What sets ChatGPT apart is that it is the first widely adopted natural language conversational AI model. With the ability to start conversations, answer questions, and compose text, poetry, and code, ChatGPT forces people and businesses to rethink the possibilities of AI.
The company (OpenAI) launched a new version of its large language model called (GPT-4) last Tuesday, which will work with (ChatGPT) bots, but for now, only the paid version (ChatGPT Plus) will work in products such as: the latest version of Bing Chat and language-learning app Duolingo, as well as an in-app virtual volunteer (Be My Eyes) to derive and analyze images for the visually impaired and accurately describe them to users.
But what is the difference between the model (GPT-4) and the previous model (eg: GPT-3.5)? Here are the five main differences between these models:
1- A GPT-4 model can see and understand images:
The biggest change in the GPT-4 model from previous versions is that it is a multimedia LLM, which means that it can understand more than one type of information entered into it.
The previous model (GPT-3) where bots operate (ChatGPT) was limited to text as it could only accept text input and generate a text response, but in the (GPT-4) model you can enter images as well as manipulate text. Forms also process them to find the information they contain.
You can just ask him to describe what is in the image and it is integrated with the application (Be My Eyes) - an application for the blind and visually impaired - which contains a tool called (Virtual Volunteer) for you to use as a description of what your phone sees when you play the image, provide text describing the image, And answer any questions about the image, and provide immediate assistance for a variety of tasks.
The GPT-4 model of the Be My Eyes video is able to describe patterns on clothing, identify plants, explain how to use machines in the gym, translate food labels, give recipes that can be processed, read maps and more photo-based tasks.
What sets Virtual Volunteer apart from other image-to-text technologies available today is that it preserves context and makes suggestions. For example, if the user submits a picture of an item in the fridge, the virtual volunteer can not only correctly identify the content, but also suggest what to do with those ingredients, because the tool can provide several recipes that can be made with those ingredients. ingredients, and send step-by-step instructions on how to make it.
But as of now, GPT-4 is still limited to text responses and cannot generate images themselves.
2- GPT-4 is hard to cheat:
Chatbots are often prone to misinformation and inaccuracies, despite the accuracy of everything these bots do these days. Not everyone who talks to her confidently can convince her to explain what an evil AI can do, play the role of an evil chatbot, or implement some little fantasy that makes the model say all sorts of weird things.
Several people have already exploited (ChatGPT) to create malicious content, in a process known as (AI-supported chatbot jailbreaking) using a variety of methods, the most prominent of which is role-playing.
So the company (OpenAI) trained the model (GPT-4) on a large number of malicious data that users themselves had sent to the bot recently, so that the new model had realistic response, portability, and rejection. to leave the protective barrier.
OpenAI claims that GPT-4 is 40% better at providing real-world answers than its predecessor (GPT-3.5) and believes that it is more advanced than its predecessor.
The company also confirmed that the model it developed (GPT-4), compared to the previous version, was able to reject 82% of attempts and questions from users seeking answers on serious topics, such as: b. Making harmful chemicals or trying to harm yourself. from GPT-3.5.
However, the company admits that it is still not immune to hallucinations, the tendency of artificial intelligence to generate wrong answers or logical errors. As usual, OpenAI co-founder and CEO Sam Altman acknowledges objections to minimizing bot errors: “It's the most compatible and robust model yet, but it still has flaws and limitations that make it look more impressive on first use than on time.” After. that."
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.
3- GPT-4 has more memory:
Large language paradigms iterate over millions of pages of web pages, books, and other text data, but when they have actual conversations with users, they are limited in what conversations they can hold to maintain context. In the GPT 3.5 model that ChatGPT is running on, this limit is only about 3000 words; It's about three or four pages of a book, so it's often not very effective for long conversations.
But the model (GPT-4) can handle more than 25,000 words, about 8 times more than the previous version, equivalent to about 50 pages of text, which is enough to run a full screenplay or short story.
This means that it can go up to 50 pages in a chat or create a text, so it remembers what you talked about on page 20 of the chat, or it can refer to events that happened there, there are 35 pages. So it can be used for long content creation, extended conversations, research, and document analysis.
4- GPT Form 4 More Languages:
English speakers dominate the world of artificial intelligence, from data and tests to research papers that use the language. But of course, the power of language paradigms applies to, and must be available in, any written language.
Open AI has achieved this feat in its GPT-4 model, demonstrating its ability to answer thousands of multiple choice questions with high accuracy in 26 languages, from Italian to Arabic to Turkish, and has been trained in 5 local Indian languages.
This pre-test for language proficiency is promising but falls short of fully supporting multilingualism; The exam standards are translated from English to make the form understandable and the multiple choice questions are not as good as the normal language. But this step is important in GPT-4 because it can be more friendly to non-native speakers.
5- GPT-4 has a different personality:
The concept of (controllability) is one of the most interesting things in artificial intelligence because it refers to the ability of the model to change its behavior when needed, and this function is present in the model (GPT-3.5) and it was used to run the ChatGPT program to persuade to act as an empathetic listener and those who use it to persuade him Playing as an evil chatbot.
The company developed this feature (GPT-4) to make it more useful. This would allow the models to perform different roles to achieve harmful results, despite the limitations companies have put into the system.
The company gave an example of this function, which consists of defining a set of instructions and instructions for the new model, allowing the new model to become a teacher for the students, and letting these instructions explain the teaching method and the simplified practical information that the model does. It should follow so that it can be presented simply and clearly.
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.