You have heard the phrase "large language model" so many times this year that it has stopped meaning anything. It is in every press release, every consultancy deck, every awkward podcast intro. So what actually is one, in plain English?
Here is the no-jargon version. Read it once, never have to ask again.
The simplest description
A large language model is a piece of software that has read an enormous amount of text and learned the patterns of how words follow each other. When you give it some text and ask it to continue, it predicts the next bit, then the next bit, and so on. That is what it does.
That sounds underwhelming. The reason it is not underwhelming is that "predicting the next bit" turns out to include things like answering questions, writing essays, summarising documents, drafting emails, suggesting code, holding conversations, and a hundred other things that look surprisingly close to thinking.
It is not, technically, thinking. But what comes out is useful enough that the distinction matters less than people argue.
How they get smart
A model gets trained on text. A lot of text. Pretty much the readable internet, plus books, plus code, plus a lot of curated content. During training, it adjusts itself, billions of times, until it gets very good at predicting what word comes next given the words that came before.
By the end of training, it has absorbed enough patterns that asking it "what is the capital of France" or "write me a polite email cancelling a meeting" both work. The first because the model has seen the answer in many forms during training. The second because the patterns of polite English emails are well-established and easy to imitate.
The model is not looking up the answer. It is generating it from what it has learned about the shape of human language.
Why some are better than others
Different models are different sizes (roughly, more parameters means more learned patterns), trained on different data (different mixes give different strengths), and tuned in different ways after training (some are nudged to be helpful, some to be honest, some to be safe).
The current frontier models from Anthropic, OpenAI, Google and Meta are all genuinely capable. The differences between them are real but smaller than the marketing suggests. Picking between them for a specific business agent is more about the right tool for the job than the abstract "best".
What models are good at
- Writing, drafting, editing, summarising.
- Answering questions where the knowledge is well-established.
- Translating between languages.
- Following instructions consistently.
- Reading and reasoning about a chunk of text you give them.
What models are not so good at
- Knowing what they do not know. They will sometimes give a confident wrong answer. This is the most important limit to remember.
- Maths beyond a certain level. Better than they used to be, still not as reliable as a calculator.
- Knowing very recent events. Models have a cut-off date for their training. Anything more recent, they have to be told.
- Acting on the world without being given tools to do so. A bare model can only generate text. It cannot send an email, look up a record, or book a meeting unless you have wired it up to do so.
Treat a large language model like a brilliant, fast graduate who has read everything ever written but has not been to your business. They are useful. They are also wrong sometimes. Build accordingly.
What this means for your agent
An AI agent for your business uses a large language model as the brain. The agent then adds:
- A clear job description (in plain technical terms, a "system prompt").
- Access to your real business data (often via RAG).
- Tools the model can use to take actions (lookups, calendar, CRM, email).
- Guardrails so it does not do things it should not.
- A way to escalate to a human when it is out of its depth.
The model itself is the engine. The agent around it is most of what makes it useful for your business.
The honest summary
Large language models are the most genuinely useful general-purpose technology to come along in a long time. They are also not magic. Used well, they save real hours and improve real customer experiences. Used badly, they make confident mistakes that cost you customers.
The work, as ever, is in the using-well bit.
If you would like to talk about how a model could be put to work in your specific business, that is what our first calls are for. Tell us what your team is stuck on, and we will be honest about whether and how AI fits.