Technology
What is a large language model?
Behind every AI chatbot is a large language model: a program that has read a vast amount of text and learned to continue it. That one skill, at a huge scale, turns into answering, summarising, translating and writing code. That same skill explains why it sometimes states wrong things with confidence.
In short. A large language model splits text into small pieces called tokens and learns, from a vast amount of text, to predict the next one. Built mostly on the Transformer design from 2017 and scaled to billions of adjustable numbers, such models can take on tasks from a few examples. Chat assistants are further trained with human feedback to follow requests. Because they produce plausible text rather than look facts up, they can be wrong with confidence; good assistants search, cite and ask before they act.
One piece at a time
Text is cut into tokens: whole words, parts of words or punctuation. The model reads the tokens so far and gives every possible next token a probability. One is chosen, added to the text, and the process repeats. A long answer is hundreds of these small steps.
The probabilities come from training. The model sees enormous amounts of text, predicts the next token, compares its guess with what actually came next and nudges its internal numbers, called parameters, to do slightly better. Repeated over a vast amount of text, that teaches it grammar, facts, styles and patterns of reasoning, all at once and all with gaps.
The Transformer
Today’s models are mostly built on the Transformer, proposed in 2017 in a paper titled “Attention is all you need”. Its key part, attention, lets each token weigh every other token in the text, so that “it” in a sentence can be tied to the right noun many words earlier. The authors also found it faster to train than the designs before it, because more of the work runs in parallel.
Bigger, and then more helpful
In 2020, the team behind GPT-3, a model with 175 billion parameters, showed that making a language model much larger greatly improved its ability to do a task from a few examples written in the prompt, without retraining. Size alone did not make a model follow what people want, though. In 2022, researchers fine-tuned GPT-3 with human feedback: people wrote example answers and ranked the model’s outputs. On their set of prompts, people preferred the answers of a 1.3-billion-parameter model trained this way to those of the 175-billion-parameter original, which has more than 100 times as many parameters. Training of this kind is how a text predictor becomes an assistant that follows requests.
Why it can be wrong with confidence
A language model produces the most plausible continuation, not a fact it has looked up. When it lacks the facts, plausible and true part ways, and the text stays just as fluent. Researchers call this hallucination; a 2023 survey describes text generated this way as prone to it, and reviews ways to measure and reduce it. In practice, assistants reduce it by searching and citing sources (what RAG is), by using tools for things such as sums and dates, and by asking you before acting.
From a model to an assistant to an agent
A chatbot is a language model you talk to. An assistant adds memory and your own information, such as your calendar. An agent can also act: send, book, buy. Each step adds usefulness and risk: AI agent vs chatbot, what an AI agent is, and prompt injection, the attack that hides instructions in what the model reads.
How Lychnos Ring uses one
Lychnos Ring runs on current language models that we choose and test, with tools for the web, your mail and your calendar in our cloud; you can also plug in your own API key for another provider. The model proposes; tested code decides what needs your yes, and anything that sends, moves, pays or deletes waits for a press of the ring. It is in development: the AI ring.
Sources, read 4 November 2026: Vaswani and colleagues, attention is all you need, 2017; Brown and colleagues, language models are few-shot learners, 2020; Ouyang and colleagues, training language models to follow instructions with human feedback, 2022 (all arXiv); Ji and colleagues, survey of hallucination in natural language generation, ACM Computing Surveys, 2023.
Questions
What is a large language model in simple terms?
A program trained on huge amounts of text to predict the next piece of text. Doing that well, at a large scale, lets it answer questions, summarise, translate and write.
Why do AI chatbots make things up?
They produce the most plausible continuation, not a looked-up fact, so they can state something false fluently. Researchers call this hallucination. Searching and citing sources, and checking answers before acting on them, reduce it.