What Is an LLM? How Large Language Models Work, Explained Simply

What does LLM stand for?

LLM stands for large language model. Each word matters:

• Language model: Google’s Machine Learning Crash Course defines a language model as one that estimates the probability of a token or sequence of tokens occurring within a longer sequence. In plain words, it predicts what text is likely to come next.
• Large: the model has a very large number of internal settings (parameters) and was trained on a very large amount of text.

ChatGPT, Gemini and Claude are products built on LLMs from OpenAI, Google and Anthropic. Meta’s Llama models are an example of open-weight LLMs that companies can download and run on their own servers.

How does an LLM predict text?

An LLM writes one token at a time. It reads everything so far, estimates how likely each possible next token is, picks one, adds it to the text and repeats.

Imagine the input “The capital of India is”. An illustrative set of next-token probabilities might look like this (these numbers are made up to show the idea, not taken from a real model):

Possible next token	Illustrative probability
“ New”	0.92
“ Delhi”	0.05
“ a”	0.01
Everything else	0.02

The model picks “ New”, then predicts the next token after “The capital of India is New”, which is almost certainly “ Delhi”. A full paragraph is hundreds of these small predictions in a row.

What makes modern LLMs good at this is the transformer design and its self-attention mechanism. Google’s course explains that older recurrent models read text token by token, while LLMs can evaluate the whole context at once. That is why a model can connect a pronoun in your last sentence to a name you mentioned three paragraphs earlier.

Because the model chooses from probabilities, the same prompt can give slightly different answers each time. That is a feature for creative writing and a reason to check factual work.

What are tokens in an LLM?

A token is the unit an LLM reads and writes. OpenAI’s help centre says a token can be a character, part of a word, a whole word or punctuation. Its rough rules of thumb for English text are:

• 1 token is about 4 characters
• 1 token is about three-quarters of a word
• 100 tokens are about 75 words

So a 1,500-word report is roughly 2,000 tokens. OpenAI also notes that other languages can have a different relationship between characters, words and tokens, so the same message in Hindi or Tamil may use a different number of tokens than in English.

Tokens matter for two practical reasons: API pricing is usually per token, and every model has a limit on how many tokens it can handle at once.

What is a context window?

The context window is the model’s working memory for one conversation or request. Anthropic’s documentation describes it as all the text a language model can reference when generating a response, including the response itself. It is different from the huge body of text the model was trained on.

Sizes vary a lot. Anthropic lists a 1 million-token context window for its recent Claude models and 200,000 tokens for some others. Google’s Gemini documentation also describes 1 million-token context windows. In consumer apps the usable amount can be smaller: OpenAI’s pricing page lists a 27K total context window for GPT Instant on the Free plan, with room for roughly 12 pages of input.

Bigger is not automatically better. Anthropic notes that as token count grows, accuracy and recall can degrade, which it calls context rot. For long documents, it often works better to give the model only the relevant sections.

What is the difference between training and fine-tuning an LLM?

People use “training” loosely. There are several different ways to shape what an LLM does, and they differ hugely in cost and effort.

Method	What happens	Who usually does it	Example use
Pre-training	The model learns language patterns by predicting tokens across a vast amount of text	AI labs with large compute budgets	Creating a new base model
Post-training	The base model is taught to follow instructions, hold conversations and behave safely, often using human feedback	AI labs	Turning a base model into a chat assistant
Fine-tuning	An existing model is trained further on a smaller, specific set of examples	Companies and developers	Matching a brand’s tone, or classifying support tickets into set categories
Retrieval (RAG)	No retraining; relevant documents are fetched and placed in the prompt at question time	Companies, developers, no-code builders	An HR bot that answers from the current leave policy
Prompting	No retraining; you shape the output with instructions and examples	Everyone	Drafting emails in a fixed format

For most business use in India, prompting and retrieval come first because they are cheaper, faster and easier to update. When a policy changes, you update the document, not the model. Fine-tuning makes sense when you need a consistent style or format at high volume. Our explainer on what RAG is in AI goes deeper on retrieval.

What are the limits of LLMs?

Hallucination

OpenAI defines hallucinations as plausible but false statements generated by language models. In its September 2025 research post, it describes asking a widely used chatbot for the title of one of its researchers’ PhD dissertation: the chatbot gave three different answers, and none was correct. OpenAI argues that standard training and evaluation reward guessing over admitting uncertainty.

Knowledge cut-off

A model only knows what was in its training data up to a certain date. Tools that add web search can look up newer information, but the model itself does not update daily.

