What Is an LLM? Large Language Models Explained Simply

By Ravindra Bagale, trainer · LinkedIn · Last updated: 19 October 2026

Generative AI course · Chapter 11 · Estimated reading time: 14 minutes

1. Suppose your phone keyboard got very, very smart

Suppose we are typing a WhatsApp message to a friend. We type I will reach in and the keyboard suggests 5, 10 and minutes. The keyboard has seen many messages, so it guesses what usually comes next.

Now imagine a keyboard that has read a huge part of the public internet, millions of books and articles, and a lot of code. It does not just guess one word. It guesses the next word, then the next, then the next, until it has written a full email, a poem or a Python program.

That is an LLM (Large Language Model):

An LLM is a computer program that has learned from a very large amount of text and writes by predicting the next piece of text, again and again.

Try it now. Type this in any AI tool:

Complete with one word only: I drink a cup of hot ___

Expected output (yours will be similar, not the same):

tea

The model picked the most likely word. In India "tea" or "chai" is very likely; "coffee" is also possible. Ask again with Give 5 possible words with how likely each one is, in simple words, and you will see the model thinks in possibilities, not in fixed facts.

2. What "large", "language" and "model" mean

An LLM does not look up a database of answers. It has learned patterns: how sentences are built, how an email usually starts, what facts often appear together. When you ask a question, it builds a new answer from those patterns. This is why the same question can give a different answer each time.

3. How an LLM is trained: 3 stages

Think of a student preparing for a job.

1. Pre-training (reading everything). The model reads a huge amount of text and practises one game: hide the next word, guess it, check, adjust. After billions of guesses, it becomes very good at language and picks up many facts along the way.
2. Fine-tuning (learning to be a helper). The model is then trained on examples of good conversations: a question and a helpful answer. Now it answers questions instead of just continuing text.
3. Feedback (learning what people prefer). People rate answers ("this one is better, this one is harmful"), and the model is adjusted to give more of the preferred kind. This is also where safety rules are taught.

After training, the model is frozen. It does not learn from your chat the way a person learns. Features like memory and projects (you used them earlier) work by adding text to your prompt, not by changing the model.

4. Parameters and model sizes

A model has parameters: numbers inside it that were adjusted during training. You can think of them as the strength of millions of small connections. Large models have billions of them.

More parameters usually means:

• better reasoning and writing,
• more knowledge,
• but slower answers and a higher cost to run.

So companies make a family of models: a small, fast one and a big, careful one.

Many apps now also have a thinking option. The model spends extra time working through the problem before it answers. You saw these in the step-by-step chapter: Think in ChatGPT, effort levels in Gemini, Extended thinking in Claude and Think Deeper in Copilot.

Worked example 1: the same question, fast vs careful

On free plans you usually cannot pick the model size by name. For example, free Gemini accounts use Auto, which sends most prompts to Flash-Lite, Google's fastest model (Google's help page, October 2026). The closest free test is to compare the fast mode with the thinking mode.

Steps:

1. Open Copilot (or ChatGPT). Keep the mode on Quick response (or the normal mode).
2. Paste this question:

A shop sells pens at ₹12 each or a box of 10 for ₹100. I need 34 pens. What is the cheapest way to buy them? Give the total.

3. Note the answer.
4. Start a new chat, switch to Think Deeper (Copilot) or Think (ChatGPT), and paste the same question.

Expected output, fast mode (yours will be similar, not the same):

Buy 3 boxes (30 pens) for ₹300 and 4 single pens for ₹48. Total: ₹348.

Expected output, thinking mode (yours will be similar, not the same):

Option 1: 3 boxes + 4 single pens = ₹300 + ₹48 = ₹348.
Option 2: 4 boxes (40 pens) = ₹400.
Option 3: 34 single pens = ₹408.
Cheapest: 3 boxes and 4 single pens, total ₹348.
Note: if you can use 6 extra pens, 4 boxes cost only ₹52 more.

Both got ₹348 here, which is correct (check: 3 × 100 + 4 × 12 = 348). The thinking answer compared all options and added a useful note. On harder questions, fast modes make mistakes more often, which is why the thinking mode exists. For simple writing tasks, the fast mode is enough and quicker.

