What is a large language model?
It answered your question correctly yesterday and confidently wrong today, on the same topic. That is not a bug. It is what prediction looks like from the outside.
Large language model (LLM)
LLM
A large language model is a neural network trained on huge amounts of text to predict the next token in a sequence. It is then tuned to follow instructions and hold a conversation. Claude, GPT and Gemini are all LLMs. A law degree is not, whatever the search results next to this page suggest.
"Large" refers to two things. First, the training data: trillions of words scraped from books, code and the web. Second, the parameter count: the internal numbers the model adjusts during training, 175 billion of them for GPT-3. Current frontier models do not publish the figure.
Training produces a system that completes text patterns, not a system that stores verified facts in a lookup table. That distinction is the whole page. It is a prediction engine. Every hallucination, stale answer and inconsistent reply we get called in to fix traces back to a client treating it like a database instead.
Prediction engine, not knowledge base
A database returns a stored row or nothing. An LLM returns whatever text is statistically likely to follow your prompt. It produces that text whether or not the facts behind it exist. Fluency does not signal accuracy, because both come from the same process.
That explains three failures buyers bring us on nearly every project. Hallucination: the model states a policy number or a case citation that sounds right and is invented. Inventing fits the pattern as well as recalling does. Staleness: the model's knowledge stops at its training cutoff, so it confidently describes a product that shipped a version ago.
A third failure, non-determinism, follows the same logic. Ask the same question twice and the wording can shift, because the model samples from likely answers instead of retrieving a fixed one. None of this makes an LLM unreliable the way a buggy script is unreliable. It needs retrieval for facts and an eval suite to catch drift, that is all.
- PromptYour text, tokenized.
- WeightsBillions of trained parameters.
- PredictionNext-token odds, repeated.
- OutputAssembled, not retrieved.
Nothing in this loop checks a fact against a source. That step gets added, or it does not happen.
Related questions
01What does LLM stand for?
LLM stands for large language model. "Large" points to the scale of training data and internal parameters. "Language model" points to the core task: predicting the next piece of text in a sequence.
02What actually happens when an LLM answers a question?
An LLM breaks text into tokens, feeds them through a trained neural network, and predicts the most likely next token, on repeat, until the answer is complete. Nothing in that loop looks up a fact in a source document. A system has to be built to fetch one first.
03Is ChatGPT an LLM?
ChatGPT is a product built around an LLM, specifically OpenAI's GPT family, with a chat interface, guardrails and tool access layered on top. The model and the product wrapped around it are different things. That is why two products can share a base model and still behave differently.
04Can an LLM know things happened after it was trained?
Not on its own. An LLM's knowledge stops at its training cutoff date. Answering about anything more recent requires giving it that information directly, through a live search tool or a retrieval system, in the prompt.
