OpenAI has launched GPT-4 – the next generation neural network behind ChatGPT. What it can do

On March 14th, OpenAI launched the AI-model GPT-4, which powers ChatGPT and is the next generation neural network. OpenAI immediately announced its integration of GPT-4 into various applications and services.

ChatGPT is the breakthrough AI bot that can communicate, generate texts of any complexity, and search for information. GPT-4 surpasses the previous version, GPT-3.5, in terms of power. The new generation not only memorizes context and adapts to user queries but also processes images and demonstrates human-level ability in a range of academic and professional tests.

Let’s take a closer look at what makes GPT-4 interesting, as developers have already started implementing it in familiar services such as the Microsoft Bing search engine and the Duolingo application.

How GPT-4 is better than the previous model

GPT-4 stands out from the familiar ChatGPT because it is a multimodal model that can process not only text but also images. This is the most noticeable difference. The neural network has learned to read images, understand their context, and work with requests based on pictures. You can send GPT-4 a photo, chart, screenshot, or meme, and then build a textual request from it. For example, you can take a photo of the contents of your fridge and suggest making a dish out of the available ingredients.

However, GPT-4 still only responds in text, and it is not capable of drawing something. Additionally, the ability to process images is currently only available to a limited number of beta testers and companies. OpenAI promises to introduce this feature more widely later.

On March 14th, OpenAI launched GPT-4 and announced its integration into various applications and services.

The language model’s primary claim to fame is ChatGPT, a breakthrough AI chatbot that can communicate, generate texts of any complexity, and search for information. Compared to the neural network that powered ChatGPT’s previous version, GPT-3.5, GPT-4 is much more powerful. The new generation not only memorizes context and adapts to user requests, but it can also process images and demonstrate human-level ability in a range of academic and professional tests.

It is worth noting what makes GPT-4 interesting, as the model has already started to be integrated into familiar services such as the Microsoft Bing search engine and the Duolingo app.

What makes GPT-4 better than the previous model

One of the most significant differences between GPT-4 and the familiar ChatGPT is that GPT-4 is a multimodal model that can process not only text but also images. This capability is a significant improvement. The neural network has learned to read images, understand their context, and work with image-based queries. You can send GPT-4 a photo, graph, screenshot, or meme and then create a text-based query from it. For instance, you can take a picture of the contents of your fridge and suggest making a dish using the available ingredients.

However, it is essential to note that GPT-4 still only responds in text format and cannot be asked to draw anything. Additionally, the ability to process images is currently limited to a small number of beta testers and companies, but OpenAI promises to deploy this feature more widely later.

Image processing in GPT-4 is an example of OpenAI’s capabilities.

User:

What’s funny about this picture? Describe the frame by frame.

GPT-4:

The picture shows three panels with packaging for a Lightning adapter.

First panel: a smartphone with a VGA adapter plugged in (a large blue 15-pin connector that is typically used for computer monitors). Second panel: packaging for a Lightning cable with an image of a VGA adapter. Third panel: a close-up of the VGA adapter with a small Lightning connector (used for charging iPhones and other Apple devices) on the end.

The humor lies in the absurdity of connecting a large and outdated VGA adapter to a small and modern port for charging a smartphone.

There are other scenarios where GPT-4’s ability to work with images is beneficial as well. During the OpenAI livestream, company co-founder Greg Brockman demonstrated this capability. He took a photo of a page in his notebook containing a sketch of a website layout. The sketch featured the name “My Joke Website” and a page with two “really funny jokes,” the funniest of which would appear when a button was pressed.

Brockman uploaded the photo to GPT-4 and asked the neural network to “Write HTML/JS code to turn this layout into a colorful website where jokes are replaced with two real jokes.” GPT-4 responded by generating code for a simple website, taking Brockman’s preferences into account.

A simple sketch of a test website by hand.

Expands the possibilities of working with text

OpenAI cautions that in everyday communication and simple tasks, the difference between GPT-3.5 and GPT-4 may be imperceptible. However, as tasks become more complex, the gap between the two becomes increasingly apparent. GPT-4 is more dependable, imaginative, and capable of handling “subtle scenarios” than its predecessor.

GPT-4 is also more adaptable in responding to the user’s request and adhering to a specific response style. The language model takes into account the context provided by the user. To accommodate this, the message limit for a text request has been raised to 25,000 words, allowing entire documents and extensive excerpts from works to be submitted to GPT-4.

It excels in difficult exams at an honors level

Indeed, GPT-4’s improved performance in handling complex tasks has resulted in impressive achievements. For instance, the language model was able to gain admission to Stanford University and win the biology Olympiad. In the American Bar exam, the neural network scored at a level comparable to the top 10% of human participants.

Results of various exams – Comparison between GPT-4 and GPT-3.5.

Performs much better with different language

According to OpenAI developers, ChatGPT performed well mainly in the English language, achieving a 70.1% accuracy score in the MMLU benchmark – a special test for language models consisting of 14,000 tasks across 57 subjects.

To evaluate GPT-4, developers translated the benchmark into 26 languages. The fourth-generation model outperformed its predecessor in “native” English in 24 cases. The OpenAI report shows that GPT-4’s performance in the Russian language is 82.7%, ranking eighth overall. GPT-4’s best result in English is 85.5%.

Does GPT-4 solve all ChatGPT problems?

Sam Altman, co-founder of OpenAI, acknowledges that the model is “imperfect and limited,” indicating that there is room for improvement. Upon initial usage, GPT-4 may seem impressive, but its limitations become apparent with active use. There are currently two main problems that have not been resolved.

The neural network still hallucinates

Language models can make errors known as “hallucinations” wherein they confidently generate false information and present it as factual. To mitigate this issue, OpenAI recommends performing additional verification on all responses.

Despite this concern, developers maintain that GPT-4 performs 40% better than its predecessor on the internal “fact-checking scale.” The neural network has made tremendous advancements in history and the exact sciences. However, it is still prone to errors when generating code and working with code in general.

Results of GPT-4 in solving tasks on the Leetcode platform.

The neural network still does not operate on up-to-date data

ChatGPT’s training dataset was restricted to the fall of 2021, which implies that GPT-3.5 was unaware of any news or events that occurred beyond that timeframe and could not integrate fresh information into its responses. This situation remains unchanged with GPT-4.

Where to try GPT-4

ChatAI recently announced the launch of GPT-4 in various services alongside the technology’s presentation. Notably, even Icelandic authorities employ the model to preserve their language, according to developers. The following services claim to support GPT-4:

  • ChatGPT – while the free version still operates on GPT-3.5, the paid Plus subscription now features GPT-4 with certain limitations. Currently, users can make no more than 100 requests within a four-hour period.
  • Bing – OpenAI and Microsoft have a close partnership, with Microsoft providing a supercomputer for GPT’s development. OpenAI’s technology has recently been incorporated into the Bing search engine, which now operates on GPT-4. Bing is the sole means to experience the neural network that has access to up-to-date global data.
  • Duolingo, the renowned language learning app, has introduced a new Duolingo Max subscription that includes two features. The first, Explain my Answer, assists in examining errors with the aid of a chatbot. The second feature, Role Play mode, allows users to interact with an AI interlocutor. Currently, this subscription is only available for Spanish and French learners.
For Duolingo with AI, they even changed the familiar appearance of the owl.
  • Poe is an app developed by the creators of the popular English-language Q&A platform Quora. The app features chatbots, including ChatGPT, which users can communicate with and ask questions. The paid subscription of the service now includes GPT-4.

We will continue to update the aforementioned list as new announcements are made. Additionally, developers can sign up for the waiting list to access the GPT-4 API and accompanying tools.

Responses