In case you missed GPT-3
Warning: this article is reserved for people who have no knowledge in data or artificial intelligence. I make some approximations in order to simplify things.
You woke up like every other morning and pulled out your phone to go on Twitter. And then, the tsunami. Thousands of tweets talking about GPT-3, OpenAI or ChatGPT. You go back to sleep.
But when you wake up again, you wonder why so many people are talking about this and what is GPT-3?
Don’t worry, I’ll explain everything to you.
Generative AI
Before getting into the heart of the matter, I need to explain what generative AI is.
Generative AI refers to a class of AI systems that are able to generate new data, such as text, images, or audio, based on a set of input data and a set of rules or objectives.
To summarize it simply, we give an algorithm a set of data that it will inspect to understand the rules and the functioning in order to generate a new output (for example a sentence or an image)
Let’s talk about GPT-3
GPT-3 is a computer program developed by OpenAI that can understand and generate human-like text. It can read and write in many different languages, and it can even write stories or answer questions that people ask it. It’s kind of like a really smart robot that can talk and write just like a person. But instead of a body, it lives in a computer and communicates through the internet.
You are probably thinking that I have summarized GPT-3 very well because I am a genius writer but you would be wrong. This definition was written by GPT-3 himself. But how? Let’s ask GPT-3 to give us more details (I will complete the results)
GPT-3 (short for “Generative Pre-trained Transformer 3”) is a language processing artificial intelligence system developed by OpenAI. It is one of the largest and most powerful language models currently available, with 175 billion parameters.
GPT-3 is what we call an LLM (Large Language Models). They are called “Large” because they have an extraordinary number of parameters. There are several types of large language models of which the two best known are the transformers (GPT-3 or BERT) and the Reccurent Neural Networks (RNN, such as LSTM or GRU).
These models are trained on huge amounts of data from all sorts of places: articles, books, Internet, etc…
A parameter is a value that determines the behavior or output of an algorithm or model. It is used to fine-tune the performance of the algorithm or model, and can be adjusted to optimize the results for a specific task or dataset.
To better understand what it is, we will look at what the parameters are in a linear regression model. Look at the picture below. You can see a scatter plot (in blue) and an affine function (in red). This red curve represents our linear regression.
Our goal with this red curve is to be able to approximate in the best way where a new point would be located if its value was in x=1 or x=2.
What can we change in our red function so that it is the most faithful to reality? An affine function is written as follows: y = ax + b. We have here two parameters: a and b on which we can play.
Without going into too much detail, we will have to define a cost function which will allow us to know if we are close to reality or not.Now try to imagine 175 billion parameters in a single AI. Incredible no ?
GPT-3 is a type of machine learning model known as a transformer, which is trained to process and generate natural language text. It does this by learning to predict the next word in a sequence of words, based on the words that came before it. The more text it processes, the better it gets at predicting what word comes next, and the more human-like its output becomes.
Transformers are quite complicated to understand. If we had to summarize what it is, it is an algorithm that takes in data and generates an output. It is called a neural network because it is inspired by the functioning of neurons in the brain.
Through this step, the transformer will modify the parameters according to what it has learned. The more data it has seen, the more it will be able to refine its model (this is an approximation, in reality it is much more complicated)
If you want to know more, here is a fascinating article on the subject (a little more technical)
GPT-3 can be used for a wide range of language tasks, such as translation, summarization, question answering, and text generation. It can even be used to generate code and perform other tasks that require a high level of language understanding.
The question that you will naturally ask me now is, how can I test GPT-3? Well, in the simplest way with ChatGPT.
ChatGPT: discuss with an AI
ChatGPT is a variant of GPT-3 that is specifically designed for conversational language tasks, such as chatbot dialogue and question answering. It is trained on a large dataset of conversational text and has a deeper understanding of the structure and patterns of human conversation.
One key difference between GPT-3 and ChatGPT is the type of data they are trained on and the tasks they are optimized for. GPT-3 is trained on a broad range of data from the internet and can be fine-tuned for specific tasks, whereas ChatGPT is specifically designed for conversational language tasks and may not perform as well on other types of language tasks.
And the success of ChatGPT was immediately felt, especially on social networks such as Twitter. And when we look at the statistics, we realize the popularity of ChatGPT.
What application in the real world?
The applications of this kind of AI will change the world. Literally. It allows to simplify many long and difficult tasks and opens the way to new applications.
