1. Home
2. Blog
3. How do Large Language Models (LLMs) actually work?

- Tuesday, December 02, 2025

Currently, in the "AI Era," many of us are experiencing a huge shift in terms of how we work and interact with technology. While many of us are still busy learning how to use an AI tool, for example, ChatGPT or Gemini, very few of us stop to ask: How do these things actually work?

Understanding the inner details of Large Language Models isn't reserved for data scientists. Knowing what happens under the hood, from terms such as Tokenization to Transformers and further to Vector Embeddings, can let you understand better the capabilities and limits of the tools you use every day.

Drawing on the insights, here is a breakdown of the magic behind the machines.

1. The Hierarchy: Where do LLMs fit?

To understand LLMs, we have to first zoom out. You can think of the technology as a set of Russian nesting dolls:

1. Artificial Intelligence: The broad umbrella of creating smart machines.
2. Machine Learning (ML): A subset of AI wherein machines learn from data.
3. Deep Learning-DL: Specialized subset of ML inspired by the human brain neural networks.
4. LLMs stand for Large Language Models: These are specific models within Deep Learning which comprehend and generate human language.

Models vs. Applications

It is important to differentiate between the Model and the Application.

• Application: ChatGPT. This is the interface you talk to.
• The Model: GPT stands for Generative Pre-trained Transformer. This is what drives the application.

Other examples include Gemini from Google (both an app and a model), Claude from Anthropic, and Llama from Meta.

2. Can I Build My Own LLM?

Technically, yes. Practically? Probably not.

With few resources, building a foundational model requires three things that are hard to come by for an individual:

• Massive Data: To train the model, you need terabytes of text data.
• Computing Power: Training these models requires thousands of high-end GPUs. Fun fact: When OpenAI launched image generation, their server load was so high it allegedly almost melted their GPUs!
• Electricity: Energy consumption for training runs is immense.

This is why major LLMs are owned by technology giants like Microsoft, Google, and Meta; they are the only ones with the infrastructure that supports them.

3. Decoding "GPT"

Let's break down the most famous acronym in AI: GPT.

• Generative: This means it creates new data. Unlike a search engine, which indexes pages already in existence and shows you a link, an LLM creates a fresh response word-by-word.
• Pre-trained: This means that it has already learned from a huge dataset. It bases all its answers on that prior knowledge.
• Transformer: The magic is here.

The Transformer Architecture

First developed at Google-also well-known for the paper "Attention Is All You Need", the Transformer is the architecture that really made modern AI possible. It processes input data to understand context and the relations between words.

4. How Computers "Read": Tokenization

If you type "Hi, how are you?" into ChatGPT, to the computer, it doesn't see English letters. All a computer knows is numbers.

Tokenization is the process that breaks down your text into smaller chunks called Tokens.

• The word "Apple" might be one token.
• A complex word could be divided into multiple tokens.
• Even spaces and punctuation become tokens.

Each token gets a unique number ID, such as "Hi" being 10, "How" being 20, and so on. These numbers, not the words themselves, are what get processed by the LLM.

5. Meaning Understanding: Vector Embeddings

But once the data is tokenized, how does the AI grasp that "King" is related to "Queen"?

This happens through Vector Embeddings. Imagine a huge 3D graph. Each word or token is plotted as a point in this space based on its meaning.

• In this 3D space, words of similar meanings appear close, such as "Cat," "Kitten," "Dog".
• Differently targeted words appear far apart.

The "King - Man + Woman = Queen" Logic: This model conceives of relationships through direction and distance: If you draw a line from Italy to Rome in this 3-D space, and then draw a parallel line starting from Germany, the model lands on Berlin. It doesn't "know" geography; it knows the mathematical relationship between those vectors is identical.

6. The "Attention" Mechanism

Finally, how does it write coherent sentences?

The Transformer uses an Attention mechanism that makes it possible for the model to look at a whole sentence and decide which words are most important to understand its context.

For instance, "Bank" in "River bank" means something different from "Bank account." The Attention mechanism takes into consideration the surrounding words, "River" versus "Account," to assign the correct meaning to "Bank."

This enables the LLM to predict the next word within a sequence quite accurately, so that the fluid human-like text we see today can be generated.

Final Thoughts

But basically, LLMs work by converting our language into numbers (tokenization), mapping those numbers in a space of meaning (vector embeddings), and using complex math - transformers/attention - to predict the best possible response.

While the engineering is complex, knowing these basics takes some of the mystery out and makes us better users of AI technology.