What Is a Large Language Model, in Plain Language
Why this matters
- Understand what the money buys and where the limits fall — before a contract is signed
- Tell apart the tasks where a model delivers from the tasks where it will fail confidently
- Ask a vendor questions that cannot be answered in generalities
What is a large language model, in plain language
A phone suggests the next word while a message is being typed. Type "Good afternoon, please find attached" and it offers "the documents." The suggestion comes from statistics: after these words, that word usually follows.
A large language model works the same way, with two differences. It learned from a body of text the size of a major library: books, articles, correspondence, manuals, code. And it holds hundreds of pages in view at once. That is why the continuation stays coherent over long stretches — out comes a whole letter, a contract review, a reply to a customer.
Nothing more complicated than a phone suggestion is happening inside. At every step, the most likely next word is selected. The model consults nothing and verifies nothing. What looks like knowledge is the statistics of the texts it read, frozen on the day training ended.
How it works
- Training. The model is shown text after text; each time, one word is hidden and the model has to guess it. Billions of such attempts tune its internal coefficients — the parameters. This takes weeks and costs millions of dollars; it is done by the model's developer, and the client company plays no part in it.
- Parsing the request. The text is cut into tokens — fragments of words; in Russian one word usually takes two or three such fragments, and cost is calculated by them. The instruction, the question, any attached documents and the conversation history all go in as a single input.
- Computing probabilities. The model estimates how to continue and ranks the options by probability: one at 60%, another at 12%, the remaining thousands at nearly zero.
- Selection and repetition. One option is chosen, appended to the text, and everything is recalculated — now including what was just appended. Piece by piece, the whole answer accumulates.
- Stopping. The model finishes when it produces an internal end marker itself or hits a set length limit.
Two properties follow that usually surprise executives. The answer to one and the same question comes out slightly different every time: the pick from the list of options carries a degree of randomness, and the level of that randomness is adjustable. And the model works identically in both cases — when it reproduces a hard fact and when it assembles a plausible guess. It does not distinguish between the two.
An example from practice
The legal department of a manufacturing company was drafting replies to routine claims from counterparties: about 60 letters a month, 40–50 minutes per letter.
The model was given a reply template, extracts from the contract and the text of the claim. A draft was ready in 4 minutes, and the lawyer spent another 10–12 minutes checking and editing it — roughly 15 minutes instead of 45. Over a month this freed up about 30 hours of a lawyer's time.
Checking the references proved to be a condition of the whole arrangement. In the first batch of 30 drafts, the model three times cited a contract clause that did not exist: the wording looked apt, the clause number was invented. Once the template was changed to require the clause to be quoted verbatim, such cases became visible at a glance.
What this gives a business
- Time on routine texts. Drafts of letters, replies, descriptions and minutes are ready in minutes. The gain is simple to calculate: how many such documents a month, and how many minutes each one takes today.
- Costs scale with volume. Payment is per token: a short request costs a fraction of a ruble, a review of a multi-page contract costs rubles. The average price of one operation is revealed by a pilot of 50–100 runs, and from there it is multiplied by monthly volume.
- The risk moves to review. The model speeds up the draft; responsibility for the result stays with the employee. A benefit calculation is honest only if it counts the time spent proofreading — that is what decides whether the saving survives.
When this can be skipped
Wherever the answer is computed from rules, a language model is redundant. Calculating a discount, validating a tax ID, routing a request by category code — that is work for ordinary code: cheaper, faster, identical results on every run.
Nor is it needed where accuracy to the last digit is required: amounts, bank details, stock balances. Numbers come from the accounting system, and the model at best restates them in human language.
A separate case: tasks with one correct answer and a high cost of error, where nobody is available to check the result. The gain from speed is eaten up by the first missed mistake.
What is worth checking
A model has to be tested on the company's own tasks and its own texts. Public rankings measure average ability on general assignments; the specifics — contracts from a particular industry, in-house abbreviations, professional jargon — do not show up there.
A working test looks like this: 50–100 real examples with the correct answer known in advance, the same set for every candidate, the error rate counted by hand. The acceptance threshold is stated by the client before the start: in a draft letter, 10% corrections are tolerable; in a reply that goes to a customer unreviewed, they are not.
Where the texts end up is a separate conversation. A public model runs on the vendor's side, and the request together with any attached documents leaves the company perimeter; the terms for storing requests are found in the vendor's contract. If such an exchange is unacceptable, the model is deployed on the company's own servers: quality is usually lower, but the data stays inside.
It is also worth confirming the training cut-off date: for the model, events after it do not exist, and asked about a recent amendment to a law, it will answer from the old version — in the same level tone.
Frequently asked questions
What is a large language model, in plain language?
It is a program that has read billions of pages of text and learned to predict the next word. Give it a beginning — a question, a letter, a section of a contract — and it builds the continuation word by word. From the outside it looks like a conversation with a person; inside, it is statistics: the model picks the most likely option based on the examples it has read.
How does a large language model differ from a search engine?
A search engine finds a document and shows it in full; the source is accountable for the text. A model composes the answer from scratch, word by word, and that answer has no source. This is why the answer looks polished and on point, yet cannot be checked against a link — not until the company's own documents have been connected to the model separately.
Why does the model invent things so confidently?
The model looks for a plausible continuation and always finds one. If the required fact was absent from the training texts, the plausible option turns out to be an invention of similar shape: a clause of law that does not exist, a made-up part number, a citation to a case that was never heard. The tone stays level: the model does not flag how confident it is. Hence the rule — facts get checked, and polished wording proves nothing.
What does "large" mean in the name?
Size is measured in parameters — the internal settings the model arrived at during training. Large models have hundreds of billions of them; compact ones, a few billion. The business implication is simple: the larger the model, the harder the tasks it can usually handle and the more every request costs. For sorting incoming mail, a large model is overkill.
Does the model know a company's internal data?
By default, no. It knows only what was in the texts at the time of training; contracts, prices and internal policies were not among them unless they were published. Company data is connected separately: the relevant fragments are inserted into the request, an arrangement known as RAG. Less often, the model is fine-tuned — but even then it picks up style and format more reliably than specific numbers.
Let’s discuss your project?
Tell us about your process — we’ll suggest where AI pays off fastest.
Related articles
What Is a Token and Why You Pay for It
A token is a chunk of text roughly three quarters of a word long: the model reads a request in tokens, and the provider bills for how many there are.
What is a context window, and why does the model "forget"?
A context window is the limit on how much text a model can see in a single request: instructions, documents, the question, and the conversation history all count together, and anything past the limit is dropped silently.