4 Crucial Things to Know about GPT-4
You should know these to use GPT-4
GPT-4 has finally been released. This model is one of OpenAI’s biggest milestones in deep learning. The biggest difference of the GPT-4 from other state-of-the-art models is that it is multi-model. This means that while ChatGPT only accepts text data, GPT-4 now accepts text and images.
To better understand this, let’s look at the example Open AI showed in the GPT-4 demo video.
The above image shows how to create a website using a visual. First, a mock-up was drawn in step 1. Next, a photo of this mockup was taken in step 2. This image was then sent to GPT-4 in step 3. The model gave the necessary codes. Finally, a colorful website was built with these codes in step 4. As you can see, even if you don’t know how to code, you can use GPT-4 to convert a paper mockup to a website.
Like ChatGPT, you can also give text as input to GPT-4. Let’s say, you want to make a game. But you don’t want to deal with writing code. Just tell GPT-4 which game you want.
Bammmm! The model gives you the codes of the game, you can make your game using these codes. Let me show you another example. Tell GPT-4 to make a snake game using Javascript and the necessary HTML and CSS codes.
GPT-4 gives you the necessary codes. If you run these codes in your browser, you can make a snake game in this way. Of course you can do much more with GPT-4. Let’s take a look at what we’ll cover in this blog.
- What exactly is GPT-4?
- GPT-4 vs GPT-3.5
- How to access GPT-4?
- GPT-4 API pricing
Let’s dive in!
What exactly is GPT-4?
You’ve seen how wonderfully GPT-4 overcomes human-level tasks. But, what is GPT-4 under the hood? GPT-4 is based on Transformers architecture developed by the Google brain team in 2017.
Transformers consists of encoder and decoder. This architecture was revolutionary, and worked great, especially in language-to-language translation tasks. That’s why Google translate does human-level translations.
There are two important models based on this architecture. One of them is BERT, and the other is GPT. BERT is mostly used for tasks like text classification, while GPT is mostly used for generating text. So, GPT-4 is based on Transformers architecture.
This model was trained with data on the Internet. You know it is not easy to deal with big data. To train the model, a supercomputer was designed within two years with Azure The GPT-3.5 model was trained last year to test this system. Think about it ChatGPT, which fascinated the whole world, was built on GPT-3.5. You can better understand the power of GPT-4.
Nice, we’ve seen the architecture behind GPT-4. To understand what you can do with GPT-4, let’s take a look at a few examples, Open AI shows on its webpage.
For example, when you give the model an image like the above, and you say define the picture panel by panel, the model understands what is in the image and explains it to you.
GPT-4 says it’s a three-panel Lightning Cable adapter package. And the model is describing the objects, in the panels one by one. Let’s take a look at this picture.
As you can see, a man is ironing in the back of the taxi. We understand that the situation in this picture is absurd. When we ask the GPT-4 what’s unusual in this picture, the model describes, what is happening in the picture.
The model first recognized the objects in the image like a human. AI models were already able to do object recognition. But, the model discovered something unusual in the picture. In other words, it noticed the picture almost like a human and makes inferences.
This is amazing. If someone had said 5 years ago that AI would recognize objects in pictures and make inferences no one would have believed him. But, now look at the level AI has come to, in such a short time.
The GPT family consists of several major language models that can generate realistic text. The previous version of GPT-4 was GPT-3.5. ChatGPT based on GPT-3.5. Let’s compare the GPT models, to better understand the GPT family.
GPT-4 vs GPT-3.5
The GPT family has accomplished the impossible. The two latest members of the GPT family are GPT-3.5 and GPT-4. Let’s compare them to understand better.
The first difference is that while GPT-3.5 only works with text data GPT-4 can also work with visual and video data apart from text data. This is the biggest difference between the two models.
As for the number of parameters, GPT-3 had 175 billion parameters. GPT-4 has 100 trillion parameters. So, this model is about 500 times larger than the GPT-3. What a large model, right?
Now, let’s compare the two models according to the conversation. In simple conversation, there is no difference between these two models, but for complex conversations, GPT-4 is excellent. GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.
So, this model works great with text data as well. Let’s move on to the exam comparison, which Open AI especially highlights.
The GPT-4 performs at the human level in academic, and professional examinations although it is less capable than a human, at solving real-world problems. If you have GPT-4 with you in an academic or professional exam you can easily score high. OpenAI researchers tested GPT-3.5 and GPT-4 on human exams. These exams consist of both multiple-choice and open-ended questions.
