A neural network trained on massive text corpora to predict the next token in a sequence, which turns out to be a remarkably general primitive for generating coherent text, answering questions, summarizing, translating, and more. Modern LLMs (GPT, Claude, Gemini, Llama) range from a few billion to over a trillion parameters and are the core engine behind every AI meeting copilot. The LLM is what turns a raw transcript into a useful recap, identifies action items, drafts follow-up emails, and answers questions about past meetings. LLMs are not databases of facts — they're probabilistic next-token predictors that have absorbed patterns from training data — which is why grounding and RAG matter so much for meeting-copilot use cases where factual accuracy about specific conversations is critical.