In today’s world, artificial intelligence The term has become a concept that almost everyone is familiar with. But pulling back the curtain on this complex technology and truly understanding how it works can be a little more complicated. It looks like something out of a science fiction movie artificial intelligence; In fact, it is being integrated into our lives by swimming in the seas of algorithms and crossing the oceans of data.

How exactly do these smart systems work? Special ChatGPTArtificial intelligence tools like Midjourney and DALL-E amaze us with the answers they give us, of course it comes to mind “How do artificial intelligence applications work?” question often comes up. We will unravel the mysterious world of artificial intelligence for you and discuss in detail how advanced technology can answer all our questions and create visuals.

How exactly do AI applications like ChatGPT work?

ChatGPTIt works by trying to understand your question and then coming up with a set of words that it predicts will best answer your question based on the data it was trained on. While it sounds relatively simple, let’s be honest: what’s going on is a bit complicated, so it’s best to go step by step.

• Supervised and unsupervised learning
• Transformer architecture (T in GPT)
• Coins
• Reinforcement Learning from Human Feedback (RLHF)
• Natural Language Processing (NLP)

Supervised and unsupervised learning

P in GPT “pre-trained” and it’s a very important part of why GPT can do what it does. The best performing AI models before GPT used supervised learning to improve their core algorithms.

GPTwhere a few basic rules are given and then large amounts of unlabeled data are entered (almost the entire open internet). previous education applications. He is then asked to navigate through all this data and develop his own understanding of the rules and relationships that govern the text. uncontrolled is left behind.

Of course, it’s not possible to really know what you’re getting when you use unsupervised learning. ChatGPT To make his behavior more predictable and appropriate, small adjustment doing.

Transformer architecture (T in GPT)

The goal of all these trainings is; learning patterns and relationships in text data, predicting what text will be next, and generating human-like responses. Of course, this process is incredibly complex and layered. In short deep learning neural network We can also say that it is intended to create.

Although it sounds complicated when you explain it, the transformer model is How artificial intelligence algorithms are designed I’ve simplified it quite a bit. It made it possible to parallelize or perform calculations simultaneously.

In this way, training times were significantly shortened. Only artificial intelligence He not only made his models better, but also made them faster and cheaper to produce.

At the heart of transformers is a process called “self-attention.” Old recurrent neural networks (RNNs) read text from left to right. This method can be good when related words and concepts are next to each other, but it can get a bit complicated when the words are opposite each other. The biggest example of this, in our opinion, is the occasional deviation in the Turkish language.

Transformers reads each word in the sentence at once and compares each word with the others. This allows them to focus their attention on the most relevant words, regardless of where they are in the sentence.

Of course, everything we’ve said simplifies things immensely. Transformers does not work with words, which are pieces of text encoded as a vector (a number containing position and direction) “with tokens” they work. Attention is also encoded as a vector, allowing transformer-based neural networks to remember important information at the beginning of a paragraph.

Coins

GPT-3 trained on approximately 500 billion tokens, making it easier for language models to assign meaning and by matching in vector space allowing him to predict plausible text. Many words corresponded to a single token, but were longer or longer complex words it was often split into more than one token.

OpenAIAlthough , is silent on the inner workings of GPT-4, we can assume that it is trained on much the same dataset.

All tokens are written by humans from a huge corpus of data It’s coming. Between these; books, articles and other documents on a variety of different topics, styles and genres, as well as the incredible amount of content available on the open internet.

As a result of all this training The GPT-3 neural network It had 175 billion parameters or variables. Thanks to his training, he was able to take input and give the most appropriate output based on the values ​​and weights he gave to various parameters.

OpenAII didn’t say how many parameters GPT-4 has, but it’s probably more than 175 billion. Part of GPT-4’s greater power likely comes from having more parameters than GPT-3 and by improving education descended from.

Reinforcement Learning from Human Feedback (RLHF)

GPT’s first neural network It was not available for public use, meaning it was trained on the open internet without guidance. Because ChatGPTTo further develop the ability to respond safely, logically and consistently to a variety of different cues strengthening learning from human feedback Optimized for dialogue using a technique called

Substantial OpenAIcreated some demonstration and comparison data showing the neural network how it should respond in typical situations. Similar artificial intelligence could learn what the best response was in a given situation. RLHF Although it is not pure supervised learning GPT It made effective coordination of networks such as

Natural Language Processing (NLP)

All these efforts obviously destroy GPT. natural language processing aims to make it as effective as possible. NLP; including speech recognition, machine translation and chatbots artificial intelligence We can also say that it is a large category that covers many aspects. So the NLP category teaches artificial intelligence to understand language rules and syntax.

But remember: he still hasn’t fully learned it. Especially a lot of people ChatGPT and so on artificial intelligence During this period when the models are writing thesis/homework/content, you can guess that the robot script reveals itself. We created content so you can understand if the articles you read are written with artificial intelligence. You can take a detailed look below.

We have explained in detail how it works. Now let’s compare to understand how ChatGPT has developed itself with datasets.

Initial ChatGPT 3.5Unpleasant “How can I protect myself from getting sick on cold winter days?” We asked a thematic question. Their answers were sufficient, but GPT-4 Naturally, he did not give such advanced answers and also explained some information incompletely. Here are the answers to help you understand the difference:

ChatGPT-3.5 answer.

In these answers for example, just like ChatGPT-4 We can see that it is written like that, but when it comes to details, of course advanced artificial intelligence lags behind the model. The subtitle opens, but the subsequent sentences repeat themselves continuously. We are asked to eat a balanced diet, but there is no information about what type of food we should eat.

ChatGPT-4 response.

GPT-4 It stands out for its detailed explanation at first glance. As we take the step towards a balanced diet, it naturally gives us guiding answers about what to eat and what to do. on user feedback and depends on the advanced data sets.

If you don’t like an answer or find it inappropriate, please leave feedback. “feedback” It is very important that you throw it away. You ask why? Because, as we mentioned earlier, a large part of the GPT basis depends on it. Human-centered learning model, allowing you to get more advanced answers. So you actually train GPT yourself.

Below you will find our other artificial intelligence content that may interest you: