A Large Language Model is an AI system trained on an enormous quantity of text so that it can understand and generate human-like language. ChatGPT, Claude, Gemini and Llama are all LLMs. So are the Indian language models being built by domestic research groups and startups.
"Large" refers to two things: the volume of text it was trained on, measured in trillions of words, and the number of internal parameters it adjusts during training, often in the hundreds of billions.
How it actually produces a sentence
An LLM does one thing repeatedly: it predicts the next token.
You type a prompt. The model reads it, and calculates a probability distribution over what should come next. It picks a token, appends it, then reads everything including that new token and predicts the next one. Again and again, until it produces a stop signal.
That is the whole mechanism. There is no plan drafted in advance, no lookup in a database of facts, no internal record of what is true. A fluent answer about Section 80C is a statistically likely continuation of your question, assembled one fragment at a time.
Understanding this explains almost every strange behaviour you will encounter with these tools.
How they are trained
Pre-training is the expensive part. The model reads a vast corpus of text from books, websites, code repositories and articles, learning grammar, facts, reasoning patterns and style purely by predicting missing text. This takes months and enormous computing resources.
Fine-tuning then shapes the raw model into something useful and safe, teaching it to follow instructions, refuse harmful requests and answer in a helpful format. Techniques such as reinforcement learning from human feedback are used here, where human raters compare responses and the model learns which style of answer people prefer.
An important consequence: the model is optimised to produce answers people rate highly, which correlates with being correct but is not the same thing. A confident wrong answer often rates better than an honest "I do not know".
The three limits that matter in practice
Training cutoff. The model's knowledge stops when its training data stops. Unless it can search the web, it will not know about the latest Budget, a rate change, or anything from the last several months. For Indian tax and regulatory questions this is a serious limitation, because the answer changes annually.
Context window. The model can only consider a limited number of tokens at once, covering your prompt, any documents you paste, and its own reply. Large context windows now run to hundreds of thousands of tokens, but in a long conversation, early messages can fall out of view.
[Hallucination](/glossary/hallucination). Because it generates plausible text rather than retrieving verified facts, it will invent citations, statistics, section numbers and case names, in the same confident tone it uses for correct information. There is no built-in signal distinguishing the two.
What LLMs are genuinely good at
- Drafting and rewriting: emails, summaries, job descriptions, first versions of anything
- Explaining a concept at whatever level you ask for
- Translating and switching register between languages, including Hindi and other Indian languages, with quality varying by language
- Summarising a long document you supply
- Structuring messy notes into an organised form
- Writing and debugging code
- Being a patient tutor that never gets bored of your third clarifying question
What they are unreliable at
- Precise arithmetic, unless connected to a calculator tool
- Current facts, rates, prices and law
- Citations, which they will fabricate convincingly
- Anything requiring certainty about a specific number that affects money
- Knowing what they do not know
Practical use in India
For everyday work, LLMs are genuinely useful for drafting client emails, summarising long PDFs, preparing study material, translating between Indian languages, and getting a plain-English explanation of a financial or legal concept before you verify it.
Two cautions worth taking seriously. First, do not paste PAN, Aadhaar, bank statements, salary slips or client contracts into free consumer tools, which often use conversations for further training and may create obligations under the Digital Personal Data Protection Act if the data belongs to someone else.
Second, treat any number the model gives you about Indian tax, interest rates or scheme limits as a hypothesis to check, not an answer. It may be a figure from two Budgets ago, delivered with complete confidence.
See prompt engineering for how to get substantially better output from the same model.