The company (OpenAI) announced yesterday its fourth-generation giant language (GPT) model, called (GPT-4), which promises a major breakthrough in the ability to generate artificial intelligence, as it is a multimedia language model (LLM) that can accept the input of images and text and produce output. Text.
The latest version of GPT-4 marks a major milestone in the field of artificial intelligence, especially in the field of natural language processing. Therefore, today in this article we will discuss the new features of the GPT-4 model and how is it different from previous versions?
First of all, what is (GPT-4)?
(GPT-4) is the latest and most advanced version of the text generation language (GPT) model introduced by (OpenAI) in 2018, then the third generation (GPT-3) in 2020 and then up to (GPT - 3.5) released. ) which I used to create a ChatGPT chatbot that was launched last November.
If you use ChatGPT, you know the previous version (GPT-3.5) was limited to text entry and creation of natural language scripts or plain scripts. But (GPT-4) is the latest version, which excels in multimedia, since it accepts both image and text input and produces text output, it is able to understand and analyze image content well, even complex images, for example: charts or images for spreadsheets And long documents, this is a huge improvement over the previous model.
The features of the new model greatly improve the capabilities of the chatbot, and the GPT-4 model can perform multiple tasks simultaneously, increasing the probability of authoritative responses by 40%, making it a more useful tool for trustworthy information in applications such as search engines.
OpenAI said it took 6 months to develop this version of the language model based on its experience with ChatGPT. The new model has proven its excellent performance. Insist not only on user needs, but also on providing facts and correct information. . Trying to take something out of context or provide an answer that goes against company standards.
The company said that with the maximum complexity of the task, it is possible to distinguish between GPT-3.5 and GPT-4, because the (GPT-4) model is more reliable, innovative, and able to handle higher levels. Instructions compared to the accuracy of the previous generation.
second; Can you try GPT-4 now?
You can try this mode while using ChatGPT, but it is currently only available to users of the paid version (ChatGPT Plus), which has a $20 monthly subscription. Please note that it is not yet available in many countries. Microsoft also confirmed that the latest version of Bing Chat is based on the (GPT-4) model in its operation.
OpenAI also announced that several companies are already using GPT-4 in their products, including: language learning app Duolingo, Khan Academy for online education, Stripe, etc.
Developers can access GPT-4 models through an API, and the company has published a queue for instant subscriptions.
third; What new features does GPT-4 offer?
1- Performance improvements:
The model (GPT-4) can handle more than 25,000 words, about eight times more than the previous version (GPT-3.5), since it only handles 3,000 words, which makes it useful for creating long content, expanding dialogs, and searching for files. and analyzed.
The GPT-4 model is designed to improve targeting, the ability to track user intent, make it more visible, and produce results that are less offensive or dangerous.
(GPT-4) has also improved performance compared to the previous version (GPT-3.5) in terms of accuracy and honesty of answers as the latest version has fewer factual or logical errors and GPT-4 improves GPT-3.5 by 40% de facto measure of performance . for open artificial intelligence.
2- Understanding the image:
The new GPT-4 model can understand even the most complex images very well. The model can accurately describe images, for example, distinguish between a VGA cable and a smartphone connected to it, as shown in the image below:
To ensure that the image input capabilities are ready for the new model, OpenAI has partnered with the application for the blind (Be My Eyes), which in a new update will receive models of the capabilities from virtual volunteers who use the new capabilities of the model to infer and analyze images, describing users in detail and what they can achieve with the components those pictures.
This feature also allows users to enter text questions or commands with images, so the text commands will prompt the GPT-4 form to perform a task or answer a question related to the image's content.
Greg Brockman, Co-Founder and Head of OpenAI, outlined some use cases for the model (GPT-4) including: the ability to read images from hand-drawn models for websites and the ability to draw images from images for websites.
Chat GPT-4 will be able to convert a hand-drawn sketch into a fully-functioning website. The future is here! 🤯 pic.twitter.com/PYIY9BQaQq
— Howard Pinsky (@Pinsky) March 15, 2023
3- Improve Orientation:
One of the notable capabilities offered by the new version (GPT-4) is controllability, the ability to change its behavior based on user requests. For example, a user can direct a form and specify a list of instructions to follow in order to respond to a request and provide various information.
The company explains this function by defining a series of instructions and instructions for the new model as a teacher for the students, so these instructions explain the method of teaching and the process of simplifying the information that the model must follow to reach the students. - Awareness in a simple and clear way.
The company says the feature will also allow developers to control how their services work and provide intelligent interactions with users by defining checklists and personas for the models their services are based on and how they interact with them.
fourth; Is GPT-4 more secure than previous versions?
The researchers incorporated more human feedback, including from ChatGPT users, to improve GPT-4 performance, according to OpenAI, which relies on 50 human experts to provide AI security feedback.
The company confirmed that compared to the previous version of GPT-3.5, the new model it developed was able to reject 82% of user attempts and questions seeking answers on serious topics such as: B: Making harmful chemicals or trying to harm yourself.
However, he also acknowledges that despite its huge new features and performance improvements, the (GPT-4) model is not foolproof, and says the model still cannot provide any information about developments after September 2021 or about learning. User issues.
The company added that GPT-4 sometimes makes simple and naive errors that seem inconsistent with its performance efficiency, such as: b. Acceptance of obvious misrepresentations by users. Sometimes he solves difficult problems the way people fail, such as finding security flaws in the code he generates himself.
here is GPT-4, our most capable and aligned model yet. it is available today in our API (with a waitlist) and in ChatGPT+.https://t.co/2ZFC36xqAJ
— Sam Altman (@sama) March 14, 2023
it is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.