For the complete documentation index, see llms.txt. Every page on this site is also served as Markdown: append `.md` to any URL, or send `Accept: text/markdown`.
Attensira Logo
Attensira
AI Glossary

Large language model (LLM)

A neural network with many parameters, trained on large amounts of text to predict the next token — and the reason an answer about your product can be fluent, confident and wrong.

Karl-Gustav KallasmaaKarl-Gustav Kallasmaa, Founder & CEOLast updated

A large language model is a neural network with a very large number of parameters, trained on large amounts of text, that generates language by repeatedly predicting the next token. Anthropic's glossary puts it as "AI language models with many parameters that are capable of performing a variety of surprisingly useful tasks" — trained on vast text data, able to generate human-like text, answer questions and summarise information.

Everything else about how these systems behave — the fluency, the confidence, the specific way they get facts about your company wrong — follows from that one mechanism plus what was done to the model after pretraining.

The mechanism, stage by stage

Tokenisation. Text is encoded into tokens before the model sees it. Anthropic describes tokens as the smallest individual units of a language model, corresponding to words, subwords, characters or bytes, and documents that a Claude token approximately represents 3.5 English characters. Your product name may be one token or five, and that is not a detail you control. See tokens.

Pretraining. Autoregressive models are pretrained on a large unlabelled corpus to predict the next word given the previous context. Anthropic notes these pretrained models "are not inherently good at answering questions or following instructions" — the raw objective is prediction, not helpfulness.

Fine-tuning and RLHF. Further training on additional data, and reinforcement learning from human feedback, turn the predictor into an assistant. RLHF ranks example outputs and pushes the model toward the higher-ranked ones. Claude, Anthropic states, has been fine-tuned and trained with RLHF.

Inference. At request time the model works inside a context window — "working memory" for that request, distinct from the corpus it was trained on. Anthropic's context windows documentation records 1M-token windows on several current Claude models, and warns in the same document that as token count grows, accuracy and recall degrade, a phenomenon it calls context rot.

The architecture underneath is the Transformer, introduced in Attention Is All You Need (submitted 12 June 2017), which dropped recurrence and convolution in favour of attention alone and reported 28.4 BLEU on WMT 2014 English-to-German and a single-model 41.8 BLEU on English-to-French after training for 3.5 days on eight GPUs.

Why the mechanism decides what you can influence

There are two entirely different routes by which a sentence about your company can appear in an answer, and they have opposite economics.

Through the weights. Something in the training corpus shaped a statistical tendency. You cannot edit it, you cannot date it, and you get no citation — the model produces plausible text, including plausible-looking URLs that were never fetched. See AI training data.

Through retrieval. A document was fetched during the request and placed into the context window. That passage exists, can be cited, and changes the moment you edit the page. This is the only route with a feedback loop short enough to work with, and it is why RAG matters more to a publisher than model choice does.

The practical consequence: optimising for an LLM is mostly not about the LLM. It is about whether a fetchable, self-contained passage of yours is the best available answer to the question at the moment it is asked.

Failure modes

Fluent wrongness. The training objective rewards likely text, not true text. A confident, well-formed, incorrect sentence is not a malfunction; it is the base behaviour when nothing grounds the request. See AI hallucination.

Fabricated citations. A URL generated from the weights looks exactly like a URL that was retrieved. The presence of a link is not evidence a page was read.

Stale claims. A parametric answer reflects a corpus frozen months earlier. Pricing, positioning and feature claims you changed last quarter can persist for as long as that model is served.

Context rot. Filling a large window is not free. Accuracy and recall degrade as the token count grows, so a retrieval step that dumps everything can perform worse than one that selects well.

Non-determinism. Anthropic states that even with temperature set to 0, results are not fully deterministic and identical inputs may produce different outputs across API calls. A single observed answer is one draw, not a measurement — which is why anything reported as a rate needs to say how many draws it came from.

Frequently asked questions

What is the model actually doing?

Predicting the next token, repeatedly, conditioned on everything already in its context.

How big is a token?

Anthropic documents roughly 3.5 English characters per token for Claude, varying by language.

Does a 1M-token window remove the need for retrieval?

No. The window is finite against a corpus that is not, and accuracy degrades as it fills.

Can I get a model to unlearn something about my brand?

Not directly. You can change what a retrieval step finds today; you cannot edit weights already shipped.

Frequently Asked Questions about Large language model (LLM)

Anthropic's glossary defines LLMs as AI language models with many parameters that are capable of performing a variety of surprisingly useful tasks, trained on vast amounts of text data, able to generate human-like text, answer questions and summarise information. The mechanism underneath is next-token prediction learned during pretraining.

Tokens are the smallest individual units a language model processes, and can correspond to words, subwords, characters or bytes. Anthropic documents that for Claude a token approximately represents 3.5 English characters, though the exact number varies by language.

The Transformer, introduced in Attention Is All You Need by Vaswani and colleagues, submitted 12 June 2017. It dispensed with recurrence and convolution in favour of attention alone, reporting 28.4 BLEU on WMT 2014 English-to-German and a single-model 41.8 BLEU on English-to-French after training for 3.5 days on eight GPUs.

No. The window is working memory for one request, not a store of your site. Anthropic's [context windows documentation](https://platform.claude.com/docs/en/build-with-claude/context-windows) records 1M-token windows on several current Claude models and notes in the same page that as token count grows, accuracy and recall degrade — a phenomenon it calls context rot. Something still has to select what goes in.

Not in the weights, and not on your schedule. What you can change is what a retrieval step finds today. A claim that lives in a retrievable, self-contained passage on a fetchable page is correctable by editing that passage; a claim baked into pretraining is not.
Share this term

Track how your brand shows up in ChatGPT, Claude, and Google AI

Attensira monitors your visibility across AI search platforms so you know exactly when and how you're being recommended.