What is a large language model, and what does it actually predict?

Defines a large language model through next-token prediction and connects that mechanism to useful language behavior and its limits.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

A large language model is a model trained on large amounts of text to predict a likely next token from the tokens already in its context. It repeats that prediction to generate a response. Training lets it capture patterns in language, code, facts, and common forms of reasoning, so next-token prediction can produce useful summaries, explanations, or transformations. It does not mean the model is retrieving a guaranteed fact from a database or executing a fully reliable reasoning procedure.

Full answer below

Practice focus

A grounded mental model of what an LLM does without equating fluent output with stored facts, reasoning guarantees, or consciousness.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
A useful definition explains the mechanism plainly, then connects it to both the surprising capability and the uncertainty an application must manage.
A large language model is a model trained on large amounts of text to predict a likely next token from the tokens already in its context. It repeats that prediction to generate a response. Training lets it capture patterns in language, code, facts, and common forms of reasoning, so next-token prediction can produce useful summaries, explanations, or transformations. It does not mean the model is retrieving a guaranteed fact from a database or executing a fully reliable reasoning procedure.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
The application sends instructions and context, the model converts them into tokens, and each generated token changes what can come next. That explains both the capability and the variability: the model can combine learned patterns in useful ways, but a plausible continuation can still be wrong. I therefore treat the output as a candidate result. For a customer-support draft I might let a person review it; for an account balance I would retrieve the value from the system of record and use deterministic code to present it.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Calling an LLM a searchable database of everything in its training data.
Treating fluent wording as proof that a claim is correct.
Explaining the model only as autocomplete and ignoring its learned representations.
Claiming the mechanism guarantees human-like understanding.

What is a large language model, and what does it actually predict?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue practice

Next interviewer turn

Tech stacks

Skill areas

Human editorial direction

Was this drill useful?

Found this answer useful?