What is a large language model?
A plain-English guide to the technology behind chat AI, from a New Zealand team building one.
How it works, in four steps
- Training. The model reads a huge collection of text and learns patterns: which words and ideas tend to follow which.
- Tuning. It is then trained further on examples of good answers, so it follows instructions and talks like an assistant.
- Your question. Your message is split into small pieces called tokens.
- The answer. The model predicts the most likely next token, adds it, and repeats until the answer is done.
What LLMs are good at
- Drafting and rewriting: emails, posts, product descriptions, policies.
- Summarising long documents and meeting notes.
- Explaining things in plain English, and answering follow-up questions.
- Brainstorming options you hadn’t thought of.
Where they go wrong
- Made-up facts. Because it predicts words, a model can invent a figure, a quote or a source that sounds right. This is often called hallucination.
- Changed details. When rewriting, it can change a price or date. See how to check AI rewrites.
- Out of date. A model only knows what was in its training text unless it can search or is given documents.
- Local gaps. Models trained mostly on overseas text can miss New Zealand law, spelling and te reo Māori. See AI and macrons.
What to look for in a chat AI for business
- Where your chats go and whether they are used for training. See does your chat AI learn from your chats?
- Who runs it, and whether it is an open or closed model.
- Honesty about limits: does it say when it isn’t sure?
- Checks on facts, especially for anything you send to customers.
NOVA’s own LLM: Ankaa
NOVA is building Ankaa, its own chat AI, in New Zealand. It is an open model trained further by NOVA and run on NOVA’s own servers. Your chats won’t be used to train anything unless you choose to share them, answers that rewrite your text pass through Gnomon to flag any changed number, price or date, and when Ankaa isn’t sure, it will say so. Ankaa isn’t ready yet; join the waitlist to try it first.
Questions
What is a large language model in simple terms?
A program trained on a very large amount of text to predict the next word. Repeating that prediction lets it answer questions and write text.
What is the difference between an LLM and a chatbot?
The LLM is the model that writes the text. A chatbot or chat AI is the app you talk to, which sends your messages to an LLM and may add search, memory or other tools.
Why do LLMs make things up?
They predict likely words rather than looking facts up, so they can produce something that sounds right but isn’t. Check important facts, or use a checker.
Is NOVA building its own LLM?
Yes. Ankaa is NOVA’s own chat AI: an open model trained further by NOVA and run on NOVA’s own servers. It isn’t released yet, and you can join the waitlist.
A chat AI built in New Zealand.
Ankaa isn’t ready yet. Join the waitlist and you can try it first.