LLM
Large Language Model
A statistical model trained to predict the next token, which makes it shockingly good at language tasks and unreliable at anything that requires guaranteed correctness.
An LLM is a model trained on enormous amounts of text to do one thing: predict the next token (roughly, the next word fragment) given everything that came before. Stack enough of that prediction together and you get fluent answers, summaries, translations, and code. That is the whole trick. It is not reasoning in the human sense, it is very good pattern completion.
This matters because it explains both the magic and the limits. An LLM has no memory of your business, no knowledge past its training cutoff, and no built-in guarantee that the fluent answer it produces is true. It will state a wrong fact with exactly the same confidence as a right one. That failure mode has a name, hallucination, and it is not a bug to be patched out, it is what next-token prediction does when it has no grounding. Treat the model as a tool, not an oracle.
Worked example of the right architecture: a company wants an assistant that answers questions about its own contracts. The naive build asks the raw model and gets confident, wrong citations. The working build wraps the model: RAG retrieves the relevant contract clauses at query time so answers come from the company’s documents rather than the model’s training data, output is validated against a schema, and anything that triggers a legal action routes through a human. The model did not get smarter; the engineering around it got serious.
The most common founder mistake is reaching for a bigger model to fix a grounding problem, or reaching for fine-tuning when the real need is retrieval. Fine-tuning changes style and format; it does not reliably inject facts, and it goes stale the moment your data changes. The honest trade-off: an LLM is the wrong tool when you need deterministic, auditable output (tax calculations, regulatory logic, anything where “usually correct” is a liability), and the right tool for drafting, classification, extraction, and search over your own data. Used in an agentic loop it can act, not just answer, which raises the stakes further. We build exactly this kind of grounded, validated system under Artificial Intelligence. Used as a source of truth, an LLM is a confident liability.