How LLMs Actually Work in Plain English (No Math, No Jargon)

• An LLM is a very large neural network trained to predict the next word in a sequence. That is the entire training objective. Everything else flows from it.
• "Large" means a model with hundreds of billions of internal parameters, trained on trillions of words of text. The scale is the magic.
• LLMs do not "know" things the way you know them. They have learned statistical patterns in language. This makes them powerful and also makes them confidently wrong.
• The chat interface you use wraps a base LLM with additional layers: instructions, conversation history, tools, and safety filters. Most of what you experience comes from these layers, not the base model.
• Understanding how LLMs work changes how you use them. They are best at pattern completion, drafting, and synthesis. They are worst at facts that are not in their training data and at multi-step reasoning.

A large language model is a mathematical function. You give it some text, it gives you back some text. The function is built from a structure called a "transformer" — a type of neural network designed in 2017 and refined in the years since.

The function has been trained. During training, the model was shown enormous amounts of text — books, articles, websites, code, conversation — and asked to predict, given a sequence of words, what word comes next. Billions of times. With billions of parameters (the internal knobs that the model adjusts during training).

After training, the model is a frozen function. It does not learn from new conversations. It does not update itself. The model you talk to in 2026 is essentially the model that finished training months or years ago (with some variations — see "fine-tuning" below).

That is the entire model. A very large function that predicts the next word.

The honest answer: from the scale.

When you have:

…something emergent happens. The model does not just predict the next word statistically. It develops internal representations of grammar, facts, reasoning patterns, code structures, multiple languages, humor, sarcasm, and a kind of general "understanding" of what text means in context.

This is the surprising part. No one programmed these capabilities. They emerged from scale. Researchers call them "emergent abilities" and they are still being characterized.

What this means practically: the model can write a poem, explain quantum mechanics, draft a contract, debug code, summarize a meeting, and translate between languages, all from the same training objective. None of that was explicitly programmed.

It also means the model sometimes invents things that fit the pattern of being true but are not. We will get to that.

• A function with enough parameters (current frontier models are 100B - 1T+)
• Trained on enough data (trillions of words)
• With enough compute (millions of GPU hours)

The raw LLM is a "next word" predictor. The chat interface is a more complex system. When you talk to ChatGPT or Claude, several things are happening:

When you have a long conversation, every new turn sends the entire conversation history back to the model. This is why very long conversations can get slow or start to lose track of early context.

• Your message is combined with a system prompt. The system prompt is a set of instructions the model follows throughout the conversation. "You are a helpful assistant," "Be concise," "Do not provide medical advice," and so on. The system prompt is set by the company running the model, and sometimes partially by you.
• The conversation history is included. The model sees your messages and its previous responses. This is what makes the conversation feel continuous.
• Tools may be called. The model can be given access to web search, code execution, file reading, image generation, and other capabilities. The model "decides" when to use them based on the conversation.
• The model generates a response. It predicts one word at a time, building up the response token by token. The system samples from the model's probability distribution to choose each word.
• Safety filters are applied. The response is checked against content policies. Outputs that violate the policies are rewritten or blocked.
• The response is shown to you.

LLMs do not actually work with words. They work with "tokens" — chunks of text that are usually parts of words or common word fragments.

"Hello" is one token. "Conversation" might be two or three tokens. "Antidisestablishmentarianism" might be five or six. The model sees tokens, not words.

Tokens matter for two reasons:

A "token" is roughly 4 characters of English text, or about 0.75 words. So 1000 tokens is about 750 words.

• Cost. LLM services charge per token, both for input and output.
• Counting. The model has a maximum context window — the number of tokens it can process at once. In 2026, this ranges from 200K to 1M+ tokens depending on the model.

The "context window" is the total amount of text the model can consider when generating a response. This includes:

If the context window is 200K tokens, that is roughly 150,000 words — about 300 pages of text. The model can "see" all of it when generating its response.

In 2026, context windows are large enough for most practical uses. But the model does not treat all of the context equally. Information at the beginning and end of the context is typically weighted more than information in the middle. This is called "lost in the middle" and is a real limitation.

This is why long conversations can lose track of things you said early on, and why you should put the most important information at the start or end of a long document you give the model.

• The system prompt
• The conversation history
• Any documents you attach
• The model's previous responses

The base model is a "next word" predictor trained on internet text. Fine-tuning is a second training step that adjusts the model for specific behavior.

Types of fine-tuning:

The model you talk to has been through several rounds of fine-tuning. The behavior you experience is a combination of the base model and the fine-tuning.

• Supervised fine-tuning (SFT). The model is shown examples of good responses and learns to imitate them. This is how chat models are taught to be helpful, harmless, and honest.
• Reinforcement learning from human feedback (RLHF). Human raters rank the model's responses, and the model is trained to produce responses that humans prefer. This is what gives Claude and ChatGPT their conversational feel.
• Constitutional AI (RLAIF). Instead of human raters, the model is graded against a set of principles (the "constitution"). Used heavily by Anthropic for Claude.
• Domain-specific fine-tuning. A company can take a base model and fine-tune it on their own data (customer support conversations, legal documents, medical records) to make it better at a specific task.