Maths and exact details

LLMs predict text rather than calculate. They can slip on long arithmetic, dates or counts unless the tool runs code or a calculator behind the scenes.

Bias and gaps

Models learn from their training data, including its biases and blind spots. Answers about Indian laws, local markets or regional languages may be weaker than answers about widely written-about topics.

Privacy

What you paste into a chatbot may be stored, and on some consumer plans may be used for training unless you opt out. Follow your company’s policy before sharing work data. Our guide on how to use ChatGPT for work covers the settings.

Where are LLMs used at work?

LLMs are most useful where work involves reading, writing or sorting text. Some common patterns in Indian companies:

Function	Typical LLM use	Human check needed
Customer support	Drafting replies, summarising tickets, multilingual chatbots	Refunds, complaints, anything sensitive
HR	Job descriptions, policy Q&A bots, interview question banks	Hiring decisions and policy wording
Sales and marketing	Emails, ad copy variations, call summaries, research	Claims, prices and brand tone
Finance and operations	Extracting data from invoices and contracts, report drafts	Every number that goes into a report
Software teams	Code suggestions, test writing, documentation	Code review and security

The bigger shift is LLMs working inside systems rather than chat windows: a model that reads an incoming email, looks up the order and drafts a reply for approval. That is where AI agents come in. Our comparison of agentic AI and generative AI explains the difference.

How do you start learning about LLMs?

1. Use two or three LLM tools daily for real tasks and notice where each fails.
2. Learn prompting properly. Our prompt engineering guide for beginners is a good start.
3. Build one small project, such as a bot that answers from a PDF, to understand context windows and retrieval in practice.
4. Learn where LLMs fit in workflows, using automation tools that connect a model to email, sheets or a CRM.

The ISS AI & Agentic Systems program starts in Week 1 with how AI actually works, including LLM foundations, token logic and context window limits, and moves on to automations, RAG pipelines and agents. It runs for 16 weeks at about 12–14 hours a week, with no coding degree needed. The fee is ₹69,999, or ₹66,499 if paid upfront.

Frequently Asked Questions

What is an LLM in simple words?

An LLM, or large language model, is an AI model trained on huge amounts of text to predict the next small piece of text. By repeating that prediction many times, it can write answers, summaries, emails and code.

Is ChatGPT an LLM?

ChatGPT is a product built on OpenAI large language models. The LLM is the underlying model; ChatGPT adds the chat interface, memory, web search, file uploads and safety features on top.

What is a token in an LLM?

A token is the unit of text an LLM reads and writes. It can be a character, part of a word, a whole word or punctuation. OpenAI estimates that 100 tokens are about 75 English words.

What is the difference between an LLM and generative AI?

Generative AI is the broad category of AI that creates content such as text, images, audio or video. An LLM is one type of generative AI model that works with text and, increasingly, other inputs like images.

Why do LLMs make mistakes?

LLMs predict likely text rather than look up verified facts, so they can produce plausible but false answers, known as hallucinations. OpenAI research argues that training and evaluation often reward guessing over admitting uncertainty.

Do I need coding to work with LLMs?

Not for most business uses. You can use LLMs through chat tools, and build bots and automations with no-code tools. Coding helps if you want to call models through an API or build custom software.

Sources and methodology

• Google for Developers, Machine Learning Crash Course: Introduction to Large Language Models: definitions of language model, tokens, self-attention. Checked September 2026.
• OpenAI Help Center, Understanding and counting tokens: token rules of thumb. Checked September 2026.
• Anthropic, Context windows: definition, sizes by model, context rot. Checked September 2026.
• Google AI for Developers, Long context: 1 million-token context window. Checked September 2026.
• OpenAI, ChatGPT pricing: context window listed for the Free plan. Checked September 2026.
• OpenAI, Why language models hallucinate (5 September 2025). Checked September 2026.
• ISS program details come from our program page and fees page.

Method: the next-token probability table is an illustration written by the ISS Editorial Team, not output from a real model. The training-methods and work-use tables are our own summaries; the linked sources support the definitions and figures.

Next steps

Try the token rule of thumb on your next long document, and notice when a chatbot starts losing track of details in a long chat. For small projects that show you understand LLMs in practice, download the free AI Projects Starter Kit on this page.

If you want live teaching from LLM basics to RAG and agents, read the AI & Agentic Systems curriculum. You can apply for free and pay only after you accept an offer. ISS does not guarantee jobs or placement.