Learn
What is an LLM?
An LLM, or large language model, is an AI model trained on huge amounts of text to predict the next small piece of text, called a token, from everything before it. That one skill lets it write, summarise, translate, answer questions and use tools. ChatGPT, Claude, Gemini and Grok are apps built on LLMs.
Updated
How does an LLM work?
An LLM works by predicting, one token at a time, what text is most likely to come next. Google's machine-learning course puts it simply: language models "estimate the probability of a token or sequence of tokens occurring within a longer sequence". Ask "What's the capital of France?" and the most likely continuation is "Paris". Repeat that prediction thousands of times and you get an email, a summary or a contract clause.
Three ideas make a language model large and capable:
- Training on vast amounts of text. The model reads a huge sample of books, websites and code, adjusting its parameters to get better at predicting the next token.
- The transformer. Since a 2017 paper titled "Attention Is All You Need", LLMs use an architecture whose self-attention lets the model weigh how every word relates to every other word in the text, rather than reading strictly left to right.
- Instruction tuning. After the first training, models are trained further on examples of following instructions and holding a conversation, which turns a text predictor into an assistant.
What are tokens and context windows?
Tokens are the units an LLM reads and writes, and the context window is how many it can consider at once. Google's course calls tokens "the atomic units of language modeling": whole words, parts of words or single characters. Google's Gemini documentation gives a useful rule of thumb: a token is about 4 characters, and 100 tokens is about 60 to 80 English words.
| Text | Roughly | Tokens, at 60 to 80 words per 100 tokens |
|---|---|---|
| A short email | 150 words | 190 to 250 |
| A one-page letter | 500 words | 625 to 830 |
| A 20-page contract | 10,000 words | 12,500 to 16,700 |
| A year of one person's sent mail | 500,000 words | 625,000 to 830,000 |
The context window holds everything the model works with in one go: in Anthropic's words, "all the text a language model can reference when generating a response, including the response itself". Claude's documentation lists windows of up to 1 million tokens, depending on the model. Anything outside the window, the model simply doesn't see, which is why what you give it matters as much as which model you pick.
What are LLMs good and bad at?
LLMs are good at working with language you give them and weak where they have to supply facts or exact logic themselves. Google's course lists hallucinations, computing costs and bias among their challenges.
| Good at | Weak at |
|---|---|
| Drafting emails, posts and documents in a given tone | Facts that aren't in their training or your prompt |
| Summarising long threads, calls and contracts | Events after their training cut-off, unless given search |
| Pulling details out of messy text: names, dates, amounts | Exact arithmetic over many numbers, unless given a calculator tool |
| Sorting and labelling: "is this a complaint or a question?" | Saying "I don't know" when they should |
| Translating and rewriting | Remembering anything between separate conversations, unless given memory |
| Writing and explaining code | Being identical every time: the same prompt can give different answers |
Most of the weaknesses have the same fix: give the model the right material and the right tools. A model that can search your documents and run a calculation is far more reliable than one answering from memory.
What's the difference between an LLM, a chatbot and an AI agent?
The LLM is the engine; a chatbot and an agent are things built around it. A chatbot, such as the ChatGPT or Claude app, wraps an LLM in a conversation. An agent gives the LLM tools, such as email, a calendar or a database, and lets it take several steps on its own towards a goal.
| LLM | Chatbot | AI agent | |
|---|---|---|---|
| What it is | The trained model | An app for talking to a model | A model using tools in a loop |
| You give it | Text | Messages | A goal |
| It gives back | Text | Replies | Completed work, plus a report |
| Example | A model family, such as OpenAI's GPT | The ChatGPT, Claude or Gemini apps | An assistant that drafts replies to every new support email |
The AI agent explainer covers agents in detail.
How should a small business use LLMs?
Use LLMs for the reading and first drafts, give them your own facts, and keep a person on anything that commits the business. In practice:
- Choose by terms as well as quality. Check whether the provider trains on your data; business plans and API access usually don't.
- Give it your material. Price lists, policies and past emails make answers specific instead of generic. This is what retrieval does automatically.
- Ask for sources. An answer that cites the email or page it came from is easy to check.
- Keep sending, paying and deleting behind approval until you've seen it get those right.
- Don't lock yourself in. Models improve every few months; keep your knowledge somewhere you can point a new model at.
How does Mojobase use LLMs?
Mojobase lets you pick the LLM: ChatGPT, Claude, Gemini, Grok, Hermes or your own setup, chosen in Settings → AI and switchable whenever something better comes along. Everything Mojobase knows about your business, such as your mail, calendar, documents, customers and Handbook, stays in Mojobase rather than in the model, so changing models loses nothing.
Before it drafts or answers, the AI reads the Handbook's business profile and searches the relevant mail, documents and records, so it works from your facts rather than its training. Content goes only to the provider running the feature, under terms that don't allow training on it. More on how the AI works.
Further reading
Questions people ask
What does LLM stand for?
Large language model. "Large" refers to the number of parameters, the adjustable numbers the model learns in training, and to the amount of text it's trained on. "Language model" means it estimates which text is likely to come next.
Is ChatGPT an LLM?
ChatGPT is an app built on OpenAI's LLMs, the GPT family. The same is true of Claude (Anthropic's models), Gemini (Google's) and Grok (xAI's): the chat app is the product, and the LLM is the engine inside it.
Why do LLMs make things up?
Because they produce likely text, not checked facts. When the answer isn't in what they were trained on or what you gave them, they can still produce something fluent and wrong, which is called a hallucination. Giving the model your own documents and asking for sources reduces it.
Do LLMs learn from my conversations?
Not during the conversation: the model's parameters don't change as you chat. Whether your conversations are later used to train future models depends on the provider and plan; business plans and API use usually exclude your data from training. Check the terms.
What is the difference between an LLM and generative AI?
Generative AI is any AI that creates content: text, images, audio or video. LLMs are the text-focused kind, though many now also read images and other inputs. Every LLM is generative AI; not all generative AI is an LLM.
How big is an LLM's memory?
Its working memory is the context window, the text it can consider at once, measured in tokens. Claude's documentation lists windows up to 1 million tokens depending on the model, which is several hundred thousand words.