Key takeaways
- An LLM generates text one token at a time, each chosen from patterns learned in training.
- Training builds the model; using it afterwards is called inference and changes nothing inside it.
- An LLM knows nothing after its knowledge cut-off unless it is given newer information.
- Fluent output is not verified output: check names, numbers and commitments.
How do LLMs work
LLMs work by repeatedly predicting a suitable next token given all the text so far.
A token is a word or part of a word. The model reads your prompt as tokens, works out through many layers how they relate and scores every possible next token. It picks one, adds it and repeats until the answer is done. Most of today's LLMs use the transformer design, whose attention mechanism lets each token weigh the tokens that came before it.
Training, inference and the knowledge cut-off
Training is when the model learns from text and its weights change; inference is when it is used to answer, and nothing in the model changes.
Training happens before release, at great cost, on a large collection of text. After that the model is fixed, so its built-in knowledge stops at a knowledge cut-off date. Anything newer, and anything private to your company, reaches it only if it is placed in the conversation or fetched by a retrieval system at the time of the question.
LLM vs generative AI
An LLM is one kind of generative AI: the kind that works with language.
Generative AI also includes image, audio and video models. Some newer models handle text and images together, so the line blurs, but when a business tool writes an email or summarises a document, a language model is almost always doing the work.
LLM limitations
The main LLM limitations are invented facts, a fixed knowledge cut-off, a limited context window and uneven reasoning.
Research on grounding, retrieval and pairing models with symbolic tools aims to reduce these problems; none of it removes them.
- Hallucination: confident, fluent statements that are not true, including invented sources.
- Stale knowledge: nothing after the cut-off unless it is supplied.
- Context limits: only a set amount of text is taken into account at once.
- Reasoning: arithmetic and long chains of logic can go wrong without warning.
- Sensitivity to wording: small changes to a prompt can change the answer.
Large language model examples in a sales workflow
A common example is turning a messy discovery-call transcript into a tidy summary for the CRM. The LLM groups the buyer's problems, names the people mentioned and lists next steps, which it does well because summarising is a language task.
Before the summary goes into a record, the seller checks the names, numbers and dates against the call, because the model may have filled a gap with something plausible. Other everyday examples are rewriting an email for a different reader, drafting answers to RFP questions from approved material and translating a message for an overseas buyer.
Where Pebbles Ai fits
Common questions
What does LLM stand for?
LLM stands for large language model. Large refers to the size of the model and of the body of text it was trained on.
Does an LLM search the internet when it answers?
Not by itself. A model answers from patterns learned in training; some products add a search or retrieval step that fetches pages or documents and passes them to the model.
Why does the same prompt give different answers?
Most products let the model pick among several likely next tokens rather than always the single most likely one. That variety helps drafting but means answers can differ from one run to the next.
About this page. Written by the Pebbles Ai team for people who run sales, marketing and growth. Last updated .
