PLAIN-ENGLISH EXPLAINER
What Is a Large Language Model?
From next-word prediction to useful assistant, the idea is simpler than the name suggests.
Large language models are the technology behind today's most capable AI assistants. This page explains what they are, how they work without the math, and why understanding the basics makes you a sharper, more effective user.
A large language model (LLM) is a type of AI trained on a vast collection of text to predict the most useful next word, phrase, or sentence given a prompt. During training, the model adjusts billions of internal parameters to capture patterns across language: grammar, facts, reasoning styles, even code. At inference time, it generates responses token by token, selecting each word based on context. The result is a system that can write, reason, translate, and answer questions without being explicitly programmed to do any of them.
CORE CONCEPTS
How a Large Language Model Works
Four ideas explain most of what LLMs do, and most of what surprises people about them.
Next-token prediction
At its core, an LLM learns to predict the most plausible next word (or token) given everything that came before. That single task, applied at enormous scale, produces surprisingly capable language.
Training on text
Before you ever talk to an LLM, it has processed a huge slice of written text: books, articles, code, conversations. It tuned its internal weights to fit the patterns in that data.
Parameters, not rules
Unlike older AI systems built on hand-coded rules, LLMs store knowledge as billions of numeric weights. No engineer wrote 'the sky is blue'; the model inferred it from patterns across millions of examples.
Emergent capabilities
Abilities like translation, summarization, and multi-step reasoning were not separately programmed. They emerged as a side effect of getting very good at predicting text at scale.
UNDER THE HOOD
Training vs. Inference: Two Very Different Phases
Training is the expensive, one-time process of building the model. A massive dataset of text is fed through a neural network repeatedly. Each time the model guesses the next word wrong, its weights are nudged to do better. After countless iterations, the weights stabilize into something that captures a broad understanding of language.
Inference is what happens when you type a message. The model takes your prompt, runs it through those fixed weights, and generates a response word by word. No learning is happening at this stage; it is applying what it already knows. This matters because the model cannot update its knowledge on the fly from your conversation.
Between raw training and a polished assistant, most models go through additional fine-tuning with human feedback. People rate responses for helpfulness and accuracy, and the model is adjusted to prefer higher-rated outputs. This is what turns a text-predictor into something that feels like a thoughtful assistant rather than a search-engine autocomplete.
HONEST LIMITS
Why LLMs Sometimes Get Things Wrong
Understanding the failure modes makes you a better user and helps you catch mistakes before they matter.
Hallucination
When an LLM lacks reliable knowledge, it may generate a plausible-sounding answer anyway. It learned that fluent, confident text is rewarded, so silence or 'I don't know' can feel unnatural to the model.
Knowledge cutoff
Training data has a cutoff date. Events, papers, or facts that emerged after that point simply don't exist in the model's weights unless it is given a tool to look things up.
Context limits
Every LLM has a context window: the amount of text it can hold in working memory at once. Information outside that window is invisible to the model while it generates a response.
Pattern-matching, not reasoning
LLMs are exceptionally good at recognizing and extending patterns. On novel problems that require genuinely new reasoning, they can struggle or confidently go wrong in ways a careful reader would catch quickly.
PRACTICAL SKILL
Why Prompting Matters More Than Most People Think
The model you talk to is fixed; its weights do not change between conversations. What you can change is the prompt: the context, instructions, and examples you give it. A vague prompt forces the model to guess your intent. A specific one narrows the space of plausible responses toward what you actually need.
This is why prompt engineering is a real skill, not a buzzword. Giving the model a role ('you are a tax attorney reviewing this clause'), a format ('respond in three bullet points'), or worked examples ('here is a good response to a similar question') systematically improves output quality. You are not tricking the model; you are reducing its uncertainty about what a good response looks like.
Understanding that the model is predicting useful next tokens also explains why it is sensitive to phrasing. Two prompts that mean the same thing to you can activate very different patterns in the model. The practical takeaway: when a response misses the mark, reframe rather than repeat.
Understanding LLMs is the foundation. Using them well is the skill.
Reading about next-token prediction is useful context. Actually practicing prompt techniques, working through real examples, and building intuition for what works is where understanding turns into ability. The Claude Academy curriculum is hands-on from the first lesson.
Frequently Asked Questions
What does LLM stand for?
LLM stands for large language model. 'Large' refers to the scale of both the training data and the number of internal parameters the model contains. 'Language model' describes what it does: model the probability of sequences of text, which in practice means predicting useful next words given a prompt.
How is an LLM different from a search engine?
A search engine indexes existing documents and retrieves the most relevant ones for your query. An LLM generates new text by predicting what a useful response would look like, drawing on patterns baked into its weights during training. Search finds; an LLM composes. Many AI assistants now combine both: the model generates while a retrieval layer gives it access to current information.
Why do LLMs hallucinate?
Hallucination happens because LLMs are optimized to produce fluent, plausible text, not to verify facts before outputting them. When the model lacks reliable signal in its training data, it may still generate a confident-sounding answer by extending the most likely-looking pattern. Checking LLM outputs against authoritative sources is good practice whenever accuracy is critical.
Can an LLM learn from our conversation?
No, not in real time. The model's weights are fixed at inference time. Your conversation may be stored and used in future training rounds depending on the provider's policy, but the model you are talking to right now is not updating as you type. Some systems simulate memory by injecting prior conversation summaries into the context window, which is not the same as true learning.
What is a token?
A token is the unit LLMs process and predict. In English, a token is typically a word or a common word fragment; a longer word might be split into several tokens. LLMs generate text one token at a time, choosing each based on accumulated context. Context windows and usage pricing are commonly described in tokens rather than words or characters.
What is the difference between an LLM and an AI assistant like Claude?
An LLM is the underlying model: the trained neural network. An AI assistant like Claude is an LLM plus a layer of fine-tuning, safety training, and often additional tools such as web search or code execution. Think of the LLM as the engine and the assistant as the finished vehicle: same core technology, shaped into something useful and safe for everyday use.
Do larger LLMs always perform better?
Not necessarily. Scale helps, but training data quality, fine-tuning technique, and the specific task matter at least as much. Smaller, well-tuned models often outperform larger ones on focused tasks. The field has moved from 'bigger is always better' toward optimizing the whole pipeline: architecture, data, alignment, and inference efficiency together.
Is prompting a skill worth learning?
Yes. Because LLMs predict useful text based on context, the framing you provide directly shapes what the model generates. Clear roles, specific formats, worked examples, and explicit constraints all improve output quality consistently. Prompting is not a workaround for a limited tool; it is how you communicate intent to a system that cannot read your mind.
Put this into practice today
Knowing how LLMs work is step one. Step two is getting good at working with them.