The features of GPT-4

The GPT-4 chatbot just unveiled by its creator OpenAI has been generating a lot of buzz, thanks to its spectacular evolution compared to its previous version. Let’s take a look at the new features of GPT-4.

Processing both text and images

The first of the revolutionary GPT-4 features is its multimodal capability, which allows it to process both text and images. The previous GPT-3.5 version only worked with written content.

In practice, if you submit an image to GPT-4 along with a question, the AI (Artificial Intelligence) is able to provide you with an answer.

Creation and synthesis at the heart of the new GPT-4 features

OpenAI has presented its robot as “more creative and collaborative” than any other existing AI system. Its knowledge base has been broadened, enabling the AI to solve more problems requiring synthesis and to produce far more sophisticated writing.

For comparison, GPT-3.5’s capacity was 3,000 words, versus more than 25,000 words for the new version. Rather than writing a short article, the new features of GPT-4 make it possible to produce much more substantial documents. Indeed, the robot not only understands words, but is also able to decipher intent through the writing style of the person asking it a question. It can also adapt its own writing to each person’s personal style.

Among the possibilities, we can mention a blog post generated from a keyword, a sentence, or an image, designing a comic strip, composing a song, summarizing a long text, and so on.

GPT-4’s ability to pass human tests

OpenAI subjected its chatbot to very demanding tests, usually reserved for humans. The GPT-4 version proves far more capable than GPT-3.5. However, the robot shows limits when it comes to demonstrating creativity, as is the case with literature.

The ideological dimension of GPT-4

While it is up to the AI to produce answers, the new features of GPT-4 were programmed by humans, which means the ideological dimension of the answers provided should not be overlooked. At a time when cancel culture and wokeism are raging, GPT-4 users must remain vigilant and never take the ideas conveyed by the robot at face value.

GPT-4’s security

For the new GPT-4 features, OpenAI carried out remarkable work on security. The goal was to address the concerns raised about security failures in the previous version. The robot is therefore better able to refuse to answer biased questions in order to prevent certain dangerous behaviors. One example would be a request for instructions on how to build a bomb.

GPT-4’s security has also been strengthened to withstand cyberterrorist attacks. However, while the rate of poor security-related outcomes has been cut in half, GPT-4 still shows vulnerabilities when it comes to malicious acts. It remains possible to code malware capable of fueling disinformation.

GPT-4 used by the Bing search engine

Microsoft’s search engine, Bing, already uses GPT-4, as confirmed by its head of consumer marketing on his blog: “We are happy to confirm that the new Bing is running on GPT-4, which we’ve customized for search.”

Concrete examples of the new GPT-4 features

Here are a few concrete examples of how the new GPT-4 features are being used.

Interpreting an image and its context

The famous American newspaper the New York Times conducted a rather conclusive experiment on the contents of a refrigerator.

• The person running the test opened their fridge door and took a photo of what was inside.
• The photo was submitted to GPT-4, along with the question “what can I make with these ingredients?”
• The AI suggested several recipes using the ingredients available in the fridge and identified by the robot.

The GPT-4 features appear particularly interesting for visually impaired people, since the AI then becomes a personal assistance service. The example cited by OpenAI concerns clothing: two shirts were placed on a bed, and the robot was asked which of the two was red. The robot proved infallible.

Generating text

Writing scripts or songs is one of the most spectacular features of GPT-4. OpenAI has put a lot of work into its chatbot’s creativity. While the text it produces isn’t usable as-is and often needs to be reworded, it does answer some of your questions and can inspire your own writing.

OpenAI also ran complex tests that the robot managed to answer. For instance, it wrote a text meeting requirements set by its requester that were, to say the least, quirky: “Explain the plot of Cinderella in a single sentence where each word must begin with the next letter of the alphabet, from A to Z, with no letter repeated.”

Answering everyday questions

GPT-4 is a paid product. To use it, you need to pay for a monthly plan. However, you can go through the Bing search engine to ask it specific questions. All you need is a Microsoft account. A simple Gmail address is enough.

Summarizing texts of more than 25,000 words

25,000 words corresponds to roughly fifty pages in a Word document. You can therefore ask the robot to read the entire text for you and then summarize it, saving you a considerable amount of time.

Answering legal questions

GPT-4 has incorporated legal jargon, which allows it to answer many highly specific and technical questions.

Learning a foreign language quickly

GPT-4 is multilingual, which broadens its applications and allows it to be used by businesses and individuals in a wide range of countries.

Duolingo — an app aimed at learning foreign languages — uses GPT-4 to let its users chat with the robot, which plays the role of a fictional conversation partner, helping them improve in the language they are learning.

Creating a video game in record time

GPT-4 knows how to code. For example, it can create video games. A test was carried out by asking it to design one of the ancestors of video games: the famous “Pong,” with two rectangles acting as paddles playing out a ping-pong match.

Thanks to the new GPT-4 features, the new version of Pong was created in under 1 minute.

Creating a Chrome extension

Need a custom Chrome extension? All you have to do is give GPT-4 your instructions, then ask it to code it.

Designing an app or website from a sketch

Interpreting an image opens the door to another particularly impressive GPT-4 feature: you draw a sketch evoking the finished project, then task the robot with building it. It will then design an app or website matching your request from scratch.

This list is of course not exhaustive, and the features of GPT-4 will keep on developing. There is no doubt that the feats of OpenAI’s robot have not finished amazing us, with a tool that keeps becoming ever more creative and versatile.