After months of waiting, it’s here: the new version of the OpenAI language model has arrived and promises to far surpass its predecessor, GPT-3.5.
Today, we are going to explain what exactly GPT-4 is, how it works and how it differs from ChatGPT.
What is GPT-4?
Although chatbots are not a new invention (the first, ELIZA, dates back to the 1960s), their ability to interact and establish a natural conversation with other people is surprising. In this case, GPT is the system in charge of making this computer-human interaction a reality. Developed by OpenAI, GPT (Generative Pre-trained Transformer) is a linguistic model that works with Artificial Intelligence, capable of reasoning and reproducing natural language. On the other hand, a model is a program that tries to replicate the behavior of a system: you introduce a series of data into the model and it allows them to be organized in a new context.
To make a simile, it is as if we were in front of a very powerful predictive keyboard. GPT is trained with large amounts of data and this helps it to predict which words are most likely to appear according to the request. For example, if we ask “What is your name?”, the most likely answer would be “My name is ChatGPT“.
Being a linguistic model, it can recognize and identify our language. In this case, GPT is also generative, so it is able to interpret the text we introduce and create a new one.
So what is GPT-4? Well, it is the latest version (as of today) of GPT. Released on March 14, OpenAI claims that it is a much more creative model and offers greater accuracy in problem solving.
Differences between ChatGPT and GPT-4
If both terms confuse you, let us explain it in a few words: basically, ChatGPT is based on GPT. Until today, the version used was GPT-3.5, the third generation of the OpenAI model. ChatGPT is just the front end: through the ChatGPT website, we enter our questions and requests, while GPT gives us the results.
In addition, ChatGPT is configured to engage in conversations. When responding to our requests, it will do it in a human way, as if we were talking to another person. This is possible because ChatGPT is able to interpret natural language.
How does GPT-4 work?
To understand how GPT-4 works, we must first explain how GPT does it. The famous OpenAI model works on the basis of neural networks (networks designed to learn tasks). For everything that we do not know how to communicate to a computer (which is programming) we use machine learning and neural networks are used for this. Basically, we show the program examples (in this case, texts) and then, we tell the neural network the points to take into account from this data. After that, the neural network finds patterns, learns from them and draws its own conclusions.
GPT makes use of Transformer technology, a neural network architecture especially focused on language understanding (in fact, it is even part of its name). However, computers are still unable to understand the context of words, as they only understand numbers.
With Transformer technology, each word is converted into a numeric vector: GPT analyzes these numbers and classifies them according to patterns. Each pattern is assigned a token and these, in turn, are marked (the token for “sky” is usually associated with “blue”, so GPT marks them to know that they go together). As a result, each token has hundreds of markers.
The truth is that it is quite a complex process and at first, it may seem difficult to understand, so we will summarize it briefly: GPT converts words into numbers and organizes them according to patterns to “understand” how they are placed.
How has GPT-4 improved?
Now that we know how it works, what does GPT-4 bring to the table? According to OpenAI, GPT-4 makes better use of the data provided to train the model. As a result, it is able to “understand” the context of conversations more deeply and is also much more creative in generating responses. Its lexicon is also more elaborate and it performs well in all types of situations (whether colloquial or formal conversations). It is also capable of generating many more words than the ChatGPT we currently have (25,000 vs. 3,000).
Its great comprehension capacity makes it offer more accurate responses, something in which it improves by far over its predecessor. GPT-4 has also acquired a new ability: the ability to respond to images. Greg Brockman, President and co-founder of OpenAI, demonstrated how GPT-4 could describe in detail a photograph from the Hubble Space Telescope. Not only that, it can also answer questions about images. Unfortunately, these features have not yet been made public.
Can GPT-4 be used now?
Although GPT-4 was introduced on March 14, there are already several applications that are making use of this new version of GPT. One of them is Bing. Microsoft has already explained that its search engine has been using GPT-4 for some time, so if you want to try it for free, just open Bing Chat (or use the Bing app for mobile devices).
Another way to use GPT-4 is by subscribing to ChatGPT Plus, the paid version of ChatGPT offered by OpenAI. For the time being, the free version that we can access through the web continues to use GPT-3.5 and is expected to continue to do so for some time. On the other hand, those who want to use the GPT-4 API to develop applications and services should join OpenAI’s waiting list.