Here, you can see the exams and the success of the models in these exams. The chart is ranked from low to high by the performance of the GPT-3.5. Blues indicate the success of GPT-3.5 and greens indicate the success of GPT-4.
If you notice GPT-4 performs better in almost all exams. The GPT-4’s success in most exams exceeds 80 percent and in some, it is close to 100. These exams are not simple, they are academic and professional exams. Note that, these exams are the tests that most people fail.
So far many models have been developed that are trained on big data. Various benchmark tests are used to compare these models. These tests, for example, measure how the model predicts word after word. Generally, these tests are in English.
In this chart, you see the scores of models in various languages. The first row shows the percentage of finding the correct answer when a random answer is chosen. The probability of finding the correct answer in a 4-choice quiz is 25 percent, right?
For example, the test score of the PaLM model developed by Google is 69 percent for the English language, 70 percent for the GPT-3.5 model, and 85 percent for the GPT-4.
The GPT-4 outperformed other models in benchmark tests. When these tests were translated into other languages and then tested, GPT-4 is great again. Ok, the GPT-4 is an excellent model, but does it have any downsides, you might ask? Let’s examine the limitations of GPT-4.
Limitations
Despite the high capabilities of this model, it has some limitations like other members of the GPT family. OpenAI researchers compared the accuracy rates of the GPT family.
There are areas on the horizontal axis and the accuracy rates of the models on the vertical axis. The closer this ratio is to 1 the more ideal humanoid responses are obtained. Look at this chart the green color indicates GPT-4.
In the tests, the GPT-4 gives more accurate outputs than other models, but unfortunately, we see that it cannot give a hundred percent correct answer. So, GPT-4 needs a little more improvement to get exact results.
In this section, we compared the GPT models and saw how powerful the GPT-4 is. We also talked about some limitations of the model. Let’s see where you can use GPT-4 now.
How to Get Access to GPT-4?
As you know, OpenAI recently introduced a paid subscription to ChatGPT. You can subscribe to ChatGPT for $20 per month. You can use GPT-4 directly if you have a subscription.
If you don’t have a subscription, another way to use this model is to sign up for the waitlist and wait your turn. I don’t know when it will be your turn. Even if it does they’ll likely offer limited use. In other words, they direct you to a paid subscription. When it’s your turn the GPT-4 model will be opened to you on the OpenAI playground page.
Another way to access GPT-4 is to use Bing. As you know, Bing is a search engine developed by Microsoft. ChatGPT was used in this engine before. Yusuf Mehdi, corporate vice president at Microsoft, announced that they changed the engine behind Bing’s preview to GPT-4. So, you can access GPT-4 using Bing Chat. But, for that, you still have to sign up for the waiting list.
In this section, we covered how to access GPT-4. Let’s take a look at the pricing of the GPT-4 API.
GPT-4 API Pricing
Developers can use GPT models in their applications. So, you can connect GPT models to your own applications as APIs. I have good news for you. You can use the GPT-4 API in your applications immediately. The ChatGPT API was released a little late. OpenAI has made GPT-4’s API immediately available.
Now let’s take a look at the GPT-4 API pricing. Note that the GPT-4 API, was named gpt-4–0314. Pricing is based on tokens. Token means a string of characters. So, you can think of the token as the smallest piece of text. It can be a token a symbol or a word. You might think that 1000 tokens roughly equal about 750 words.
Here, you can see the prices of the models according to the token usage. For example, 2133 tokens equal about 1600 words. You can write a medium blog post with this word count. It costs $0.128 for completion and $0.064 for the prompt. The price for the GPT-3.5 is $0.004. You can see the prices of other models here. Note that these models are not as powerful as the GPT-4.
Let’s look at the highest number of tokens. This is equivalent to about 1 million words. You can write Harry Potter books with this word count. So, writing Harry Potter books with GPT-4 costs $43 by using the prompt.
GPT-4 for the prompt is 14 times more expensive than ChatGPT, and 29 times more expensive to complete. So, is it worth the difference? It depends on you. You saw the power of GPT-4 in the charts. This model gives more accurate and more nuanced outputs. The decision is yours. This is how the GPT-4 API charges.
Wrap-Up
GPT-4 is the latest state-of-the-art model developed by OpenAI. This model accepts both text and visual data. In this blog, we saw the power of GPT-4 and explored its limitations. Next, we learned how to access the model and finally, we talked about GPT API pricing.
That’s it. Thanks for reading. Let’s connect YouTube | Medium | Twitter | Instagram.
