Technology

Large Language Model (LLM)

A large language model is a transformer-based neural network trained on web-scale text to predict the next token given the preceding context. The architecture is mechanically simple, predict the next piece, repeat. The consequence is not: at sufficient scale the model produces fluent, contextually coherent output from a short prompt. GPT, Claude, Gemini, and Llama are specific instances of the pattern.
Reviewed by 7wData
Mentioned Hallucination (in LLMs) Prompt Injection

Why it matters

LLMs are the substrate the current AI wave is built on. Most “AI governance” conversations in 2026 are LLM governance conversations dressed in broader language. Four architectural facts produce every downstream risk class: output is probabilistic next-token prediction (not deterministic), it depends entirely on training data, it is bounded by a fixed context window, and there is no built-in fact verification. From those you get hallucination, prompt injection, training-data provenance disputes, and model-output non-reproducibility.

Where you’ll encounter it

Three concrete contexts. A vendor pitches “AI-powered” and the right follow-up is “which LLM, and which provider hosts it”, because that answer decides most of your data-residency and model-risk questions. A model-risk team asks for an LLM inventory and discovers the count is an order of magnitude higher than the official one, because every SaaS tool quietly added one. A customer asks whether the LLM was trained on their data, and an honest answer separates training, fine-tuning, and inference-time context. The pitfall underneath all three: teams treat LLM behaviour as software-deterministic when it is fundamentally probabilistic.

Part of the 7wData AI Glossary. Tracking how concepts like this move in the expert conversation: daily signals at ins7ghts.com.