What Is an LLM (Large Language Model)? A Plain-English Guide

How does an LLM work?#

An LLM works by breaking text into small pieces called tokens and predicting the most likely next piece, again and again, until a response is complete. It learns those predictions from huge amounts of text during training. Further training then shapes the raw predictor into an assistant that follows instructions and answers questions helpfully.

Step 1: pretraining on text#

Anthropic's documentation describes pretraining as training a language model on a large, unlabelled collection of text so that it learns to predict the next word, given the text that came before. The model stores what it learns as parameters: numerical settings adjusted during training. "Large" refers to the number of parameters and the amount of training data. The 2020 research paper "Language Models are Few-Shot Learners", for example, trained eight sizes of model ranging from 125 million to 175 billion parameters.

Step 2: the Transformer design#

The model in the 2020 paper is built from Transformer layers, a design first proposed in the 2017 research paper "Attention Is All You Need". The Transformer is based solely on attention mechanisms, which let the model weigh how each part of the text relates to the other parts. Its authors reported that it was more parallelisable and took significantly less time to train than the designs it replaced.

Step 3: shaping it into an assistant#

A pretrained model is good at continuing text but, according to Anthropic's glossary, is not inherently good at answering questions or following instructions. Fine-tuning (further training on extra data) and reinforcement learning from human feedback (RLHF) refine it. In RLHF, people rank example responses and the model is trained to prefer responses like the higher-ranked ones.

Step 4: answering your request#

When you type a prompt, the model reads it as a series of tokens and generates a reply one token at a time. Everything it can see while answering sits in its context window, which Anthropic describes as the model's "working memory", separate from the data it was trained on. Our guide to tokens and context windows explains why that matters.

What can an LLM do well at work?#

An LLM is most useful for language tasks where a person can quickly judge the result: drafting, summarising text you provide, rewording for a different reader, brainstorming and pulling key points into a structure. It is weakest where accuracy depends on facts it has not been given, such as current figures, internal policies or legal details.

The 2020 paper found that larger models could perform new tasks from only a few examples or simple instructions given as text, without any retraining. In practice, that is why a well-written prompt with one or two examples often gets better results than a vague request.

The last row is the one to remember. An LLM always produces an answer, whether or not it has the facts.

What are the limits of an LLM?#

The main limits are that an LLM can state false things confidently, can only see what fits in its context window, gives varied answers to the same prompt and knows only what was in its training data. None of these is a fault in one product. They follow from how these models generate text.

• Confident errors. A NIST risk profile for generative AI uses the term confabulation for confidently stated but erroneous or false content, known colloquially as hallucination. NIST explains it as a natural result of how generative models work: they predict the next token based on patterns in their training data, which can produce accurate output but also false or inconsistent output. Our guide to AI hallucination covers how to reduce it.
• Limited view. The model can only use what is in its context window. A larger window allows longer prompts, but Anthropic notes that more context is not automatically better, because accuracy and recall can degrade as the token count grows.
• Variation. A setting called temperature controls how much randomness is allowed. Lower temperatures give more conservative output, but Anthropic notes that even at zero, results are not fully deterministic.
• Fixed knowledge. The model learned from a set of training data. Unless a tool supplies newer material, such as through retrieval-augmented generation (RAG), it cannot know about later events or your internal documents.

LLMs come in many sizes, and a small language model (SLM) can be enough for narrow tasks. To weigh the trade-offs, read when a smaller model is the better choice.

What might an LLM look like in a Singapore workplace?#

Picture a six-person accounting practice in Tanjong Pagar where staff spend time drafting routine client emails and reading long, publicly available guidance notes. An approved LLM tool could help produce first drafts and plain-English summaries, provided someone checks each output and no client data goes into an unapproved service.

