A large language model (LLM) is a neural network pretrained on a very large body of text to predict the next token (a word or word fragment) in a sequence, usually then further trained to follow instructions, and used to generate, summarize, translate, and classify text.
Most current LLMs are built on the transformer architecture introduced by Vaswani et al. in 2017. Because an LLM generates output by statistical prediction, its output is not guaranteed to be correct. NIST AI 600-1 calls confidently stated false output confabulation, known colloquially as “hallucination,” and notes it is “a natural result of the way generative models are designed.”
History
The phrase appears in “Large Language Models in Machine Translation” (Brants et al., EMNLP-CoNLL 2007), which described n-gram models trained on up to 2 trillion tokens, before the transformer-based models the term usually refers to today.
Sources
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, section 2.2, NIST, July 2024. Retrieved October 7, 2026.
- Attention Is All You Need, Vaswani et al., 2017. Retrieved October 7, 2026.
- Large Language Models in Machine Translation, Brants, Popat, Xu, Och, and Dean, EMNLP-CoNLL 2007. Retrieved October 7, 2026.
See also
Go deeper
Articles
- Protect Production SQL Databases from AI/LLM Agentic SQL Query Risks February 4, 2026
- Use AI to Describe Images as a Background Job in Ruby on Rails January 23, 2026
- Local AI Setup with Ollama and Nvidia GPU on Ubuntu Linux December 17, 2025