What is an LLM?

The three words, in the order that matters

Take the name apart backwards.

Model. A model is a function that turns an input into an output, with the rules learned from data rather than written by a programmer. The weather model takes today's readings and outputs tomorrow's forecast. A language model takes text and outputs a probability for what comes next.

Language. The input and the output are text, chopped into tokens, which are pieces of words. What are tokens in AI covers the unit; the short version is that the model reads and writes in fragments of roughly three quarters of a word.

Large. The model has a very large number of parameters, the internal numbers that training adjusts. GPT-2 in 2019 had 1.5 billion. GPT-3 in 2020 had 175 billion. The catalogue this site prices in 2026 includes a Qwen model with 2.4 trillion parameters, of which about 95 billion are active on any one token. Large is doing a lot of work in the name, and the count has grown more than a thousandfold in seven years.

So an LLM is a very big next-token predictor. Every other property people attribute to it, the reasoning, the tone, the confident mistakes, follows from that one job.

An LLM is not a chatbot, and the difference is the whole product

ChatGPT is an app. GPT is the model inside it. Claude is both a family of models and the app Anthropic wraps around them. Gemini is the same story at Google. What is an AI chatbot covers the app layer; this page is about the engine.

The distinction stops being academic the moment you learn the engine is swappable. The same chat window can send your message to GPT, Claude, Gemini, DeepSeek or Llama, and the answers differ in tone, cost, speed and accuracy. A multi-model workspace like Whizi is exactly that: one window, 280+ models behind it, and a picker to choose which one answers. The app is the steering wheel. The LLM is the engine, and there are dozens.

How an LLM is trained, in two stages

Stage one, pretraining. The model reads an enormous corpus of text and, for every position, tries to guess the next token. Each wrong guess nudges the parameters a little. Meta reported that Llama 3 was pretrained on more than 15 trillion tokens, which, at the roughly 100,000 tokens of a 75,000-word novel (three quarters of a word per token), is the text of roughly 150 million books. After this stage the model is a superb autocomplete and a poor assistant: ask it a question and it may continue with three more questions, because that's what text on the internet often does.

Stage two, post-training. The model is then shown examples of good answers and rated by people on which of two replies is better, a process called reinforcement learning from human feedback, or RLHF. This is where the assistant behaviour comes from: answering the question in the format you asked for, and refusing some requests. It is coaching, not new knowledge. The facts still come from stage one.

Two things follow from the two stages. The model's knowledge stops at the date its pretraining data ends, which is why every model has a knowledge cutoff. And nothing in either stage installs a fact-checker: the model learned to produce likely text and then to produce likeable text. Truth is a frequent side effect of both, not a guarantee. Why AI gets things wrong takes that further.

Why an LLM predicts instead of looking things up

A search engine retrieves. It has an index of pages and returns the ones that match. An LLM has no index. Everything it read during pretraining was compressed into its parameters, and at answer time it regenerates text from that compression, one token at a time, picking each token from a probability distribution.

This is why an LLM can write a working recipe for a dish it has never seen named, and also why it can produce a court citation that doesn't exist. Both are plausible continuations. The model has no separate mechanism that checks whether a plausible continuation is a real one.

The fix that products use is to bolt retrieval on: search the web or your files first, then hand the model the passages and ask it to answer from those. That technique is called RAG, and what is RAG explains it. Web search inside a chat app is RAG with the internet as the source. The model is still predicting; it's just predicting from better material.

What the price buys inside a family, measured

Every family ships a ladder: a small model, a middle one, a flagship. The Artificial Analysis Intelligence Index, a composite score across public reasoning, coding and knowledge benchmarks, puts a number on each rung, and the leaderboard lists a blended price per million tokens next to it. The four ladders below were read on 2026-09-17.

OpenAI buys 15 points for 18 times the rate, Anthropic 35 points for 36 times, Qwen 19 points for 16 times. DeepSeek breaks the pattern: V4 Flash outscores V4 Pro at less than half the price. Two things follow. Inside a family, the last few points are the expensive ones, so for a short factual question the small model gives the same answer as the flagship. And across families, price is a weak guide to quality: DeepSeek V4 Flash at $0.27 per million tokens scores two points above Claude Sonnet 5 at $5.09. That is the single most useful thing to know about LLMs as a buyer: the model is a dial, not a fixed cost. Whizi charges by the same logic, one credit for a message on the house model, 20 for Claude Opus 5, and as much as 100 for the priciest rows, which how Whizi credits work lays out per model.

The honest caveat is that a composite index is an average over tasks. For a long chain of reasoning the flagship earns its price; for a rewrite it doesn't. For a comparison of what each family is actually good at, which AI model should I use is the decision page.

What an LLM cannot do, in one list

Everything below is a consequence of the architecture, not a bug that the next release fixes.

• It cannot know anything after its training cutoff unless retrieval hands it the text.
• It cannot count reliably on long inputs, because it works in tokens and probabilities, not in a loop over items.
• It cannot verify its own output; a second pass by the same model is the same predictor guessing again.
• It cannot remember you between conversations unless the product stores notes and pastes them back in; that memory is a product feature, not a model one.
• It cannot see the whole conversation forever. Every model has a context window, the maximum text it can consider at once, and what is a context window explains what happens when you hit it.

The practical consequence: treat an LLM as a fluent, occasionally wrong colleague, and keep the verification on your side of the desk. The loop that turns a model into something that acts on its own is a different thing, and what is an AI agent draws that line. For anything with a cost attached, ask a second model. Two different predictors agreeing is weak evidence; two disagreeing is a strong signal to check.