Large language model (LLM)
LLM · foundation model
ひとことで
A large language model is a neural network with billions of parameters, trained on large text corpora to predict the next token in a sequence. That single objective, at sufficient scale, yields the ability to answer questions, write code, translate and summarise — none of which were trained for directly.
A large language model is trained on one task: given a sequence of tokens, predict the next one. Repeat that across an enormous corpus and the model's parameters come to encode grammar, facts, reasoning patterns, code conventions and a great deal of accumulated human writing style.
Most modern LLMs use the transformer architecture, whose attention mechanism lets every position in the input influence every other. That is what makes long-range coherence possible and what makes computation grow with sequence length.
Training runs in stages. Pre-training on broad text produces raw capability. Post-training — supervised fine-tuning on demonstrations, then reinforcement learning from human or AI feedback — turns a text predictor into something that follows instructions and refuses harmful requests.
Two properties matter in practice. The model is stateless: it remembers nothing between requests except what you put in the context window. And it is probabilistic: the same prompt can yield different outputs, and confidence in its tone carries no information about correctness.
よくある質問
- Does an LLM understand what it writes?
- It has no beliefs or intentions in the human sense. It models statistical structure in language well enough that its output is often correct and useful — and confidently wrong when the structure points the wrong way.
- What makes a language model 'large'?
- Parameter count, training data volume and training compute, all far beyond earlier models. There is no formal threshold; the term is comparative and its baseline keeps moving.