Divyanshu’s Blog

Understanding LLMs (The Brains Behind AI)

Welcome to my blog 👋

If you’ve ever wondered how ChatGPT seems to “understand” you, this post is for you.

Behind tools like ChatGPT, Gemini, Claude, and Copilot are Large Language Models (LLMs), the powerful AI systems that can understand and generate human language.

Let’s explore what makes them so capable, and how they’re evaluated to ensure they work reliably.

🏁 What Is an LLM?

A Large Language Model (LLM) is an AI system trained to understand and generate human language.

You’ve already interacted with many: ChatGPT, Gemini, Claude, Copilot. They’re all LLMs.

Think of an LLM as a “statistical brain” that predicts what word (or token) should come next in a sentence, based on everything it’s seen so far.

Input:   "Artificial intelligence is"

Model:   → predicts "changing"
          → then predicts "the"
          → then predicts "world"

Output:  "Artificial intelligence is changing the world."

🧩 The “Large” in LLMs

The “large” refers to:

  • The amount of data it’s trained on (hundreds of billions of words)
  • The number of parameters: Think of parameters as tiny “knobs” or “dials” that the model tunes during training to adjust how it understands language patterns.

GPT-4 and Gemini have hundreds of billions of parameters!

            -------------------------------------
Data  >>>>    billions of sentences, books, web  
            -------------------------------------
                       ↓  ↓  ↓  ↓  ↓
               billions of internal parameters

Each parameter adjusts slightly during training to help the model understand relationships between words, meaning, and context.

⚙️ How an LLM Works?

Let’s simplify the pipeline into 3 stages:

-------------------------
  Stage 1: Tokenization    
-------------------------

Input text is broken into smaller units (tokens).

"Hello world!" → ["Hello", " world", "!"]


-------------------------------------
  Stage 2: Encoding (Understanding)  
-------------------------------------

Each token becomes a vector (a list of numbers that captures meaning).

"Hello" → [0.1, 0.9, -0.3, 0.4, ...]

These vectors go into a neural network (Transformer) that learns relationships.

----------------------------------
  Stage 3: Decoding (Generation)  
----------------------------------

The model predicts the next token based on previous ones:

Input: "The sky is"
→ predicts "blue"
→ output: "The sky is blue"

That’s how every chatbot answer is built, one token at a time, extremely fast.

🧠 Why LLMs Are So Capable?

Because they’ve seen almost everything humans have written online and they’ve learned patterns of reasoning, style, and structure.

They don’t “think” like humans, they “guess” extremely well based on context. That’s why they can:

  • Write essays, code, and stories
  • Answer factual questions
  • Translate languages
  • Simulate conversation

⚖️ Why Evaluate LLMs?

LLMs are powerful, but they’re not perfect.

They can:

  • Make up facts (hallucinate)
  • Be biased
  • Miss key context
  • Give inconsistent answers

So, just like humans are tested in exams, LLMs must be tested before being trusted. That’s where evaluation techniques come in.

🔍 Evaluating LLMs Is Hard

Unlike a simple math test where answers are either right or wrong, LLMs produce subjective, creative, or contextual outputs.

Example:

Prompt: "Write a paragraph about the moon."

Output A:

The moon is Earth's only natural satellite, influencing tides...

Output B:

A glowing companion of the night sky, the moon inspires poets...

Which one is “better”? There’s no single correct answer. That’s why evaluation is hard.

So researchers use multiple approaches:

  • Human judgments (slow, expensive)
  • Automated metrics (fast, limited)
  • Hybrid or model-based evaluations (modern standard)

🧭 The Types of Evaluation

We can group LLM evaluation into three main families:

-----------------------------
  1️⃣ Human Evaluation    
     - Humans rate outputs   
     - Very reliable         
     - Slow & costly         
-----------------------------

-----------------------------
  2️⃣ Automated Metrics     
     - Scores via formulas   
     - Fast & reproducible    
     - May miss nuance       
-----------------------------

--------------------------------
  3️⃣ Model-based Evaluation 
     - LLM judges another LLM
     - Scalable & modern       
     - Needs careful prompt   
--------------------------------

✍️ Summary:

Large Language Models are statistical engines trained on vast text data to predict the next word.

They don’t “think”, they pattern-match extremely well.

Their creativity and inconsistency make evaluation challenging, leading to sophisticated testing techniques.

Key takeaway diagram:

Prompt → Tokenize → Encode → Generate → Evaluate → Improve

Thanks for reading till the end! 🙏

I hope this post helped you get a clearer picture of how LLMs actually work behind the scenes.

We’re just getting started! In the upcoming posts, we’ll explore how these models are judged and improved.

See you in the next post 👋

← Back to all posts