The new GPT-4 AI model from OpenAI just made its big debut and is already being used to power things like a virtual volunteer for the blind and an improved language learning bot in Duolingo. But what’s different about GPT-4 from older versions like ChatGPT and GPT-3.5? Here are the most important ways that these two well-known systems are different.
ChatGPT was first called GPT-3.5, which means that it is a little bit more advanced than GPT-3. It’s not a version of OpenAI’s large language model, though. Instead, it’s a chat-based interface for whatever model it uses. The ChatGPT system, which has become very popular in the last few months, used to only work with GPT-3.5, but now it can also be used with GPT-4.
Table of Contents
What is Chat GPT 4?
GPT-4 is a “multimodal” tool, which means that it can make content in response to both image and text commands. With this new version of Chat GPT, it’s easy to work together and come up with new ideas. It can also create, edit, and improve writing projects for both creative and technical reasons, like writing screenplays and songs or learning a user’s writing style. It means that a text generator with good analytical skills and the ability to form opinions can act as a legal advisor one minute, a chemistry expert the next, and so on.
Features of GPT-4
The new version did better than the old one on both the U.S. bar exam and the Graduate Record Examination (GRE). The president of OpenAI, Greg Brockman, showed how GPT-4 can also be used to help people figure out their taxes. The demo showed that it could take a photo of a hand-drawn mock-up of a simple website and turn it into a real website. People who are blind or have low vision can use the Be My Eyes app. The app will use GPT-4 to power a tool for virtual volunteers.
- The most advanced system in OpenAI is GPT-4. It gives answers that are safer and more in-depth.
- Based on how GPT-4 is used in the real world, it will be checked and updated often.
- GPT-4 sped up safety work because it was smart and could do what it was told.
- GPT-4 helps make training data for classifier iterations, which are used to train, evaluate, and keep an eye on things.
- GPT-4 was trained on AI supercomputers from Microsoft Azure, and its infrastructure lets people from all over the world use it.
- With the help of more than 50 experts, OpenAI’s latest step to scale up deep learning, including AI safety and security, was GPT-4.
- Based on our own tests, GPT-4 is 40% more likely than GPT-3.5 to answer with facts and 82% less likely to answer with information that isn’t allowed.
- GPT-4 can solve hard problems because it has more general knowledge and more ways to solve problems.
What Are the Differences Between ChatGPT and ChatGPT 4?
GPT-3.5 can only be told what to do by text, but the most recent version of the large language model can also be told what to do by an image. This lets it see things in a picture and figure out what they are. GPT-3.5 answers can’t be longer than about 3,000 words, but GPT-4 answers can be more than 25,000 words long.
GPT-4 is 82% less likely than its predecessor to respond to requests for content that isn’t allowed, and it scores 40% higher on some tests of accuracy. It will also let people who make AI decide how it sounds and how much it talks. One big difference between Chat GPT-4 and Chat GPT-3 is that Chat GPT-4 can understand pictures.
This is because Chat GPT-4 is multimodal, which means that it can understand information in many different ways, from words to pictures. On the other hand, Chat GPT-3 could only receive and send text, which limited how it could be used. GPT-4 is on ChatGPT Plus, and programmers can use it as an application programming interface (API) to make software and services.
Even though GPT-4 can accept both text and images, ChatGPT Plus members and software developers will only be able to use the text-input feature. The image-input feature will have a list of people who want to use it. The default model for your account is GPT-3.5. For $20 per month, you can change it to GPT-4.
And a user who chooses the subscription plan will get answers faster, get new features first, and get updates more often. Sam Altman, the CEO of OpenAI, said that GPT-4 was “most capable and aligned” with human values and goals, but that “it is still broken.” GPT-4 stopped collecting most of its data in September 2021, so it doesn’t know much about what happened after that.
Also, it doesn’t get smarter as time goes on. Open AI says that Chat GPT-4 can process up to 25,000 words at once, which is eight times more than Chat GPT-3. This will make it better able to handle larger documents, which could make it much more useful at work in some situations.
Chat GPT-4 is 82% less likely than Chat GPT-3 to respond to requests for content that isn’t allowed, and it is 40% more likely to give answers that are based on facts. On the other hand, OpenAI users have given GPT-4 a huge number of malicious prompts over the last year or so. Keeping these things in mind, the new model is much better than its predecessors in terms of being true to facts, being easy to steer, and not going outside of guardrails.
What Are the Limitations of ChatGPT 4?
OpenAI says that GPT-4 has the same problems as its previous versions and is less capable than humans in many real-world situations. Hallucinations, which are wrong answers, have been a problem for GPT-4 and many other AI programs. OpenAI said that when GPT-4 works with a human editor, it can compete with human propagandists in many ways.
It gave an example of how GPT-4 came up with suggestions that seemed reasonable when asked how to get two people to disagree. Sam Altman, the CEO of OpenAI, said that GPT-4 was most capable and aligned with human values and goals, but “it is still broken.” In September 2021, GPT-4 stopped collecting most of its data, so it doesn’t know much about what happened after that. Also, it doesn’t get smarter as time goes on.
OpenAI says that GPT-3.5, which ChatGPT used, was a “test run” of a new architecture for training. They learned from that and used what they learned to make the new version, which was “unprecedentedly stable.” Also, they knew more about what it could do, so there were less surprises. GPT-4 can handle both text and image inputs, but ChatGPT Plus users and software developers will only be able to use the text-input feature.
The public has not yet been able to use the image-input feature. The plan to subscribe came out in February, and each month it costs $20. It lets users get answers faster and see new features and improvements before anyone else. GPT-4 is what makes Microsoft’s Bing AI chatbot and some of the paid features of Duolingo work.