Large Language Model (LLM)
Mô hình Ngôn ngữ Lớn (LLM)
A deep learning model trained on massive text datasets to understand and generate human language.
A Large Language Model (LLM) is a type of neural network trained on hundreds of billions — sometimes trillions — of words sourced from the web, books, code repositories, and other text. Through this training process the model learns statistical patterns of language: which words follow which, how ideas connect, what makes a sentence grammatical, and much more.
**What it actually does**
At inference time (when you're talking to it), an LLM predicts the most likely next token given everything it has seen so far. "Token" is roughly a word-piece; a 1,000-word essay is about 1,300 tokens. This simple prediction task, applied billions of times over, produces surprisingly coherent and useful text.
**Why it matters**
LLMs power tools like ChatGPT, Claude, Gemini, and Copilot. They can draft emails, write code, summarize documents, translate languages, answer questions, and reason through problems — all from a single model without task-specific retraining.
**Scale is the key variable**
Bigger models trained on more data tend to show emergent capabilities — abilities that smaller models lack entirely, like multi-step reasoning or following complex instructions. GPT-4, Claude 3.5, and Gemini 1.5 Pro are examples of frontier LLMs.
**When you'd use one**
Directly: API calls for text generation, chat, classification, summarization. Indirectly: almost every AI product today runs on an LLM under the hood.
**Pitfalls**
LLMs hallucinate — they generate plausible-sounding but factually wrong content. They have knowledge cutoffs, so recent events are unknown. They can reflect biases present in training data. And they're expensive to run at scale compared to traditional software.