5. The model and the app are different things

Learners often say "ChatGPT is a model". Not exactly.

• The model is the brain: GPT, Gemini, Claude.
• The app is the house around it: the chat box, history, file upload, voice, web search, memory, projects.

This matters for you in 2 ways:

1. The same app can give better answers on a paid plan because a bigger model is inside, not because the app changed.
2. Features like web search are part of the app, not the model. The model alone knows only what it learned in training. The app can search the web and give the results to the model. You will learn this in the web search chapter.

6. Tokens: how an LLM reads text

An LLM does not read letters or whole words. It reads tokens: small pieces of text. A common word is often 1 token; a long or rare word is split into 2 or 3 pieces.

Worked example 2: count tokens in a sentence

We counted tokens with OpenAI's free tiktoken library (encoding o200k_base, used by recent OpenAI models) on 19 October 2026. These are real counts:

Notice:

• The space usually belongs to the word after it (send).
• Numbers and symbols use many tokens.
• The same invoice sentence written in Hindi (Devanagari script) took 16 tokens, double the English. Indian-language text usually uses more tokens, so long Hindi or Marathi chats fill the model's memory faster.

Try it yourself: open OpenAI's free Tokenizer page (platform.openai.com/tokenizer), paste a sentence, and see the pieces in colour. Other companies use different tokenizers, so the count for Gemini or Claude will be a little different.

A rough rule from OpenAI for English: 1 token is about 4 characters, or about ¾ of a word. So 100 tokens is about 75 words. The next chapter is all about tokens and why they decide how long a chat the AI can remember.

7. Knowledge cutoff: the model's "last day of school"

Pre-training stops on some date. After that, the model has not read anything new. That date is its knowledge cutoff. If you ask about something that happened after it, the model alone does not know, and it may guess.

Modern apps solve this with web search: the app searches and gives fresh results to the model. So to see the cutoff, turn search off or ask the model to answer from memory only.

Worked example 3: the knowledge cutoff test

Steps:

1. Open a new chat in any tool. If there is a web search button, make sure it is off.
2. Paste:

Answer only from your training, without searching the web:
1. What is your knowledge cutoff (the last month and year you have information about)?
2. Name the most recent major event you know about.
3. What is today's date? If you do not know, say so.

3. Then start a new chat, turn web search on, and ask: What are 3 top news headlines in India today? Give the source for each.

Expected output, search off (yours will be similar, not the same):

1. My training data goes up to around [a month and year in the past].
2. The most recent major event I know about is [an event from that time].
3. I do not have access to today's date unless the app gives it to me or you tell me.

Expected output, search on (yours will be similar, not the same):

1. [Headline] – Source: [news site], [date]
2. [Headline] – Source: [news site], [date]
3. [Headline] – Source: [news site], [date]

What this shows: the model alone lives in the past; the app with search can look at today. Some apps quietly add today's date to your prompt, so the model may "know" the date but still not know today's news. A model can also be wrong about its own cutoff. Google's Gemini help page itself warns that Gemini can misrepresent how it works. Treat the self-reported cutoff as a rough guide.

8. What LLMs are good and bad at

The last row on the left side and the "counting letters" row on the right side both come from tokens. The model sees unbelievable as 3 pieces, not 12 letters, so "how many e's are in unbelievable?" can go wrong.

9. Compare the 4 tools on one idea

Plans and model names change often. Check each company's pricing page before you pay for anything.

10. Beginner mistakes and quick fixes

11. Try at home

Recap

Run the knowledge cutoff test in your favourite tool today. Once you have seen the model's "last day of school" with your own eyes, you will never trust a price or a date from it without checking again.

Previous chapter: 25 ChatGPT Prompts for Daily Work (Copy and Paste, with Expected Output)

Next chapter: Tokens and Context Window Explained: Why AI Forgets Long Chats

Ravindra Bagale, trainer: linkedin.com/in/ravindra-bagale. The shop example is made up for learning. Token counts were measured with tiktoken 0.14.0 (o200k_base) on 19 October 2026. Model names and plan features were checked on the official OpenAI, Google, Anthropic and Microsoft pages on 19 October 2026 and can change. AI answers change every time, so your output will be similar, not the same.