A large language model learns statistical patterns from billions of sentences, then predicts the next word in a sequence one step at a time. Chain enough of those predictions together and you get essays, code, summaries, or direct answers. The word large refers to both the training corpus and the internal parameters, which now number in the billions. The models behind ChatGPT, Claude, and Gemini all fit this description.
For a founder, an LLM is usually an engine you rent through an API rather than something you build from scratch. You send text in and get text back, paying roughly by the amount of text that moves each way. That billing shape rewards concise prompts and disciplined output, so trimming instructions can noticeably lower your monthly bill.
Three limits are worth designing around: a fixed knowledge cutoff, a maximum context window, and a habit of stating wrong facts with total confidence. A model that never saw your internal documents cannot answer questions about them unless you feed that data in at request time. Planning for these constraints early saves painful rework later.
Model choice comes down to balancing quality, speed, and cost. Smaller models reply faster and cost less, which suits high-volume features like autocomplete or tagging. Larger models reason more reliably on messy work such as contract review. Many teams route easy requests to a cheap model and save the expensive one for genuinely hard cases.