A hallucination is when the model generates text that is fluent and confident but factually wrong. The classic example: an LLM confidently cites a legal case that does not exist, with a plausible-sounding name, court, and citation.

Hallucinations happen because the model is not "looking up" facts. It is generating text that fits the pattern of being a fact. The model has no internal notion of "true" or "false." It has a notion of "what would naturally come next given what came before."

If you ask "Who was the first woman to win the Nobel Prize in Physics?" the model will probably answer correctly (Maria Goeppert Mayer, 1963) because that is well-represented in the training data and the answer is a common pattern.

If you ask "What did Maria Goeppert Mayer say in her 1962 lecture about quantum tunneling?" the model might invent a plausible quote because it has the pattern of "scientist saying something in a lecture" but not the specific content. The model cannot tell you it does not know. It produces a fluent, confident, plausible, and possibly wrong answer.

This is the central limitation of LLMs. They are powerful and unreliable in exactly the way that is hardest to detect: the wrong answers sound like the right answers.

The use cases where LLMs excel:

The use cases where LLMs fail or are dangerous:

The right mental model: the LLM is a very capable assistant that is sometimes confidently wrong. Use it for tasks where you can verify the output, where a draft is enough, or where the cost of being wrong is low.

• Drafting. Emails, memos, outlines, first drafts. The model produces plausible text quickly; you edit for accuracy and voice.
• Summarization. Condensing long documents into shorter summaries. The model is good at identifying key points.
• Brainstorming. Generating options, alternative framings, names, ideas. You select what resonates.
• Translation and rewriting. Especially between languages and registers (formal to casual, verbose to concise).
• Code assistance. Generating boilerplate, explaining code, finding bugs in code you already have.
• Synthesis. Combining multiple documents into a coherent answer. The model is better at synthesis than most humans.
• Pattern matching. "Look at these 50 customer complaints and group them by theme."
• Factual recall without verification. Always check important facts against a real source.
• Multi-step reasoning without scaffolding. Complex logic (multiplication of large numbers, tracking many variables) is unreliable.
• Decisions that require accountability. The model cannot be held responsible for what it says. You are responsible for what you sign, what you publish, what you act on.
• Anything where the wrong answer is worse than no answer. Medical advice, legal advice, financial advice, safety-critical decisions.

Prompt engineering is the practice of writing inputs that get the model to produce better outputs. The model is sensitive to:

In 2026, prompt engineering is less of a specialized skill than it was in 2023. The models are better at understanding natural language. But the basic principles still hold: clear, specific, with examples when possible.

• Specificity. "Write a 200-word product description for a noise-cancelling headphone for remote workers, in a friendly but professional tone" is much better than "Write a product description."
• Examples. Showing the model what you want is more reliable than describing it.
• Role prompts. "You are a senior copywriter" sets a context the model will follow.
• Step-by-step instructions. "First, list the key themes. Then, draft an outline. Then, write the introduction" produces better output than "Write this."
• Constraints. "Use no more than 200 words" or "Avoid these phrases" or "Cite the source for each fact."

An agent is an LLM that can take actions in the world, not just produce text. The agent has access to tools (web search, file operations, code execution, email send) and decides which to use based on the user's goal.

"Book me a flight to Berlin" — a chatbot will tell you how to book a flight. An agent will check your calendar, search flights, compare prices, and either book it or ask for confirmation.

In 2026, agents are increasingly common but still imperfect. They work well for narrow, well-defined tasks. They struggle with ambiguous goals, multi-step dependencies, and any task where the cost of failure is high. The honest assessment: agents are good at "do this specific thing in this specific way" and bad at "figure out what to do."

An LLM is a very large function trained to predict the next word. Trained at scale, it develops emergent capabilities that look like understanding but are really sophisticated pattern matching. The chat interface is a wrapper that makes the function feel conversational.

This explains both the power and the failure modes. The model is good at tasks that involve pattern completion on text — drafting, summarizing, brainstorming, translating, code assistance. It is bad at tasks that require reliable fact recall or accountability for its outputs.

Use the model accordingly. Drafts, not final answers. Synthesis, not authority. Speed, not certainty. The technology is remarkable. The right mental model is "capable assistant who sometimes makes things up" — not "source of truth."

• ChatGPT vs Claude vs Gemini in 2026: Which to Use for What
• What AI Can and Can't Do in 2026: Setting Realistic Expectations
• Prompt Injection in 2026: What It Is and Why It Matters Even If You're Not Technical
• How to Spot AI-Generated Content in 2026: Text, Images, Video, and Audio
• Deepfakes in 2026: How to Spot Them and When to Trust What You See