Large Language Model (LLM)
The Sorting Hat of language models — probabilistic, trained on everything, occasionally wrong about which house you belong in.
A Large Language Model (LLM) is a neural network with billions of parameters trained on text from the internet, books, code, and other sources to learn the statistical patterns of language. During training, the model learns to predict the next token (roughly, a word fragment) given the preceding context — a deceptively simple objective that, at sufficient scale and data quality, produces emergent capabilities including reasoning, summarization, translation, code generation, and nuanced question answering. The resulting models are not databases that retrieve stored facts but probabilistic text generators that produce the most likely continuation of any given input — which is why they can be creative and flexible, and also why they can hallucinate confidently incorrect information.
Modern LLMs are the foundation of virtually every AI product released since 2020. They power ChatGPT, Claude, Gemini, and all consumer AI assistants; they're the reasoning engine inside AI agents, copilots, and code editors; they generate the text in AI writing tools and summarize documents in AI research tools. The key dimensions that differentiate LLMs are: scale (number of parameters and training tokens), context window (how much text the model can process in a single request), capability profile (coding, reasoning, multilingual), and safety properties (how the model handles harmful or misleading requests). Different models have different strengths — Anthropic's Claude models, OpenAI's GPT series, and Google's Gemini models each optimize across these dimensions differently.
For B2B teams evaluating AI products and tools, understanding that LLMs are probabilistic generators rather than knowledge retrieval systems is essential for appropriate deployment. LLMs should be trusted for tasks where some creative variation is acceptable (drafting, brainstorming, summarizing), augmented with grounding mechanisms (RAG, tool use, structured data) for tasks requiring factual accuracy, and never deployed unsupervised for consequential decisions that depend on verified facts. For content and video production teams, LLMs offer genuine productivity gains in scripting, research synthesis, and metadata generation — with the caveat that human review remains essential for factual accuracy and brand voice consistency.
Related terms
- Prompt Engineering— Asking the Mirror of Erised exactly what you need, not what you want — the difference between useful AI and a wish gone wrong.
- Fine-Tuning— Training a model on your specific data — Hermione studying twelve targeted textbooks versus winging it from general knowledge.
- Hallucination— When the AI confidently states something false — the Palantír showing what Sauron wants you to see, not what's real.
- Context Window— How much the AI holds in working memory — the Pensieve has infinite capacity; LLMs are still catching up.