$0.06 / 1K prompt tokens 2.$0.12 / 1K completion tokens	$20/mo
Function	Functional for normal use: make plain text requests (image input is still in limited alpha)	Limited access to 32,768 token context (about 50 pages of text) versions	Limited text input and generation

GPT-4 History

OpenAI’s founder, Sam Altman, mentioned in January 2023 that GPT-4 would be unveiled sometime in 2023.He also clarified that the rumors of GPT-4 training with 1 quadrillion parameters were false. (Reference: interview with StrictlyVC)

US Representatives Don Beyer and Ted Lieu confirmed to The New York Times that Altman visited Congress in January 2023, showcasing GPT-4 and its improved “safety controls” compared to other AI models. (Reference: The New York Times)
On February 8, 2023, Microsoft released a new version of Bing based on GPT-4.
During a digital kickoff event called “AI in Focus,” held by Microsoft Germany, it was announced on March 9, 2023, that GPT-4 would be released the following week. (Reference: German media, Heise Online)
On March 14, 2023, OpenAI officially announced the release of GPT-4. Contrary to previous rumors, it does not include video input functionality, and the training parameter count was not disclosed.