Read in Korean

🗝️
When reading lectures or articles related to GPT, it can often be confusing as the principles and usage methods are mixed together. This can create unnecessary difficulties for learners. To effectively utilize GPT, it is important to clearly distinguish and understand the principles from the usage methods.
In this blog, we will first summarize the principles behind how GPT generates responses. In the next blog, we will cover practical usage methods and effective prompt writing techniques.

⚙️ Exploring the Operating Principles of the GPT Engine – Understanding the Process of Response Generation
🎯 Utilizing GPT: Secrets to Maximizing Output with Effective Prompts – How to Write Effective GPT Prompts

The Generation Principle of GPT: How Does AI Create Answers?

GPT is not simply repeating pre-entered data; it is an AI that understands the flow of conversation and learns patterns to generate appropriate responses. Just like a skilled writer can write in various styles, GPT grasps the intent of questions and formulates answers in the most suitable manner. But how exactly does it operate? This article will explain the generation principles of GPT in an easy-to-understand way.

1 Probability Model for Predicting the Next Word

GPT operates by probabilistically calculating “What will the next word be?” For example, if a sentence starts with “I am today,” there are several possibilities for the next word, such as ‘feeling good,’ ‘having lunch,’ or ‘the weather is cloudy.’

The AI learns from vast amounts of data to calculate the probability of each word appearing and selects the most natural word to complete the sentence. Like rolling a die, there is no fixed answer, and it can generate different responses each time based on this probabilistic model.

This ability results from a three-step learning process that GPT undergoes. It improves its performance through pre-training, supervised learning, and reinforcement learning. In the pre-training phase, the AI absorbs a vast amount of publicly available text data from the internet, building a foundation for understanding language structure and context. This process enhances its ability to grasp the flow of words, sentences, and entire texts. Subsequently, through supervised and reinforcement learning, it strengthens its ability to follow user instructions accurately and generate useful responses.

2 In-Context Learning: AI Learns on the Spot!

GPT does not only generate answers based on trained data; it also has the ability to learn on the spot when provided with new contexts. This is called “In-Context Learning.” It is similar to a person with excellent reading comprehension who can understand a text they have never seen before by grasping its context.

For example, if asked to “classify the reviews of this product as positive or negative,” the AI can find patterns in existing data to provide an answer. However, if a few examples are added, the likelihood of generating the desired response increases significantly.

Example:

Review: “This product is excellent. The battery life is long.” → Positive

Review: “This product is too expensive. I do not recommend it.” → Negative

Input: “This product has a good design but disappointing performance.”

Output: “Neutral”

By adding a few examples like this, the AI learns and can generate more

refined answers. This is akin to a tutor showing a student several examples to teach them how to solve problems.

3 Learning Patterns, But Not Truly Understanding

GPT generates responses based on patterns rather than thinking like a human. For instance, when asked to solve a math problem, it can provide an answer using previously learned formulas and patterns, but it does not actually think logically. This is similar to how a calculator can quickly process complex calculations without understanding their meaning.

Therefore, it may have limitations in problems requiring complex reasoning or logical consistency. However, its ability to quickly analyze large amounts of data and suggest optimal answers is remarkable.

There is a way to further enhance GPT’s capabilities: “Chain-of-Thoughts (CoT) prompting.” CoT encourages the AI to show the process of solving a problem step by step. For example, in response to the question, “A is greater than B, and B is greater than C. Is A greater than C?” it would respond, “A > B, B > C, so A > C. Therefore, A is greater than C.”

Recent models can perform this reasoning process internally without directly showing it to the user. When a question is posed, a message saying “Thinking…” indicates that the AI is generating an answer using the CoT method.

4 The More Examples, the More Accurate the AI Becomes

GPT generates more accurate answers when provided with more clues. This is similar to saying that the AI is “quick-witted.” Even with just a few examples, the AI can adjust the style and content of its responses accordingly. It is like a seasoned actor who perfectly understands and portrays a character based on small hints from the director.

For example, to get GPT to write in a specific style, providing a few examples of the desired style is effective.

Example:

Example 1: “This product offers great value for money and is recommended for beginners.”

Example 2: “It features a sleek design and reasonable price.”

Input: “Describe the features of this laptop.”

Output: “It boasts powerful performance and lightweight portability.”

By providing many examples like this, the AI is more likely to generate responses in the desired direction.

5 Need the Latest Information? Utilize Retrieval-Augmented Generation (RAG)!

GPT may struggle to provide accurate answers regarding the latest information or specific internal company data not included in its training data. To overcome this limitation, the “Retrieval-Augmented Generation (RAG)” technology can be utilized.

RAG involves searching for relevant information using search engines like Google or Naver based on the user’s question and then providing this information to GPT to generate an answer. This is akin to a skilled journalist researching and analyzing various materials to write an article.

For example, if asked, “What indoor experience spaces in Seoul are suitable to visit with my child this weekend?” RAG would first retrieve relevant search results and then allow GPT to generate an answer based on this information.

Summary of GPT’s Generation Principles:

Probabilistic word prediction model → Calculates the probability of the next word and selects the most appropriate one.

In-Context Learning → Generates answers using existing trained knowledge plus on-the-spot learning ability.

Pattern-based understanding → Learns patterns based on data but does not think logically.

Sample reflection ability → The more examples provided, the higher the accuracy of the answers.

Retrieval-Augmented Generation (RAG) → Utilizes the latest information and external knowledge.

Understanding the principles of GPT allows for more effective utilization. To obtain desired answers, it is important not only to ask questions but also to provide appropriate examples and context. Additionally, when the latest information is needed, utilizing technologies like RAG is a good approach.