This is an illustrative scenario, not a case study. A sensible way to start would be:

1. Pick low-risk tasks. Drafting reminder emails from a template and summarising public guidance the team already reads.
2. Give the model the source. Paste the guidance text into the prompt and ask for a summary based only on that text, rather than asking from memory.
3. Define the output. For example: "Five bullet points for a busy client, no jargon, flag anything with a deadline."
4. Check before sending. A staff member compares each summary with the original and fixes dates, amounts and wording.
5. Protect client details. Client names, identification numbers and financial details stay out unless the tool is approved for them. Under Singapore's Personal Data Protection Act (PDPA), personal data is data about an individual who can be identified from that data, or from that data and other information the organisation has or is likely to have access to.

The time saved, if any, would depend on the task and on how much checking each draft needs. The firm would only know by trying it on real work and keeping a simple record.

What are the common mistakes with LLMs?#

The most common mistakes are treating an LLM like a search engine, giving it too little context, trusting fluent wording as proof of accuracy and pasting in data the tool is not approved for. Most can be avoided with a clear prompt, supplied source material and a habit of checking before anything is sent.

• Asking for facts from memory. Give it the source text instead, and ask it to answer only from that text.
• Writing one-line prompts. State who the output is for, what it is for, what to include and what format you want.
• Mistaking fluency for accuracy. Well-written text can still be wrong. NIST warns that confident responses can lead users to believe and act on false content.
• Accepting invented references. NIST notes that output can include confabulated logic or citations. Open every source the model names.
• Ignoring data rules. Use only tools your organisation has approved, and keep personal or confidential data out of the rest.
• Using one long chat for everything. Older instructions compete with newer ones. Start a fresh conversation for a new task.

What to learn next#

Once you know what an LLM is, the next steps are understanding why it sometimes gets facts wrong, how its working memory is measured in tokens, and when a smaller model is enough. Each topic below builds directly on this guide and helps you use AI tools with more confidence at work.

• Browse more guides for everyday users in AI for users.
• See where this topic fits in the full AI learning pathway, from first prompts to governance.
• If you want to build something with an LLM yourself, the GenAI Builder Sprint is a 2-day classroom course where adults plan, build, test and launch a simple AI app. No previous AI or programming experience is needed, although formal entry requirements apply.

Sources#

1. Glossary, Anthropic (Claude Platform Docs), accessed 8 November 2026
2. Attention Is All You Need, arXiv, accessed 8 November 2026
3. Language Models are Few-Shot Learners, arXiv, accessed 8 November 2026
4. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, National Institute of Standards and Technology (NIST), accessed 8 November 2026
5. Context windows, Anthropic (Claude Platform Docs), accessed 8 November 2026
6. PDPA Overview, Personal Data Protection Commission (PDPC), accessed 8 November 2026

Frequently asked questions#

What is a large language model (LLM)?+

An LLM is an AI language model with many parameters, trained on vast amounts of text data. It can generate human-like text, answer questions and summarise information. It works by predicting the next piece of text from everything that came before.

How does an LLM learn?+

First it is pretrained on a large collection of text to predict the next word. A pretrained model is not naturally good at following instructions, so it is then refined with further training such as fine-tuning and reinforcement learning from human feedback (RLHF).

Does an LLM look up facts when it answers?+

Not by default. It predicts likely text based on patterns in its training data and what is in its context window. That is why it can produce confident but false content, which NIST calls confabulation and which is known colloquially as hallucination.

Why does the same prompt give different answers?+

LLMs choose from likely next words, and a setting called temperature controls how much variety is allowed. Lower temperatures give more conservative output. Even at the lowest setting, results are not fully deterministic, so identical prompts can produce different answers.

What is an LLM good for at work?+

Drafting first versions, summarising long text you provide, rewording for a different reader, brainstorming options and extracting key points into a table. In each case a person should check the result before it is used.

Can I paste client or staff details into an LLM tool?+

Only into a tool your organisation has approved for that data. Under the PDPA, personal data is data about an individual who can be identified from that data, or from that data and other information the organisation has or is likely to have access to.