Generative AI and Large Language Models — How They Work
한국어 English
A large language model is trained to predict the next piece of text over very large corpora, and generates output by repeating that prediction. Most current models are built on the transformer architecture introduced in 2017. Because they predict plausible text rather than retrieve verified facts, fluent output and correct output are not the same thing.
Verified facts
- A large language model is trained to predict the next token in a sequence, and produces text by applying that prediction repeatedly.
- Most contemporary large language models are based on the transformer architecture, introduced in a 2017 paper.
- Model outputs are generated from statistical patterns learned during training rather than looked up from a verified database.
Reported, not confirmed
Statements from a single source, or where sources disagree- Because generation is based on plausibility rather than verification, models can produce confident statements that are factually wrong — commonly called hallucination. Fluency is not evidence of accuracy.
- Capability generally increases with model size and training data, but the relationship is not simple and larger models are not uniformly better at every task.
Timeline
- 2017The transformer architecture is introduced, becoming the basis for most later language models.
Frequently asked
Does an LLM look up facts?
Not by default. It generates text that is statistically plausible given what it was trained on. Some systems add retrieval from external sources, but the underlying model itself is predicting, not consulting a reference.
Why do these models state wrong things so confidently?
Confidence in the output reflects how plausible the text is as language, not how well-supported the claim is. The model has no separate mechanism that marks a sentence as verified, so a fabricated detail can read exactly like a correct one.
What is a token?
The unit a model reads and generates — typically a word fragment rather than a whole word. Text is split into tokens before processing, which is why model limits are usually stated in tokens rather than characters.
Are bigger models always better?
Not uniformly. Scale tends to help, but the relationship is not simple, and a larger model can still be worse at a specific task than a smaller one tuned for it.
Sources
Source titles are kept in their original languageThis English edition covers the verified facts and sourced claims. The full document — with sentence-level source references, timeline, FAQ and the complete revision history — is maintained in the Korean edition.