Here is what Sequoia Capital says in this article:
“Just as mobile unleashed new types of applications through new capabilities like GPS, cameras and on-the-go connectivity, we expect these large models to motivate a new wave of generative AI applications.”
I will give you several examples of uses of generative AI.
Allow developers to code faster
GPT-3 revolutionizes the world of development. No need to bother coding certain functions yourself when writing a sentence, the AI takes care of generating it for us.
The example below is basic but allows you to see the power of generative AI. Of course, it’s not perfect yet but it’s already exceptional.
Moreover, if we have an error in our code and we can’t find it, we can turn to this kind of AI to tell us what to do or even correct the error for us.
Create video content
It will soon be possible to create from scratch or from a video to be able to change what we see with a simple prompt.
An example is better than 1000 words. So I invite you to watch the video below (you’ll see that it still requires some technical knowledge but I have no doubt that it will soon be easier).
Besides, you will hear about DALL-E and I have to explain it to you.
DALL-E (pronounced “dolly”) is a deep learning model developed by OpenAI that is capable of generating images from text descriptions, using a neural network trained on a dataset of text-image pairs. The model is trained to generate images by processing a text prompt and generating an image that represents the content described in the text. For example, given the text prompt “A two-story pink house with a white fence and a red door”, the model might generate an image of a pink house with a white fence and a red door.
Write about a specific topic
What better than ChatGPT to answer this question itself?
The following paragraph is generated by ChatGPT
To use GPT-3 to write about a specific topic, you would typically provide the model with a prompt or set of prompts that outline the main points you want the text to cover. For example, you might provide a prompt that outlines the main points you want to make in an article about the benefits of renewable energy.
Once the prompt is provided, GPT-3 will generate text that is relevant to the topic and that follows the structure and style of the prompt. The generated text will be coherent, well-written, and free of errors, making it an effective tool for generating high-quality content quickly and efficiently.
Overall, GPT-3 is a powerful and useful tool for generating text on a wide range of topics. It can be used by writers, marketers, and other professionals to produce high-quality content quickly and efficiently, without the need for extensive research or writing time.
Provide access to data for all
Accessing data in the corporate world is often a headache. One needs to have a solid knowledge of data science in general but also of the structure of the data in the company.
I created Vazy a solution using GPT-3 which solves this problem. You just have to ask a question in any language, GPT-3 will translate it into SQL (the database language) and will display the data in a few seconds. No need to waste hours looking for the data you need.
Here is a video of what it looks like.
The product is still in development (GPT-3 is only a small part of the project, we also manage the display and the quality of the data and guide the user in its use) but if you are interested, join the waitlist.
GPT-4: better and better
You can imagine that OpenAI does not intend to stop at GPT-3 as we see the crazy potential of the model and the enthusiasm of the community. So they are working on GPT-4, a new evolution of the model that should bring down GPT-3.
According to many informants on Twitter, GPT-4 should be populated with more than 100 TRILLION parameters. 100 TRILLION.
To give you an idea of the difference with GPT-3, here is a picture that compares the size of the number of parameters.
So what will the future of generative AI be made of?
Conclusion
First of all I hope that this article has helped you to understand a bit more what generative AI, LLM, GPT-3 or ChatGPT were.
I would like to finish by warning you against two approaches that we see quite regularly expressed on social networks, especially Twitter.
GPT will soon be able to replace some jobs completely
I think this is the wrong approach with the current state of generative AI. It should be seen as a great help for humans in their daily tasks and I think we can hypothesize that a hybrid job will appear in the coming years. The human does not work alone, nor the machine, but the knowledge mix of both. On a level even more important than what we experience with the Internet.
GPT is an AGI
Artificial general intelligence (AGI) is a type of artificial intelligence that is capable of understanding or learning any intellectual task that a human being can.
AGI is the Holy Grail that every AI researcher dreams of achieving.
However, there is a very simple way to understand that GPT is still very very far from becoming an AGI. If it had been the case, GPT could have already discovered things in math, physics or computer science that no human had discovered before.
Nevertheless, GPT and the work of OpenAI is bringing us closer and closer to this AGI. The field of generative AI exceeds Moore’s Law predictions and I personally believe that by 2030 we will reach AGI.
And our world will change forever. The greatest upheavals ever encountered by humans will occur and a profound change in society will take place.
I can’t wait to see this.
If you liked the article and you want to please me, don’t hesitate to go on vazydata.com and join the waitlist ;)
