Large language model (LLM)

A large language model, or LLM, is an AI model trained on broad language data to predict and generate text. It’s the language engine inside products such as ChatGPT, Claude, and Gemini, and it can help you read, classify, summarize, and draft across many kinds of work.

But being able to write an answer isn’t the same as knowing your firm’s current records or making sure a job finished correctly. The model needs the right information and checks around it before you can depend on it in client delivery.

How does a large language model produce an answer?

An LLM breaks text into tokens, which can be whole words, parts of words, or punctuation. It turns those pieces into numbers and processes them through learned settings called parameters, which shape what it predicts next.

A parameter isn’t a database row holding one fact. The model uses all those learned settings to estimate the next token, adds it to the answer, and repeats with the longer sequence until the response ends.

“Attention Is All You Need,” submitted June 12, 2017, introduced the Transformer architecture behind most modern LLMs. Its attention mechanism lets the model weigh relationships among tokens in the available input.

Think about the word “bank” in a proposal for a financial client and in a landscaping brief about a river. The surrounding words help the model choose which pattern fits, rather than treating the word as having one meaning everywhere.

During training, the model’s parameters change as it works through broad language data and improves its predictions. The GPT-3 paper, submitted May 28, 2020, tested a 175-billion-parameter model on more than two dozen tasks, including translation and question answering.

The same model handled different tasks through instructions and examples in its input. It didn’t need a new training run for each request.

What is the difference between an LLM and an AI application?

The model is one part of the application. The application decides what you can send, which model receives it, what tools it may use, how conversation history is stored, and what happens to the result.

Products such as ChatGPT, Claude, and Gemini add an interface and may add files, web access, saved preferences, safety controls, and tool calls around one or more models. Those layers can change your experience without changing the underlying model.

An AI workflow adds a fixed path, using the model for defined steps and sending exceptions to an owner. An AI agent has a goal and allowed tools, with room to choose actions, so it also needs permissions, testing, monitoring, and a way to recover failures.

Stanford’s August 18, 2021 foundation-model definition describes models trained on broad data at scale that can be adapted to many later tasks. They are building blocks for applications, and the application determines what a particular use can do.

What changes when you write a prompt?

A prompt supplies instructions and context for the current request. You can name the task, attach sources, show examples, ask for a format, and set limits the model should follow.

The GPT-3 researchers called requests with no example zero-shot, with one example one-shot, and with a few examples few-shot. The model’s parameters didn’t change during those requests, even though the examples could change the answer.

That distinction helps separate several terms you’ll hear together. The context window limits how much input and generated output the model can consider at once, while memory saves selected information for later runs.

Retrieval-augmented generation fetches outside records and adds them to the current context. Fine-tuning changes model behavior through more training, which is different from putting a useful example into this request.

A long conversation can look like the model is learning because earlier messages affect later answers. That works while the application keeps those messages available or brings selected details back from memory, without changing the base model’s parameters.

Why can an LLM sound sure and still be wrong?

The model produces text from learned patterns, and factual support is a separate question from whether those words read well. A clean sentence can still have no source behind it.

NIST’s July 2024 Generative AI Profile connects confabulation, false or internally inconsistent output, to that statistical design. The same prediction process can produce accurate material or a confident falsehood, with more exposure in open-ended work and tasks needing deep knowledge of the subject.

Suppose you ask for a summary of a client’s market with web access switched on. A paragraph can come back with a study title, date, and percentage that look credible together, but if the study doesn’t exist, the work still fails.

That’s a hallucination, and its polished wording can make it harder to notice. You need to check the claim, not reward the sentence for sounding professional.

Grounding ties claims to source material you can inspect, and an eval tests the work against a written standard. Neither guarantees every answer is true, but they give you checks beyond “this sounds right.”

What must a business add around the model?

Name the task, choose the source of truth, supply the needed context, limit permissions, check completion, and name the person who owns exceptions. Those decisions belong to the firm, and buying a stronger model doesn’t make them for you.

Take a client-intake summary. The LLM can extract requirements from an approved call transcript, but the surrounding system has to confirm the client, attach the current scope, block unsupported pricing promises, store the summary in the right record, and send missing information to somebody who can decide.

Without those connections, a general chatbot remains an individual search and drafting tool. It can save time, but it cannot use your firm’s current truth unless you supply it.

Document the process first, so the model gets the source, rule, and stopping point your team uses. If two managers resolve the same exception differently and nobody records why, there’s no stable standard for the model to follow.

And give the system less room to act when being wrong costs more. A draft someone can check in thirty seconds needs a different boundary from a client promise or a deleted record, so delegate based on whether you can check the result and undo the action.

Does a firm need to train its own LLM?

I would test the first use with an existing model, one task, approved records, a clear prompt, and a small eval made from real cases. Compare finished results and review effort across more than one model before changing the model itself.

Fine-tuning becomes worth considering when tests keep showing a stable behavior problem that prompts, examples, retrieval, and workflow rules don’t fix. Private facts alone are usually something to retrieve for the current run, rather than put into the model through training.

A more capable model may make fewer mistakes on difficult work and cost more per request. A cheaper model may be enough for a narrow classification task, so compare what each accepted result costs, including the checking.

Use the smallest model that passes your own eval at the quality the task requires. That choice can differ between steps in the same workflow.

Where does an LLM sit on the Delivery Model Ladder?

Direct use of an LLM belongs mainly at Stage 1, Enhanced, on the Delivery Model Ladder. A person opens the application, prepares the input, checks the answer, and moves it to the next system.

Stage 2 begins when the workflow supplies approved context, calls the model for a defined task, checks the result, and carries it forward. The change is in how delivery runs, not just how quickly someone gets an answer.

Stage 3, AI-native, needs the wider operating model, shared context, ownership, testing, and recovery. A larger LLM doesn’t create those things by itself, and that’s part of the Production Gap.

When is an LLM the wrong tool?

Use ordinary software where the answer must follow an exact rule, such as arithmetic, fixed database lookups, and stable routing tables. An irreversible account change also needs an enforced permission and decision rule, not another plausible response from a model.

The LLM can still help with the messy language around that exact step. It might extract invoice fields from an email while ordinary code calculates tax and validates the total.

That division assumes you can name the source of truth and test the output. If you can’t do either yet, keep the work manual until those parts exist.